Automatic information extraction in the AI chip domain using gated interactive attention and probability matrix encoding method.

Artificial intelligence (AI) that utilizes neural networks (NNs) has a broad range of applications. However, NNs necessitate significant amounts of computation and data storage, which imposes considerable hardware demands and drives the need for AI chip research. To address this need, AI chip researchers must delve into extensive literature for entity information on hardware issues, architectures, optimizations, performances, and applications. Nevertheless, the influx of relevant papers has far surpassed their ability to read and absorb all the information. In this article, automatic information extraction in the AI chip domain using entity recognition techniques is conducted to alleviate the burden of paper reading for AI chip researchers. Our approach involves creating a manually annotated dataset called the ACER dataset to support automatic information extraction from literature in the field of AI chips. To address the challenge of recognizing entities with complex structures and lacking explicit features, the GIA-PME model is proposed, which utilizes a gated interaction attention mechanism and probability-matrix encoding. Our proposed approach enhances entity cognition by utilizing the keyword sequence of each entity type as prior knowledge. It also incorporates a dedicated embedding to learn the eigenvectors of entity structures. Finally, it combines the prior knowledge with eigenvectors using a designed gated interactive attention mechanism to assist with recognition. In addition, the proposed probability matrix encoding is used to detect nested entities to avoid information loss. Experimental results show that our GIA-PME model achieves the best performance compared to existing models, thus improving the strict/relaxed F1-score to 70.9/81.7 (+5.3 ∼ 9.1) on the ACER dataset. [ABSTRACT FROM AUTHOR]