Metric Learning in Multilingual Sentence Similarity Measurement for Document Alignment

Proceedings of Recent Advances in Natural Language Processing, pages 1150–1157 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_129 1150 Metric Learning in Multilingual Sentence Similarity Measurement for Document Alignment Charith Rajitha, Lakmali Piyarathne, Dilan Sachintha, Surangika Ranathunga Department of Computer Science and Engineering, University of Moratuwa, Katubedda 10400, Sri Lanka. [rajitha.16,lakmali.16,dilansachintha.16,surangika]@cse.mrt.ac.lk Abstract Document alignment techniques based on multilingual sentence representations have re- cently shown state of the art results. How- ever, these techniques rely on unsupervised distance measurement techniques, which can- not be fined-tuned to the task at hand. In this paper, instead of these unsupervised distance measurement techniques, we employ Metric Learning to derive task-specific distance mea- surements. These measurements are super- vised, meaning that the distance measurement metric is trained using a parallel dataset. Us- ing a dataset belonging to English, Sinhala, and Tamil, which belong to three different language families, we show that these task- specific supervised distance learning metrics outperform their unsupervised counterparts, for document alignment. 1 Introduction Document alignment is an important precursor to build parallel corpora from noisy data sources. Doc- ument alignment is also useful in multilingual Infor- mation Retrieval, as well as for tasks such as false news detection across different languages. Tradi- tionally, document alignment has been achieved by metadata-based methods (Resnik, 1998, 1999) and translation based methods (Uszkoreit et al., 2010; Dara and Lin, 2016). Metadata-based meth- ods rely on exploiting the meta information of the selected data sources, which may be inconsistent across different sources. On the other hand, trans- lation based methods assume the availability of a Machine Translation (MT) system. To overcome these issues, recent research has exploited the use of multilingual sentence represen- tations (multilingual sentence embeddings) such as LASER (Artetxe and Schwenk, 2019)1. Here, vector representations are derived for documents in both source and target languages. Then, for 1https://github.com/facebookresearch/LASER a given document in the source side, the most similar counterpart is identified from the target side. Euclidean distance and cosine distance are used in existing document alignment systems (Uszkoreit et al., 2010; El-Kishky and Guzm´an, 2020). However, these similarity metrics cannot be fine-tuned for the selected task or data at hand. The alternative is to use Metric Learning, which focuses on constructing a problem-specific distance metric automatically from data (de Vazelhes et al., 2019). Metric Learning-based distance measurement techniques have been successfully employed in image classification and image identification tasks (Pacchiardi et al., 2021). In this paper, we apply two such Metric Learning algorithms on multilingual sentence represen- tations to identify similar document pairs in a bilingual setting. We experimented with Sparse High-Dimensional Metric Learning (SDML) (Qi et al., 2009), and Information Theoretic Metric Learning (ITML) (Davis et al., 2007) algorithms. To the best of our knowledge, this is the first work to exploit the use of Metric Learning with respect to multilingual sentence representations for the purpose of document alignment. El-Kishky and Guzm´an (2020)’s system was used as the baseline. This research can be considered as the best in this line of research at the moment. They used …

Content truncated to 500 words