Unregulated Chinese-to-English Data Expansion Does NOT Work for Neural Event Detection

Proceedings of the 29th International Conference on Computational Linguistics, pages 2633–2638 October 12–17, 2022. 2633 Unregulated Chinese-to-English Data Expansion Does NOT Work for Neural Event Detection Zhongqiu Li, Yu Hong∗, Jie Wang, Shiming He, Jianmin Yao, Guodong Zhou School of Computer Science and Technology, Soochow University, China {leosanqiu, tianxianer, jawn233, smhelpai}@gmail.com {jyao, gdzhou}@suda.edu.cn Abstract We leverage cross-language data expansion and retraining to enhance neural Event Detection (abbr., ED) on English ACE corpus. Machine translation is utilized for expanding English training set of ED from that of Chinese. How- ever, experimental results illustrate that such strategy actually results in performance degra- dation. The survey of translations suggests that the mistakenly-aligned triggers in the expanded data negatively influences the retraining pro- cess. We refer this phenomenon to “trigger falsification”. To overcome the issue, we apply heuristic rules for regulating the expanded data, fixing the distracting samples that contain the falsified triggers. The supplementary experi- ments show that the rule-based regulation is beneficial, yielding the improvement of about 1.6% F1-score for ED. We additionally prove that, instead of transfer learning from the trans- lated ED data, the straight data combination by random pouring surprisingly performs better. 1 Introduction We tackle ED, a task of recognizing trigger words (triggers for short) that signal different types of events (Ahn, 2006). For example, the trigger “meet- ing” in (1) signals the CONTACT-MEET event. (1) “We have the transcript of the meeting.” Trigger: meeting Type: CONTACT-MEET The current study of ED, in general, utilizes neu- ral classification models for determining the event types in the word-by-word manner, including the predefined ACE1 event types and Non-trigger type. Correspondingly, supervised learning is ap- plied to pursue the semantic-level distributed rep- resentations of words, so as to provide perceptible evidence for decoding event types. ∗Corresponding author. 1https://catalog.ldc.upenn.edu/LDC2006T06 他 还 将 前往【Transport】 沙特阿拉伯 (He) (also) (will) (travel to) (Saudi Arabia) He will also travel【Transport】to【Transport】Saudi Arabia Alignment Transmission Chinese ED instance Translation Figure 1: A case study of trigger designation and class- label transmission which is bridged by word alignment. Expanding the training data is able to enhance the ED-oriented neural classification models. It is because there are a larger amount of knowledge and diverse pragmatic phenomena can be introduced into the supervised learning process, as claimed in different tasks of natural language processing. We childishly embrace this method, conducting expansion by translating Chinese ED corpus and pouring translations into the English training set, where word alignment (Sabet et al., 2020) is used for designating triggers in the translations. What is beyond our expectation is that, however, cross-language data expansion actually results in performance degradation. We survey the transla- tions which were used for expansion, and observe that a large number of low-quality instances were involved. The survey also suggests that the falsified triggers in such instances probably misled neural models during supervised learning. We provide an example in Figure 1 where Chinese event mention is taken from the publicly-shared ACE corpus of Chinese ED, and word alignment is carried out for …

Content truncated to 500 words