Multi-Task Dynamic Joint Training for Aspect-Base-Sentiment-Analysis

Abstract

In Aspect-Base-Sentiment-Analysis (ABSA), a single sentence may contain multiple aspect words, making it challenging for the model to find appropriate sentiment words for each aspect word from the text. Therefore, this paper proposes a Multi-Task Dynamic Joint Training (MDJT). To learn better feature representations, we apply a text feature extractor and a graph feature extractor to extract semantic features and graph-structured features of text, respectively. In this paper, we use the multi-task frame to joint two tasks dynamically: the main task is the sentiment classification task (Polarity Task), which is used to classify the sentiment of each aspect word; the auxiliary task is the opinion word task (Opinion Task), whose purpose is to extract text. The features are most likely to become opinion words and the accuracy of the main task sentiment classification are further improved by optimizing the loss value of this task. Experiments on several benchmarking collections illustrate that our proposed model has comparable effectiveness to a range of state-of-the-art models, and they further demonstrate that both syntactical information and multi-task join dynamically and properly.