AI Reveals How Verbs Shape Sentence Structure

Language is full of subtle patterns that shape how we construct sentences, and a new study using artificial intelligence has shed light on one of these long-debated connections. Researchers have developed a computational approach to explore how the meanings of verbs influence their grammatical behavior, specifically focusing on intransitive verbs that don't take a direct object. This work revisits a classic linguistic question: whether verbs describing agentive actions (like 'play') tend to appear in one syntactic structure, while those describing telic events with an endpoint (like 'freeze') appear in another. , published in a preprint, support this link but reveal that human intuitions about these properties can be misleading, highlighting the value of AI-driven s in linguistics.

The key finding is that interpretable dimensions—computational axes created from seed words—can predict the syntactic behavior of verbs more accurately than human ratings. The researchers set up these dimensions for agentivity and telicity by selecting positive and negative seed words, such as {think, you, he, she, causally} versus {affected} for agentivity, and {result, effect, completely, fully, eventually} versus {still, ongoing, being, acting} for telicity. They projected verb embeddings onto these axes to score how agentive or telic each verb is, then compared these scores to syntactic acceptability ratings from a previous study by Kim et al. (2024), which measured how well verbs fit into unaccusative constructions like prenominal participles (e.g., 'the frozen lake'). showed that the seed-based dimensions were superior predictors, with agentivity decreasing and telicity increasing acceptability in these constructions.

Ology involved using word embeddings from GLoVe, pre-trained on Wikipedia and Gigaword, to create interpretable dimensions. The researchers optimized the seed words by testing which ones improved the fit to the syntactic data through linear regression models. They then employed mixed-effects Bayesian ordinal regression models to analyze the relationship between the dimensions and the syntactic ratings, accounting for inter-subject variation. For comparison, they also fitted models using human ratings for agentivity and telicity from Kim et al.'s study, which asked subjects to rate verbs on scales like intentionality and event completion. Additionally, the team introduced a new ranking loss model to fit dimensions to human animacy ratings, allowing them to extrapolate simulated ratings for verbs based on noun data.

Analysis of , as shown in Figures 1 and 2 of the paper, demonstrated that the seed-based model achieved a better fit to the observed syntactic ratings than the ratings-based model. A leave-one-out analysis revealed that the seed-based predictor improved expected log predictive density by 755.4 compared to a null model, while the rating-based predictor improved it by only 232.2. For telicity specifically, the seeds model outperformed the ratings model significantly, with improvements of 575.2 versus 69.2, and test-retest reliability was higher for seed-based measures. The researchers attributed this to human difficulties in rating telicity for verb types in isolation, as evidenced by high inter-speaker variation and context-dependent ratings in other studies. For agentivity, the seeds model also performed better, but the difference was less pronounced, and further analysis linked this to the seeds capturing broader animacy features like the ability to move, rather than just intentional behavior.

Of this research are twofold. Theoretically, it supports the correlation between agentivity/telicity and the unergative/unaccusative distinction in syntax, suggesting that verb meaning does influence grammatical structure. ologically, it shows that interpretable dimensions can overcome limitations of human rating tasks, particularly for abstract properties like telicity that are hard to evaluate out of context. This approach could be applied to other linguistic phenomena where semantic properties are elusive, offering a tool for more robust analysis. For everyday readers, it underscores how AI can uncover hidden patterns in language that even human intuition might miss, potentially aiding in natural language processing applications and deepening our understanding of grammar.

However, the study has limitations, as noted in the paper. Both the seed-based dimensions and human ratings are computed at the verb type level, which may not capture token-level variations where context matters. For example, the verb 'break' was poorly predicted due to its transitive use confounding agentivity scores. The researchers suggest that future work should focus on token embeddings to directly analyze verb occurrences in specific constructions, but current s for interpretable dimensions in token space are underdeveloped. Additionally, the seed selection process relied on optimizing fit to syntactic data, which might not fully represent the semantic concepts, and the animacy analysis was based on noun ratings extrapolated to verbs, potentially introducing noise. These constraints highlight the need for further ological advances to fully disentangle the syntax-semantics interface.