A Study of Multi-label Event Types Recognition on Chinese Financial Texts

Abstract

Event extraction is a technique that automatically extracts key event elements from large-scale texts. In classic event extraction process, recognizing event types is earlier than extracting argument roles in the whole event extraction task. But in practice, multiple-label problems are often encountered (that is, one event sentence corresponds to multiple event types). In order to solve this problem, this paper introduces Binary-Relevance, Classifier-Chain, MLkNN and many other multi-label classification strategies from the perspective of problem transformation and algorithm adaptation, trying to find the best classification method to adapt to our Chinese financial corpus. The experimental results show that the Adaboost method based on single-layer decision tree with Classifier-Chain is the best strategy for the task of recognizing event types in this paper. The micro-F1 score and average-precision value of this strategy are 8.91% and 12.46% higher than the baseline strategy (Binary-Relevance + SVM) respectively. At the same time, this method achieves the lowest value on the three indicators of Hamming-Loss, Coverage-Error and Ranking-Loss. In addition, the results also show: (1) Classifier-Chain strategy is better than Binary-Relevance strategy when the classifiers are the same; (2) Under the same problem transformation strategy, the Adaboost method performs best, followed by KNN and the worst case is SVM; (3) If only single classifier is allowed, the MLkNN strategy based on algorithm adaptation is better than other strategies based on problem transformation.