Abstract
Using the word as a basic unit may undermine Chinese event detection model’s performance because of the inaccurate word boundaries generated by segmentation tools. Besides, word embeddings are contextual independent and cannot handle the polysemy of event triggers, which may prevent us from obtaining the desired performance. To address these issues, we propose a BiLSTM-CRF (Bidirectional Long Short-Term Memory Conditional Random Field) model using contextualized representations, which regards event detection task as a character-level sequence labeling problem and uses contextualized representations to disambiguate event triggers. Experiments show that our proposed method sets a new state-of-the-art, which proves Chinese characters could replace words for the Chinese event detection task. Besides, using contextualized representation reduces the false positive case, which verifies that this kind of representation could remedy the weakness of the word embedding technique. Based on the results, we believe that character-level models are worth exploring in the future.
Highlights
Event Extraction is a basic task in information extraction field and proposed by the Automatic Content Extraction(ACE)program [1], which has broad application prospects
For the shortage of word embedding, researchers have proposed some contextualized word embedding techniques which learn their representations based on their contexts, such as Context2Vec [10], ELMo [13], and BERT [14], We explore the following two ideas to incorporate contextualized into the BiLSTM-CRF model: 1) We concatenate character embeddings and contextualized representations
3) We find that compared with the model without contextualized representations, the model using these representations can reduce the false positives; either character or contextualized representations cannot help the model to detect unseen triggers
Summary
Event Extraction is a basic task in information extraction field and proposed by the Automatic Content Extraction(ACE). Is constant in different contexts and cannot help the model to discriminate different meanings of a word [9] [10] This shortage hinders performance gains of current Chinese ED neural models because the same event trigger may express different event types. For the shortage of word embedding, researchers have proposed some contextualized word embedding techniques which learn their representations based on their contexts, such as Context2Vec [10], ELMo [13], and BERT [14], We explore the following two ideas to incorporate contextualized into the BiLSTM-CRF model: 1) We concatenate character embeddings and contextualized representations. Our contributions are as follows: 1) We propose a BiLSTM-CRF model for event detection task, which incorporates contextualized representations. We will use the term ‘‘representation’’ interchangeable with the term ‘‘embedding’’ in the following paper
Sign-in/Register to view highlights & summary 
Published Version (
Free)
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call