Why is "Chicago" Predictive of Deceptive Reviews? Using LLMs to Discover Language Phenomena from Lexical Cues

Technical Summary: "Why is 'Chicago' Predictive of Deceptive Reviews?"

Problem Statement
Deceptive online reviews mislead consumers and undermine trust in marketplaces. While machine learning classifiers can effectively distinguish deceptive reviews from genuine ones using large datasets, the features they rely on are often subtle, fragmented, and unintuitive to humans. For instance, classifiers may identify specific words like "Chicago" as highly predictive of deception in hotel reviews. However, without an explanation of why these words are predictive, users may struggle to interpret the model's logic, potentially leading to confusion or a loss of trust in the system. The core challenge is translating these unintuitive, machine-learned lexical cues into human-understandable language phenomena that are grounded in data rather than being mere artifacts or hallucinations.

Methodology: A Conjecture-then-Validate Framework
The authors propose a framework where Large Language Models (LLMs) act as interpreters to bridge the gap between predictive lexical cues and human-understandable explanations. The process involves three main stages:

1. Feature Identification: Using a logistic regression classifier trained on a dataset of Chicago hotel reviews (800 genuine, 800 deceptive), the authors identified the most salient predictive words (e.g., "Chicago," "luxury," "seemed") that appeared unintuitive to laypeople.
2. Conjecture Generation: The authors prompted LLMs (including GPT-5-mini, Haiku 4.5, Nova Pro, and Gemini-2.5-flash) to generate natural language explanations ("conjectures") for why these specific words are predictive of deception or genuineness. The prompt asked the LLM to identify underlying language phenomena or psycholinguistic patterns associated with the words.
3. Algorithmic Validation: To ensure these conjectures were not hallucinations, the authors evaluated them empirically:
   • Predictiveness (RQ1): They tested whether providing these conjectured phenomena as context improved an LLM's ability to classify reviews as genuine or deceptive compared to a zero-shot baseline.
   • Generalizability (RQ2): They trained classifiers on Chicago data and tested them on hotel reviews from other cities (Houston, New York, Los Angeles). They compared performance using standard unigram TF-IDF features versus features derived from the LLM-conjectured phenomena.
   • Source Comparison (RQ3): They compared phenomena conjectured from predictive words against phenomena conjectured from the LLM's prior knowledge or from in-context learning (sampled reviews).
4. Human Evaluation: A crowdsourced study (N=220) and an in-person lab study (N=8) assessed how these explanations affected user reasoning and trust. Participants were asked to judge review credibility with either only predictive words highlighted or with the words accompanied by the LLM-conjectured phenomena.

Key Results

• Predictive Power (RQ1): LLMs provided with the conjectured phenomena consistently outperformed zero-shot baselines in classifying deceptive reviews. This indicates that the generated phenomena capture real, detectable patterns in the data rather than being entirely fabricated. However, even with these phenomena, LLM performance remained lower than a simple logistic regression classifier using raw text features.
• Generalization (RQ2): Classifiers using phenomena-based features demonstrated superior generalization to out-of-distribution data (reviews from other cities) compared to classifiers using only unigram features. For example, while the word "Chicago" is a strong predictor for Chicago reviews, it is not predictive for New York reviews. In contrast, the phenomenon "deceptive reviews tend to name-drop hotel and city names" remained a valid predictor across different cities.
• Source of Phenomena (RQ3): Phenomena conjectured based on the specific predictive words identified by the classifier were significantly more effective for deception detection than phenomena conjectured based solely on the LLM's prior knowledge or a small sample of reviews. This suggests that the classifier's distilled signals anchor the LLM's reasoning toward subtle, data-specific patterns that are difficult to infer otherwise.
• Human Impact: In user studies, participants provided with phenomena-based explanations shifted from guessing based on isolated words to reasoning about higher-level patterns. Crucially, users did not blindly accept the explanations; they compared them against their own heuristics. This led to "calibrated trust," where users selectively relied on the model's reasoning rather than accepting or rejecting it outright.

Significance and Contributions
The paper claims to make two primary contributions:

1. Framework for Interpretability: It proposes and validates a "conjecture-then-validate" framework that successfully translates unintuitive, machine-learned lexical cues into empirically grounded, human-interpretable language phenomena. This approach moves beyond simply highlighting predictive words to explaining the underlying linguistic patterns.
2. Bridging the Gap in Trust: The work demonstrates that while feature importance explanations (highlighting words) identify signals, they often fail to explain why those signals matter. By generating explanations that connect these signals to broader patterns, the approach improves user understanding and supports calibrated trust in NLP systems.

The authors emphasize that their goal is not to build a superior deception detector, but to evaluate the feasibility of using LLMs to make model-identified features interpretable. They note that while the generated phenomena are empirically grounded, they may represent simplified interpretations rather than complete causal mechanisms, and future work is needed to assess the faithfulness of these explanations more rigorously.