Joint Inference: a Statistical Approach for Open Information Extraction

In recent decades, natural language processing has great progress. Better model of each sub-problem achieves 90% accuracy or better, such as part-of-speech tagging and phrase chunking. However, success in integrated, end-to-end natural language understanding remains elusive. It is mainly due to the systems processing sensory input typically have a pipeline architecture: the output of each stage is the input of the next, errors cascade and accumulate through the pipeline of naively chained components, and there is no feedback from later stages to earlier ones. Actually, later stages can help earlier ones to process. For example, the part-of-speech tagger needs more syntactic and semantic information to make this choice. Previous researches tend to ignore this. So errors in each step will propagate to later ones. In current, a number of researchers have paid attention to this problem and proposed some joint approaches. But they do not perform Open Information Extraction (Open IE), which can identify various types of relations without requiring pre-specifications. In this paper, we propose a statistical modeling such unified consideration known as joint inference, which is based on Markov logic and can perform both traditional relation extraction and Open IE. The proposed modeling significantly outperforms the other Open IE systems in terms of both precision and recall. The joint inference is efficient and we have demonstrated its efficacy in real-world Open IE detection tasks.