Integrating structured and unstructured knowledge sources for domain-specific chatbots

Sarkar, Rajdeep
The increasing demand for customer support in various industries and the popularity of con versational interfaces has necessitated the development of chatbots. The availability of domain knowledge has enabled chatbots to understand and reason over complex concepts and con texts of specific domains enabling richer communication and engagement with customers. The integration of external knowledge is especially important in customer service, education and healthcare where accurate information is vital. Knowledge-grounded chatbots can source external knowledge from either unstructured sources such as Wikipedia articles or structured sources like knowledge graphs. A significant hurdle is to effectively integrate such knowledge sources into the chatbots for enhanced user experi ence. Structured knowledge sources such as knowledge graphs are rich sources of information. However, an issue that arises during the integration of structured knowledge in chatbots is the need for careful consideration in constructing domain-specific subgraphs and selecting rel evant knowledge. In contrast to structured data, utilising unstructured sources for knowledge grounded chatbots presents different challenges, as it requires annotated data that is grounded in the domain, which necessitates domain expertise. It is crucial to develop systems that can adapt to available knowledge without the need for expert annotated datasets. Therefore, careful con sideration must be given when selecting external knowledge sources for knowledge-grounded chatbots. Additionally, chatbots should be explainable, and their behaviour should be under standable to users. Furthermore, chatbots should leverage these knowledge sources effectively for informative and fluent response generation. This dissertation aims to present effective solu tions for addressing the aforementioned problems during the integration of external knowledge into chatbots. This dissertation proposes frameworks to effectively integrate external structured knowledge as well as unstructured knowledge into chatbots. We study the impact of by constructing con textually relevant subgraphs from a knowledge graph. Additionally, it presents frameworks to fuse unstructured knowledge into chatbots for question-answering without requiring manual annotation of datasets. Furthermore, it suggests using path traversal on a knowledge graph con ditioned on chat context semantics for explainability, as the paths can convey context changes in a chat. Finally, a response generation methodology is proposed for generating informative, fluent, and coherent responses for knowledgeable response generation. The integration of structured and unstructured knowledge is vital for developing effective chatbots, especially in domain-specific scenarios. Careful consideration must be employed for selecting the knowledge sources to ensure that these chatbots make optimal utilisation of the knowledge sources. Additionally, these systems must be explainable so that the behaviour is understandable to its users. The frameworks proposed in this thesis offer effective solutions for integrating structured and unstructured knowledge into chatbots while addressing the issues around explainability. The proposed response generation methodology also demonstrates the capacity of generating knowledge-grounded responses.
NUI Galway
Publisher DOI
Attribution-NonCommercial-NoDerivs 3.0 Ireland