Unsupervised Training of Automatic Dialogue Systems for Customer Support
Aigul Nugmanova, Irina A. Chernykh, Anna Bulusheva, Yuri N. Matveev · 2019
Automated dialog systems for customer support have recently become a popular area of research in the field of natural language processing. The traditional (supervised) approach for training the dialog model for customer support includes: (1) building a tree of topics (states), (2) finding query examples for each state, (3) training the classifier. This is an expensive, labor-intensive process. In contrast, we trained our dialog model in an unsupervised way, avoiding the need for labeled corpora. Our dialog model is retrieval-based, and its architecture incorporates a siamese network similar to the Dual Encoder. Similar to other retrieval-based dialog models, our model attempts to find the best response out of provided ones for the current dialog history. The research we performed provided the opportunity to develop a customer service solution - software that acts as an intelligent assistant, suggesting a list of relevant prompts to the operator in real time.