Do LLMs Think Fast and Slow? A Causal Study on Sentiment Analysis (2404.11055v2)

Abstract: Sentiment analysis (SA) aims to identify the sentiment expressed in a text, such as a product review. Given a review and the sentiment associated with it, this work formulates SA as a combination of two tasks: (1) a causal discovery task that distinguishes whether a review "primes" the sentiment (Causal Hypothesis C1), or the sentiment "primes" the review (Causal Hypothesis C2); and (2) the traditional prediction task to model the sentiment using the review as input. Using the peak-end rule in psychology, we classify a sample as C1 if its overall sentiment score approximates an average of all the sentence-level sentiments in the review, and C2 if the overall sentiment score approximates an average of the peak and end sentiments. For the prediction task, we use the discovered causal mechanisms behind the samples to improve LLM performance by proposing causal prompts that give the models an inductive bias of the underlying causal graph, leading to substantial improvements by up to 32.13 F1 points on zero-shot five-class SA. Our code is at https://github.com/cogito233/causal-sa

• The paper demonstrates that incorporating causally informed prompts significantly enhances sentiment analysis performance via fast (C2) and slow (C1) thinking.
• It applies the peak-end rule to distinguish causal directions between review content and sentiment, achieving a 32.13 point improvement in F1 score in zero-shot scenarios.
• The study bridges psychological theories with LLM mechanistic interpretability, providing actionable insights for refining machine learning models in sentiment analysis.

Insights from Natural Language Processing: Unpacking the Causal Relationships in Sentiment Analysis

Introduction to the Study of Causal Relationships in Sentiment Analysis

This paper introduces a novel approach to sentiment analysis (SA) by integrating causal discovery with traditional prediction tasks to enhance the performance of LLMs. By acknowledging two possible causal hypotheses—either the sentiment influences the review content (C2), or the review content generates the sentiment (C1)—the research investigates the applicability of psychological theories like the peak-end rule to classify causal relationships in SA data.

Causal Discovery in Sentiment Analysis

Problem Setup and Causal Hypotheses

Drawing from well-established psychological findings, this paper treats SA as unveiling the causal direction between a review (X) and its sentiment (Y). Two primary hypotheses are considered:

1. Causal Hypothesis C1 (Slow Thinking): Here, the review primes the sentiment, representing a reasoned response typical of slow cognitive processing.
2. Causal Hypothesis C2 (Fast Thinking): Conversely, the sentiment primes the creation of the review, indicative of rapid, instinctual cognitive reactions.

To identify the causal direction in real-world datasets (like Yelp, Amazon), the paper applies the peak-end rule, categorizing reviews into C1 and C2 based on how closely the overall sentiment score approximates the average versus the peak and end sentiments.

Implications for Sentiment Analysis Using LLMs

Predictive Performance Enhancements

Upon determining the predominant causal direction of data samples, causal mechanisms were implemented to guide LLMs through tailored causal prompts, significantly enhancing sentiment analysis efficacy. Noteworthy gains include substantial improvements in F1 score, around 32.13 points, in zero-shot scenarios across five classes of sentiment.

Mechanistic Understanding by Models

The paper also explores if LLMs, when directed with causally aware prompts, can genuinely grasp the underlying causal dynamics. Through mechanistic interpretability methods like causal tracing, the paper reveals the degree to which these models attend to components in sentiment-laden texts in alignment with learned causal structures (C1 or C2).

Observations and Future Directions

Findings suggest differential capabilities of LLMs in capturing the essence of causal dynamics applied through new prompting strategies. While models showed improved performance in alignment with psychological theories when proper causal prompts were used, there remains potential for deeper understanding and usage of these cognitive processing theories in machine learning frameworks. The exploration paves the way for enriched models that more closely resemble nuanced human cognitive and emotional processes.

Conclusions

This research marks a significant stride in bridging psychological insights with machine learning, particularly in the domain of sentiment analysis. By leveraging causal discovery grounded in psychology, the paper not only enhances the predictive performance of LLMs but also enriches our understanding of how complex, realistic datasets can be approached from a causally-informative perspective. Future explorations could expand these insights to multilingual datasets or incorporate more intricate causal models involving additional variables like contextual or demographic factors.

The broad applicability and the potential for fine-tuned, causally aware models suggest a promising direction for future NLP applications, extending beyond sentiment analysis to other areas where understanding the directionality of influence is crucial.