Document-Level Supervision for Multi-Aspect Sentiment Analysis Without Fine-grained Labels (2310.06940v1)

Abstract: Aspect-based sentiment analysis (ABSA) is a widely studied topic, most often trained through supervision from human annotations of opinionated texts. These fine-grained annotations include identifying aspects towards which a user expresses their sentiment, and their associated polarities (aspect-based sentiments). Such fine-grained annotations can be expensive and often infeasible to obtain in real-world settings. There is, however, an abundance of scenarios where user-generated text contains an overall sentiment, such as a rating of 1-5 in user reviews or user-generated feedback, which may be leveraged for this task. In this paper, we propose a VAE-based topic modeling approach that performs ABSA using document-level supervision and without requiring fine-grained labels for either aspects or sentiments. Our approach allows for the detection of multiple aspects in a document, thereby allowing for the possibility of reasoning about how sentiment expressed through multiple aspects comes together to form an observable overall document-level sentiment. We demonstrate results on two benchmark datasets from two different domains, significantly outperforming a state-of-the-art baseline.

Introduction

The importance of aspect-based sentiment analysis (ABSA) is evident given the explosion of user-generated textual content. However, the conventional approach to ABSA requires extensive fine-grained annotations which render it impractical in many applications. This research proposes a novel Variational Auto-Encoder (VAE) based topic modeling technique that leverages document-level sentiment ratings for ABSA without the need for fine-grained labels on aspects or sentiments. The significance of this work lies in its ability to deduce multiple aspects and sentiments within a document using only the overarching sentiment score, thereby presenting a compelling solution for analyzing user feedback efficiently.

Methodology

The proposed model departs from traditional topic models by using document-level sentiment scores instead of aspect-level annotations. It infers topic distributions within documents through a VAE, where the input to the encoder is the token embeddings from a pretrained transformer. Significantly, the researchers freeze the transformer weights during training, which is crucial for model performance. Topics are then associated with aspects or sentiments, and the model employs a pooled sentiment representation for each aspect to forecast the overall sentiment. This framework permits extracting multiple aspects and their sentiments from a single document, forging a link between latent aspect-sentiment configurations and observable document-level sentiments.

Evaluation

For evaluation, the paper harnesses datasets from the restaurant and laptop domains, comparing results to JASen, which represents the state-of-the-art in unsupervised ABSA. Quantitative results showcase superior performance for both aspect detection and aspect-sentiment pairing, outperforming the baseline by significant margins across both domains. Additionally, they qualitatively demonstrate the model's ability to discern topically relevant terms for various aspects and sentiments, further underscoring the efficacy of the proposed approach.

Conclusion & Outlook

This work addresses a crucial gap in ABSA research by providing a robust model that operates without granular annotations, circumventing the cost-intensive and laborious process of data labeling. The effectiveness of the model in real-world datasets indicates its promise for practical applications. The paper concludes by suggesting enhancements to the aspect-sentiment detection accuracy, positing the inclusion of a minimal set of labeled examples to guide the model. Such continuations of the research are likely to further optimize the trade-off between the need for labeled data and the desire for comprehensive, nuanced sentiment analysis.