KiPT: Knowledge-injected Prompt Tuning for Event Detection

Proceedings of the 29th International Conference on Computational Linguistics, pages 1943–1952 October 12–17, 2022. 1943 KiPT: Knowledge-injected Prompt Tuning for Event Detection Haochen Li1 , Tong Mo1, Hongcheng Fan1, Jingkun Wang1, Jiaxi Wang1, Fuhao Zhang2 and Weiping Li1,∗ 1Peking University, Beijing, China 2Chinese Academy of Surveying and mapping, Beijing, China {haochenli, wjk, wangjxi, fanhongcheng}@pku.edu.cn {wpli, motong}@ss.pku.edu.cn, zhangfh@ac.cn Abstract Event detection aims to detect events from the text by identifying and classifying event trig- gers (the most representative words). Most of the existing works rely heavily on complex downstream networks and require sufficient training data. Thus, those models may be struc- turally redundant and perform poorly when data is scarce. Prompt-based models are easy to build and are promising for few-shot tasks. However, current prompt-based methods may suffer from low precision because they have not introduced event-related semantic knowl- edge (e.g., part of speech, semantic correlation, etc.). To address these problems, this paper proposes a Knowledge-injected Prompt Tuning (KiPT) model. Specifically, the event detection task is formulated into a condition generation task. Then, knowledge-injected prompts are constructed using external knowledge bases, and a prompt tuning strategy is leveraged to op- timize the prompts. Extensive experiments in- dicate that KiPT outperforms strong baselines, especially in few-shot scenarios. 1 Introduction Events describe state changes of participating enti- ties. The Event Detection (ED) task is one of the essential tasks in the Information Extraction field. Event triggers are the most representative words or phrases in events, and they are usually composed of verbs or nouns. There is a one-to-one correspon- dence between events and event triggers, so the ED task is equivalent to identifying and classifying event triggers. The ED task has a wide range of applications, providing helpful information for downstream tasks such as text summarization, auto summa- rization, machine question and answer (QA), etc. Meanwhile, with the vigorous development of In- ternet news and social media, ED has become a *Weiping Li is the corresponding author. practical approach for extracting information from massive texts. Therefore, the ED task has attracted increasing attention with great academic and ap- plied value in recent years. Most current ED models use a pre-trained lan- guage model to build complex downstream net- works (including CNN, RNN, GCN, etc.) These methods perform well on public datasets, but they rely heavily on the fine-tuning strategy to train their downstream networks and introduce massive ex- tra parameters. However, due to the scarcity and uneven distribution of the annotated data for ED, these methods may cause severe overfitting prob- lems. Further, these methods may perform poorly in data-scarce scenarios because their extra param- eters cannot be fully optimized. Recently, prompt-based learning methods is a new trend in natural language processing. Re- searchers verified that pre-trained language models already have enough knowledge, so the complex downstream networks are unnecessary in many cases. Prompt-based learning methods make full use of the information in pre-trained language mod- els by constructing prompts to guide the language model to solve NLP tasks. Specifically, prompt- based methods …

Content truncated to 500 words