Exploiting dependency information to improve biomedical event detection via gated polar attention mechanism

Abstract

This paper tackles the task of biomedical event detection, which includes identifying and categorizing biomedical event triggers. We find that the current biomedical event detection models driven by dependency fail to benefit more distinct improvement from the existing manual dependency embeddings. Here an interpretable hypothesis for the problem above is, that the model using manual dependency embeddings may suffer from low dependency information density (named as dependency weakness) and diffusion of noises from sparse dependency items (called as sparsity diffusion). We argue that dependency representation learning is more effective than the existing manual dependency embeddings, which can reduce dependency weakness and sparsity diffusion. In this work, we first confirm the hypothesis above and then propose to explicitly apply dependency representation learning and triple context representation learning for the biomedical event detection task via gated polar attention mechanism. In specific, we systematically investigate our model under the gated polar attention mechanism. Experimental results demonstrate that our approach outperforms the recent state-of-the-art methods and achieves the best F-score on the biomedical benchmark MLEE dataset.