An event detection method based on hierarchical topic-driven self-attention mechanism

Abstract

The invention discloses an event detection method based on a hierarchical topic-driven self-attention mechanism, comprising the following steps: (1) constructing topic-aware document representation vectors and word representation vectors; (2) performing sequence encoding on candidate event mentions (3) Establish a hierarchical self-attention model; (4) Predict event types; the method first uses a variational autoencoder to infer the document-topic distribution and topic-word distribution of the document, and calculates the topic-aware document representation vector and word representation vectors; then splicing the topic-aware document representation vectors and the word embeddings of candidate event-triggered words, using Bi‑LSTM for sequence encoding, and obtaining an intermediate representation containing general global information through a document-level self-attention model, and then The intermediate representation and the topic-aware word-level representation vector are spliced ​​together, and the final representation of the word is obtained through Bi-LSTM and word-level self-attention model. Finally, the event detection result is obtained through the fully connected layer and softmax normalization.

Description

technical field [0001] The invention relates to the technical field of information extraction in natural language processing, in particular to an event extraction technology, in particular to an event detection method based on a hierarchical theme-driven self-attention mechanism. Background technique [0002] In recent years, with the continuous development of information technology, the amount of information on the Internet is also expanding. Therefore, how to use automated tools to accurately extract the information that users are interested in from massive information has become an urgent problem to be solved. In this context, information extraction technology has become a particularly important research direction. Since the late 1980s, information extraction technology has begun to develop, which is mainly due to the convening of the Message Understanding for Comprehension (MUC). It was initiated and funded by the US Defense Advanced Research Projects Committee. It was...

AI Technical Summary

Problems solved by technology

Although this can indeed bring valuable disambiguation basis for the trigger word, it may also introduce some noise information, which will exacerbate the ambiguity of the trigger word and even interfere with the detection.

To alleviate this problem, some works select different document representations for each sentence through attention mechanism, which directly enhances the specific semantic representation of each word in the sentence, however, they only consider the document where the sentence is located and ignore Valuable clues provided by relevant documentation

Images