Unsupervised dialogue intent detection via hierarchical topic model

Proceedings of Recent Advances in Natural Language Processing, pages 932–938, Varna, Bulgaria, Sep 2–4, 2019. https://doi.org/10.26615/978-954-452-056-4_108 932 Unsupervised dialogue intent detection via hierarchical topic model Artem Popov1,2, Victor Bulatov1, Darya Polyudova1, Eugenia Veselova1 1Moscow Institute of Physics and Technology 2Lomonosov Moscow State University popov.artem.s@yandex.ru, viktor.bulatov@phystech.edu, darya.polyudova@phystech.edu, veselova.er@phystech.edu Abstract One of the challenges during a task- oriented chatbot development is the scarce availability of the labeled training data. The best way of getting one is to ask the assessors to tag each dialogue according to its intent. Unfortunately, performing la- beling without any provisional collection structure is difficult since the very notion of the intent is ill-defined. In this paper, we propose a hierarchical multimodal regularized topic model to ob- tain a first approximation of the intent set. Our rationale for hierarchical mod- els usage is their ability to take into ac- count several degrees of the dialogues rel- evancy. We attempt to build a model that can distinguish between subject-based (e.g. medicine and transport topics) and action-based (e.g. filing of an application and tracking application status) similari- ties. In order to achieve this, we divide set of all features into several groups ac- cording to part-of-speech analysis. Vari- ous feature groups are treated differently on different hierarchy levels. 1 Introduction One of the most important goals of task-oriented dialogue systems is to identify the user intention from the user utterances. State-of-the-art solutions like (Chen et al., 2017) require a lot of labeled data. User’s utterances (one or several for a dia- logue) have to be tagged by the intent of the dia- logue. This is a challenging task for a new dialogue collection because the set of all possible intents is unknown. Giving a provisional hierarchical col- lection structure to assessors could make the intent labeling challenge easier. The resulting labels will be more consistent and better suitable for model training. Simple intent analysis is based on empirical rules, e.g. “question” intent contains phrase “what is # of #” (Yan et al., 2017). More universal and robust dialogue systems should work without any supervision or defined rules. Such systems can be implemented with automatic extraction of the semantic hierarchy from the query by multi-level clustering, based on different semantic frames (ca- pability, location, characteristics etc.) in sentences (Chen et al., 2015). In our work intents represent a more complex entity which combine all intentions and objectives. Many previous works take advantage of hier- archical structures in user intention analysis. In paper (Shepitsen et al., 2008) automatic approach through hierarchical clustering for document tag- ging is used. However, this approach does not take advantage of peculiar phrase features, such as syn- tax or specific words order. Syntactic parsing of intention was applied in (Gupta et al., 2018) to decompose client intent. This hierarchical repre- sentation is similar to a constituency syntax tree. It contains intentions and objects as tree elements and demands deep analysis of every sentence. At- tempt to extract subintents along with main in- tent can be found in …

Content truncated to 500 words