Improving Concept Representations for Short Text Classification

Short text classification is a challenging task in Natural Language Processing (NLP). Compared with documents, short texts are more sparse and ambiguous due to lack of context. In order to overcome this difficulty, recent work tends to take an approach of combining neural language models with external linguistic resources such as knowledge bases. In this kind of approaches, concepts obtained from a knowledge base are usually mapped to an implicit space and represented as a vector. However, how to effectively represent concepts in a neural language model is not well studied yet. Hence, in this study, we construct several formulae for concept embedding and compare them in a short text classification task. Our experimental results show that utilizing proper concept embeddings can slightly improve the performance of short text classification.