An intent recognition model supporting the spoken expression mixed with Chinese and English

Abstract

Intent recognition is one of the most essential foundations as well as a very challenging task for language understanding, especially for spoken language. As spoken text is short, and lack of full context. Moreover, it may mix multi-language forms. These non-standard spoken expressions further lead to the shortage of text information. In consideration that sparse text information seriously affects the effect of intention understanding, a multi-feature fusion-based intent recognition model for the bilingual phenomenon mixed with Chinese and English is proposed. Combining word2vec and multilingual wordNets with the same synset_id (synonym set id), the model can mask the differences between different languages. Meanwhile, it can enrich the information representation of the spoken text by fusing the word intention features with the context-dependent features represented by transformer as well as the word frequency features. To verify the correctness and effectiveness of the model, extensive experiments were conducted on a real online logistics customer service platform and SMP2018-ECDT dataset. The results show that our model is superior to other models. And it improves the accuracy of intent recognition in logistics data by 20% compared with that of transformer.