Terry Winograd [l]

[COMMENTARY: This discussion was planned to be chapter 10 in my volume Linguistic Theory: The Discourse of Fundamental Works. London: Longman, 1991. But it was removed party because the publisher objected (as always) to the length of the book, and partly because Winograd himself took strong exception to some of my comments. A special point of contention was his sudden switch of allegiance to Heidegger, whose fervent commitment to National Socialism in Hitler’s Germany, unbeknownst to Winograd, was just being made public by Hugo Ott’s biography Martin Heidegger: Unterwegs zu seiner Biographie (Frankfurt am Main: Campus, 1988). I release the chapter now that Winograd has moved on to other things.]

10.1 Terry Winograd's work is an obvious choice for demonstrating how ‘artificial intelligence’, or ‘AI’ as it is usually called, makes use of ‘research on linguistic theory’ (LAP 6). He was a pupil of Halliday's, and the breadth of his concerns is attested by the titles of his books: Understanding Natural Language (1972, hereafter UN), Language as a Cognitive Process (1983, hereafter LC), and Understanding Computers and Cognition (1986, hereafter UC). In his works linguistics is central, whereas it is disregarded in many AI works on ‘natural language’ (e.g. Raphael 1968), and vigorously opposed in some (e.g. Schank, Goldman, Rieger & Riesbeck 1975).

10.2 Winograd's 1972 dissertation, an implementation of Halliday's ‘systemic grammar’ in a computer program, was an undisputed landmark. It was deemed ‘sufficiently general and important’ to be printed in full as ‘an entire issue’ of a prestigious journal, Cognitive Psychology (UN vii), and soon after as a book. Instead of merely parsing sentences into trees like the ‘Mitre’ program (Zwicky et al. 1965), Winograd's program both embodied a linguistic model and ‘simulated’ a conversational partner, a ‘robot’ that engaged in ‘English dialogue’ with a human (UN 1, x) (10.11f). This banner achievement fuelled hopes that AI would not only enable computers to use natural language instead of programming languages, but would reveal basic properties of language and mind by means of simulation and real-time interaction.

10.3 Winograd's next book project was to be a mammoth survey of ‘models of language’, both in ‘linguistics’ and ‘computer science’ (LC viif), thus integrating natural and formal languages into one perspective. After ‘nearly ten years’, during which the writing ‘underwent fission into volumes’ (LC vii, xi) a 640-page tome on ‘Syntax’ appeared, wherein he backed away from ‘systemic grammar’ toward more ‘generative’ approaches. A projected companion, volume on meaning’, however, did not appear, and Winograd now tells me (by letter) that he has come to consider it ‘an impossible project’, and his ‘thinking has gone in a different direction’ (cf. 10.53). Just how different can be seen from his latest book, co-authored with Fernando Flores, which emphatically questions the ability of AI to model the understanding of natural language. A strange career on the face of it, but a close reading of Winograd's works reveals some strategic continuity and canny policy behind the shifts and manoeuvres.

10.4 Winograd's dissertation illustrated ‘a newly developing paradigm’ ‘grown up from working with computers’ (UN ix) (cf. 10.40). For inquiring ‘what kind of process could be going on to produce’ ‘the highly complex and organized behaviour’ of ‘language’, ‘computers and computer language give us a formal metaphor’ to ‘model the processes and test the implications of our theories’. Although ‘our models’ remain, incomplete’ and, their ‘connection’ ‘with the processes going on in the human mind’ ‘is not yet clear’, they do offer ‘a clear framework for thinking about’ how people ‘understand and respond to natural language’. Major ‘reasons for writing such programs’ included: ‘increasing the ability of computers to communicate with people’, ‘clarifying what language is and how it works in human communication’; and ‘discovering the basic principles underlying, intelligence’ (UN ixf). ‘A usable language-understanding system’ encourages us to ‘make all of our knowledge’ about ‘the entire language process’ ‘explicit’ and thus affords ‘a rigorous test for linguistic theories’ and a means for ‘making new theories’ (UN 2) (cf.10.14, 46, 73). And since ‘language is one of the most complex and unique of human activities, understanding its structure may lead to a better understanding of how our minds work’, above all in ‘areas which involve integrating large amounts of knowledge into a flexible system’ (cf. 10.42, 54, 56, 72; 11.4, 21). Such prospects signal a sly ambivalence in the title of the book: we might better ‘understand natural language’ (as a human capacity) by seeing how a computer goes about ‘understanding’ samples of it (as English dialogue).

10.5 Accordingly, Winograd set the scope of the project very wide. If ‘a person’ ‘makes full use of his knowledge and intelligence to understand’ ‘a sentence’, then an adequate model demands that we ‘deal in an integrated way with all aspects of language': ‘syntax, semantics, inference’, ‘knowledge, and reasoning’ (UN x, 1). A model of ‘intelligence needs a highly structured and coordinated body of knowledge rather than a set of separate uniform facts or axioms’ (UN 142) (cf. 10.26, 39, 48, 71; 11.21). ‘Adding new information’ requires ‘understanding its relationship’ to ‘whatever is already there': ‘a problem-solving activity rather than a clerical one’ (cf. 10.10, 30, 73. 11.12f, 25, 51). A computerized ‘language understander needs to have an interpreter’ for ‘deciding how to use’ ‘each new sentence’, ‘checking for consistency’, ‘creating new data or types of data in storage, modifying theorems’, and so on.

10.6 At that stage of his work, Winograd took pains to refute the commonplace charge that models of entities which ‘exist in the speaker's and hearer's minds’ are ‘mysterious’ and ‘meaningless’ (UN 26) (cf. 4.8f. 5.10; 7.9ff; 8.24; 10.39; 12.38). He stressed the capacity of the ‘structure of concepts which is postulated’ to be ‘manipulated’, ‘obeyed, answered, and added to’ ‘within the computer’. He drew a ‘comparison to the use of “forces” in physics’ (cf. 7.16, 32; 6.62; 12.49, 59. 13.11). ‘We have no way of directly observing a force like gravity, but by postulating Its existence, we can write equations describing it and relate’ them to ‘events’ ‘Similarly, the “concept” representation of meaning is not intended as a direct picture of something which exists in a person's mind’, but ‘a fiction that gives us a way to make sense of data and to predict actual behaviour’, and thus to ‘gain a better understanding of language use’. For example, we can ‘justify using concepts’ if the ‘system’ is ‘thereby enabled to engage in dialogs that simulate in many ways the behaviour of a human language user’ (cf. 11.88; 13.33).

10.7 Within ‘the field’ of ‘language understanding systems’, Winograd distinguished four basic types’ (UN 34). First, ‘special format’ systems were ‘designed’ for 5 ‘particular subject matter’, and treated as ‘relevant’ only the ‘information In a sentence’ that fit the ‘format’ (UN 34f). Naturally, they lacked ‘flexibility’ and were ‘minimally concerned with the complexities of language’, but their, restricted domain’ and ‘special-purpose heuristics’ could produce ‘impressive results’. Second, ‘text-based systems’ had ‘a body of text stored directly under’ an ‘indexing scheme’; ‘an English, sentence input’ was ‘interpreted as a request to retrieve a relevant sentence or group of sentences from the text’. Third, ‘limited logic systems used a ‘more formal notation’ in place of ‘English sentences in the base of stored knowledge’, thereby ‘freeing simple information’ from one ‘specific way of expressing it in English’ (UN 36). Then the ‘bulk of effort’ got invested in ‘translating’ between the ‘English input’ and the ‘assertion format’ (cf. 13.51). The ‘limited logic’ consisted in ‘some mechanisms for accepting more complex information and using it to deduce the answers to more complex questions’ by means of ‘inference’ (UN 36f).

10.8 Fourth, ‘General deductive systems’ had their ‘knowledge’ ‘expressed in some standard mathematical notation (such as first order predicate calculus)’, so that, the work logicians have done on theorem proving could be utilized’ (e.g. Robinson 1965) (UN 38) (cf. 10.78). ‘The procedure is uniform’ rather than ‘suited to a particular subject’; and ‘if any proof is possible’, ‘the procedure will eventually find it’, but ‘may take a very long time’. This method provided ‘a way to present complex information as data rather than embedding it into the inner workings of the language-understanding system’ (UN 38f). The user’ could ‘describe a body of knowledge’ ‘in a “neutral” way’, and the ‘system was applicable to any subject’. But such systems had ‘a low level of practicality’ and ‘tremendous problems of efficiency’. ‘Predicate calculus’ has ‘a serious deficiency’; since first-order logic is a declarative rather than’ a procedural ‘language, specifying how to do something’ is difficult’ without addling ‘strategy information’ (cf. 11.14ff. 13.17). ‘An important part of a person's knowledge concerns, how to go about figuring things out’, deploying ‘large sets of heuristics and procedures for solving problems at different levels of generality’ (cf. 9.14, 17, 10.5, 10, 15, 23, 29). Without this guidance, any ‘large set of axioms, even well below the number needed for really understanding language’, gets it ‘bogged down in searching for a proof’. Indeed, in ‘a truly uniform system, the theorem prover is forced to “rediscover and world” every time it answers a question’, and ‘every goal forces it to start from scratch, looking at all the theorems in the data base’ (UN 41). Thus, an emphasis on ‘abstract logical properties’ like ‘consistency and completeness’ ‘may in fact be bad’ if the system tries to ‘prove’ something which is ‘false’, it will never ‘give up’ until it ‘exhausts every possible way’.

10.9 Surmounting the drawbacks of these four types of system called for a more judicious balance between ‘procedural representations, which embody knowledge in the program, and declarative representations, which emphasize the structure of the stored knowledge’ (LC 18) (cf. Winograd 1975). ‘New programming techniques’ were ‘needed’ for ‘using procedural information’ but ‘expressing it in ways which did not depend on the peculiarities and special structure of a particular program or subject of discussion’ (UN 40). So ‘procedural deductive systems’ were developed whose ‘language was goal-oriented’ and not ‘concerned about the details of interaction among procedures’. The ‘advantages’ were ‘a flexible control structure’ and ‘a uniform notation’ that allowed ‘adding new theorems without relating them to others’ (UN 41; cf. LC 380; 10.48). These systems could ‘handle simple assertions’ as well as ‘complex information’ ‘expressed as procedures’, including knowledge of how best to go about attempting a proof’, e.g., by ‘trying theorems in a particular order’, or taking them from a ‘specified set’, or ‘setting up a subgoal’ with ‘arbitrarily complex calculations’ (UN 40f). The ‘PLANNER language’ used in Winograd's own model for various tasks, such as ‘representing knowledge’, ‘manipulating meanings’, ‘making deductions’, and so on (UN 4, 6f, 40f, 108-17, 126-38, 152), was a system of this type (cf. Hewitt 1973).

10.10 To attain the desired breadth, Winograd's system had an assembly of components. These included ‘a parser, a recognition grammar of English, programs for semantic analysis, and a general problem-solving system’ (UN 1). ‘Heuristic programs’ were supplied for ‘using syntactic, semantic, contextual, and physical knowledge’, and for ‘reasoning about the subject’ being ‘discussed’ (UN 1, x) (cf. 10.5). The system also had ‘a detailed model of a particular domain’ (a ‘world’) and even a ‘model of its own mentality’, so that it could ‘remember and discuss plans and actions as well as carry them out’ (UN 1, 4; cf. 10.12, 16, 22ff, 25, 28, 44, 69). Thus, ‘the system had some understanding of its own motives’ and could ‘remember what it did, not how the request was worded’ (UN 13).

10.11 The system created ‘a simulated robot with an arm and eye’ (UN x, 2, 117f). The ‘robot’, whose name was ‘SHRDLU’,[2] could ‘manipulate objects’, namely ‘blocks’ of various sizes and shapes, ‘on a table': a limited task, but still requiring substantive knowledge about ‘size, shape, colour, and location’ in ‘three dimensions’, and about ‘spatial relations between objects’, such as ‘support’ (UN x, 2, 118f). ‘The events In the world are actions taken by the robot’ ‘displayed’ ‘on a video screen’ ‘in real time so that the human can get visual feedback’; but they could ‘theoretically be sent directly to a physical robot system’ (UN 121, 6f, LC 391, UC 109). ‘The robot grasps by moving its hand directly over the centre of an object’ and ‘turning on a magnet’ the ‘object’ is then ‘moved along with the hand’ and ‘ungrasped’ when ‘there is something supporting’ it (UN 121).

10.12 The robot's chief talent was to carry on ‘an interactive English dialogue’ with a person, ‘accepting information’, ‘answering questions’, and ‘executing commands’ in the world of ‘blocks’ (UN 1, x). ‘Where running compiled, the system was fast enough to carry on a real time discourse’; ‘each sentence took from 5 to 20 seconds to analyse and, respond to, and the display’ could ‘move at the speed of a real arm’ (UN 7f). SHRDLU's ‘knowledge about the properties of particular physical world’ and its ways of ‘achieving goals in this world’ were ‘designed less as a realistic simulation of a robot than to give the system a world to talk about’ (UN 6). SHRDLU's ‘memory of motion events’ was used ‘to reconstruct the scene’, and ‘the data base’ was made ‘current’ ‘whenever an object was moved’, here again for purposes of discourse (UN 125, 119f).

10.13 Just as for Halliday ‘the grammar is the central processing unit of a language’, ‘the main coordinator of the language understanding process’ in Winograd's system was ‘the Grammar’ (IF xxxiv, UN 5) (cf. 9.23, 32, 37). ‘The grammar’ ‘was based on’ and ‘adapted’ from early versions (1967-68, 1970) of Halliday's ‘theory of systemic grammar’, ‘which emphasizes the limited and highly structured sets of choices made in producing a syntactic structure and abstracting the features that are important for conveying meaning’ (UN 3, 16f, 47 62) (cf. 10.13, 22, 66f; 9.3, 19, 21, 81). ‘A system’ is ‘a set of mutually exclusive features’ ‘among which we will choose one’ when ‘its entry condition’ is ‘satisfied’ -- in ‘the simplest case (and most common)’, the presence of a single other feature’ (UN 19) (9.19).'These choices are represented by ‘syntactic features attached to all levels’ according to ‘a set of system networks explicitly describing their logical independence’ and ‘interaction’, instead of ‘a “deep structure” tree’ (UN 3, 16) (cf. 10.19f, 23, 26, 32).[3] This ‘systemic analyses’ enables ‘a coherent outline of the way English is structured, rather than concentrating on describing particular linguistic phenomena in detail’ or striving to be ‘definitive’ (UN 3f). ‘The exact way the choices are “realized” in the final form is a necessary but secondary part of the theory’ (UN 16).

10.14 ‘The Grammar’ was ‘written to handle’ the ‘three basic ranks of units, the clause, the group, and the word’ (UN 5, 17) (cf. 9.33ff- 13.29). ‘In this analysis, the, word is the basic building block’ and ‘is not chopped into hypothetical bits’, but is rather viewed as ‘exhibiting features’ (like “infinitive”’) (cf. 13.30). A further operational ‘distinction’ is made, though not a ‘sharp’ one’, between ‘function words’, those, defined’, in terms of their function in sentence structure’, and ‘content words’, those which do not ‘presume detailed knowledge of the syntax of the language’ (UN 30). ‘The definition of each word’ is a ‘complex program describing its different uses’ and designed ‘to be run at the appropriate time in the semantic analysis’. Having ‘the procedures actually run in an integrated language understanding system’ offers ‘a strict test for the representation of the meaning of a word’ (UN 30f).

10.15 ‘Standard functions’ are used for ‘simple cases’, and ‘special operations’ for ‘complex cases’, e.g. ‘a heuristic program for understanding back-references’ (UN 30) (cf. 10.23, 78). Thus, ‘by making the formalism for specifying grammars a programming language, we enable the grammar to use special tools to handle complex constructions and irregular forms’ (UN 22) (cf. 13.40). ‘We can set up programs to define’ ‘idioms’ or ‘words like "and" and "or"‘; with these, ‘processing can be interrupted at any point and other computations (either semantic or syntactic) can be performed before going on’ (UN 21) (cf. 10.22f).

10.16 The ‘groups’ were divided’ into ‘noun group, verb group, preposition group, and adjective group’ (UN 17) (cf. 9.57, 75-81). ‘Each group’ ‘has a particular function in conveying meaning: noun groups describe objects; verb groups carry complex messages about the time and modal (logical) status of an event or relationship; preposition groups describe simple relationships; and adjective groups convey other kinds of relationships and descriptions’ (UN 18, 29). In this respect, ‘English syntax’ is ‘good at conveying’ ‘basic elements’ of a ‘person's "model of the world"’ (UN 28) (cf. 13.24). ‘Each group’ also has ‘slots for the words’, e.g., for ‘determiner’, ‘number, adjective, classifier, and noun’ in ‘a noun group’ (9.77f).

10.17 ‘Finally, the top rank’ and ‘primary unit of discourse’ is the clause’, which ‘is the most complex and diverse unit of the language’ (UN 18, 47) (9.44). It can appear as, "Question", "Declarative", or "Imperative", as "passive" or "active", and can be used to express relationships and events involving time, place, manner, and many other aspects of meaning’ as well as the ‘focus of attention and emotion’. Also, ‘its structure indicates’, what the speaker wants to emphasize’ or, question’, and signals ‘the purpose of a utterance’ (UN 47, 49). ‘The sentence’, in contrast, ‘is more a unit of discourse and semantics than a separate syntactic structure’ or ‘unit’, and can be dealt with in terms of its ‘conjoining’ ‘c1auses’ (UN 18, 47; cf. 9.82; 13.54). Due to, the ‘phenomenon called rankshift’ -- ‘one of the basic principles of systemic grammar’ -- a ‘sentence’ may not have ‘a simple three layer structure’ in which ‘clauses are made up of groups, which are in turn made up of words’, but one in which ‘a group’ ‘contains other groups’, or a ‘clause’ forms ‘part of other clauses’ or of ‘groups’ (UN 18, 50, 52) (cf. 13.54).[4]

10.18 ‘Each unit has associated with it a set’ of ‘features, which are of primary significance in conveying meaning’ and are ‘related by a definite logical structure’ (UN 19) (cf. 10.14f, 23, 26, 32). For example, ‘word classes can be divided into subclasses by the features assigned to individual words’, according to ‘arbitrary decisions as to whether a distinction between groups of words should be represented by different classes or different features of the same class’ (UN 66) (cf. 13.27). ‘In our dictionary, we simply list all syntactic features the word has for all of the classes to which It can belong’, ‘the verb having the most complex network of features of any word’ (UN 66, 70). ‘Each group can also exhibit features just as a word can’, e.g. ‘"singular"‘ or "definite''‘ for a ‘noun group’, or "negative''‘ for a ‘verb group’ (UN 18).[5]

10.19 The ‘functions a syntactic "unit" can have’, such as ‘Subject and Object’ in ‘a transitive clause’, can be approached by considering ‘which features of a syntactic structure are important to conveying meaning and which are just a by-product of the symbol manipulations needed to produce the right word order’ (UN 20f). ‘In most current theories, these features and functions are implicit in the syntactic rules’ (or "'deep structures"‘), and ‘there is no attempt to distinguish significant features from the many other features we could note about a sentence and which are also implied by the rules’ (cf. 10.67). ‘Systemic grammar’, however, uses ‘realization rules’ to ‘relate the set of features to the actual surface structure’, thereby handling what ‘would be done by transformations in transformational grammar’ (cf. 9.10, 9[3], 9[20]; 10.64). For ‘recognition rather than generation’, ‘interpretation rules’ act as ‘the Inverse of realization rules’, ‘looking at a pattern, identifying Its structure, and recognizing its relevant features’ (UN 21f) (cf. 13.57).

10.20 Obviously, Winograd was not following the ‘recent linguistic theories’ that ‘consider syntax as a proper study devoid of semantics’ (UN 16) (cf. 4.15; 5.61f 7.56; but see 10.52ff). There, ‘language is viewed as a way of organizing strings of abstract symbols, and competence is explained’ with ‘symbo1-manipu1ating rules’ (10.40, 43, 77). Such a method may ‘describe in great detail’ how ‘sentences are put together’, but can provide only ‘the most rudimentary and unsatisfactory accounts of semantics’. Winograd's SHRDLU system’, in contrast, thematically favours ‘the interaction between syntax and semantics’ (UN 68f; cf. UN 3, 5f, 21f 29, 31f, 44, 47, 69, 73, 76, 89, 93, 101f, 127, 131f). The flexibility of writing a grammar as a program’ makes it easier both to handle the comp1exties of English syntax and to combine the semantic analysis of language with the syntactic analysis in an intimate way’ (UN 89). The ‘ability to integrate semantics with syntax is particularly important in handling discourse, where the interpretation of a sentence’ ‘may depend in complex ways on the preceding discourse and knowledge of subject matter’ (UN 22) (cf. 5.57- 9.16- 11.86, 91).

10.21 Therefore, ‘the language process is not segmented into the operation of a parser followed by the operation of a semantic interpreter; rather, the process is unified, with the results of semantic interpretation being used to guide the parsing’ (UN 22f) (cf. 7.49; 11.3, 34, 77). This tactic was justified not merely on theoretical, but on operationa1 grounds, as a matter of efficiency. Without ‘semantic’ guidance, a purely syntactic parser expends time and resources on possible structurings allowed by grammar but easily recognized as yielding ‘contradictory’ or ‘meaning1ess interpretations’ (UN 127, 131). The ‘dire danger of combinatorial explosion’ [6] can thus be met by ‘doing the interpretation continuously’ ‘for each phrase’ and ‘immediate1y 1ooking into our memory to see which interpretation is meaningful In the current context of discourse’ (UN 32) (cf. 10.65. 11.79).

10.22 ‘Since the parser users systemic grammar, the semantic programs can look directly for syntactic features such as ‘'Passive"‘ ‘to make decisions about the meaning’ (UN 29). These programs being able to ‘work separately, there is no need to wait for a complete parsing before beginning semantic ana1ysis’. Besides, ‘any semantic program has full power to use the deductive system’, and can ‘ca11 the grammar to do a special bit of parsing before going on’. ‘The semantic ana1ysis’ is thus both "bottom-up"‘, in that ‘each structure is analysed as it is parsed; and ‘"top-down"‘ in that the ‘specia1ist programs’ allow it to anticipate what ‘larger structure’ is to be ‘ana1ysed’ (UN 29f, 22) (cf. 10.59, 62 10 [11], 10[15]; 11.13, 19, 25, 32, 55, 59, 73, 77f, 95, 97- 13.32).

10.23 Furthermore, we can ‘use the overall context in determining the plausibility of a particular interpretation’ (UN 32). Winograd proposes that ‘semantic theory can account for three different types of context’. ‘Loca1 discourse context’ ‘covers the discourse immediately preceding and Is important to semantic mechanisms like pronoun reference’ (e.g. for "'How many of them were there then?''‘). Overall discourse context’ covers ‘general subject matter’ (e.g. for "The group didn't have an identity''‘ in ‘discussing mathematics or sociology’). Finally ‘know1edge about the world’ ‘affects our understanding of language’ (e.g. for "The city councilmen refused the demonstrators a permit because they ‘''advocated revolution"‘) (UN 33) (cf. 11.20). In Winograd's system, most of this discourse knowledge is called on by semantic specialists and by particular words such as "one", "it", "then", "there", etc. (UN 33) (cf. 10.29, 31). These ‘specialists are programs’ designed to ‘work separately’, to be ‘expert In looking at particular’ data’, and to ‘create parts of a complete description of the meaning of a sentence by building complex list structures’ (UN 33, 29, 126).

10.24 So broad a view of semantic theory reveals ‘the importance of a comprehensive representation of meaning’ (UN 3). Winograd envisions ‘three different levels’ at which ‘semantic theory must describe relationships’ (UN 28). The ‘first’ level should ‘define the meaning of words’, in the ‘limited sense’ of a ‘formal description attached to a word which allows it to integrated into the system’; the ‘forma1ism’ should ‘allow users to add to vocabulary in a simple and natural way’, and should ‘hand1e quirks and idiosyncrasies of meaning, not just well-behaved standard words’ (UN 28) (cf. 10.15; 13.40). The second ‘level’ should ‘relate meanings of groups of words in syntactic structures’. The third level should ‘describe how the meaning of a sentence depends on context': both on its ‘linguistic setting’ and on its ‘rea1-wor1d setting’.

10.25 ‘A system of semantic features’ [7] is a further means to control the ‘semantic interpretations of any phrase or sentence’ (UN 3, 6, 32, 76, 127, 131) (cf. 10.14f, 23, 32; 13.30). ‘A network’ of ‘features’ can be ‘kept on property lists and used for an initial phase of semantic analysis’ (UN 6). ‘The features subdivide the wor1d of objects and actions into simple categories’ and ‘rough classes such as "animate", " inanimate", " physical", "abstract", etc. (UN 6, 31) (cf. 7.69; 10.18f, 22, 31). This approach is claimed to be a ‘complete representation of meaning’ or a full-blown, semantic theory’, as Katz and Fodor (1964) imagined (UN 31f) (cf. 5.76. 7.67, 77). ‘There is no self-contained set of "primitives" from which everything else can be defined’ (cf.11.24; 13.59). Yet for ‘useful purposes’ within a ‘model of the world’, we can ‘consider some concepts as ‘"atomic"‘, with a ‘meaning’ that is not a ‘combination of other more basic concepts’ (UN 26f). What counts as ‘atomic’ is decided not by some ‘logical status’ or ‘sharp dividing line’, but by ‘distinctions’ that fit ‘the needs of the particular language communication’ and by the consensus underlying the ‘close similarity between the models held by speaker and listener’.

10.26 Although ‘there has never been a clear definition of what the field of "semantics" should cover’, attempting to program computers to understand natural language has clarified what a semantic theory has to do’ (UN 28). ‘A notation for representing certain kinds of meaning’ can be devised with ‘a pragmatic approach’ rather than ‘a deep philosophical’ one (UN 25f) (cf. 11.3, 22, 40; 13.51). ‘Language ‘is inextricably enmeshed in the knowledge’ ‘people have about the world': ‘not a neat collection of definitions and axioms, complete, concise, and consistent’. ‘Rather, it is a collection of concepts designed to manipulate ideas’, and is liable to be, incomplete, highly redundant, and often inconsistent’ (UN 26f) (cf. 5.86; 9.110; 10.5, 8, 39, 71; 11.24). ‘Definitions are circular, with each concept depending on the other concepts’. ‘The meaning of a concept’ might indeed ‘depend on entire knowledge of the speaker, not just the kind’ of knowledge ‘in a traditional dictionary’ (cf. 4.14. 5.28; 10.90; 11.20).

10.27 The ‘primary goals’ ‘of a semantic theory’ should not be "'the number and content of the readings of a sentence"‘, or potential "anomalies''‘ and ‘"paraphrase relations"‘ (UN 33) (cf. 7.61). All these are only ‘by-products of the analysis made possible by a more complete semantic theory’ (UN 33f) and can be handled in an operational way. An ‘anomaly’ occurs if ‘the system produces no possible interpretations’; ‘two sentences are paraphrases if they produce the same representation to the internal formalism’; and so on. Such instances too ‘depend on the entire range of in the internal formalism, ways in which language communicates meaning’ in ‘the interaction of context with understanding’, and not just ‘on a restricted subset such as the logical relations of markers’ (cf. 7.63; 10.62). .

10.28 To ‘gain flexibility and power’, ‘knowledge can be represented in the form of procedures written in special languages’, ‘rather than tables of rules or lists of patterns’ (UN If) (cf. 10.9). Then ‘each piece of knowledge’ ‘can call directly on any other’ ‘in the system’ (cf. 10.22). And ‘each definition is a program’ able to ‘examine the description’ and ‘produce an appropriate meaning relative to the object being described’ (UN 129, 143). So in Winograd's system, ‘both the semantic knowledge and the definitions of individual words were in the form of programs’ which ‘interpret’ ‘words’ in terms of ‘definitions’ in ‘the dictionary’ (UN 3, 6) (10.14).

10.29 In addition, we can ‘allow each predicate’ ‘to have associated with it a program which knows how to evaluate its "priority" in any given environment’ (UC 130). We might select ‘a single number’ or ‘a complex heuristic program which takes into account the current state of the world and the discourse’. Also, we can ‘include mechanisms for carrying along with each semantic structure an accumulated plausibility rating’. ‘As a semantic structure is built, it takes on the sum of the plausibilities of its components’ (UN 150). For a ‘pronoun’ like "'it"‘, say, ‘a special heuristic program’ can ‘look into the discourse for all the different things it might refer to and assign a plausibility value to each, according to’ ‘its position in syntactic structure’, ‘the form of its determiner’, and so on (UN 158, 160) (cf. 10.23, 31). Where appropriate, the ‘decision’ about ‘which is better’ can wait until ‘the end’ of ‘the sentence’, or, as a ‘last resort’, the system can ‘ask for clarification’ (UN 158; cf. UN 150).

10.30 The ‘novelty’ of Winograd's method was to ‘approach semantics’ ‘as a practical problem of relating words and syntactic structures to a comprehensive logical formalism within a specific problem-solving system’ (UN 5). The, semantics’ was built for a ‘detailed analysis of linguistic structures to extract an expression of their meaning’ CUN 3). This design fits Winograd's view -- one he later repudiated -- of ‘the process of understanding language as a conversion from a string of sounds or letters to an Internal representation of meaning’ (UN 23) (cf. 10.79). ‘To do this’, the ‘system must have some formal way to express its knowledge’ and to, represent the "meaning" of a sentence In this formalism’, e.g. in terms of "objects'‘, " relations", and "properties"‘ (UN 23f, 27).[8] ‘The formalism must be structured so the system can use Its knowledge In conjunction with a problem-solving system’ to ‘make deductions, accept new information, answer questions, and interpret commands’ (UN 24). (10.4, 8, 12).

10.31 As we have seen, Winograd's system foregrounds the process of ‘understanding’, as the book's title promised. The ‘production’ of ‘discourse’ in contrast, received only 11mlted treatment under, four aspects’ (UN 163). The system made, patterned responses’, either ‘fixed’ (like "0K''‘ or "'I understand"‘) for specific situations’, or ‘more complex’ for ‘"filling in the blank" with a phrase’ ‘from the input’ (like "'Sorry, I don’ t know the word ____"‘ or "'I'm not sure what you mean by _____"‘) (UN 163f). The latter kind of response, which may ‘involve manipulating the determiners of the phrase’, 1s useful when ‘the system cannot figure out what is referred to’ or tries to ‘handle ambiguity’ (UN 163f). ‘Answering questions’ was done according to the ‘types of response people expect’, with ‘no attempt’ ‘at full sentences, since people rarely answer questions’ that way (e.g., ‘"which block is in the box?" -- "the red one"‘). ‘Naming objects and events’ was done by consulting the ‘features’ like ‘colour and size’ of ‘objects’ (e.g. ‘"a large green cube"‘), or noting which ‘object’ was ‘involved’ in which ‘event’ (e.g. "'put a large red cube on the table"‘) (UN 166ff).[9] Finally, ‘fluent discourse’ was sought with the aid of ‘three’ ‘devices’ : ‘combining identical descriptions’ to ‘avoid redundancy’ (e.g. to get "'three small red cubes"‘); ‘using substitute nouns’ ( e.g. ‘"a large one"‘); and ‘using "it" and "that"‘ when ‘referring to the same object more than once’ (UN 163, 168f) (cf. 10.23, 29). These various tactics help to prevent ‘awkward and stilted responses’ that might be ‘at times incomprehensib1e’ (UN 168). Some responses may seem ‘verbose’ in providing ‘extra information’, but this too ‘usually gives a natural sounding answer’ and an ‘intelligent’ impression of ‘telling the questioner information he would be interested In knowing, even when he doesn't ask for it explicitly’ (UN 140).

10.32 My summary should give some idea of Winograd's ‘SHRDLU’ system. As is typical of AI work, the main thrust was to get the system running on a computer well enough so that It could carry out reasonably human-like dialogues without either making ridiculous mistakes or breaking down, i.e., stopping, exploding, or going into endless loops. A natural language system that runs successfully will seem ‘intelligent’, albeit by ‘artificial’ means, thanks to the ‘intimate connection between ‘intelligence and language’ (UC 107) (cf. 10.43, 75, 104). Admittedly, the system embodied a limited set of choices for a very simple ‘world’, and its impressive success depended on extensive restrictions on its design (cf. 10.73, 88).

10.33 Winograd's next book adopted a much wider scope by covering a whole range of possible designs and systematically considering them in relation to ‘the major directions in linguistics’ (LC 3). Although ‘the study of linguistics may be as old as language itself, and current linguistic science can trace its origins back at least as far as the Sanskrit grammarians’, Winograd picks out some ‘major turning points at which the focus of study changed and linguists felt they had finally arrived at the "real" Issues of language’ (LC 6). If we follow ‘the philosophy of science’ rather than the ‘popular view’, we see that the ‘history’ of a ‘science’ does not manifest a ‘linear progress’ of ‘theories’ ‘getting better and better, explaining more phenomena, making more accurate predictions, and becoming more elegant’. Instead, ‘the scientist is faced with a complex interconnected world’ and has to ‘select the questions to be asked’ and ‘determine what kinds of answers will be considered acceptable’ (LC 6f) (cf. 13.1).

10.34, Periods of normal science’ show ‘widespread agreement’, and ‘the foundations’ and ‘basic assumptions’ ‘are taken for granted’ (LC 7). ‘Science in this state’ is ‘operating within a paradigm’, i.e., a ‘social structure’ and a ‘conceptual framework’ ‘of methods and biases about what deserves study and how it can be described’ (LC 7, UN ix, UC 24). ‘Progress’ consists in ‘the details of the theories’ being ‘worked out’, while ‘problems that cannot be explained within the current paradigm’ are ‘ignored, or excuses are found’ (LC 7). ‘But gradually, as the current phenomena become overstudied, more scientists move toward the difficult areas, and the shortcomings of the whole framework’ become ‘apparent’. ‘Finally’, ‘a radically different paradigm’ is proposed to supplant ‘the current standard’, and a ‘heated debate’ ensues. Though most attempts to establish new paradigms’ are ‘rejected’, ‘a few’ ‘revolutions’ are ‘successful’, because the ‘difficulties of the old theories are eliminated’ and ‘areas previously unexplored are now opened’. In return, ‘old issues appear to be less relevant’ and are ‘dropped from consideration’. ‘The new paradigm becomes normal science’, enters, positions of academic power’, ‘gets formalized into textbooks’, and persists until ‘the cycle is repeated’. Because ‘practitioners’ are ‘rarely’ ‘converted’, ‘the cycle’ typically requires ‘enough time for younger scientists to replace the old establishment’ (10.103).

10.35 Thomas 5. Kuhn (1970) ‘used the so-called "hard sciences" as his examples’, e.g., ‘astronomy, chemistry, and physics’, but his ‘concept of scientific revolution applies even more convincingly in the "human sciences", such as psychology, linguistics, and sociology, where the revolutions are more frequent and radical in throwing away all that came before, and where there are few technological applications’ as a ‘measure of progress’ (LC 8 cf. UN ix; 10.106; 13.4). There, ‘a scientist looking for a new paradigm is strongly affected by the other sciences currently enjoying successful deve1opment’ (LC 8). ‘Either consciously or unconsciously, these are viewed as a model’, leading to ‘a metaphorical imposition of their ideas’. ‘Linguistics’ ‘has been especially open’ to ‘using the hard sciences as bases for analogy’ in this fashion (cf. 13.11).

10.36 Winograd accordingly proffers a ‘survey of linguistic history’ as ‘a series of metaphors’ (LC 8). The oldest ‘metaphor’, and still ‘the predominant view’ ‘in our society’, is ‘linguistics as law’ (LC 9). ‘Prescriptive grammar’, though ‘long rejected by scientists studying actual language use’, envisions ‘a set of rules that must be followed’ in order to gain a ‘place in the social structure’ (i.a., cf. 4.5f, 86; 8.4).'The main concern’ Is ‘correctness or purity of the language’, and ‘the linguist is to act’ ‘as judge and policeman’. ‘Current theories of linguistics reject this metaphor’ and refuse to ‘fight’ ‘evolution’ or to ‘force the conventions of one social class onto the rest of society’. Even so, Winograd allows that "'correct grammar"‘ might help ‘people’ ‘to function within the social structure’, and ‘a practical theory of linguistics can provide a basis for language teaching’ (LC 9, 26f). His book says nothing about how this could be done, but it is the only one in my survey to comment upon constructions that have traditionally bothered English teachers and purists: ‘split infinitives’, ‘dangling prepositions’ and "who"/"whom''‘ distinctions’ (LC 249, 481, 245) (cf. 3[51).[101

10.37 ‘In the nineteenth century, a paradigm for linguistics’ was provided by ‘comparative philology’ (LC 9) (cf. 2.5f. 6.5; 8.6, 15, 40; 12.20, 90f). The metaphor here was ‘linguistics as biology’, ‘in the style of natural history’, notably ‘Darwin’s theory of evolution’ (8.6; 12.17). ‘Just as biologists developed taxonomies of organisms’, ‘linguists’ used ‘comparisons of structures’ to ‘classify’ ‘data into a complete phylogenetic tree’ of ‘languages’ (LC 9f) (2.5; 4.73; 12.91). This work illustrated the ‘puzzle-solving activity’ common’ in ‘normal science, with ‘pieces’ being ‘fitted together’ as proof of ‘being on the right track’ and of ‘progress being made’. Gradually, ‘phi1o1ogists’ ‘exhausted the body of well-known languages’ and, turned to more remote languages reported by anthropologists’. But enthusiasm waned as ‘cataloguing’ became ‘tedious’ and ‘few satisfying general principles were discovered’.

10.38 The next ‘revolution’ ushered in ‘structural’ or ‘descriptive 1inguistics’, accompanied by, a shift of focus from the family of languages to the structure of the single language’ (LC 10f). ‘This paradigm’ ‘was, strongly influenced’ by ‘behaviourism’, which ‘dominated American psychology’, insisted on ‘objective scientific experiment’, and abjured all reference to ‘mental processes’. Even though behaviourism arose from animal biology, Winograd diagnoses the dominant metaphor to have been ‘linguistics as chemistry’ (cf. 13.12). ‘The analysis of language data was modelled after a positivist view’ of how to use ‘experimental techniques to rigorously determine underlying structure’, just as ‘a chemist’ ‘determines the set of molecules’ in a ‘complex substance’ and the ‘basic elements’ In ‘those molecules’ (cf. 5.28). ‘The analogy to chemistry is closest in the way sounds are organized into words’ (cf. 7.16). ‘Every language has a mall set’ of ‘phonemes’ amenable to ‘discovery procedures’, even for ‘a language not known to the scientist’ (13.26). But ‘the set of meaning elements or morphemes in a language is much larger and less well-structured’, and ‘the same methods’ were even ‘less satisfying’ ‘in the study of syntax’ (13.27f). Nonetheless, ‘many different languages were described’ within ‘as wide a range as possible’, including ones ‘outside modern Western society’, again with the aid of ‘anthropologists’ (5.2).

10.39 Next came the ‘generative’ paradigm: ‘linguistics as mathematics’ (LC 11) (7.44- 13.15, 17f). ‘Empirical methodology’ was ‘rejected’ and ‘observab1e behaviour’, was neglected in favour of ‘the intuitions of native speakers’, their ‘tacit knowledge’, and ‘the underlying faculty’ to ‘create and understand sentences’ (7.14, 24f). This ‘theory must postulate mental structures and processes’ in absence of ‘techniques for observing what goes on in the mind of the language user’; yet ‘Chomsky argued that, linguistics could study the abstract mental structures without indulging in unprovable speculations’ (LC 12) (cf. 7.10, 15, 27). His notion of ‘competence, an abstract characterization of a speaker's knowledge of a language’, is ‘close1y related to the notion of proof in mathematics’. ‘We can think of mathematics as a "language" of formulas’, ‘symbols’, ‘axioms, and rules of inference’. ‘Mathematics is not the study of how people invent such expressions or what goes on in their minds then they read or try to prove them’. Rather, ‘its goal is to produce a set of rules and formal mechanisms that precisely determine which ones are true. The measure of success for axiomatization (laying down the rules and operations)’ ‘lies in elegance and economy’. Similarly, ‘generative linguistics views language as a mathematical object and builds theories’ ‘very much like sets of axioms and inference rules’. ‘A sentence is grammatical’ if ‘some derivation’ ‘demonstrates that its structure is in accord with the set of rules, much as a proof demonstrates the truth of a mathematical sentence’ without ‘describing how a mathematician sets out to generate’ a ‘proof’.

10.40 The final ‘paradigm’ in Winograd's history is his own, the ‘computational’ one, with its metaphor coming from the ‘computation’ done with ‘stored program digital computers’ (LC 13). ‘The computer shares with the human mind the ability to manipulate symbols and carry out complex processes’, but its ‘workings are completely open to inspection and study’ and to ‘experiments in building programs, knowledge bases’, and ‘precise’ ‘mode1s of mental processing’ (10.5). Therefore, ‘we can try to explain the regularities among linguistic structures as a consequence of the computations underlying them’ (13.31).

10.41 Winograd's history ends with no indication of whether and how far this latest paradigm has displaced the ‘structural’ and ‘generative’ ones, as a Kuhnian viewpoint would suggest it must have done. Instead of the absolute rupture wherein the science ‘throws away all that came before’ (LC 8; 10.36), Winograd prefers to absorb and conserve, thereby making the book more suitable for ‘graduate work’ ‘in linguistics’ (LC vii) -- with the specific objective, I think, of retraining linguists for jobs in the computer industry or at least for joint research with computer scientists. [11] He concentrates not on how the earlier paradigms may have led to crisis and stagnation, but how they might be made computationally feasible. He merely notes, without trying to adjudicate, the many ‘debates’ arising in linguistics over generative transformational grammars (LC 133, 152, 166, 170, 173, 182, 188, 233, 319, 557, 563, 572, 574, 582).

10.42 The book therefore pursues a delicate compromise. The opening chapter promises ‘a book about human language’ ‘as a process of communication’, ‘motivated’ by the ‘questions’ like ‘what knowledge must a person have to speak and understand language?’ and ‘how is the mind organized to make use of this knowledge in communicating?’ (LC 1). But most of the book sidesteps these questions by simply collapsing the distinction between natural and formal languages and presenting, on fairly even-handed terms, a wide array of ‘grammars’ from linguistics, mathematics, algebra, automaton theory, and computer programming. A substantial level of formality and abstraction is maintained, and commentary and criticism are couched more in computational terms than in linguistic ones. Winograd serenely hopes that ‘in long run the technical material presented’ ‘will retain its usefulness’ by ‘fitting in new ways into our interpretation of language as a human phenomenon’ (LC ix) (cf. 11.5).

10.43 A deft though tricky manoeuvre is to draw the borderlines in just such a way as to include computation and standard linguistics and to exclude other fields of concern, including some raised before in UN. Winograd now sees a ‘high degree of commonality between the "generative"‘ and ‘"computational" paradigms’, and suggests ‘they may be seen as two variants of a single "cognitive paradigm", (LC 20). He admits the computational side disagrees when it maintains that ‘the structure of language is to be derived from the structure of processes’ (versus the division of ‘competence’ from ‘performance’)[12] and that ‘the knowledge structures and processes for dealing with language are to a large degree shared with other aspects of intelligence’ (versus the ‘distinct language faculty’) (LC 21, 151, 186f) (cf. 7.12, 26). But he says these ‘differences’ are contained ‘within an ‘overall area of agreement': that ‘the proper domain of study is the structure of the knowledge possessed by an individual’, and that ‘this knowledge can be understood as formal rules concerning structures of symbols’ (LC 20, 273).

10.44 Winograd does seem uncomfortable about the term ‘cognitive’, using it rarely in the book and mainly in connection with the aspirations of generative linguistics (LC 133, 149, 164, 175, 177, 186), plus the approach that simply seized the term by calling itself ‘cognitive grammar’ (cf. Lakoff and Thompson 1975) (LC 252, 311, 578). ‘Cognitive psychology’ gets just one mention, in regard to a trite finding of list-learning research (primacy and recency effects) (LC 505). Even in the 1986 volume, ‘cognitive psychology’ is anachronistically charged with admitting only ‘well-controlled stimuli’ and ‘patterns of recurrence’ in ‘experiments with rats in mazes, nonsense syllable memorization’ and the matching of geometrical figures’, and with neglecting ‘models of memory, attention and inference’ (UC 114, 25). That Winograd was simply unaware of research like that surveyed in Ch. 11 is hard to Imagine. More likely, he saw a potential threat first to the generativist outlook preserved here In LC, then to the phenomenological one adopted in UC.

10.45 The ‘principles’ Winograd enunciates for the ‘cognitive paradigm’ again accord with generative linguists in rejecting the ‘primary’ status of ‘social interaction’ along with ‘the text itself as the central focus’ (LC 20) (cf. 9.3; 10.81; 13.20). This move marginalizes both the Hallidayan ‘systemic grammar’ used earlier in UC and the ‘phenomenology’ advocated in UC (cf. LC 21, 273, 278f), as if his own most characteristic work were outside the ‘cognitive paradigm’. When expounding its ‘relevance to the study of language as a cognitive process’, Winograd does grant ‘systemic grammar’ a ‘deep cognitive significance’ and contemplates ‘pushing’ it ‘into the cognitive or generative paradigm’ by ‘moving to a mechanically applicable rule system’ (LC 280), but no such step is taken. The prospects are blurred even more when he decides that ‘the phrases "computational" and "cognitive processing" will be used interchangeably’, the one for ‘computational details’ and the other for ‘the general approach to modelling human language’ (LC 22). Surely this correspondence is precisely what needs to be established (cf. 7.78; 13.45).

10.46 Incongruously, the ‘basic model of communicative processing’ in the opening chapter features ‘communicative goals’, ‘effects to be achieved, information to be conveyed, attitudes to be expressed’, and ‘actions’ and ‘reactions’ to be ‘caused’ ‘on the part of the ‘comprehender’ (LC 13) (cf. 10.97f). ‘The producer must map this multi-dimensional collection of goals onto a sequence of sounds’ or ‘marks’ (LC 14). Aside from ‘tone and vocal gesture the message is forced into a linear form’; ‘a variety of mechanisms’ ‘merge multiple messages into a single structure that enables the comprehender to perform the reverse process, inferring the original goals and messages from the information received’ (cf. 7.83; 10.18, 47, 65; 11.15, 81; 12.47, 13.57). Hence, ‘language provides’ ‘information resources that can be manipulated by the producer: the choice of words, the structure of phrases, and the patterns of emphasis producer and intonation’ (cf. 9.34, 37, 12.43, 72; 13.24). Of course, ‘communication’ is, reflexive’, because ‘the design of an utterance depends critically on the producer's expectations’ about ‘the comprehender's’ ‘knowledge of the language’, ‘the world’, and ‘the situation’ (LC 14f). All these precepts resemble those of UN far more than those of the rest of LC (cf. 10.9f, 12, 23ff 26, 29).

10.47 ‘Most work’ in ‘linguistics’ is ‘on spoken language, since it is more fundamental than written language’, and Winograd's ‘basic model’ follows suit, even though ‘most computational models deal only with typewritten character sequences’ (LC 15, 13; cf. 10.50). The ‘model of processing’ offered as ‘a first approximation’ ‘when an utterance is heard’ foresees ‘sound patterns’ being ‘created’ and ‘then analysed to produce syntactic structures, which are in turn used to form appropriate representations’ (LC 15f). ‘The production of an utterance’ ‘goes in the opposite direction -- from representation to syntax to sound’ (LC 16). ‘This general model corresponds’ to that of ‘analytic philosophers of language drawing on ideas dating back to Aristotle and beyond’. However, the model, is, distinguished’ by, its focus on the description of processes that explicitly manipulate formal structures’ -- a ‘central’ ‘idea’ for ‘a11 areas of computer science’ and thus also for ‘the computationa1 paradigm’ of ‘linguistic description’. ‘Artificial intelligence’ in particular makes the ‘basic assumption’ that the ‘representations operating in mental processes’ ‘can be formally described as data structures like those of a computer’ (LC 15) (cf. 10.4, 24, 30, 69. 73, 76-79).

10.48 In another manoeuvre congenial to linguists, the ‘traditional’ way of ‘representing’ ‘linguistic structures’ in ‘forma1isms corresponding to different levels of structure, such as sounds, words, and phrases’, is said to reflect ‘a natural series of levels common to all human languages’ (LC 16) (cf. 4.71; 5.34f. 7.45; 8.51f- 9.30; 13.29). Equally welcome is the idea that ‘the organization at each level is to some extent independent of how it relates to the levels around it’ (cf. 10.51; 13.27). ‘This stratification is a natural organization for any complex mapping process’ ‘into stages, each of which is simple’ (cf. 9.30- 13.57). ‘Computational models are also based on the, idea that a complex process can be decomposed into a collection of simpler processes, each operating to some extent independently’. ‘The advantage of modularity’ is that ‘the system is flexible and expandable’, and ‘the effects’ of ‘changes’ ‘can be localized’ (cf. 10.9). ‘Integration of the levels’ remains feasible If we have ‘a uniform definition for the structures that the components accepted as inputs and produced as outputs’ (but cf. 10.8f). Winograd thus obtains ‘a simple stratified model of language comprehension’ (with ‘production’ ‘operating in the opposite direction’), in which ‘the knowledge of language is made up of rules for manipulating different levels of structure’ (LC 16f) (Fig. 10.1). The model looks like the usual linguistic level scheme, except for the added component of ‘reasoning’ running on ‘deductive’ and ‘lnferentia1 rules’.

-- INSERT FIGURE 10.1 HERE --

10.49 Winograd concedes ‘this model is wrong in its suggestion that levels and processes can be separated’, but ‘the simplification’ ‘has served as a basis for the design of computer systems’ and ‘psycho1og1ca1 models’ (LC 17). Also, the lines and arrows In the model suggest that each level interacts only with adjacent ones by ‘operating with the results of the one above It and producing structures for the one below it’, which does not always work, e.g., for ‘stock phrases or idioms’ ‘related directly to meaning’ rather than ‘ana1yzed’ into ‘syntactic structures’ (LC 17, 19) (cf. 2.61; 4.60; 5.32, 54; 7[34]; 9.93; 13.28). Indeed, any model ‘depicting the understanding process’ as a ‘strict sequence’ of ‘stages’ is ‘wrong’ (LC 19). ‘Many experiments have demonstrated that in listening to speech we often use knowledge of the expected meaning’ to ‘analyse the sounds’ and ‘decipher the words’ (LC 19f) (11.77; 13.57). Moreover, ‘a system organized in sequential stages is "brittle": ‘a problem at the beginning’ can make ‘the whole process break down’. So we should ‘increase the flexibility’ by ‘separating the processing sequence from the structural levels’. ‘We can think of the assigned structures all being written on a blackboard and the component processes’ (or ‘know1edge sources) reading’ from it or writing’ on it; ‘a process’ is ‘not limited to reading only what was written by its upper neighbour’, but can use any, information’ ‘avai1ab1e’.

10.50 Yet ‘since the simpler stratified model has been the most developed, much of the material presented in the book’ is ‘based on it’ (LC 20). ‘In particular, most of the material on syntax will assume’ ‘a separate component called the parser, which operates on words that have already been recognized and found in a dictionary, and which produces syntactic structures for use in a separate phase of semantic analysis’ (cf.10.10, 21, 47, 11.33f, 77; 13.32). The ‘advantages and problems of increasing intercommunication’ ‘in a non-stratified process model’ are viewed as a topic in ‘current research’, but aired only briefly for ‘stratificational’ and ‘systemic grammar’ (LC 20, 299-303). Otherwise, most ‘interactions’ are either ‘left ‘implicit in the rules’ or seen as potential ‘comp1exities’ and ‘difficult to control’ (LC 151f, 305, 310, 389).

10.51 Readers of Winograd's dissertation may be surprised to find him now calmly ‘adopting the classical linguistic method of studying syntax independently’ (LC vii) (cf.10.20). This manoeuvre is justified by pointing to the ‘broad self-contained literature on syntactic theory and technique that needs to be mastered before the Interaction between syntax and meaning can be understood’ (LC viii). Winograd professes no ‘theoretical commitment’ to ‘the autonomy of syntax hypothesis’, which postulates, a relatively independent body of phenomena that can be characterized by syntactic rules without considering other aspects of language or thought’ (LC vii, 21, 151) (7.57- 9.2). ,

To interpret the ‘hypothesis’ ‘as a belief that meaning is unimportant in the study of language’ is dismissed as a ‘mistaken caricature’ (LC 151), though Chomsky's own gibe that meaning is no more relevant than ‘the hair color of speakers’ (SS 93; 7.56) hardly leaves room for further caricature. For Winograd, at any rate, ‘the essential claim is that an analysis of the structure of language’ ‘will be best achieved by finding the structure of each component separately and then understanding their interactions (LC 151). ‘A complex system’ can be ‘treated’ as ‘nearly decomposable’ by assuming that ‘the interactions among components are much less crucial than the independent functioning within each one’ (LC 152). ‘Having thus dismissed (for the time being) questions of linguistic processing’, ‘sound, and meaning, the linguist can ask "what is the nature of the syntactic structure of a language?"‘

10.52 No doubt these moves are steered by the pressures of composing a separate volume on ‘syntax’. In practice, however, the volume continually refers to ‘correspondences’ between the ‘syntactic’ (or ‘grammar’, ‘parser’, etc.) and the ‘semantic’ (or ‘meaning’), as well as to systems that utilize them.[13] Some systems try to go ‘directly’ to the semantics without any thorough syntactic analysis, such as ‘semantic grammar’ (Wilks 1973) (LC 51, 260, 363, 374) or ‘conceptual dependency’ (Schank et al. 1975) (LC 74, 318, 367, 402, 405) (cf. 10.1; 11.4, 34; 13.53). Also, several post-transformational approaches to syntax and grammar have markedly increased the role of semantics, notably ‘case grammar’, ‘generative semantics’, and ‘Montague grammar’ (LC 180, 185, 257, 311-26, 348, 530, 561, 560-64, 575, 582; MSF 92, 95). Even ‘transformational theory’, long driven by a ‘subliminal’ ‘desire to reflect meaning’, now reveals a clear ‘trend to shift work from the syntactic’ ‘to the semantic component’; and Winograd ends his overview of ‘directions in transformational grammar’ with the prediction that ‘the theme for the coming years will be the exploration of semantic formalisms and their integration into the grammar’ (LC 180, 299, 494, 526, 577f, 581).

10.53 The interaction between semantics and syntax would doubtless have been more thoroughly treated in the ‘volume on meaning’ so frequently promised in LC (LC viii, 21, 25, 37, 88, 151, 282, 318, 326, 359, 361, 396, 490, 514, 581). Winograd planned to cover the following: relations among ‘artificial intelligence’, ‘philosophy of language’, and ‘pragmatics’; ‘issues of representation, meaning, and language use’, including ‘quantification and reference’, ‘the context of action and knowledge’; ‘the, organization’ of ‘discourse’; ‘the use of a data base in semantic analysis’; ‘the use of frames in representation languages’; the ‘overlap’ of ‘transitivity’ ‘with semantic problems’; reasoning mechanisms’ and, integrated question-answering systems’; and ‘computer aided instruction’ (UN viii, 318, 151, 361, 514, 282, 21, 396, 326, 490, 37, 359).

10.54 In addition to the postponement of semantics, other ‘1imitations of the approach’ are acknowledged, which ‘to a large degree’ are ‘common to the computational and generative paradigms, following from the basic cognitive orientation’ and the view of ‘language as a process going on in the mind of an individual’ (LC 28). ‘We lose sight of the social dynamics of language use’ and the, social interaction’, from which’ ‘language’ ‘takes its meaning’, as well as the ways ‘linguistic devices’ can ‘establish personal power relationships’ and ‘social distinctions of rank and status’. We cannot tell ‘why a particular dialect is adopted’ or ‘how dialect differences play a role’ in ‘group identity and cohesiveness’ (LC 29; cf. LC 181). We discount ‘the central problem’ of ‘language acquisition’ without which there could be no ‘body of language knowledge’ (LC 19). We do not treat the ‘evocative aspects of language’ in ‘the tradition of a culture’, such as ‘the emotional dimensions’ of ‘literature’ (LC 29). We pass over the ‘historical aspects of language’; ‘though all change takes place as a result of language acts by individuals, the relevant patterns are not visible at that level’ (cf. 2.45; 3.57; 4.81, 4[6]). And finally, we do not consider ‘the social effects’ and ‘political’ impact of ‘applications’, e.g. of ‘creating computer therapists or judges’ (LC 29f).

10.55 Moreover, ‘for the most part the book does not deal with specific applications or the problems’ of ‘making practical use of linguistic theories’ (LC 22) (13.60). Nonetheless, Winograd briefly reviews the issues. First and foremost, naturally, is the ‘major goal’ of ‘current linguistics’ to describe language with the formality and precision needed for computer implementation’; prospects of ‘resources’, ‘support’ and ('Air Force’) ‘funding’ are raised. Reciprocally, ‘the computer has opened many new possibilities for linguistics’ and enabled ‘the development of specialized artificial languages’ (LC 22f). ‘Machine translation’, which has ‘focused on the syntactic structures of language’ and on ‘computer forms’ for ‘bilingual dictionaries’ and which has not been ‘successful’ in ‘fully automatic high quality translation’, has now sparked ‘renewed interest’ (LC 23, 358f, i.r; but see 10.106). Further domains of, cf. UN 41f; ‘human-machine interaction’ ‘in natural language’ include ‘information retrieval’ (for ‘a "library of the future"‘), ‘text retrieval’, ‘question answering’,[14] ‘explanation systems’ , ‘expert systems’ , and ‘speech understanding systems’ (i.e. ‘converting spoken sounds to written text)’, along with ‘text analysis’, ‘knowledge engineering’, ‘natural language’ ‘front ends’ ('the interface seen by a user of a computer system’) ‘for data base systems’, ‘computer aided instruction’, and ‘aids to text preparation’ such as ‘word processors’, ‘spelling checkers’, and ‘dictation systems’ (LC 23-26, 359-61).

10.56 In addition, the ‘potential’ of ‘the computational approach to language’ -- this time ‘shared with’ ‘structural and generative linguistics’ -- extends to such ‘practical aspects’ as ‘psychology’ (compare ‘psycholinguistics’), ‘language therapy’ for ‘deficits’ and ‘aphasia’ (compare ‘neurolinguistics’), ‘effective communication’ (compare ‘rhetoric’, ‘general semantics’, and ‘preservers of the "purity" of language"‘, all three ‘not held in high regard by academic linguists’), ‘designing languages that are easier to understand and learn’ (compare ‘Esperanto’), ‘teaching language skills In elementary and secondary schools’, and ‘teaching’ ‘translation’ to ‘humans’ (LC 26ff). These listings resemble Halliday's, who is in fact cited (LC 27; cf. 9.111), just when Winograd was de-emphasizing ‘systemic grammar’ for not being a properly ‘cognitive approach’ ‘rooted’ in ‘mathematics or formal logic’ (LC 273) (cf. 10.39, 45). Curiously sanguine too is the idea that ‘cognitive linguistics’ as ‘a theoretical framework for dealing with grammatical complexities’ could enable us to ‘understand how language works’ and thus bring us ‘a long way toward understanding how the rest of the mind works in reasoning, learning, and remembering’ (LC 27f) (cf. 13.21f).

10.57 As if in compensation for the various exclusions and limitations, the concept of ‘syntax’ expands far beyond Winograd's definition, namely ‘the part of linguistics that deals with how the words of a language are arranged into phrases and sentences and how components like prefixes and suffixes are combined to make words’ (LC 35; cf. LC 11). Instead, the term subsumes the construction and organization of every class of formal symbol system. Winograd's presentation of formalisms and grammars resembles Pike's treatment of the physiology of utterance in two ways: its monumental thoroughness, and its uncertain relevance for essential qualities of human language (cf.5.44). ‘Much of the material’ is ‘an explanation of techniques for structuring data and program in computers’, designed ‘to develop the student's mastery of the concepts of computational processing’ (LC 13). These ‘techniques’ are offered not as ‘precise theories of human language use, but rather building blocks from which theories can be constructed’.

10.58 Accordingly, we are painstakingly shown the qualities and uses of ‘grammars’, ‘patterns’, ‘schemas’, ‘registers’, ‘records’, ‘agendas’, ‘stacks’, or ‘buffers’. We find out about ‘rules’, ‘predicates’, ‘categories’, ‘classes’, ‘elements’, ‘variables’, or ‘variants’. We watch the construction of ‘charts’ or ‘tables’ with ‘edges’ and ‘vertices’, of ‘networks’ with ‘states’ and ‘arcs’, of ‘trees’ with ‘nodes’ and ‘leaves’, and of ‘roles’ with ‘slots’ and ‘fillers’. We are initiated into the issues of ‘computational implementation’, such as ‘system engineering, compi1ing’, ‘bookkeeping’, ‘pattern-matching’, ‘nesting’, ‘arc ordering’, ‘backtracking’, and ‘automatic emptying’, along with ‘classification’, declaration’, ‘invocation’, ‘enumeration’, ‘activation’, ‘deactivation’, ‘iteration’, ‘continuation’, and ‘termination’. We see numerous ‘figures’ for ‘procedures’ to ‘test’ or ‘match a pattern’, and to ‘recognize’, ‘generate’, or ‘parse a sentence’ by means of a ‘network’, ‘chart’, or ‘tree’ -- and even to ‘create a grammar’.

10.59 And virtually no species of grammar gets overlooked. The main types treated are, context free grammars’, ‘transformational grammars’, ‘augmented transition network grammars’ and ‘feature and function grammars’ (this last type including ‘systemic’ and ‘case’ grammars). These types differ both in the notations used and in the ‘general issues’ of ‘design’ for applying them as programs, e.g., ‘formal power’, ‘uniformity of processing’, ‘separation’ ‘of levels’, or ‘flexibility’ versus ‘precision’ (LC 89f, 114). Already for ‘purely syntactic, uniform, precise, context-free parsing’, ‘strategies differ’ along ‘three major dimensions': (a) ‘parallel versus sequential treatment of alternatives’, i.e. ‘keeping track’ of them ‘simultaneously’ versus ‘trying’ just one and ‘backtracking when its choices lead to failure’, (b) ‘top-down’ or ‘goal directed processing’ ‘versus bottom-up’ or ‘data directed processing’, i.e. ‘looking at rules for the desired top-level structure (usually a sentence) and seeing what constituents would be needed’, versus ‘beginning with the words’ and trying to ‘combine’ them into a ‘constituent’;[15] and (c) ‘systematic’ ‘choice of nodes to expand (in a top-down procedure)’ ‘or combine’ (in a bottom-up procedure’), either in a ‘directional’ manner, ‘moving’ ‘in one direction (usually left to right)’, or in a ‘size-oriented’ manner, ‘taking chunks of increasing size’, or in a ‘mixed’ manner doing some of both (LC 90f) (cf. 11.13, 19, 25, 32, 55, 59, 73, 77, 79, 13.44).[16]

10.60 As these ‘dimensions’ indicate, a computational framework allows more elaborate and operational comparisons between grammars than do other frameworks. We can see this clearly in the comparison between ‘transformational grammars’, which ‘take the process of abstract derivation as a starting point’, and ‘augmented transition network (ATN) grammars’, which ‘take the process of parsing’ and put It in ‘the clearest formulation’ (LC 195). ATNs ‘are currently one of the most common methods of parsing natural language in computer systems’ and have ‘served as the basis for psycholinguistic theories and experiments’. The ‘formalism is clear enough to be grasped and followed easily’, yet can ‘deal with complex phenomena’ (LC 196). ‘Instead of’ ‘ru1es’, we have ‘labelled networks’ in which ‘each’ is labelled with a word, a lexical category, or a syntactic category’. In a ‘simple transition network, an arc whose label is a word or lexical category can be traversed if it matches a single word of the Input’. But we can make the ‘network’ ‘recursive’ by using ‘composite syntactic categories’ as ‘labe1s’; then ‘the arc is traversed by matching a sequence of input symbols’, and ‘a network’ appears within the ‘network’; for example, we could have ‘an arc labelled NP’ to ‘recognize’ the entire, ‘constituent’ of a noun phrase (LC 197). Hence, we can ‘adapt a schema from grammar rules to networks’; the method of ‘choosing a rule’ and testing whether’ it ‘spans a contiguous sequence of constituents’ ‘is replaced by choosing an arc’ and ‘testing whether a network accepts a contiguous sequence of constituents’ (LC 199).

10.61 ‘Augmentation’ is the technique of ‘adding conditions and actions associated with the arcs of a network’ (LC 204). ‘The conditions restrict the circumstances under which an arc can be taken’, e.g., by stipulating ‘specia1 properties of the word or constituent to be matched’ or of the ‘constituents that have already been found’; the ‘actions perform feature-marking and structure-building operations’ (LC 204f, 208). ‘Conditions and actions make use of registers’, ‘each having a name and storing some information’, such as, roles and features’. ‘Whenever an arc Is taken, the associated action is carried out, causing the contents of registers to be set’ (LC 208). ‘A configuration’ is a ‘temporary structure’ that ‘includes the register table for the constituent being built, along with the current network, state, and position’ (LC 210).'As with other syntactic formalisms’, ‘ATNs’ have ‘a "basic theory" and then an ever-growing collection of changes and extensions’ to ‘cover more of the data’ (LC 244).'For practical uses’, ‘one picks out a subset of structures’ to ‘handle and simply ignores the others’, ‘hoping’ ‘the subset will be "habitable"‘ and ‘convenient to converse with’ (LC 245).

10.62 Through ‘the use of registers’ in ATNs, ‘features’ can be ‘associated with whole networks’ and need ‘not appear in the dictionary’, whereas, transformational grammars’ ‘add special grammatical markers’ and ‘have them manipulated by transformations’ (LC 210, 216; cf. 7.63f, 72). Also, ‘the ordering or conditions and actions’ in ATNs can do ‘the same work as the ordering of transformations’ in a ‘cyc1e’ (LC 220, 229). ‘In an ATN’ grammar, ‘the parser will work left-to-right and top-down’, so ‘the action that sets a register will affect’ only ‘conditions on arcs to the right’. This design ‘removes flexibility’, e.g. by impeding ‘bottom-up’ operation’, but in return offers ‘a relatively simple and elegant way for the interactions between different syntactic phenomena’, e.g. ‘subject-verb agreement’ and ‘pass1ve voice’ (LC 220, 222). Hence, ‘phenomena’, related to the rule-ordering mechanisms, in, transformational grammar’ get treated in, a somewhat more natural’ manner, ‘since the order is imposed by the order of the phrases in the sentence rather than by a less Intuitively guided choice of ordering for derivation rules’ (LC 222; cf. 7.54). Finally, ‘global hold registers’, which ‘be1ong to the sentence being parsed as a whole’ and are ‘expanded and contracted as nets are entered and left’, can keep ‘items on hold’ until needed, and thus handle ‘long-distance dependencies’, ‘one of the most complex phenomena of syntax’ and ‘a major motivation for many of the mechanisms of transformational grammar’, e.g. the ‘movement transformation’ (LC 232ff).

10.63 Comparing grammars this way on computational grounds illustrates the operational criteria entailed in choosing and designing a grammar (13.31). In general, ‘scientists’, including ‘linguists’, ‘are motivated to find’ ‘simpler’ ‘explanations’ (LC l84). But computationally, the striving for ‘simplicity’ can involve disadvantages in regard to ‘power’, ‘efficiency’, ‘effectiveness’, ‘precision’, ‘appropriateness’, ‘selectivity’, ‘ambiguity’, and ‘arbitrariness’ (cf. LC 73, 108, 361, 526, 326, 371, 75, 323, 532). For instance, ‘context free grammars’ have been ‘widely used because of their simplicity’, but cannot ‘handle the full complexity of natural language’ (LC 72, 361, 383, UN 42) (cf. 7.48, 73f. 13.39). Like ‘phrase structure grammars’, they can work only on some ‘simplified subset’ of English’ and must ‘lose simplicity’ to ‘gain power’ (UN 42, LC 73). ‘Simplicity’ also enhanced the appeal of the transformational approach (7.36f, 40, 50f) and has been a motive for various revisions, but has played a very ‘mixed role’ in ‘the generative grammar literature’ (LC 148, 162, 173, 319, 168, 578, 183). Theorists called for a ‘specific and restrictive formalism’, yet these ‘are generally also more complex’ (LC 183f, 175; cf. 7.22, 66, 94). Studies of ‘psychological reality’ showed that the measure of ‘simplicity’ and ‘comp1exity’ in terms of ‘number of transformations’ of a, sentence’ did not fit the observed ‘ease of comprehension’ (LC 177f). The idea that a ‘language learner, like the scientist, prefers parsimony’ and ‘simplicity’ has also failed to produce ‘precise measures’ (LC 184).

10.64 So Winograd proposes further ‘desiderata’ for ‘choosing among competing grammars’ (LC 327). First, ‘correspondence with meanings’ is essential to ‘mathematics and programming languages’, and, as we saw (10.52), is frequently invoked despite the book's overriding concern with syntax. Second, ‘perspicuity’ is attained for a grammar by making ‘the facts it expresses about language directly visible in the form’. The ‘consequences’ of ‘a transformational rule that moves an abstract marker’ in ‘a tree structure’ ‘are implicit in its interactions with other rules’ and ‘may be incomprehensible on inspecting the rule’ (cf. Woods 1970) (LC 327, UN 45) (cf. 10.19, 67). In contrast, in ‘context-free grammars’ ‘every rule is also a pattern’ (LC 327). ‘Recursive transition networks’ have the same ‘transparency’ ‘every path through the arcs’ is ‘a pattern of constituents’; but this ‘perspicuity is no longer guaranteed’, when ‘augmentations such as conditions, actions, and registers are added’. ‘In systemic grammars’, ‘the effects of particular realization rule can be seen as a direct statement of the form of the constituent to which it applies, but because the rules can interact’ they are less ‘perspicuous’ than ‘context-free rules’. ‘Functional grammars’ are now ‘returning to simpler rules which are more immediately structural’.

10.65 Third, ‘nondirectionality’ is attained by ‘using the same set of rules in both the generation process and the parsing process’ (LC 328) (cf. 10.46; 13.57). ‘Simple patterns and context-free grammars’ are ‘neutral’ in just this way. But ‘a transformational grammar "runs" only in the direction of generating and is quite difficult to apply’ to ‘parsing’, it gets bogged down in ‘combinatorial explosion’ because it ‘tries to reproduce the deep structure of a sentence while doing surface structure recognition’ (Woods 1970) (LC 328, UN 31, 42, 45) (cf. 7.83; 10.21; 11.3, 81).'Simple transition networks are’ ‘nondirectional’, but in ATNs ‘the’ use of conditions and actions forces a left-to-right ordering, since registers can be accessed only after they are set’ (LC 328) (cf. 10.59, 62). ‘Lexical-functional grammars’ try to ‘modify the ATN formalism to get rid of order dependencies’.

10.66 Fourth, ‘efficiency’ is attained by limiting the ‘resources used by a given procedure’, such as ‘time’ and ‘storage’ (LC 114, 111). ‘The theory of computation deals with the gross order of efficiency': ‘linear, logarithmic, polynomial, exponential, etc.’ ‘with respect to the length of the input’ (LC 376). ‘In context-free parsing,’ for example, ‘the maximal efficiency’ is ‘a time proportiona1 to the cube of the length of the input if the grammar allows ambiguities or to its square if the grammar is unambiguous’ (LC 114). But this ‘formal efficiency’ ‘is not the same’ as ‘practical efficiency’, which depends on ‘how the procedure is implemented using particular data structures in a particular programming language on a particular machine’ (LC 114f). Two ‘implementations’ might be formally ‘equivalent,’ ‘but one might be tens or hundreds of times faster’ (LC 376).[171 ‘Over the years,’ people ‘have learned to sort out those sources of inefficiency’ ‘due to a particular choice of how structures are stored or accessed from those’ ‘inherent in the type of procedure’ (LC 114). We now have ‘many ways that procedures can be modified to make them more efficient while producing the same results’ (LC 109). For instance, we can have ‘the grammar’ ‘compiled’ into another ‘form’ (LC 376) (cf. 10.12, 71). Or, we can ‘set things up so a particular computation can be done once and used in many places’ (LC 109). Or, we can insert, pre-computations, in which auxiliary tables or indexes are created to avoid computation steps as the parser runs’ (LC 115).

10.67 Fifth and finally, ‘multiple dimensions of patterning’ are needed to cover the various ‘aspects of language structure’ (LC 328). The ‘dimensions’ concern ‘form’, ‘function’, and ‘features’ (LC 50, 205, 211, 289, 292). Also, ‘syntactic resources’ are structured along ‘dimensions of organization’, including ‘choice of items’ ('classification’ of ‘elements’), ‘sequential arrangement’ ('grouping’), and ‘function’ ('relation of one element or group to another’ in terms of ‘roles’) (LC 274f).[18] ‘Transformational grammar’ gets a low rating here, because the ‘different dimensions’ are not ‘expressed’ ‘explicitly’, but ‘represented in the structure of the phrase marker at different points in the derivation’ (LC 328) (cf. 10.19, 62, 64). ‘Systemic grammar’ gets a high rating, however, because ‘constituent structure is analysed’ in a greater variety of ‘different dimensions’ and, is described in terms of units that correspond to meaningful elements and that are less broken up than those of transformational grammar’ (LC 276) (cf. 9.33). Also, ‘since systemic grammar is not centred’ on, formal rules’ (i.e., is not ‘generative in the strong sense’, 10.56), ‘descriptive structure’ can more easily ‘deal with multiple dimensions of analysis’; the ‘organization’ of ‘a text’ ‘as a speech act’, ‘a logical proposition’, a configuration of ‘cohesion relationships’, or a ‘theme and information structure’ (LC 278) (cf. 9.49f, 55ff, 89-96).[19]

10.68 These desiderata plainly do not coincide to enforce the choice of just one grammar or formalism, but confront us with trade-offs (cf. 13.1). For example, ATN grammar and systemic grammar, which Winograd had combined in his own SHRDLU program, are less ‘perspicuous’ than ‘context-free grammar’, but superior in regard to ‘dimensions’ and ‘meanings’. Besides, the formalisms are often ‘equiva1ent’ enough to mimic each other.[20] For example, ‘in general it is possible to find ATN correlates of the transformational constraints without significantly more or less complication’ (LC 244).Thus, criteria of notation and design are inconclusive. The definitive criterion therefore ought to be the one raised in the first book: in which system is the ‘operation’ ‘closer to the actual operations humans use in understanding language’ (UN 43)? But, as we saw (10.42ff), this criterion is much less central in LC than in UN, apparently to enhance the impartiality of the discussion and to maintain a cordial alliance with the generativists.

10.69 As if to make amends, the next book (UC) places so strong an emphasis on human knowledge and processing that all formalisms are put in question. Already in LC, Winograd had warned: ‘many critics of artificial intelligence argue that much of our skill of using language is not in the nature of formal rules’ and that ‘the ability to use language’ cannot be ‘exp1ained by any formal characterization analogous to data structures of computers or the rules of formal logic’ (LC 29). His own alliance with those ‘critics’ is signalled when he refers us to his ‘forthcoming’ book whose ‘theoretical framework’ ‘exp1ains why the current work on AI cannot provide a basis for understanding and modelling the full range of human language understanding’ (LC 32). This volume, co-authored with Fernando Flores,[21] who was trained in management science and had been ‘Minister of Economy and Minister of Finance in the government of Salvatore Allende’, asserts flat out that ‘one cannot construct machines that either exhibit or successfully model intelligent behaviour’; ‘computers cannot understand language’; and, ‘computers will remain incapable of using language in the way humans do’ (UC xi, 11f, 107).

10.70 To demystify the computer and to debunk the ghost in the machine, Winograd and Flores enumerate the various ‘levels’ upon which a computer might be viewed (UC 86-90). ‘The physical machine’ is a ‘network’ of ‘wires, integrated circuits, and magnetic disks’ ‘operating according to the laws of physics’ in ‘patterns of electric and magnetic activity’ (UC 86f). ‘At the bottom’ are ‘basic elements’ like ‘strands of copper and areas of semiconductor metal’ on ‘wafers of silicon crystal’. ‘The 1ogica1 machine’ is composed of ‘logical abstractions such as or-gates, inverters, flip-flops’, ‘multiplexers’, and ‘address decoders’; ‘voltages’ serve to ‘represent a logical "true"‘, or "false''‘. ‘The abstract machine’ Is a ‘single sequential processor which steps through a series of instructions’; ‘logical patterns’ ‘of trues and falses are interpreted as representing a higher-level symbol such as a number or a character’ (UC 88). ‘Most descriptions of computers’ are at this ‘level’, which Is ‘usually the lowest level at which programmer has control’.

10.71 Next, the ‘high-level language’ is ‘based on more complex symbol structures, such as lists, trees, and character strings’ (UC 88). ‘A compiler or interpreter converts a formula’ ‘into a sequence of operations for the abstract machine’, and ‘complex mathematical operations’ can be done in ‘a single step’. Finally, the, representation scheme’ ‘uses the symbol structures of a high-level language to represent facts about the world’ (UC 89). A ‘fact’ can be ‘encoded’ ‘as a series of manipulations on a data base or as the addition of a new proposition to a collection of axioms’ (cf. 10.50). Winograd and Flores stress ‘the complexity that lies between an operation’ in ‘a program’ and ‘the operation of the physical computing device’ (UC 89f). ‘There is no intelligible correspondence between operations at distant levels’; ‘computer systems’ ‘can exhibit many levels of representation, each of which is understood independently of those below it’. However, a more integrated view is needed to allocate ‘resource use’ in ‘implementation’ (e.g. ‘speed’, ‘storage’) and to handle ‘breakdowns generated by the lower levels’ (UC 91). ‘Some programmers argue’ that ‘the program should be written at the level’ where ‘resources’ ‘can be directly described’ (e.g. ‘real time control processes’ ‘in assembly language’), but ‘in practice, programs are often initially designed without taking into account the lower level, and then modified to improve performance’.

10.72 Having presented the computer as a ‘tower of levels’, Winograd and Flores inquire ‘why anyone would consider that computers could be intelligent’, any more than ‘a clock or an adding machine’ (UC 89, 93) The reason is the ‘apparent qualitative leap’ to "mind-like'‘ qualities’, which is actually an ‘effect’ of ‘quantitative’ dimensions along which computers ‘differ in degree’ from other machines (UC 95, 93). They have ‘apparent autonomy’, being able to ‘carry out comp1ex sequences of operations without human intervention’ (UC 94). They have ‘complexity of purpose’, able to ‘provide’ a wide ‘range of services’. They have ‘structural plasticity’, allowing us to ‘build mutable higher-level structures’ ‘on a relatively fixed underlying structure’. And they have ‘unpredictability’ because we cannot know ‘how a program will act short of running’ or ‘simulating it’ (UC 95).

10.73 Or, ‘intelligence’ is attributed because computers can be used for ‘problem-solving behaviour': ‘a process of search for a sequence of operations that will lead to a solution point’ (Newell & Simon 1972) (UC 95) (cf. 10.5, 10, 30; 11.97f).[22] But for all such cases, Winograd and Flores argue that the intelligence and reasoning belong not to the computer, but to those who program it (cf. UC 85, 97, 123f, 131, 165, 178). ‘The programmer’ ‘characterizes the task environment’, ‘generates the systematic domain’, ‘designs the formal representation’, ‘sets its structures in correspondence with the structure available on the computer’, and ‘implements the search procedure’ (UC 96). ‘The computer’ can only ‘operate’ according to ‘the formal representation’ and ‘the rules of the formal system’ (UC 96f). Whereas ‘the programmer acts within a context of language, culture, and previous understanding’, ‘the program is forever limited to working within the world determined by the programmer's explicit articulation of possible objects, properties, relations’.

10.74 The focus on machine construction and programming forms only one part of a wide attack on ‘artificial intelligence’ and its projects for ‘natural language understanding’. Here, Winograd's typical concern for large issues now scales such heights as to make even Chomskyan universalism seem parochial (cf. 7.18ff). The new book undertakes to ‘address the fundamental questions of what it means to be human’, ‘what it means for something to exist’, and ‘what it means to know’ and ‘understand’ (UC 7, 13, 70, 30, 72, 119). And the answers are far from what ‘common sense’ suggests (UC 30, 43, 46). An assault is mounted on the very mainstream of Western science and thought, whose basic views and assumptions are decried (17 times) as ‘naive’ (UC 8, 30, 40f, 46, 50f, 60f, 69n, 71f, 135, 149).

10.75 Like Winograd's other books, the latest one is ‘deeply concerned with the question of language’ (UC 17). ‘Linguistic action’ is declared ‘the essential human activity’, "'the central feature of human existence is its occurrence in a linguistic cognitive domain"‘ (cf. Maturana 1970) (UC 7, 51) (cf. 10.4, 32, 43; 13.22). The thesis that ‘we create our world through language’ leads to a ‘radical recognition': ‘nothing exists except through 1anguage’ (UC 11, 68) (but see 12.60). Of course, similar claims have been made for centuries by philosophers (e.g. Heidegger, 12C21]), who tend to be more at ease with language than with existence, But here, the claim strategically implies that if computers can't understand language, they can't understand anything about the world either.

10.76 The chief target of attack is ‘the rationalistic tradition’, which, emphasizes "information", "representation", and "decision making"‘, and prizes ‘particular styles of conscious rationalized thought’ (UC 8). ‘This tradition has been the mainspring of Western science and technology, and has demonstrated its effectiveness most clearly in the "hard sciences", those that explain the operation of deterministic mechanisms whose principles can be captured in formal systems’ (UC 14) (cf. 10.30, 39, 43, 47). It ‘underlies both pure and applied science’ and ‘has greatly influenced’ ‘linguistics and cognitive psychology’, plus ‘management theory and cognitive science’, because it is ‘regarded’ as the ‘paradigm of what it means to think and be intelligent’ (UC 16). It ‘finds its highest expression in mathematics and logic’, which ‘are taken as a basis for formalizing what goes on when a person perceives, thinks, and acts’; and it is deemed ‘self-evident that this is the right or even the only approach to serious thinking’ (UC 14, 16) (cf. 10.8, 39- 13.15, 17). Winograd and Flores now propose the ‘tradition’ be, above all in ‘current, re-examined and challenged as a source of understanding’, thinking about computers and their impact on society’ (UC 14, 26). They ‘attempt’ to ‘reveal the blindness it generates’, and ‘argue that’ it ‘needs to be replaced’ ‘if we want to understand human thought, language, and action, or to design effective computer tools’ (UC 17, 26).

10.77 ‘One cornerstone’ of the, rationalistic tradition’ is a ‘correspondence theory’ of ‘language as a system of symbols’, composed into patterns that stand for things in the world’ (UC 19, 17) (cf. 10.20, 40, 43).[23] Here, ‘the content of words’ ‘denotes’ ‘objects, properties, relationships’ ‘in the world’; ‘what a sentence says’ is ‘a function of words it contains’ and their ‘structures’; and ‘sentences say things about the world’, and are ‘either true or false’ (UC 17). ‘More formal studies of semantics’ ‘examining meaning from a formal analytical perspective’, however, rarely seek ‘formal answers to the problem’ of ‘correspondence’ (UC 17f). Such ‘questions’ are ‘taken as unproblematic’ or ‘pushed aside’ as ‘too difficult and open-ended’ (UC 15). Instead, one ‘looks at the relations among the meanings’ of ‘words, phrases, and sentences’ without ‘reference either to act of uttering the words or to the states of affairs they describe’ (UC 18).

10.78 ‘It is assumed that each sentence in a natural language’ can be matched with ‘one or more’ ‘interpretations in a formal language, such as first-order predicate calculus’ (UC 18) (cf. 10.8). ‘The study of meaning’ proceeds by ‘translating sentences’ into ‘formal structures’ and applying ‘logical rules’. ‘Truth theoretic’ ‘systems of rules’ are expected to allow for ‘translating’ without losing the ‘essence of the meaning’, for ‘determining’ ‘the meanings of formulas’ from ‘the meanings of parts and the structures by which those parts are combined’; and for ‘interrelating the truth conditions for different formulas’ (UC 19) (cf. ‘7.82; 13.18, 59). Naturally, ‘the fundamental’ ‘sentence is the indicative’ ‘stating that a certain proposition is true; its meaning’ depends on ‘the conditions in the world under which it would be true’ (cf. 9.74).[24] Finally, ‘the meanings of the items being composed should be fixed without reference to context in which they appear’ (cf. 7.73, 79; 11.2, 36, 40). Some ‘obvious’, and ‘exceptions’ are ‘recognized’, such as ‘pronouns’, ‘place and time adverbs’, ‘tenses’, ‘but the central theory of meaning (semantics) deals’ with ‘literal meaning’, as not context-dependent’ (cf. 12.68).

10.79 In a similar vein, ‘rationalistic theories of mind all adopt’ some ", representation hypothesis"': ‘that thought is the manipulation of representation structures in the mind’ (UC 20; cf. UN 3, 23; LC 16, 18, 186; UC 8, 78, 85f, 89, 96). ‘Though not specifically linguistic’, these ‘are treated as sentences in an "internal language", ‘connected to the world’. Using a, compatible’, approach’, ‘information processing psychology’ assumes’ that, all cognitive systems are symbol systems’ and ‘achieve their intelligence by symbolizing external and internal situations and events and by manipulating those symbols’ with one ‘basic set of underlying’ ‘processes’ (UC 25). Hence, ‘a theory of cognition can be couched as a program’ -- i.e., ‘a formal system’ having ‘variables’ and ‘generating predictions about the behaviour (outputs) of some naturally occurring system’ --- which ‘when run in the appropriate environment will reproduce the observed behaviour’ (10.4, 6). ‘The computer’ ‘enables the scientist to deal with more complex theories’, and ‘AI’ ‘programs ‘patterned after human thought and language’ offer a handle on, phenomena that do not have the obvious limitations of the sparse experimental situations of cognitive psychology’ (cf. 10.44).

10.80 Winograd and Flores take just the contrary view. They find it ‘naive’ to think that ‘language conveys information about an independent reality’ (UC 50). ‘Words correspond to our intuition about ‘'reality'‘ because our purposes in using them are closely aligned with our physical existence in a world and our actions within it’ (UN 61) (cf. 5.68; 6.12; 8.33; 13.24). ‘But the coincidence is the result of our use of language within a tradition’. And ‘in using language we are not transmitting information or describing an external world’ that ‘defines’ ‘the meaning of words and sentences’ (UC 50, 61). Only if we ‘stick to the rather idealized isolated sentences used’ ‘in philosophy books’ does it ‘seem plausible to ground the meaning of words in a language-prior categorization’ (LC 61, MSF 97f). ‘As soon as we look at real situated language, the foundation crumbles’.

10.81 So we need to emphasize not ‘the mental dimension’ of ‘the cognitive paradigm’ in LC, but the social, because both language and cognition are fundamentally social (UC 60f) (cf. 10.43ff, 54; 13.20). To ‘use language’ is to ‘create a cooperative domain of interactions’ (UN 50, MSF 101). The view of ‘cognition’ proposed here depends ‘critically’ on the ‘work of Humberto R. Maturana, a Chilean neurobiologist’ (UC 10, 38).[25] His studies of ‘visual’ ‘perception’ in primates’ (mainly ‘frogs’) led him to conclude that ‘the nervous system’ is ‘a generator of phenomena, rather than a filter on the mapping of reality’ (UC 42) (cf. 4.10, 14, 18f- 5.27f. 8.21, 23). He ‘described the, nervous system as a closed network of interacting neurons’ that ‘does not have "inputs’ and "outputs"‘; it is ‘perturbed by structural changes in the network itself’, which in turn ‘trigger changes’ in ‘the relative activity’ of ‘neurons’ (i.r.). ‘The structure of the system’, specifies what structural configurations of the medium’ (the ‘environment’)[26] ‘can perturb it’ and thus ‘determines a domain’ or ‘space of possible effects the medium could have’ (UC 42f, i.r.). This interaction between system and medium is termed ‘structural coup1ing’ and provides an alternative to ‘extreme’ ‘behaviourist descriptions’ of ‘stimuli and response’ ‘without reference to the structure of organism’ but only to ‘the patterning of events’ (UC 45f, 48).[27] The term figures conspicuously in Winograd and Flores's own account of understanding and knowing (cf. UC 10, 45, 47ff, 61, 72, 104, 119; 10.83).

10.82 Maturana ‘rejected’ ‘information processing as the basis for cognition’ and averred that ‘"living"‘ itself ‘"is a process of cognition"‘ (UC 46). He ‘sought to explain the origins of all phenomena of cognition in terms of phylogeny (species history) and ontogeny (individual history) of living systems’ (UC 44, 46) (cf. 9.12). He defined ‘an autopoetic system’ as, "a network of processes of production, transformation, and destruction of components"‘ whose ‘"interactions and transformations"‘ ‘"continuously regenerate"‘ and, "constitute the network"‘. ‘A cognitive explanation’ ‘deals with the relevance of action to the maintenance of autopoesis in a phenomena1 domain’, i.e., with the ‘relevance of the changing structure of the system to behaviour that is effective for its survival’ (UC 47). Therefore, ‘"learning is not a process of accumulation of representations of the environment"‘ but of "'transformation of behaviour through continuous change in the capacity of the nervous system to synthesize it"‘ (UC 45).[28]

10.83 When ‘structural coupling’ arises from ‘perturbations generated’ by ‘other organisms’, the resulting, interlocked patterns of behaviour’ ‘form a consensual domain’ (UC 48). The ‘behaviours involved are both arbitrary’ ‘because they can have any form as long as they’ ‘trigger perturbations’, and ‘contextual because their participation’ Is ‘defend only with respect to the interactions that constitute the domain’ (UC 48f).'Maturana extended the term "linguistic" to include any’ ‘behaviour in a consensual domain’. But even in the ordinary sense, ‘human language is a clear example of a consensual domain, and the properties of being arbitrary and contextual have’ ‘been taken as its defining features’ (cf. 13.27). Thus, ‘language’ gets redefined as ‘a patterning of "mutual orienting behaviour", not a collection of mechanisms in a "language user'‘ or a "semantic" coupling between linguistic behaviour and non-linguistic perturbations experienced by the organism’. From there, Maturana concluded ‘that language is connotative and not denotative, and that its function is to orient the orientee’ within a ‘cognitive domain, and not to point to independent entities’. So, we were told (10.80), ‘the basic function of language is not the transmission of information or the description of an independent universe’, but ‘the creation of a consensual domain of behaviour between linguistically interacting systems’ (UC 50).

10.84 The ability to ‘talk about a world’ Is reserved for ‘observers': "'human beings"‘ ‘who can generate distinctions in a consensual domain’ and ‘"operate"‘ ‘''as if external to (distinct from) the circumstances"‘ (UC 50) (cf. 5.9). Even then, ‘a statement made by an observer to another’ ‘is grounded not in an external reality but in the consensual domain shared by those observers’. ‘Properties of things (in fact, the recognition of distinct things at all) exist only as operational distinctions in a domain’ ‘specified by an observer’ (UC 51). ‘We speak as if there were external things and properties’, but ‘this is an inescapable result of using language’ and ‘not an ontological claim’. ‘This idea that all cognitive distinctions are generated by an observer’ unites Maturana with ‘gestalt psychology’ (e.g. Kohler 1929) and ‘recent work in systems theory and cybernetics’ (e.g. von Foerster [ed.] 1974; Pask 1976) (UC 51, 38fn) (cf. 11.4).

10.85 A second and more surprising witness in the case against rationalism and AI is Martin Heidegger, lauded ‘as the modern philosopher who has done the most thorough, penetrating, and radical analyses of everyday experience’, where, radical’ seems to have its etymological meaning: ‘lying at the root of much of that other philosophers have said’ and ‘of our own orientation’ (UC 9). He concentrated on ‘phenomenology, the philosophical examination of the foundations of experience and action (LC 9). He rejected ‘the separation of subject and object’ in favour of a ‘more fundamental unity’; for him, ‘existence is interpretation’ and vice-versa (UC 31). ‘His philosophy is based on a deep awareness of everyday 11fe’, and ‘the issues’ are ‘difficult because they are concea1ed by their "ordinary everydayness", (UC 34). [291 Instead of following Heidegger's critique of language in Unterwegs zur Sprache (1959), conceptions are appropriated from Sein und Zeit.[30l

10.86 ‘Heidegger's discussion of "thrownness" and "breakdown"‘ is used as evidence that ‘models of rationalistic problem-solving do not reflect how actions are really determined and that programs based on such models are unlikely to prove successful’ (UC 12; cf. 10.5, 30, 69, 73). ‘Thrownness’ is ‘the condition of understanding in which our actions find some resonance or effectiveness in the world’ (UC 33). ‘Imagining you are chairing a meeting’ is offered as an ‘example’ or ‘metaphor’ (UC 34, 36). There, ‘you cannot avoid acting’ (UC 34).'You cannot step back and reflect on your actions’. ‘The effects of actions cannot be predicted’, so ‘you cannot count on rational planning’ to ‘achieve your goals’ (UC 34f). ‘You do not have a stable representation of the situation’ but at most ‘fragmentary, possibly contradictory’ ‘pieces’. ‘Every representation is an interpretation, so ‘even in the post-mortem, your description’ ‘is hardly an objective analysis’ ‘subject to proof’; ‘there is no ultimate way to determine that any one Interpretation Is really right or wrong’, and ‘the people whose behaviour is in question may well not be in touch with ‘their own deep motivations’.

10.87 For Winograd and Flores, the ‘study of Heidegger reveals the central role of breakdown in human understanding’ and ‘activity’ (UC 165f). ‘Breakdown’ is the moment when ‘objects and properties’ ‘that constitute the domain of action for a person’ and are ‘not inherent in the wor1d’, ‘emerge’ and ‘become present-at-hand’ (UC 36, 166). Such ‘a breakdown is not a negative situation to be avoided, but a situation of non-obviousness in which the recognition that something is missing leads to unconcealing (generating through our declarations) some aspect of the network of tools’ we are ‘using’ (UC 165).[31] And if ‘the objects and properties that constitute the domain of action for a person are those that emerge in breakdown’, then , breakdowns play a fundamental role’ in ‘the analysis of a human context of activity’ (UC 166). ‘A breakdown reveals the nexus of relations necessary’ for the ‘task’. When a person is, hammering, the hammer as such does not exist’; ‘it is part of a background of readiness-to-hand’ ‘taken for granted without explicit recognition or Identification of the object’ (UC 36). ‘The hammer presents itself only when there is some kind of breaking down’, e.g. if ‘it breaks or slips from grasp’ or simply ‘cannot be found’. ‘Observers’ ‘may talk about the hammer’, ‘but for the person engaged in the thrownness of unhampered hammering, it does not exist’. The same case is made for a, computer': ‘my hand and arms, a keyboard, and many complex devices that mediate between it and a screen -- none of this equipment is present for me except when there is a breaking down’ (UC 36f). Evidently, something unusual is meant here by ‘exist’ and ‘present’.[32]

10.88 ‘Whenever we treat a situation as present at hand, analysing it in terms of objects and their properties, we thereby create a blindness’ (UC 97). ‘Our view is limited to what can be expressed in the terms we have adopted’. This ‘is necessary and inescapable: reflective thought is impossible without the kind of abstraction that produces blindness’. At best, we can become ‘aware of the limitations that are imposed’.

10.89 Armed with these diverse conceptions and arguments from ‘biology, hermeneutics, and phenomenology’, whose ‘unity [33] lies in the elements of the tradition they challenge’, Winograd and Flores set out to displace ‘the rationalistic tradition’ with ‘a new orientation’ relating to ‘the fundamental questions of what it means to exist as a human being, capable of thought and language’ (UC xii, 70) (cf. 10.74). ‘Cognition’ is not ‘a distinct function ‘separated from the rest of the activity of the organism’, like ‘respiration’ (UC 70). Nor can ‘cognition’ be ‘characterized’ or ‘exp1ained’ in terms of a ‘knowing "subject"‘, or of ‘menta1 states’, ‘mode1s’, and ‘operations’, or of ‘representations, concepts, and ideas’, or of ‘detached reflection’, ‘gathering information’, ‘manipulating symbols’, and ‘reasoning’ (UC 71-74). All such accounts are beset by ‘naiveté’ and ‘blindness’. Instead, ‘cognition’ is ‘a pattern of behaviour relevant to the functioning of the person or organism in the world’ (UC 71) -- a wide definition covering humans as well as ‘babies’, ‘frogs’, and ‘worms’ (UC 52, 36, 38, 46, 49, 103, 105), but not computers.

10.90 ‘Understanding’ is also redefined so as to place it far beyond the reach of computers. In Winograd's first book, ‘understanding language’ was described as ‘converting from a string of sounds or letters to an internal representation of meaning’ (UN 23). Now, ‘understanding’ is described as ‘a commitment to carry out a dialog within the full horizons of speaker and hearer in such a way that new distinctions emerge’ (UC 124). ‘Any individual, in understanding his or her world, is continually involved in activities of interpretation’, which ‘depends o