An Automatic Labeling Method for Subword-Phrase Recognition in Effective Text Classification

要旨

Text classification methods using deep learning, which is trained with a tremendous amount of text, have achieved superior performance than traditional methods. In addition to its success, multi-task learning (MTL for short) has become a promising approach for text classification; for instance, a multi-task learning approach employs the named entity recognition as an auxiliary task for text classification, and it showcases that the auxiliary task helps make the text classification model higher classification performance. The existing MTL-based text classification methods depend on auxiliary tasks using supervised labels. Obtaining such supervision signals requires additional human and financial costs in addition to those for the main text classification task. To reduce these costs, this paper proposes a multi-task learning-based text classification framework reducing the additional costs on supervised label creation by automatically labeling phrases in texts for the auxiliary recognition task. A basic idea to realize the proposed framework is to utilize phrasal expressions consisting of subwords (called subword-phrase) and to deal with the recent situation in which the pre-trained neural language models such as BERT are designed upon subword-based tokenization to avoid out-of-vocabulary words being missed. To the best of our knowledge, there has been no text classification approach on top of subword-phrases, because subwords only sometimes express a coherent set of meanings. The proposed framework is novel in adding subword-phrase recognition as an auxiliary task and utilizing subword-phrases for text classification. It extracts subword-phrases in an unsupervised manner, particularly the statistics approach. In order to construct labels for effective subword-phrase recognition tasks, extracted subword-phrases are classified for document classes so that subword-phrases dedicated to some classes can be distinguishable. The experimental evaluation of the five popular datasets for text classification showcases the effectiveness of the involvement of the subword-phrase recognition as an auxiliary task. It also shows comparative results with the state-of-the-art method, and the comparison of various labeling schemes indicates insights for labeling common subword-phrases among several document classes.