Chinese Event Extraction via Graph Attention Network

Event extraction plays an important role in natural language processing (NLP) applications, including question answering and information retrieval. Most of the previous state-of-the-art methods were lack of ability in capturing features in long range. Recent methods applied dependency tree via dependency-bridge and attention-based graph. However, most of the automatic processing tools used in those methods show poor performance on Chinese texts due to mismatching between word segmentation and labels, which results in error propagation. In this article, we propose a novel character-level Chinese event extraction framework via graph attention network (CAEE). We build our model upon the sequence labeling model, but enhance it with word information by incorporating the word lexicon into the character representations. We further exploit the inter-dependencies between event triggers and argument by building a word-character-based graph network via syntactic shortcut arcs with dependency-parsing. The architecture of the graph minimizes error propagation, which is the result of the error detection of the word boundaries in the processing of Chinese texts. To demonstrate the effectiveness of our work, we build a large-scale real-world corpus consisting of announcements of Chinese financial news without golden entities. Experiments on the corpus show that our approach achieves competitive results compared with previous work in the field of Chinese texts.