UAlberta at SemEval-2020 Task 2: Using Translations to Predict Cross-Lingual Entailment

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 263–269 Barcelona, Spain (Online), December 12, 2020. 263 UAlberta at SemEval-2020 Task 2: Using Translations to Predict Cross-Lingual Entailment Bradley Hauer, Amir Ahmad Habibi, Yixing Luan, Arnob Mallik, Grzegorz Kondrak Alberta Machine Intelligence Institute, Department of Computing Science University of Alberta, Edmonton, Canada {bmhauer,amirahmad,yixing1,amallik,gkondrak}@ualberta.ca Abstract We investigate the hypothesis that translations can be used to identify cross-lingual lexical entailment. We propose novel methods that leverage parallel corpora, word embeddings, and multilingual lexical resources. Our results demonstrate that the implementation of these ideas leads to improvements in predicting entailment. 1 Introduction In this paper, we discuss the University of Alberta systems for SemEval-2020 Task 2: Predicting Multilin- gual and Cross-Lingual Lexical Entailment (Glavaˇs et al., 2020). We focus on the subtask of cross-lingual binary lexical entailment (LE). Vyas and Carpuat (2016) define this task as “the task of detecting whether the meaning of a word in one language can be inferred from the meaning of a word in another language.” They note its potential applications to machine translation, question answering, as well as to cross-lingual inference and entity linking. LE is related to hypernym detection, with the former being more general (Upadhyay et al., 2018). Our principal objective is to provide evidence for the hypothesis that translations are useful in predicting cross-lingual entailment. It has been observed in prior work on cross-lingual lexical semantics that translations may be broader in meaning than the original text (Bentivogli and Pianta, 2000; Rudnicka et al., 2012). In particular, translations may represent concepts entailed by the translated concept. For example, from the English phrase “you gave me the bottle”, and its Italian translation “mi hai dato il contenitore”, it can be inferred that bottle entails contenitore (“container”).1 We are interested in leveraging this phenomenon to perform unsupervised LE prediction. Our use of bitexts, word embeddings, and multilingual wordnets builds upon prior work. Qiu et al. (2018) observe that similar words share entailments, and so semantic similarity can be used to detect additional entailment pairs. Cross-lingual word embedding similarity has been used by Hauer et al. (2017) to identify translations, and by Hauer et al. (2019) to detect frequent word senses. Mehdad et al. (2011) leverage bitext alignment for textual entailment. The principal contribution of this paper is the presentation and evaluation of LE prediction methods that leverage: (1) translation mining for entailment classification; (2) monolingual word embeddings for expanding the set of entailment pairs; and (3) multilingual lexical resources for improving translation alignment. To the best of our knowledge, we are the first to apply these ideas to cross-lingual LE prediction, and demonstrate that each of them improves LE prediction performance, especially in a low-resource setting. 2 Methods In this section, we outline our bitext-based approach to predicting cross-lingual entailment. Section 2.1 describes our base method, BITEXT, which mines entailment pairs from word alignments in a bilingual parallel corpus. Section 2.2 describes an enhanced method, VECTORS, which identifies additional This work is licensed under a …

Content truncated to 500 words