2. Psychological and Computational Models of Sentence Processing

Last time, I argued that there are substantive open questions about whether the theoretical constructs of formal linguistics play any role in the psychological processes underlying language use. Let’s now address those questions.

When people talk about “the psychological reality of syntax”, there are (at least) two importantly different types of psychological state that they might have in mind. One of them is what I call mental phrase markers (MPMs)—representations of the syntactic structure of incoming linguistic stimuli, constructed in the course of on-line comprehension, and also whatever states play an analogous role in language production. MPMs are relatively transient states; their “lifespan” is typically measured in milliseconds. The other type of state is what I’ll call mental syntactic principles (MSPs)—representations of the general rules, principles, or constraints that jointly constitute a grammar for a language. MSPs are standing structures, architectural features of the human parsing mechanism.  When the language faculty is inactive, i.e., when no comprehension or production processes are taking place, these structures are, so to speak, dormant.  In the jargon of contemporary metaphysics, we can say that they are dispositional, rather than occurrent.

In principle, MPMs can be psychologically real while MSPs are not. Alternatively, as I shall argue, both can be useful psychological constructs, but MPMs are best seen as explicit (subpersonal) representations, whereas there is little or no evidence that MSPs are psychologically real in that sense.


Much of the work in psycholinguistics and computational linguistics takes it for granted that MPMs are constructed in the course of comprehension. But this assumption is not without its detractors, in philosophy, psycholinguistics, and artificial intelligence:

“[L]anguage use [is] a fairly brute-causal associationist process … rather than a process involving metalinguistic representations of the syntactic and semantic properties of linguistic expressions.”  (Devitt, 2006: p. 220)

“[N]o independent level of syntactic representation is constructed, operated on, or output by the language analysis process.” (Schank and Birnbaum, 1984: p. 220)

“[I]t isn’t even clear that we do parse, in the sense of constructing an explicit and complete representation of the structure of a sentence in the course of comprehending it. The only language tasks that we really know humans are able to perform are those we can directly observe, including comprehension, production, repetition, and so forth. That mapping from an acoustic signal to a message occurs via deep structure, or some other hierarchical syntactic representation, is merely a convenient assumption—a pervasive and often useful one, but an assumption none the less. In advancing our understanding of the human language processing system, we may do well to question this assumption.” (Rohde, 2002: p. 6-7) … There seems to be little direct evidence that we construct a representation of a syntactic parse tree while comprehending.” (p. 18)

So my first task is to show that MPMs are psychologically real.  Here are some relevant lines of evidence:

  1.  Neurolinguistics:  EEG studies using the violation paradigm have found that early left anterior negativity is elicited by, and only by, syntactically ill-formed phrases. Semantic and pragmatic violations elicit different EEG signatures (Bornkessel-Schlesewsky and Schlesewsky, 2009). And, more recently, MEG data have revealed that the brains of competent speakers respond to sequences of words in ways that track not only their acoustic, syllabic, and prosodic features, but also phrase-level groupings (Ding, et al., 2016).  These and other studies show that the human sentence-processing mechanism constructs distinctly syntactic representations.

2.  Structural priming:  In producing language, people tend to employ the syntactic structures that they recently produced or comprehended. Priming studies that make use of this fact allow researchers identify some of the representations that people construct when processing language (Pickering and Ferreira, 2008). The behavioral data can be used not only to demonstrate the psychological reality of mental phrase markers, but also to determine their content to a degree of precision that ERP studies cannot yet achieve.

3.  “Garden-path” processing:  Linguistic input is rife with ambiguity. The language processing system is remarkably effective in selecting the correct resolution of such ambiguities. When it fails to do so, the anomaly shows up in behavior—e.g., extended fixations on a crucial part of a sentence, in eye-tracking studies (Rayner, Carlson, and Frazier, 1983), or a modulation of reaction times in cross-modal priming studies (Nicol and Swinney, 1989). Three principles of ambiguity resolution, Minimal Attachment, Late Closure, and the Minimal Chain Principle, form the foundation of many psychologically plausible parsing models (Frazier, 1979; DeVincenzi, 1991). Taken together these principles predict and explain a vast range of behavioral and neurocognitive data. The principles themselves are not represented, nor even embodied, in the mind/brain, but they make ineliminable reference to the construction and manipulation of mental phrase markers. Competing accounts of the same data appeal to statistical information in the input, but the relevant frequencies are, once again, defined over mental representations of phrase structure.

A fourth argument for the psychological reality of MPMs is indirect:  I claim that no known model of language processing can explain the available data concerning human parsing preferences without positing MPMs.  Computational models that eschew MPMs have been developed in the classical AI tradition (e.g., Schank and Birnbaum, 1980).  More recently, Devitt (2006) tentatively endorses an account on which comprehension does not involve constructing mental phrase markers but, instead, maps acoustic input directly into thoughts, whose syntactic structure is assumed to be more-or-less the same as the structure of public-language sentences.  I argue that both models are unworkable in light of the available evidence, and that Devitt’s argument for “brute-causal” models rests on an untenable distinction between mental representations and “mere responses” on the part of an organism.

Having argued for the psychological reality of MPMs, I claim that their construction and manipulation can only be accomplished by a mechanism that either explicitly represents a grammar or embodies it. A system that “embodies” a grammar does not store a set of rules and principles in an explicit data structure (Stabler, 1983) and does not “access” or “read” them during real-time operations. Rather, the rules of the grammar are “hardwired” into the causal structure of the system, in such a way as to guide the construction of MPMs. In order to be an instance of embodiment, in the sense that I intend here, this hardwiring must also meet a condition that is stronger than the mere ability to process inputs of a certain type—i.e., stronger than mere “conformity” to a rule. For every rule or principle of the grammar, the hardwired system must have a unique causal mechanism that mediates the computation of all the syntactic representations that are in the domain and range of that rule or principle (Davies, 1995). A language-processing system can conform to a particular grammar without embodying it in this sense. Embodiment is weaker than representation, but stronger than conformity, in ways that are open to empirical test.

The embodiment/representation distinction is important, but making progress on the psychological reality issue requires being as clear as possible about the relationships between it and five other distinctions: conscious/nonconscious, personal/subpersonal, implicit/explicit, declarative/procedural, and occurrent/dispositional. These are often treated by theorists in philosophy and psychology as approximately co-extensive. I argue that they’re in fact mutually orthogonal, and that failure to notice this is the source of widespread confusion about related issues.  

Particularly significant is the difference between the personal-level conceptual episodes that are constitutive of sapience and the subpersonal states involved in language processing (e.g., MPMs). The distinguishing marks of a subpersonal state—inexpressibility in speech, limited interaction with beliefs and desires, and inaccessibility to consciousness, inter alia—are interrelated in theoretically significant ways, which deserve more scrutiny, particularly with respect to the normative assumptions that we make in ascribing such states. 

Likewise, we must distinguish beliefs–i.e., dispositions to have occurrent thoughts or judgments when the occasion arises, or to process various occurrent thoughts and judgments in specific ways.–from thoughts/judgments, which are states that we come to be in occurently, either by way of perceptual processing or in the manifestation of dispositional beliefs in inference. Some philosophers assume that the occurrent/dispositional distinction lines up with either the conscious/nonconscious distinction or with the personal/subpersonal distinction.  Neither of these assumptions has any merit.  The folk-psychological practices of predicting and explaining behavior rest on the ascription of both thoughts and beliefs to whole persons, not to subpersonal mechanisms. Moreover, intentional states can occur both consciously and nonconsciously, so being an occurrent state is neither a necessary nor a sufficient condition for being a conscious state, and being a dispositional state is neither necessary nor sufficient for being a nonconscious state.

These conceptual clarifications pave the way for a detailed account of the co-evolution of syntactic theory and computational models of language processing.  In the final chapters of Psychosyntax, I lay out the details of context-free grammars and the Earley and CYK parsing algorithms (Jurafsky and Martin, 2008), as well as probabilistic models of language processing, and examine the psychological plausibility of parsers that employ transformational grammars, Augmented Transition Networks (Wanner and Maratsos, 1978), Government and Binding Principles (Johnson 1989), and Minimalist feature-checking constraints (Weinberg 1999; Harkema, 2001). I sketch the grammar underlying each of these models and discuss their commitment to one of a range of positions on the psychological reality issue. Along the way, I flesh out the idea that parsing is a species of natural deduction (Johnson 1991; Shieber, Schabes, and Pereira, 1993).  Within the parsing-as-deduction framework, a grammar can be treated as a set of declaratively represented axioms, which must be accessed and inserted as a “step” in the deduction.  But it can also be seen as a set of embodied inference rules, in accordance with which the deduction proceeds. At present, I argue, there are no decisive reasons for thinking that grammars are declaratively represented as data structures in the mind/brain, rather than embodied as hardwired procedural dispositions.

If it turns out that MSPs are embodied, rather than explicitly represented, then they can be thought of as a subpersonal analogue of standing beliefs. MPMs would then be subpersonal analogues of perceptual thoughts or judgments.  To illustrate the use and plausibility of this analogy, let’ quickly s run through two psychological explanations—one pitched at the personal level and another at the subpersonal level.

At the personal level—from the “intentional stance”, so to speak—we tell stories such as the following: “Ever since he was a kid, Jeff has believed that spiders are dangerous.  So when he just now saw several spiders crawl out of a box in his attic, he reasoned that there are probably more dangerous bugs in the box.”  In this story, the perceptual judgment regarding the presence of spiders interacts with a standing belief to yield an occurrent thought as a conclusion. The process of syntactic parsing, though it takes place subpersonally, may have roughly the same structure.  To illustrate, consider what happens when the human sentence processing mechanism (HSPM) encounters the sentence fragment, Have the sick soldiers…  The HSPM has two standing belief-like states that embody two principles: (i) a sentence-initial auxiliary verb serves to introduce a question; (ii) a subsequent noun phrase will be the subject of the clause.  As a result of encountering the first word of (2), it constructs a mental phrase marker that represents ‘Have’ as an inverted auxiliary verb.  From this MPM and the two standing MSPs, the HSPM “concludes” that that the subsequent noun phrase will be the subject of the clause.

 

Given this analogy between the belief/judgment contrast at the personal level and the MSP/MPM contrast at the subpersonal level, we can address  the psychological reality of MSPs by asking the following question:  Are beliefs, construed in the manner above, psychologically real?  That is, can we attribute a psychological reality to a “mere procedural disposition”?  Can we do so simply on the strength of the fact that it’s a disposition to give rise to an occurrent representation?

Let’s imagine a scenario where we learn that there is nothing more to having the belief that red strawberries are ripe than having the disposition to infer ‘This is ripe’ directly from ‘This is a red strawberry’, ‘It’s red’ directly from ‘It’s a ripe strawberry’, and so forth.  I think we would still grant that the standing state underlying this cluster of dispositions is “psychologically real”.  The procedural nature of this state does not hamper our ability to explain and predict a person’s behavior by positing it.  And given that we have, as yet, no way of identifying the state in neurophysiological terms, it seems that the predictive and explanatory leverage that we derive from classifying it in intentional or semantic terms is enormous.  Moreover, when we do eventually look at the neurophysiological details, we expect—in accordance with the view of embodiment that I endorsed above—to find that there is a common causal factor in the brain that accounts for the aforementioned inferences.  For these reasons, I see no compelling grounds for resisting the ascription of psychological reality to the dispositional states of belief. These remarks carry over, I suggest, to the subpersonal case.  Mental syntactic principles may well turn out to be nothing more than embodied procedural knowledge—a systematic disposition to move directly from one type of mental phrase marker to another.  This would not, however, impugn their explanatory value as theoretical posits.