Data augmentation for aspect-based sentiment analysis

Guangmin Li

doi:10.1007/s13042-022-01535-5

2022-05-18

PRE

DOI：10.1007/s13042-022-01535-5

原文链接

原文求助

摘要

En 中文

In recent years, deep learning has been widely used in the field of natural language processing (NLP), achieving spectacular successes in various NLP tasks. These successes are largely due to its capability to automatically learn feature representations from text data. However, the performance of deep learning in NLP can be negatively affected by a lack of sufficiently large labeled corpus for training, resulting in limited improvement in performance. Data augmentation overcomes this small data problem by expanding the sample size for the classes of data in the training corpus. This paper introduces the data augmentation for aspect-based sentiment analysis (ABSA), a classical research topic in NLP that has been applied in various fileds. The study aims to enhance the classification performance of ABSA through various augmentation strategies. Two specific augmentation strategies are presented, part-of-speech (PoS) wise synonym substitution (PWSS) and dependency relation-based word swap (DRAWS), which augment data using PoS, external domain knowledge, and syntactic dependency. These strategies are evaluated through extensive experimentation on four public datasets using three representative deep learning models-aspect-specific graph convolutional network (ASGCN), content attention-based aspect-based sentiment classification (CABASC), and long short-term memory (LSTM) network. Compared with the results without data augmentation, our augmentation strategies achieve a performance gain of up to 11.49% on Macro-F1, with the lowest gain being 2.9%. The experimental results demonstrate that the proposed data augmentation strategies are very useful for training deep learning models on small data corpus.

Keyword:

Sentiment analysis

Text classification

Data augmentation

Deep learning

Dependency syntax