Comparing Human and Large Language Model Interpretation of Implicit Information¶
Conference: ACL 2026 arXiv: 2604.17085 Code: Available (link in paper) Area: Knowledge Graph / Implicit Information Understanding Keywords: Implicit Information Extraction, Knowledge Graph, Human-AI Comparison, Reasoning Verification, Temporal Analysis
TL;DR¶
This paper proposes the Implicit Information Extraction (IIE) task and a three-stage LLM pipeline (information extraction → reasoning verification → temporal analysis), building structured knowledge graphs to represent implicit textual meaning. Crowdsourced human comparisons reveal LLMs are more conservative in socially-rich contexts but humans are more conservative in short factual contexts.
Method¶
Key Designs¶
-
ATOMIC-Based Implicit Reasoning Types: Guides LLMs to systematically infer implicit information through structured reasoning types: preconditions, postconditions, participant intentions, emotional reactions, perceived attributes.
-
Reasoning Verification (Self-Critique + Correction): Model reviews each implicit triple for textual support, with up to 3 correction rounds.
-
Nested Triples (RDF Reification-Inspired): Handles subordinate clauses and modal verbs through recursive nesting.
Key Experimental Results¶
- Humans agree with most LLM-extracted triples but consistently suggest substantial supplements — indicating limited coverage of LLM implicit reasoning
- LLMs are more conservative in socially-rich contexts; humans are more conservative in short factual contexts
- Temporal reasoning is a weak point for LLMs
Highlights & Insights¶
- Formalizing implicit information understanding as a knowledge graph construction task provides a quantitatively comparable framework
- The context-dependent conservatism finding offers new perspective for understanding human-AI differences
Rating¶
- Novelty: ⭐⭐⭐⭐
- Experimental Thoroughness: ⭐⭐⭐⭐
- Writing Quality: ⭐⭐⭐⭐
- Value: ⭐⭐⭐⭐