Team Innovators at SemEval-2022 for Task 8: Multi-Task Training with Hyperpartisan and Semantic Relation for Multi-Lingual News Article Similarity

Proceedings of the 16th International Workshop on Semantic Evaluation (SemEval-2022), pages 1163 - 1170 July 14-15, 2022 ©2022 Association for Computational Linguistics Team Innovators at SemEval-2022 for Task 8: Multi-Task Training with Hyperpartisan and Semantic Relation for Multi-Lingual News Article Similarity Nidhir Bhavsar* Navrachana University Vadodara, India 18103488@nuv.ac.in Rishikesh Devanathan* IIT Patna Patna, India rishikesh_2001cs85@iitp.ac.in Aakash Bhatnagar* Navrachana University Vadodara, India 18124526@nuv.ac.in Muskaan Singh, Petr Motlicek Speech and Audio Processing Group, IDIAP Research Institute, Switzerland (msingh,petr.motlicek)@idiap.ch Tirthankar Ghosal ÚFAL, MFF Charles University, Czech Republic ghosal@ufal.mff.cuni.cz Abstract This work represents the system proposed by team Innovators for SemEval 2022 Task 8: Multilingual News Article Similarity (Chen et al., 2022). Similar multilingual news articles should match irrespective of the style of writ- ing, the language of conveyance, and subjective decisions and biases induced by medium/outlet. The proposed architecture includes a machine translation system that translates multilingual news articles into English and presents a mul- titask learning model trained simultaneously on three distinct datasets. The system lever- ages the PageRank algorithm for Long-form text alignment. Multitask learning approach allows simultaneous training of multiple tasks while sharing the same encoder during training, facilitating knowledge transfer between tasks. Our best model is ranked 16 with a Pearson score of 0.733. We make our code accessible here1 1 Introduction Over the last decade, English has been one of the most dominant languages on the internet. However, the number of non-English websites is rapidly in- creasing. From 2001-to 11, online use of English increased at a slower rate than that of Spanish, Chi- nese, etc. Approximately 4 billion people connect to the internet every day, but only half of them ac- cess web pages written in English (Pimienta, 2009). This creates a severe problem regarding verifying the integrity of the documentation because most systems in use are enhanced with English as the medium of information delivery. 1https://github.com/rdev12/Multilingual-News-Article- Similarity * First three authors have equal contribution SemEval has conducted similar tasks on sentence-level semantic textual similarity in the past. The SemEval 2017 Task 1 (Cer et al., 2017) dealt with finding the similarity between sentence pairs of both monolingual and cross-lingual nature. ECNU (Tian et al., 2017) was the best model that used feature engineering and deep averaging net- work (DAN) (Iyyer et al., 2015). Furthermore, language-agnostic representation for sentence similarity described in Tiyajamorn et al. (2021) uses meaning embedding to estimate the cross-lingual sentence similarity without using human annotations. It improves the performance of any pre-trained multilingual sentence encoder, even in low-resource languages, with a few thousand parallel sentence pairs. Inspired by Ham and Kim (2021), we explored the concept of semantically aligned multilingual sentence embedding. It covers the biases induced by monolingual similarity evaluation and multilin- gual sentence retrieval to generate language-aware embeddings. This method aligns semantic struc- tures across different languages and uses a teacher network to distill the knowledge of pivot languages, thus achieving state-of-the-art STS 2017 multilin- gual corpora. Our model is inspired by a multitask training approach using a BERT-based encoder. We …

Content truncated to 500 words