The signal is coming from inside the noun phrase! Tracking semantic proto-role inferences during sentence processing

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Expert, quick

Summary

Researchers introduce Generalized Contextual Decomposition for Transformers (GCD-T), a novel LLM feature attribution method. This technique is specifically designed to investigate how large language models infer semantic proto-role properties, such as Instigation, which represent decomposed semantic roles between a predicate and its argument. GCD-T probes which specific parts of a sentence enable these models to make such inferences. The study's findings, derived from applying GCD-T to LLMs, are then compared with human inferences to understand the mechanisms underlying semantic role processing in both artificial and natural intelligence. This comparison aims to highlight how models attribute and understand complex semantic relationships within sentences.

Key takeaway

For NLP Engineers focused on LLM interpretability, understanding how models infer semantic roles is crucial. This research introduces GCD-T, a method you can use to pinpoint which sentence components drive proto-role property inferences. Applying such attribution techniques helps you diagnose model behavior, compare it against human linguistic processing, and ultimately build more robust and transparent language models.

Key insights

GCD-T is a new LLM attribution method for tracking semantic proto-role inferences within sentence processing.

Principles

• Semantic roles decompose into proto-role properties.
• LLM feature attribution can reveal inference mechanisms.
• Comparing LLM and human inferences is valuable.

Method

Generalized Contextual Decomposition for Transformers (GCD-T) attributes LLM features to specific sentence parts, revealing their contribution to proto-role property inferences. This method then compares these LLM inferences with human judgments.

In practice

• Analyze LLM semantic role understanding.
• Compare model and human linguistic processing.
• Improve interpretability of LLM outputs.

Topics

• LLM Feature Attribution
• Semantic Proto-roles
• GCD-T
• Sentence Processing
• Model Interpretability
• Human-AI Comparison

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• Probing Semantic Alignment, Lexical Invariance, and Syntactic Influence in LLM Metaphor Processing — LLMs struggle with deep metaphor comprehension, often misinterpreting concepts and relying on lexical associations over context or full syntactic understanding.
• Revisiting Semantic Role Labeling: Efficient Structured Inference with Dependency-Informed Analysis — A new SRL framework offers explicit predicate-argument structure and 10x faster inference with modern encoders.
• A Cross-Genre Analysis of Discourse Relation Signaling in the GUM Corpus — The paper analyzes discourse relation signaling across diverse genres in the GUM Corpus.
• Beyond Surface Forms: A Comprehensive, Mechanism-Oriented Taxonomy of Indirect Linguistic Encoding for LLM-Based Coded Language Detection — A mechanism-oriented taxonomy of indirect linguistic encoding significantly enhances LLM-based coded language detection.
• A Scalable Tool for Measuring Manner and Result Verbs in Developmental Language Research — A RoBERTa classifier, trained on LLM-generated annotations, accurately distinguishes manner and result verbs at scale.
• Do Language Models Know Theo Has a Wife? Investigating the Proviso Problem — LLMs align with human judgments on the proviso problem but use shallow pattern matching, not deep pragmatic reasoning.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.