Unsupervised Natural Language Inference Using PHL Triplet Generation

Findings of the Association for Computational Linguistics: ACL 2022, pages 158 - 168 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Sememe Prediction for BabelNet Synsets using Multilingual and Multimodal Information Fanchao Qi1∗, Chuancheng Lv2,5∗, Zhiyuan Liu1,3,4, Xiaojun Meng6, Maosong Sun1,3,4†, Hai-Tao Zheng2,5 1Dept. of Comp. Sci. & Tech., Institute for AI, Tsinghua University, Beijing, China Beijing National Research Center for Information Science and Technology 2Shenzhen International Graduate School, Tsinghua University, China 3Institute Guo Qiang, Tsinghua University, Beijing, China 4International Innovation Center of Tsinghua University, Shanghai, China 5Peng Cheng Laboratory 6Huawei Noah’s Ark Lab {qfc17,lvcc21}@mails.tsinghua.edu.cn Abstract In linguistics, a sememe is defined as the minimum semantic unit of languages. Se- meme knowledge bases (KBs), which are built by manually annotating words with sememes, have been successfully applied to various NLP tasks. However, existing sememe KBs only cover a few languages, which hinders the wide utilization of sememes. To address this is- sue, the task of sememe prediction for Ba- belNet synsets (SPBS) is presented, aiming to build a multilingual sememe KB based on BabelNet, a multilingual encyclopedia dictio- nary. By automatically predicting sememes for a BabelNet synset, the words in many lan- guages in the synset would obtain sememe annotations simultaneously. However, previ- ous SPBS methods have not taken full advan- tage of the abundant information in BabelNet. In this paper, we utilize the multilingual syn- onyms, multilingual glosses and images in Ba- belNet for SPBS. We design a multimodal in- formation fusion model to encode and com- bine this information for sememe prediction. Experimental results show the substantial out- performance of our model over previous meth- ods (about 10 MAP and F1 scores). All the code and data of this paper can be obtained at https://github.com/thunlp/MSGI. 1 Introduction A word is the smallest unit of language that can stand on its own (O’Grady et al., 1997), but its meaning can be further divided into smaller com- ponents. In linguistics, a sememe is defined as the minimum semantic unit (Bloomfield, 1926). It is believed by some linguists that the meanings of all the words in any language can be decomposed of a limited set of language-independent sememes, ∗Equal contribution †Corresponding author. Email: sms@tsinghua.edu.cn husband human economize family male spouse word sense sememe "married man" "carefully use" Figure 1: Sememe annotations of the English word “husband” in HowNet. For succinctness, we only show the English notations of sememes, although sememes have both English and Chinese notations in HowNet, e.g., family|家庭. which is equated with the idea of semantic primi- tives (Wierzbicka, 1996). Sememes are implicit in words and hence cannot be utilized in natural language processing (NLP) directly. To tackle this challenge, Dong and Dong (2006) manually defined about 2, 000 sememes and used them to annotate over 100, 000 English and Chinese words, whereupon a sememe knowledge base called HowNet was established. Figure 1 gives an example of sememe annotations in HowNet. HowNet is a special lexical knowledge base (KB). Different from other lexical KBs like Word- Net (Miller, 1998), …

Content truncated to 500 words