Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Reasoning Implicit Sentiment with Chain-of-Thought Prompting

Hao Fei, Bobo Li, Qian Liu, Lidong Bing, Fei Li, Tat-Seng Chua

TL;DR

This work tackles implicit sentiment analysis by requiring multi-hop reasoning to uncover latent aspects and opinions. It introduces THOR, a Three-hop Reasoning prompting framework that sequentially elicits aspect, opinion, and polarity via chain-of-thought prompts, aided by self-consistency and supervision-based revisions. Empirical results on SemEval14 ISA show THOR with Flan-T5-11B achieving ~6% F1 gains in supervised settings and THOR with GPT-3 175B delivering ~50% F1 gains in zero-shot scenarios, with benefits amplifying as model size grows. The approach demonstrates strong improvements over prior baselines and highlights the potential of scalable CoT reasoning for nuanced NLP tasks, while noting limitations for mid-sized LLMs and annotation-related challenges.

Abstract

While sentiment analysis systems try to determine the sentiment polarities of given targets based on the key opinion expressions in input texts, in implicit sentiment analysis (ISA) the opinion cues come in an implicit and obscure manner. Thus detecting implicit sentiment requires the common-sense and multi-hop reasoning ability to infer the latent intent of opinion. Inspired by the recent chain-of-thought (CoT) idea, in this work we introduce a Three-hop Reasoning (THOR) CoT framework to mimic the human-like reasoning process for ISA. We design a three-step prompting principle for THOR to step-by-step induce the implicit aspect, opinion, and finally the sentiment polarity. Our THOR+Flan-T5 (11B) pushes the state-of-the-art (SoTA) by over 6% F1 on supervised setup. More strikingly, THOR+GPT3 (175B) boosts the SoTA by over 50% F1 on zero-shot setting. Our code is open at https://github.com/scofield7419/THOR-ISA.

Reasoning Implicit Sentiment with Chain-of-Thought Prompting

TL;DR

This work tackles implicit sentiment analysis by requiring multi-hop reasoning to uncover latent aspects and opinions. It introduces THOR, a Three-hop Reasoning prompting framework that sequentially elicits aspect, opinion, and polarity via chain-of-thought prompts, aided by self-consistency and supervision-based revisions. Empirical results on SemEval14 ISA show THOR with Flan-T5-11B achieving ~6% F1 gains in supervised settings and THOR with GPT-3 175B delivering ~50% F1 gains in zero-shot scenarios, with benefits amplifying as model size grows. The approach demonstrates strong improvements over prior baselines and highlights the potential of scalable CoT reasoning for nuanced NLP tasks, while noting limitations for mid-sized LLMs and annotation-related challenges.

Abstract

While sentiment analysis systems try to determine the sentiment polarities of given targets based on the key opinion expressions in input texts, in implicit sentiment analysis (ISA) the opinion cues come in an implicit and obscure manner. Thus detecting implicit sentiment requires the common-sense and multi-hop reasoning ability to infer the latent intent of opinion. Inspired by the recent chain-of-thought (CoT) idea, in this work we introduce a Three-hop Reasoning (THOR) CoT framework to mimic the human-like reasoning process for ISA. We design a three-step prompting principle for THOR to step-by-step induce the implicit aspect, opinion, and finally the sentiment polarity. Our THOR+Flan-T5 (11B) pushes the state-of-the-art (SoTA) by over 6% F1 on supervised setup. More strikingly, THOR+GPT3 (175B) boosts the SoTA by over 50% F1 on zero-shot setting. Our code is open at https://github.com/scofield7419/THOR-ISA.
Paper Structure (20 sections, 11 figures, 2 tables)

This paper contains 20 sections, 11 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Three-hop Reasoning Framework
  3. Chain-of-Thought Prompting
  4. Step 1.
  5. Step 2.
  6. Step 3.
  7. Enhancing Reasoning via Self-consistency
  8. Reasoning Revising with Supervision
  9. Experiments
  10. Setups
  11. Results on Supervised Fine-tuning
  12. Results on Zero-shot Reasoning
  13. Influence of Different Model Sizes of LLMs
  14. Improving ChatGPT with ThoR
  15. Failure Analysis
  16. ...and 5 more sections

Figures (11)

  • Figure 1: Detecting the explicit and implicit sentiment polarities towards targets. Explicit opinion expression helps direct inference, while detecting implicit sentiment requires common-sense and multi-hop reasoning.
  • Figure 2: An illustration of our ThoR framework for three-hop reasoning of implicit sentiment.
  • Figure 3: Influences of LLM scales.
  • Figure 4: Comparisons between GPT3&ChatGPT on randomly-selected 50 ESA and 50 ISA instances.
  • Figure 5: Error analysis.
  • ...and 6 more figures