Comparing Human and Large Language Model Interpretation of Implicit Information

1. Problem Statement

While Large Language Models (LLMs) excel at generating human-like text, their ability to interpret implicit information (meaning not explicitly stated but inferable) differs from human communication. Human communication relies on "interpretive cooperation," where readers actively construct meaning based on context, shared knowledge, and social cues. The paper addresses a critical gap: there is little research specifically focused on extracting implicit information from text using LLMs, nor is there a systematic comparison of how LLMs and humans differ in their inference strategies, strictness, and coverage.

The authors define the task as Implicit Information Extraction (IIE), aiming to build a structured representation of a model's understanding of a context sentence and compare it against human consensus.

2. Methodology

The authors propose a fully automated, model-agnostic pipeline designed to extract, validate, and structure implicit information into a Knowledge Graph (KG). The pipeline operates via few-shot prompting and does not require fine-tuning or access to model parameters, making it suitable for black-box LLMs.

The pipeline consists of three main stages:

A. Information Extraction

The model processes a context sentence to extract:

1. Entities: Tagged with a specific ontology (e.g., person, organization, object).
2. Explicit Triplets: Relational triplets $(Subject, Relation, Object)$ directly supported by the text.
3. Implicit Triplets: Triplets deduced from the text using commonsense reasoning. These are categorized by inference types adapted from the ATOMIC taxonomy (e.g., pre-conditions, post-conditions, intents, emotional reactions).
   • Handling Complexity: The pipeline uses RDF reification to handle subordination and aspectual verbs (e.g., "Chris stopped talking" becomes a nested triplet where the action is the object), allowing for recursive structures.

B. Inference Validation

To improve precision, the pipeline acts as its own critic:

1. Inference Challenge: The model evaluates each implicit triplet to determine if it is "deducible" or "unsupported." If unsupported, it must provide a justification.
2. Inference Correction: If a triplet is discarded, the model attempts to correct it to make it valid. If correction is impossible, the triplet is rejected.
3. Duplicate Removal & Explanation: Redundant triplets are removed, and the model is asked to provide the explicit premises that support its implicit inferences.

C. Temporal Analysis

The final stage analyzes the temporal relationships between extracted triplets:

1. Classification: Triplets are labeled as Events (occurring situations) or States (conditions).
2. Temporal Grounding: Absolute time references are extracted where available.
3. Relation Extraction: Pairwise temporal relations are determined using a taxonomy: before, after, while, or none. To mitigate hallucination, the model evaluates pairs in both orders, and a relation is only assigned if the results are consistent.

3. Experimental Setup

• Models Evaluated: Mistral Large 2 and GPT-4o mini.
• Datasets:
  • SocialIQA: 15 sentences focusing on social interactions and commonsense reasoning.
  • COPA (Choice of Plausible Alternatives): 15 sentences focusing on causal reasoning, generally shorter and more fact-oriented.
• Human Evaluation:
  • Participants: 205 university students (SocialIQA) and 75 MTurk workers (COPA).
  • Task: Annotators performed tasks mirroring the pipeline: classifying triplets (Factual, Deducible, Wrong), reviewing model discard decisions, classifying events/states, judging temporal relations, and adding missing triplets.
• Metrics: Model-Human Agreement (MHA), Cohen's $\kappa$ for inter-rater reliability, and analysis of coverage gaps (additions vs. removals).

4. Key Results

A. Agreement and Coverage

• Moderate Agreement: Overall MHA is moderate (ranging from ~55% to 75% depending on the task).
• Coverage Gap: Humans consistently propose significantly more triplets than the models. The median number of human additions was 13–16 per sentence, compared to 5–12 generated by models. This indicates current LLMs have limited coverage in implicit extraction.
• Hallucinations: Post-validation hallucinations were rare (0% in most cases), suggesting the validation stage effectively filters out unsupported inferences.

B. Strictness and Inference Drivers (RQ1 & RQ2)

• Context-Dependent Strictness:
  • SocialIQA (Rich Context): Models were more conservative than humans. They frequently labeled triplets as "Wrong" that humans considered "Factual" or "Deducible." Humans were more willing to infer social nuances and intents.
  • COPA (Fact-Oriented): The gap narrowed. Humans became stricter, often rejecting speculative inferences in short, fact-heavy contexts. Models and humans showed more aligned label distributions here.
• Inference Types: Models under-generated specific inference types, particularly pre-conditions, post-conditions, and attributes, which humans frequently added.

C. Temporal Reasoning (RQ3)

• Event/State Classification: High agreement between models and humans (>86%).
• Temporal Relations: This was the weakest area (lowest MHA, ~33–60%).
  • Models were highly conservative, frequently selecting "No clear relation" when temporal cues were implicit.
  • Humans readily inferred event ordering from commonsense sequences, whereas models struggled to chain implicit temporal cues.

D. Human Consensus (RQ4)

• Human consensus was statistically significant and reliable across all tasks.
• Disagreement was concentrated in subjective steps (judging inference support and recovering implicit temporal structures), suggesting these areas have multiple plausible interpretations even for careful annotators.

5. Key Contributions

1. Task Definition: Formalized Implicit Information Extraction (IIE) as a distinct task from standard Open Information Extraction (OIE) or Natural Language Inference (NLI).
2. Pipeline Architecture: Developed a robust, black-box compatible pipeline that extracts, validates, and structures implicit knowledge into a two-tier Knowledge Graph (relational + temporal).
3. Empirical Comparison: Provided the first large-scale, quantitative comparison of LLM vs. human implicit reasoning, revealing that LLMs are more conservative in social contexts while humans become more conservative in fact-oriented contexts.
4. Identification of Bottlenecks: Highlighted that temporal reasoning and coverage of specific inference types (like intents and pre-conditions) are current limitations for LLMs compared to human cognition.

6. Significance

This work is significant for several reasons:

• Interpretability: It sheds light on the "black box" of how LLMs process implicit meaning, revealing that they do not simply mimic human interpretive cooperation but often apply stricter, literal filters.
• Trust and Safety: Understanding where models fail to infer implicit social cues or temporal logic is crucial for deploying LLMs in high-stakes fields like law, journalism, and scientific research.
• Future Directions: The findings suggest that future IIE systems need to be tuned for specific domains (social vs. factual) and that current models require better mechanisms for recovering implicit temporal structures and commonsense pre/post-conditions.

The paper concludes that while LLMs are powerful tools for information extraction, they currently lack the full "interpretive cooperation" capability of humans, particularly in socially rich contexts where implicit meaning is dense and nuanced.