A Multi-Semantic Fusion Framework for Chinese Natural Language Understanding

Chinese natural language understanding (NLU) is a very challenging task, due to the diversity of semantics in the Chinese language. Most existing methods consider the influence of Chinese word segmentation for the task. However, most of them ignore the significant impact of homophones and synonyms on Chinese, which inevitably leads to semantic information loss. Therefore, how to effectively represent the semantic information of Chinese has become the most crucial challenge in Chinese NLU tasks.To address this issue, we propose a novel Multi-Semantic Eusion Framework (MSFF) that incorporates information on proper nouns, homophones, and synonyms via words, characters, pinyin, and parts of speech. In addition, we present a crossentropy based loss function for model optimization. We extend the CAIS++ and ECDT-NLU++ datasets to validate the efficiency of the proposed model. Extensive experiments show that MSFF outperforms state-of-the-art methods on two benchmark datasets, which confirms the capability of the proposed model in the Chinese NLU task.