[PAPER] Self-Consistent Reasoning-based Aspect-Sentiment Quad Prediction with Extract-Then-Assign Strategy

Mar. 1, 2024

Jieyong Kim et al.

 

1. ASQP (Aspect Sentiment Quad Prediction) task

- predicts 4 core factors in a sentence (aspect category, aspect term, opinion term, sentiment polarity)

e.g. "The food was delicious but the service was poor"

-> (aspect category, aspect term, opinion term, sentiment polarity)

-> (food, food, delicious, positive), (service, service, poor, negative)

 

2. Proposed strategy for ASQP task: Extract-Then-Assign

- 2-step reasoning strategy that mimics human cognition

 

3. Proposed framework for ASQP task: SCRAP (Self-Consistent Reasoning-based Aspect quad Prediction)

- predicts aspect-sentiment quads based on the Extract-Then-Assign reasoning.

 

1) Extract-Then-Assign Reasoning

(1) first extracts aspect term, opinion term from an input sentence

(2) infers the corresponding aspect category and sentiment polarity

 

2) Reasoning generation with LLM: using few-shot Cot prompting

 

3) Supervised Fine-Tuning the seq-to-seq LM

(1) x = sentences, y = generated reasoning + ground-truth quadruplets

 

4) Self-Consistent Quad Prediction

- consolidates multiple outputs based on self-consistency voting

 

4. Contributions

- SCRAP outperforms baseline methods = SOTA

- Extract-Then-Assign reasoning offers interpretability for quad prediction.