YNU-HPCC at SemEval-2022 Task 2: Representing Multilingual Idiomaticity based on Contrastive Learning

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 211 - 216 July 14-15, 2022 ©2022 Association for Computational Linguistics YNU-HPCC at SemEval-2022 Task 2: Representing Multilingual Idiomaticity based on Contrastive Learning Kuanghong Liu, Jin Wang and Xuejie Zhang School of Information Science and Engineering Yunnan University Kunming, China liukh@mail.ynu.edu.cn, {wangjin, xjzhang}@ynu.edu.cn Abstract This paper will present the methods1 we use as the YNU-HPCC team in the SemEval-2022 Task 2, Multilingual Idiomaticity Detection and Sentence Embedding. We are involved in two subtasks, including four settings. In subtask B of sentence representation, we used novel approaches with ideas of contrastive learning to optimize model, where method of CoSENT was used in the pre-train setting, and triplet loss and multiple negatives ranking loss func- tions in fine-tune setting. We had achieved very competitive results on the final released test datasets. However, for subtask A of idiomatic- ity detection, we simply did a few explorations and experiments based on the xlm-RoBERTa model. Sentence concatenated with additional MWE as inputs did well in a one-shot setting. Sentences containing context had a poor perfor- mance on final released test data in zero-shot setting even if we attempted to extract effective information from CLS tokens of hidden layers. 1 Introduction Meaning of sentence could be captured by com- positionality of word representations. However, there widely exists potentially idiomatic phrases in different languages, which are multiword expres- sions (MWEs) with idiomatic and literal meanings. Therefore, representation of idiomatic phrases is not directly compositional. A previous study has shown that the representation of idiomatic phrases by contextual models, such as ELMo (Peters et al., 2018), BERT (Devlin et al., 2019) and some of its variants, are not accurate(Garcia et al., 2021). It will be challenging to represent the MWEs cor- rectly in the downstream tasks (Tayyar Madabushi et al., 2021). SemEval-2022 Task 2, Multilingual Idiomaticity Detection and Sentence Embedding (Tayyar Madabushi et al., 2022), involves English 1The code of this paper is available at: https: //github.com/lkh-meredith/SemEval2022_ Task2_YNU-HPCC (EN), Portuguese (PT), and Galician (GL). This task includes two subtasks and each subtask con- tains two settings: • Subtask A: Determining whether a sentence contains an idiomatic expression. This is a binary classification task with two settings of zero-shot and one-shot. Zero-shot setting means that idiomatic phrases (MWEs) in train- ing examples are completely disjoint to those in development, evaluation, and test sets. In a one-shot setting, MWEs appearing in devel- opment and test sets include in the training sentences. • Subtask B: Outputing the correct Semantic Text Similarity (STS) scores between sen- tence pairs whether or not either sentence con- tains an idiomatic expression. The STS scores represent semantical similarity between two sentences ranging from 0 (least similar) to 1 (most similar). This is a regressive task with two setting of pre-train and fine-tune. In the pre-train setting, models require to be trained on any semantic text similarity dataset with- out idiom. The fine-tune setting should use provided training sets included MWEs. The remainder of this paper …

Content truncated to 500 words