Unifying Structure Reasoning and Language Model Pre-training for Complex Reasoning

Recent knowledge enhanced pre-trained language models have shown remarkable performance on downstream tasks by incorporating structured knowledge from external sources into language models. However, they usually suffer from a heterogeneous information alignment problem and a noisy knowledge injection problem. For complex reasoning, the contexts contain rich knowledge that typically exists in complex and sparse forms. In order to model structured knowledge in the context and avoid these two problems, we propose to unify structure reasoning and language model pre-training. It identifies four types of elementary knowledge structures from contexts to construct structured queries, and utilizes the box embedding method to conduct explicit structure reasoning along queries during language modeling. To fuse textual and structured semantics, we utilize contextual language representations of knowledge structures to initialize their box embeddings for structure reasoning. We conduct experiments on complex language reasoning and knowledge graph (KG) reasoning tasks. The results show that our model can effectively enhance the performance of complex reasoning of both language and KG modalities.