Patterns of Priming in Production: Lexical, Semantic and Structural Alignment in Language Model Generation
Language models exhibit structural priming in production, where preceding structural context influences sentence completion choices Priming effects are strongest in semantically coherent sentences, with lexico-semantic coherence boosting structural alignment Inverse frequency effects observed: greater relative increase in less frequent double-object datives, but larger absolute increase in more frequent prepositional-object constructions Structurally primed completions show elevated lexico-seman
Analysis
TL;DR
- Language models exhibit structural priming in production, where preceding structural context influences sentence completion choices
- Priming effects are strongest in semantically coherent sentences, with lexico-semantic coherence boosting structural alignment
- Inverse frequency effects observed: greater relative increase in less frequent double-object datives, but larger absolute increase in more frequent prepositional-object constructions
- Structurally primed completions show elevated lexico-semantic repetition, indicating cross-level alignment
- Evidence supports that structural priming in LMs operates across syntactic, lexical, and semantic representation levels simultaneously
Why It Matters
This research bridges a critical gap between psycholinguistic findings on structural priming in human comprehension and their manifestation in AI language model production. For AI practitioners, understanding these alignment patterns is essential for controlling generation behavior, improving prompt design, and building more predictable and interpretable language systems. The findings also inform debates about whether LMs genuinely acquire linguistic structure or merely simulate it through statistical patterns.
Technical Details
- Controlled sentence-completion experiments focused on dative construction alternations (double-object vs. prepositional-object) to test structural priming in LM production
- Compared primed conditions against baseline conditions to measure relative and absolute priming magnitudes across construction types
- Analyzed the interaction between structural priming and lexico-semantic coherence, finding that coherence amplifies priming effects
- Examined repetition patterns in primed completions, demonstrating that structural alignment co-occurs with lexical and semantic alignment
- Published on arXiv (2609.04484) in the Computation and Language (cs.CL) and Artificial Intelligence (cs.AI) categories
Industry Insight
- Prompt engineering strategies should account for structural priming effects, as prior context can systematically bias generation toward specific syntactic patterns, affecting output consistency in production systems
- The cross-level alignment phenomenon suggests that interventions targeting lexical or semantic consistency may indirectly influence syntactic choices, enabling more holistic control over generation behavior
- As LMs are deployed in high-stakes language tasks, understanding priming mechanisms is critical for detecting and mitigating unwanted bias propagation through structural repetition in conversational and generative pipelines
Disclaimer: The above content is generated by AI and is for reference only.