Knowledge Graph Population from Conversations

Abstract

This thesis describes the design of a system that is capable of the generation of a Knowledge Graph (KG), referred to as Knowledge Graph Population (KGP), from conversations, specifically with elderly people. While this system still follows a traditional KGP approach with Entity Recognition (ER), Entity Linking (EL) and Relation Extraction (RE), we propose novel approaches for: (1) Annotating Wikipedia with Named Entities exhaustively in order to extract datasets for a variety of NLP tasks. (2) An interactive system to test and visualize the output of a KGP system based on a conversation. (3) Fast and accurate dataset creation for RE. (4) Dataset creation as well as training procedures for partially annotated ER datasets. (5) KG Question Answering (KGQA) using RE datasets for KGP instead of solely relying on dedicated KGQA datasets, not necessarily suited to the KG at hand. A KG consists of nodes, corresponding to distinct entities, and edges, corresponding to semantic relations that link entities to each other. The main concern is composed of three topics: Family/Friends, Health, and Nutrition. We chose these since they are the main topics we think elderly people are interested in during daily conversations. The KGP follows a pipeline consisting of three systems: (1) ER, (2) EL and (3) RE. ER is concerned with finding entities of interest, mainly named entities and entities related to the topics Nutrition and Health. The EL system aims to link mentions of entities found by the ER system to their corresponding KG nodes or create a new one if not existing. The RE system then tries to find evidence in utterances that link entities to each other. In order to test the resulting KG, a KG Question Answering (KGQA) system is used to translate natural language questions into structured queries to retrieve answers from the KG. In addition, future tasks, such as response generation, could take advantage of such a KG in order to continue a conversation by creating informative utterances using the gained knowledge from previous utterances from the current or previous conversations in a structured way.