The bleaching of syntax
András Kornai
At the time of the linguistic wars, both sides believed themselves to be fighting for a valuable prize, the crown jewels of grammar, syntax. The generative semanticists saw syntax as an assembly line that creates well-formed surface propositions from the underlying raw units of meaning, while Chomsky saw it more like a transportation system whereby the meaning already present at LF is delivered from speaker to hearer in language-specific ways ways determined by a genetically encoded language acquisition device.
Either way, syntax was seen central the same way an earlier generation of linguists, Sapir, Whorf, Pike, considered morphology to be central for shaping thought and form alike. But what if syntax is extremely limited, effectively restricted to LFG-like f-structure, with c-structure dictated by finite-depth recursion, CFGs that can be compiled out to regular expressions? In this talk we present such a `bleached' version of syntax essentially devoid of the very functions either or both camps of the warring linguists aimed at. Consider:
- Snow is white
- Meat is murder
By and large, any good syntactic parser of English will analyze (1A) and (1B) as copulative sentences that identify the nominal (+N -V) constituent preceding the (+N -v) one following it. But (1A) is uncontroversial, whereas (1B) has a strong moral import, offering powerful rhetorical support for a vegetarian stance. The A/B minimal pair is evidently not powered by syntax but by lexical information. The mainstream syntacto-semantic approach, which concludes, after a great deal of computation, that snow' ⊂ white' and meat' ⊂ murder' has very little traction over the interesting cases, even though it does something superficially interesting for at least three professors flunking at most seven students in more than one subject.
In this talk, we well sketch a research program half of which is already put in practice by Anthropic's approach to word meaning (see Scaling Monosemanticity). We suggest that classical syntactic features (typically with morphological exponents) like case, both deep and surface, correspond to deltas of such lexical features, instructions for particular linking operations that are operating on the (static) lexical representations in the manner of tagmemic slot-filling, LFG's f-structure, or type-logical grammar's variable binding term operations. We present a calculus in terms of tensor products that will be familiar to the current generation of computational linguists working on LLMs, but perhaps less familiar to generative syntacticians.