Proposition-Level Clustering for Multi-Document Summarization

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 169 - 174 July 10-15, 2022 ©2022 Association for Computational Linguistics SUBS: Subtree Substitution for Compositional Semantic Parsing Jingfeng Yang†⋆ Le Zhang‡⋆ Diyi Yang† † Georgia Institute of Technology ‡ Fudan University jingfengyangpku@gmail.com zhangle18@fudan.edu.cn dyang888@gatech.edu Abstract Although sequence-to-sequence models often achieve good performance in semantic pars- ing for i.i.d. data, their performance is still inferior in compositional generalization. Sev- eral data augmentation methods have been pro- posed to alleviate this problem. However, prior work only leveraged superficial grammar or rules for data augmentation, which resulted in limited improvement. We propose to use sub- tree substitution for compositional data aug- mentation, where we consider subtrees with similar semantic functions as exchangeable. Our experiments showed that such augmented data led to significantly better performance on SCAN and GEOQUERY, and reached new SOTA on compositional split of GEOQUERY. We have publicly released our code at https: //github.com/GT-SALT/SUBS . 1 Introduction Semantic parsing transforms natural language ut- terances to formal language. Because meaning rep- resentations or programs are essentially composi- tional, semantic parsing is an ideal testbed for com- positional generalization. Although neural seq2seq models could achieve state-of-the-art performance in semantic parsing for i.i.d. data, they failed at compositional generalization due to lack of reason- ing ability. That is, they do not generalize well to formal language structures that were not seen at training time. For example, a model that observes at training time the questions “What is the popula- tion of the largest state?” and “What is the largest city in USA?” may fail to generalize to questions such as “What is the population of the largest city in USA?”. This leads to large performance drops on data splits designed to measure compositional ⋆Equal Contribution. Jingfeng Yang proposed subtree substitution data augmentation for compositional semantic parsing, implemented augmentation and LSTM/BART parsers, and ran SCFG/GECA baselines. Le Zhang induced span trees and ran span-based semantic parsing baseline. generalization (compositional splits), in contrast to the generalization abilities of humans. To improve compositional generalization in se- mantic parsing (compositional semantic parsing), prior work focused on incorporating inductive bi- ases directly to models or data augmentation. From the model perspective, some work used neural- symbolic models (Chen et al., 2020), generated intermediate discrete structures (Herzig and Be- rant, 2020; Zheng and Lapata, 2020), or conducted meta-learning (Lake, 2019). From the data perspec- tive, Jia and Liang (2016) proposed to recombine data with simple synchronous context-free gram- mar (SCFG), despite not for compositional gener- alization. Andreas (2019) used some simple rules for data augmentation, where tokens with the same context were considered as exchangeable. Such techniques are still limited since they only lever- aged superficial grammars or rules, and failed when there are linguistically rich phrases or clauses. To fill this gap, we propose to augment the train- ing data of semantic parsing with diverse compo- sitional examples based on induced or annotated (semantic and syntactic) trees. Specifically, …

Content truncated to 500 words