Semi-supervised Domain Adaptation for Dependency Parsing with Dynamic Matching Network

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 1035 - 1045 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Semi-supervised Domain Adaptation for Dependency Parsing with Dynamic Matching Network Ying Li, Shuaike Li, Min Zhang Institute of Artificial Intelligence, School of Computer Science and Technology, Soochow University, China yingli_hlt@foxmail.com, skli20@stu.suda.edu.cn minzhang@suda.edu.cn Abstract Supervised parsing models have achieved im- pressive results on in-domain texts. However, their performances drop drastically on out-of- domain texts due to the data distribution shift. The shared-private model has shown its promis- ing advantages for alleviating this problem via feature separation, whereas prior works pay more attention to enhancing shared features but neglect the in-depth relevance of specific ones. To address this issue, we for the first time apply a dynamic matching network on the shared-private model for semi-supervised cross- domain dependency parsing. Meanwhile, con- sidering the scarcity of target-domain labeled data, we leverage unlabeled data from two as- pects, i.e., designing a new training strategy to improve the capability of the dynamic match- ing network and fine-tuning BERT to obtain domain-related contextualized representations. Experiments on benchmark datasets show that our proposed model consistently outperforms various baselines, leading to new state-of-the- art results on all domains. Detailed analysis on different matching strategies demonstrates that it is essential to learn suitable matching weights to emphasize useful features and ignore useless or even harmful ones. Besides, our proposed model can be directly extended to multi-source domain adaptation and achieves best perfor- mances among various baselines, further veri- fying the effectiveness and robustness. 1 Introduction Dependency parsing aims to capture syntactic and semantic information over input words via a de- pendency tree. As depicted in Figure 1, given an input sentence s = w0w1 . . . wn, a dependency tree is defined as d = {(h, m, l), 0 ≤ h ≤ n, 1 ≤ m ≤ n, l ∈ L}, where (h, m, l) is a dependency from the head word wh to the modifier word wm with the relation label l ∈ L. Recently, supervised neural models have achieved significant improve- ments in dependency parsing (Chen and Manning, $ 好 萌 ， 好 可爱 ！ $ very cute , very lovely ! root adv punc sasubj adv punc Figure 1: An example of a dependency tree which is from the target-domain product comment (PC) data. 2014; Andor et al., 2016; Kiperwasser and Gold- berg, 2016; Dozat and Manning, 2017; Li et al., 2019a). Particularly, Dozat and Manning (2017) propose a BiAffine parser and achieve good results on various languages. In order to obtain better performance, supervised parsing models rely on sufficient in-domain train- ing data. However, the parsing accuracy degrades significantly when the training data is from out-of- domain that has a large gap between the in-domain data.The main reason can be attributed to differ- ent feature distributions between source and target domains. Thus modeling the relevance of these distributions becomes the key challenge for cross- domain dependency …

Content truncated to 500 words