Linguistic Productivity in Large Language Models: Models Coerce, but do not Preempt

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Linguistic Productivity in Large Language Models".

Jane: Here's the gist of the paper. The gist: Large Language Models recognize and can reproduce with nonce words constructional productivity (entrenchment) in cases of coercion, but they do not extend negative evidence to novel language, and statistical preemption does not enable models to avoid overgeneralization of patterns that are semantically felicitous, but never observed in data.

Tom: Usage-Based Theoretical Framework Usage-based theories posit that creative productivity of the structures of language is both bolstered and constrained by two distinct frequency signals: entrenchment, stemming from high frequency usage, and preemption, stemming from having never observed a particular linguistic structure in a context where one might expect that structure to appear The success of Large Language Models (LLMs) in a variety of tasks can be understood to signal the validity of the assumption that exposure to the patterns of language in the form of text alone is sufficient to build a functional model of language in which one need not posit grammatical rules apriori Usage-based grammatical theories, such as Construction Grammar (CxG) (e.g., Goldberg (1995); Croft (2001); Bybee (2010)), similarly posit that language usage engenders natural language learning such that frequency of exposure supports entrenchment and generalization of the structures of a language Under many of such theories, the usage-based development of grammar eschews the need to posit syntactic rules: instead of learning and memorizing lexical items and a limited number of syntactic rules guiding their grammatical combination, speakers acquire and learn un-analyzed holophrases of language that they are most frequently exposed to Therefore, frequency of exposure to a particular construct (token of a particular construction type) of language, it is theorized, plays an important role in the development of grammar In contrast to Generative Grammar theories that posit speakers only memorize minimal information (lexical items and their meanings, a small, reusable set of syntactic rules guiding the grammatical combination of those items), usage-based theories posit that speakers store, classify, and cluster a tremendous amount of information with each token of linguistic experience Under a Construction Grammar approach, these tokens of linguistic experience are form-meaning pairings that may be words, phrases, or even sub-parts of words, known as constructions As a result of storing an array of contextual information with each linguistic exposure, speakers build a grammar beginning with frequently heard linguistic forms associated with frequent experiences In this way, frequency supports the entrenchment of certain structures and patterns of language However, it is clear that there are many constructions of language that are not infinitely productive, and in fact there are some patterns that are semiproductive, wherein the productivity is seemingly idiosyncratic For example, the English Adjectivephrase construction is highly productive: the red ball; that big building; a fun party However, there is a class of English adjectives that are not productive extensions of this pattern:?the asleep cat;?that aloft balloon;?an ashamed look There is, however, a competing viable alternative for expressing that a particular thing has a particular characteristic in English: the ball is red; the cat is asleep Thus, there is negative evidence blocking the usage of a-adjectives in the Adjective-phrase construction, and positive evidence supporting the usage of a-adjectives in the Predicate-adjective construction Therefore, under a usage-based, CxG view, generalization of linguistic structures operates under the influence of positive evidence and negative evidence Coercion Coercion was first introduced by Pustejovsky (1991), who describes “type coercion” as a formal semantic operation in which a syntactic context requires a particular semantic type, and if the filler doesn’t match that type, it gets coerced into the required type via mechanisms in the lexical entry In this research, we adopt the definition of coercion outlined in Michaelis (2004), who brings a CxG lens to coercion, and posits the Override Principle: “If a lexical item is semantically incompatible with its morphosyntactic context, the meaning of the lexical item conforms to the meaning of the structure in which it is embedded” Thus, constructional semantics will win out given a clash between a construction and a lexical item Whereas Michaelis (2004) focuses primarily on aspectual coercion, we focus on the coercive process that enables type shifting from container to contents, described in detail in Radden and Kövecses (1999) as a common metonymic conceptual relationship Thus, we commonly see metonymies which target the contents via the container: This is an excellent bottle (referring to the wine within the bottle) Notice that interpretation and deployment of coercive constructions requires a fair amount of world knowledge that supports conceptual clusters such as containers and contents, but also possessor for possessed, and part-whole metonymy This is consistent with a cognitive view of word meaning in which concepts are not defined according to necessary and sufficient conditions, but instead by clusters of metaphors, where “each metaphor highlights certain aspects of the concept and implicitly hides others” Thus, coercion facilitates an atypical, but plausible interpretation of a lexical item in a particular constructional context Furthermore, coercion is one process supporting the productive generalization of constructions to novel, previously unseen instantiations Statistical Preemption While psycholinguistic literature has demonstrated that more frequent forms are more likely to serve as the basis for productive generalization (Bybee and Eddington, 2006) through processes like coercion, what prevents speakers from overgeneralizing?

Jane: Experimental evidence points to the effect of statistical preemption: An inference that speakers make from repeatedly hearing form B, when a semantically and pragmatically appropriate expression A could have been used, that B is the appropriate formulation while A is not (Suttle and Goldberg, 2011) This process has been experimentally demonstrated in studies showing that this is how children eschew overgeneralization of certain morphological Cxns As discussed earlier, consider, for example, the English past tense ending “-ed” on highfrequency irregular verbs: Having repeatedly heard “went” in the contexts where “go-ed” might be expected is negative, indirect evidence that “go-ed” is not appropriate (Aronoff, 1976; Kiparsky, 1982) The process of statistical preemption at the phrasal level is more complex because it is not clear what is semantically and pragmatically similar enough to preempt another phrasal form Nonetheless, compelling experimental evidence for statistical preemption has come from several psycholinguistic studies, including (Boyd et al., 2012) as well as (Theakston, 2004) Theakston (2004) concludes that entrenchment operates independently from preemption in a task where both children and adults are asked to provide grammaticality ratings for sentences with argument structure errors, involving both high and low-frequency verbs Our research most directly draws upon the experimental evidence of Boyd and Goldberg (2011), who discuss the phenomenon of blocking aadjectives (e.g., awake, aware, ashamed, awash) in the pre-nominal position of the English Adjectivephrase construction (e.g.,?The awake child) In this research, speakers track the fact that although adjectives can generally sit in either the prenominal position of the English Adjective-phrase construction (e.g., the blue bucket) or in the post-copula position of the Predicate-adjective construction (e.

Title and authors: Tom: Let's start with the title and authors of this paper, "Linguistic Productivity in Large Language Models: Models Coerce, but do not Preempt". It’s by Claire Bonial, Claire Benet Post, Laura Michaelis, and Harish Tayyar Madabushi.

Jane: That sounds like a deep dive into how language actually works. I wonder what the title itself hints at about the main discovery.

Lu: The title is really telling us that models can coerce things—meaning they force a word to fit a certain pattern—but they don't use preemption signals, which means they can't effectively block themselves from making mistakes when those patterns aren't in the data.

Meng: So the main point is that coercion is something these models get pretty good at doing when forced into a context, but preemption isn't a reliable guard against overgeneralization.

Lalam: Exactly. It points to the fact that while models learn constructions through exposure, they don't use those negative evidence signals we talked about in human language to stop themselves from making things up that sound plausible.

The paper's summary: Tom: So, what does the actual summary of "Linguistic Productivity in Large Language Models: Models Coerce, but do not Preempt" actually say? It sets up the framework of entrenchment and preemption as the two main signals for creativity.

Jane: They show that LLMs can reproduce constructional productivity when they are under coercion, which is one way speakers extend language to new uses, like a container being interpreted as its contents.

Lu: But then they test the second signal, statistical preemption, and find that models don't use this negative evidence to constrain their output from overgeneralizing patterns that aren't observed in the data.

Meng: So even though we know in human language we block those overly broad generalizations based on things that never happen, the AI doesn't do it automatically.

Lalam: That’s the key finding. The experiments show models acquire constructional categories, but they don't use them as preemption signals to avoid errors when those structures are novel and haven't been seen before.

The paper's improvements: Tom: Now, the authors suggest some ways to improve this understanding of how LLMs handle language productivity. They propose leveraging both positive and negative evidence in linguistic generalization.

Jane: That means they think we should be looking at both what the model sees and what it *doesn't* see when it tries to create a new sentence structure.

Lu: They present two experiments: one tests coercion, showing how models interpret semantics of coercive constructions, like "I drank the bottle," where the construction forces a meaning onto the lexical item.

Meng: And then they have the statistical preemption experiment, which looks at whether models can attend to that negative evidence—the context where a structure is viable but never occurs.

Lalam: The paper suggests these two experiments show how models learn to generalize and how they fail to avoid overgeneralization when the structure is novel.

Conclusion: Tom: So, wrapping up the paper "Linguistic Productivity in Large Language Models: Models Coerce, but do not Preempt", the main implication is that LLMs are good at following rules when forced into a pattern, but they lack an internal mechanism to use statistical preemption to stop themselves from generating patterns that sound okay but never actually appear in the training data.

Jane: It means we need to think about how these models learn structure not just from what's there, but also by recognizing what's missing. That’s a big shift in how we view language learning in AI systems.

Lu: Yeah, it suggests that for true generalization in language, we need to bake in that mechanism where models actively look for negative evidence to constrain their choices.

Meng: From a deployment standpoint, this means if we want reliable generation on novel inputs, just relying on the model's sheer size or training data exposure isn't enough; we need a better way to handle those constraints.

Lalam: The paper shows that without preemption, models can still get away with creating grammatically plausible but contextually impossible sentences if the coercion works in their favor.

Tom: So we see they can coerce things into place, but they don't use preemption to stop them from overgeneralizing when those patterns aren't in the data. That’s the main point of "Linguistic Productivity in Large Language Models: Models Coerce, but do not Preempt".

Jane: That’s a fascinating look at the limitations of current generative models. We'll be back next time to talk about something completely different.