Relation construction for aspect-level sentiment classification

Highlights

• This paper is the first attempt to extract aspect relations as an auxiliary task for aspect-level sentiment analysis.

• We propose a novel relation construction multi-task learning network that can capture aspect relations.

• Extensive experiments on several benchmark datasets show the superiority of the proposed model.

Abstract

Aspect-level sentiment classification aims to obtain fine-grained sentiment polarities of different aspects in one sentence. Most existing approaches handle the classification by acquiring the importance of context words towards each given aspect individually, and ignore the benefits brought by aspect relations. Since the sentiment of one aspect can be deduced through their relationship according to other aspects, in this paper, we propose a novel relation construction multi-task learning network (RMN), which is the first attempt to extract aspect relations as an auxiliary classification task. RMN generates aspect representations through graph convolution networks with a semantic dependency graph and utilizes the bi-attention mechanism to capture the relevance between the aspect and the context. Unlike conventional multi-task learning methods that need extra datasets, we construct an auxiliary relation-level classification task that extracts aspect relations from the original dataset with shared parameters. Extensive experiments on five public datasets from SemEval 14, 15, 16 and MAMS show that our RMN improves about 0.09% to 0.8% on accuracy and about 0.04% to 1.19% on F1 score, compared to several comparative baselines.

Introduction

As an important Nature Language Processing (NLP) task, sentiment classification is widely applied in online review analysis [1], text mining [2] and emotion recognition [3]. Document-level [4], [5] and sentence-level sentiment analyses [6], [7] aim to predict the overall sentiment of the input text, which can be considered as coarse-grained sentiment. Given a set of aspects in one sentence, aspect-level sentiment analysis aims at identifying sentiment polarities towards these specific entities [8]. For example, in the sentence “The menu is limited but almost all of the dishes are excellent”, although the overall sentiment of the sentence is neutral, the polarity of “menu” is negative while the one for “dish” is positive.

For the aspect-level sentiment analysis, traditional methods mainly constructed sentiment lexicon dictionary [9] or adopted a feature-based SVM [10] for classification. However, such kind of lexicon dictionary and feature-based methods are labor-intensive. Recently, with the ability to handle complex structure sentences, neural network-based approaches have become the mainstream of sentiment classification. Besides, the attention mechanism has been proved effective for capturing potential semantic relations between the aspect and the context [8], [11], [12]. Apart from the attention mechanism, Zhang et al. [13] applied Graph Convolutional Networks (GCNs) to extract syntactical dependencies among the context, and presented its advantages when handling long sentences.

However, the above studies all associated a single aspect with its contextual words individually, and ignored the benefits brought by multiple relations among aspects. One aspect can be better classified when considering its aspect relations. For instance, as shown in Fig. 1, given the sentence “Nice food. The price is reasonable although the service is poor.”, we can obtain that “food” and “price” are in positive semantic while “service” is negative. Meanwhile, three relations are extracted based on three aspects, namely “food-price”, “food-service” and “price-service”. Suppose that the semantic word “reasonable” is masked, and the polarity of “price” is unknown. If taken relations “food-price” and “price-service” into consideration, we can easily deduce that the polarity of “price” is negative. Therefore, for the sake of better aspect-level sentiment analysis, two challenges should be addressed: 1) how to model relations among aspects; and 2) how to utilize these aspect relations for the aspect-level sentiment classification.

For the first challenge, several methods [14], [15] adopted the attention mechanism or GCNs over aspects to capture potential relations. However, they exaggerated aspect information since the aspect with neutral polarity may generate noises if considered. To address this issue, we only consider explicit aspect relations (“similar” and “opposite”) to avoid neutral aspect noises. Specifically, we model aspect relations by the subtraction between two aspect representations, then auto-annotate their labels based on aspect polarities. For the second challenge, we construct an auxiliary task named relation-level classification, and train the basic aspect classification task and the auxiliary task simultaneously in a multi-task framework to explore the benefits of the aspect relations. Our contributions are summarized as follows:

• To the best of our knowledge, this is the first attempt to extract aspect relations as an auxiliary task for aspect-level sentiment analysis, and we explore the benefits brought by aspect relations.

• We propose a novel relation construction multi-task learning network that utilizes the bi-attention mechanism to capture bidirectional semantic information between the context and the aspect, and we adopt aspect disagreement regularization to better identify aspect-specific features from overlapped representations.

• Extensive experiments on several benchmark datasets validate the effectiveness of the proposed model compared to several comparative baselines, and show the ability to handle relatively small datasets.

The rest of this paper is organized as follows: Section 2 reviews related works. The detailed modules of the proposed method are presented in Section 3. Experimental results and further analyses are shown in Section 4. Section 5 concludes the paper and discusses the future research.