Domain-independent Intent Extraction from Online Texts

Abstract

Identifying user’s intents from texts on online channels has a wide range of applications from entrepreneurship, banking to e-commerce. However,
intent identification is not a simple task due to the intent and its attributes are various and strongly depend on the domain of data. If the number of intent domains increases, the number of intent’s attributes will get bigger. As a result, the complexity of intent extraction task grows up significantly. Additionally, when a new domain comes, it involves considerable physical efforts to define specific labels for intent and attributes for that domain. Hence, it would be much better to come up with a new method for extracting user’s intents which is not
dependent on a specific domain. In our research, we study the problem of domain-independent intent identification from posts and comments crawled from social networks and discussion forums. We present ten general labels, i.e. labels do not depend on a specific domain, and utilize them when extracting intent and its related information. We also propose a map between general labels and domain-specific labels. We extensively conduct experiments to explore the efficiency of using general labels compared to specific labels in extracting user’s intents when the number of intent domains increases. Our study is conducted on a medium-sized dataset from three selected domains: Tourism, Real Estate and Transportation. In term of accuracy, when the number
of domains grows, our proposal achieves significantly better results than domain-specific method in identifying user’s intent.