Knowledge graph construction and completion using pre-trained language models
Abstract: One of the important tasks in Artificial Intelligence (AI) field is the task of quick and efficient information extraction from various sources, such as news, scientific articles, videos or photos. The resulting information can be stored in different structural forms, such as tables, arrays, lists, graphs, etc. But no less important is the correct interpretability of these data and completion of the data structures with information that could be inferred from context. This thesis project presents an approach that constructs a graph data structure (knowledge graph) from unstructured text without human supervision using pre-trained language models (such a BERT or GPT-2). In addition, the language model is fine-tuned to predict relationships between entities that are not explicitly specified in the text. At the first step, the algorithm accepts an unstructured text as an input, and generates a set of facts (head; relation; tail), where head and tail are entities, and relation is a relationship between entities extracted from the text. At the second step, the fine-tuned language model predicts relations between entities logically inferred from the context but not explicitly mentioned in the text. The results show that our approach could be practical and convenient in real life tasks.