Modeling Span-Level Interactions with Boundary Probabilities for Targeted Sentiment Analysis

Anonymous

Abstract: Aspect extraction and sentiment prediction are two important tasks for targeted sentiment analysis.Recently, some span-based methods have gained great attention to capture the associations between the two tasks in this domain, where they first extract aspects by detecting aspect boundaries and then predict the span-level sentiments. Most existing studies on modeling the inter-task interactions either share the input representation of the two tasks at the encoding layer, or approximate the output representation of the two tasks at the task layer. Both of them focus on modeling sentence-level correlations between these two tasks, leading to insufficient inter-task feature interactions. Since the aspect-level features are also crucial to connect these two tasks, thus, different from previous approaches, in this paper, we propose to model the span-level interactions with boundary probabilities (SIBP) to explicitly consider the inner correlations for these two tasks. Specifically, we use the predicted boundary probabilities of aspects to generate all possible spans as input of the sentiment prediction module, such that the sentiment information can be backpropagated into the boundary detection process in a fully differentiable manner. Further, we devise an alternate learning strategy to take the best of both tasks between predicted aspects and real aspects. This strategy not only guides sentiment prediction more properly but also improves computational efficiency. Moreover, to predict the boundary probabilities of the aspects more accurately, we design a semantic compatibility mechanism. Finally, we conduct extensive experiments on three real-world datasets to demonstrate the model's superiority.

Paper Type: long

Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining