RVISA: Reasoning and Verification for Implicit Sentiment Analysis

TL;DR

Implicit sentiment analysis is challenged by the absence of explicit polarity cues. RVISA couples a decoder-only LLM to generate structured, three-hop rationales and an encoder-decoder backbone fine-tuned with multi-task learning and an answer-based verification signal to ensure reasoning quality. The approach achieves state-of-the-art ISA performance on Restaurant and Laptop datasets, withGPT-3.5-turbo-driven rationales providing the strongest guidance. By uniting generation and structured reasoning across model families, RVISA offers a scalable path to reliable implicit sentiment inference.

Abstract

With an increasing social demand for fine-grained sentiment analysis (SA), implicit sentiment analysis (ISA) poses a significant challenge with the absence of salient cue words in expressions. It necessitates reliable reasoning to understand how the sentiment is aroused and thus determine implicit sentiments. In the era of Large Language Models (LLMs), Encoder-Decoder (ED) LLMs have gained popularity to serve as backbone models for SA applications, considering impressive text comprehension and reasoning ability among diverse tasks. On the other hand, Decoder-only (DO) LLMs exhibit superior natural language generation and in-context learning capabilities. However, their responses may contain misleading or inaccurate information. To identify implicit sentiment with reliable reasoning, this study proposes RVISA, a two-stage reasoning framework that harnesses the generation ability of DO LLMs and the reasoning ability of ED LLMs to train an enhanced reasoner. Specifically, we adopt three-hop reasoning prompting to explicitly furnish sentiment elements as cues. The generated rationales are utilized to fine-tune an ED LLM into a skilled reasoner. Additionally, we develop a straightforward yet effective verification mechanism to ensure the reliability of the reasoning learning. We evaluated the proposed method on two benchmark datasets and achieved state-of-the-art results in ISA performance.

Figures (6)

• Figure 1: Different LLMs demonstrate diverse reasoning abilities for implicit sentiment analysis. It is conducive to detecting implicit sentiment polarities by explicitly inferring sentiment elements as rationale but verification is required to ensure reliability.
• Figure 2: Reasoning promptings applying to sentiment analysis. Left: commonly used prompting modes. Right: three-hop prompting for ISA.
• Figure 3: The overview of proposed two-stage reasoning framework RVISA. Left: rationale generation stage leveraging DO LLM to prepare effective rationales and corresponding answer-based verification signals. Right: multi-task fine-tuning stage to train an ED backbone model as an enhanced reasoner with additional explanation tasks along with verification supervision.
• Figure 4: The impact of diverse rationales and different model sizes on implicit F1 score. The dashed horizontal line represents the best result of THOR rerun with the Flan-T5-XXL(11B) model on the implicit dataset.
• Figure 5: Error analysis for two datasets with rationales generated by GPT-3.5-turbo. The error ratio here refers to the proportion of the number of error types to the total number of error instances.