This research investigates structural priming in language model production, focusing on how prior structural context impacts sentence completion. It addresses whether priming effects observed in comprehension also occur during generation, where models sample from multiple continuations.
Controlled experiments on dative constructions reveal that language models are susceptible to structural priming, especially in semantically coherent sentences. The priming effect varies with construction frequency, showing a larger absolute increase in more common structures like prepositional objects.
The findings demonstrate that priming is amplified by lexico-semantic coherence and that primed completions tend to repeat lexical and semantic elements. This suggests that priming operates across multiple linguistic levels, influencing syntactic, lexical, and semantic alignment during generation.
Understanding these effects can inform the development and evaluation of language models, particularly in tasks requiring syntactic consistency and semantic coherence.
Source: https://arxiv.org/abs/2609.04484
