[Paper Review] Weakly-Supervised Aspect-Based Sentiment Analysis via Joint Aspect-Sentiment Topic Embedding

Jiaxin Huang, Meng Yu+3 more|arXiv (Cornell University)|Oct 13, 2020

Sentiment Analysis and Opinion Mining32 references4 citations

TL;DR

This paper proposes JASen, a weakly-supervised method for aspect-based sentiment analysis that jointly learns sentiment-aspect topic embeddings in word embedding space using only a few keywords per aspect/sentiment. By leveraging embedding-based predictions for pre-training and self-training convolutional neural networks, the model achieves state-of-the-art performance, outperforming baselines by 7.4% and 5.1% in F1-score on average for aspect and sentiment classification, respectively.

ABSTRACT

Aspect-based sentiment analysis of review texts is of great value for understanding user feedback in a fine-grained manner. It has in general two sub-tasks: (i) extracting aspects from each review, and (ii) classifying aspect-based reviews by sentiment polarity. In this paper, we propose a weakly-supervised approach for aspect-based sentiment analysis, which uses only a few keywords describing each aspect/sentiment without using any labeled examples. Existing methods are either designed only for one of the sub-tasks, neglecting the benefit of coupling both, or are based on topic models that may contain overlapping concepts. We propose to first learn joint topic embeddings in the word embedding space by imposing regularizations to encourage topic distinctiveness, and then use neural models to generalize the word-level discriminative information by pre-training the classifiers with embedding-based predictions and self-training them on unlabeled data. Our comprehensive performance analysis shows that our method generates quality joint topics and outperforms the baselines significantly (7.4% and 5.1% F1-score gain on average for aspect and sentiment classification respectively) on benchmark datasets. Our code and data are available at https://github.com/teapot123/JASen.

Motivation & Objective

• To address the challenge of aspect-based sentiment analysis without requiring large-scale labeled data.
• To jointly model sentiment and aspect information by learning fine-grained joint topic embeddings in the word embedding space.
• To improve both aspect extraction and sentiment classification by capturing shared semantics in terms like 'semi-private' that signal both ambience and positive sentiment.
• To overcome limitations of topic models, such as topic overlap and instability, by enforcing distinctiveness in learned joint topics.
• To enable effective self-training of neural classifiers using embedding-based predictions on unlabeled data.

Proposed method

• Learn joint topic embeddings for each ⟨sentiment, aspect⟩ pair in the shared word embedding space using regularization to enforce topic distinctiveness.
• Train topic and word embeddings jointly on in-domain corpora, modeling the distribution of user-provided keywords over all joint topics.
• Derive embedding-based predictions for unlabeled reviews using the learned joint topic representations.
• Pre-train convolutional neural networks (CNNs) using these embedding-based predictions as supervision.
• Refine the CNNs via self-training on unlabeled data using high-confidence predictions from the model.
• Use a joint learning objective that encourages the model to associate keywords with specific ⟨sentiment, aspect⟩ pairs while maintaining semantic coherence in the embedding space.

Experimental results

Research questions

• RQ1 Can joint sentiment-aspect topic embeddings in the word embedding space improve performance in weakly-supervised aspect-based sentiment analysis?
• RQ2 Does learning distinct, coherent joint topics from only a few keywords lead to better generalization than separate modeling of aspects and sentiments?
• RQ3 Can embedding-based predictions effectively pre-train neural models and be further refined through self-training on unlabeled data?
• RQ4 How does the proposed method compare to existing topic model-based and neural network-based approaches in terms of F1-score for aspect and sentiment classification?
• RQ5 Can the model capture fine-grained, context-specific terms (e.g., 'semi-private') that signal both aspect and sentiment simultaneously?

Key findings

• The proposed JASen model achieves a 7.4% average F1-score gain over baselines for aspect classification on benchmark datasets.
• The model records a 5.1% average F1-score gain over baselines for sentiment polarity classification.
• The model generates coherent term clusters that describe joint topics, such as 'semi-private' and 'date' for the ⟨good, ambience⟩ topic.
• Ablation studies show that removing joint topic learning significantly degrades performance, confirming its importance.
• The model correctly distinguishes between similar aspects like 'os' and 'support' in the same domain, unlike models that conflate them.
• Case studies show that the model's predictions are often more contextually accurate than gold-standard labels in ambiguous cases, such as preferring 'ambience' over 'location' for reviews describing atmosphere.