Domain-Specific Sentence Encoder for Intention Recognition in Large-Scale Shopping Platforms.

Intelligent dialog robot is now a primary necessity in online shopping, since it can significantly improve the service quality of the online merchants with fewer customer service staffs. And the core functionality of such robots is to identify the customers’ intention in order to provide better service. In this paper, we propose a domain-specific sentence encoder (DSSE) method for real-time online intent identification that hierarchically consists of a representation module and a classification module. We construct a large corpus using the dialog text from e-commerce platforms. We train the representation module of DSSE with the dialog corpus of all kinds of goods and then fine-tune the downstream multi-layer perceptron (MLP) classification module on the corpus of a specific kind of goods. Our model can be easily extended to new kinds of goods as it only needs a little training corpus to fine-tune the small MLP module on new goods chatting dialog text. In addition, our model has quick inference speed and relatively low resource requirement, as well as a higher intent identification accuracy score than BERT. Specifically, our model achieves 85.95% accuracy score and 81.62% F1 score. The inference time per-batch (size = 350) is 33 ms with CPU and only 5.8M parameters are loaded into the RAM.