A Study of Event Elements Extraction on Chinese Bond News Texts

摘要

Acquiring related information automatically from massive financial texts is helpful to investors. In order to do it, in this paper an event elements extraction framework based on a sequence labelling method is studied and applied to bond news texts. According to the pre-defined event type framework, the labelling task is tackled with the Conditional Random Field (CRF) model. A total of 40 feature templates (28 atomic templates and 12 compound templates) are constructed from the five dimensions including word, part of speech, dependency grammar, location and compound features. The experimental results verify the difference of extraction performance under different feature selection strategies and show that the setting with the features of all dimensions achieves a promising result.