A Data Bootstrapping Recipe for Low-Resource Multilingual Relation Classification

Proceedings of the 25th Conference on Computational Natural Language Learning (CoNLL), pages 575–587 November 10–11, 2021. ©2021 Association for Computational Linguistics 575 A Data Bootstrapping Recipe for Low­Resource Multilingual Relation Classification Arijit Nag IIT Kharagpur arijitnag@iitkgp.ac.in Bidisha Samanta IIT Kharagpur bidisha@iitkgp.ac.in Animesh Mukherjee IIT Kharagpur animeshm@cse.iitkgp.ac.in Niloy Ganguly IIT Kharagpur, Leibniz University Hannover niloy@cse.iitkgp.ac.in Soumen Chakrabarti IIT Bombay soumen@cse.iitb.ac.in Abstract Relation classification (sometimes called ‘ex­ traction’) requires trustworthy datasets for fine­ tuning large language models, as well as for evaluation. Data collection is challenging for Indian languages, because they are syntacti­ cally and morphologically diverse, as well as different from resource­rich languages like En­ glish. Despite recent interest in deep gen­ erative models for Indian languages, relation classification is still not well­served by pub­ lic data sets. In response, we present IndoRE, a dataset with 21K entity­ and relation­tagged gold sentences in three Indian languages, plus English. We start with a multilingual BERT (mBERT) based system that captures entity span positions and type information, and pro­ vides competitive monolingual relation clas­ sification. Using this system, we explore and compare transfer mechanisms between lan­ guages. In particular, we study the accuracy­ efficiency tradeoff between expensive gold in­ stances vs. translated and aligned ‘silver’ in­ stances. We release the dataset for future re­ search.1 1 Introduction Relation classification (sometimes called relation ‘extraction’ or RE) is the task of identifying a se­ mantic relation (from a catalog of canonical re­ lations) that holds between two nominal entities in text. Related to semantic role labeling, it is a critical capability for question answering and knowledge graph (KG) augmentation. The key challenge in RE is the diversity of textual expres­ sions of canonical relations. WikiData, among the largest public KGs, has a hundred million entities with aliases but only thousands of well­ instantiated canonical relations, with relatively sparse textual descriptions of relations, particu­ larly in low­resource languages. 1https://github.com/NLPatCNERG/IndoRE Low­resource languages, including many In­ dian languages, neither have well­tuned standard NLP pipelines, nor unsupervised corpora as large as resource­rich languages. Not only are they of­ ten morphologically distant from resource­rich lan­ guage families, but even their syntax for relation expression differs. E.g., while subject­predicate­ object (infix) is more common in English, subject­ object­predicate (postfix) is more common in Indic languages: “Ravi went to Delhi” (En) vs. रिव िद- ल्ली गया था/“Ravi Delhi went­did” (Hi). Word order is more relaxed: िदल्ली गया था रिव/“Delhi went­did Ravi” is acceptable; only the entity types help ori­ ent the predicate. Verb particles are more common than in English: “Talk to him” (En) becomes उससे बात करो/ “him­to talk do” (Hi) or তার সােথ কথা বলুন/“him with speech talk” (Bn). Transliteration and code switching are rampant: “I called him yes­ terday” (En) translates to “ मैंने कल उसे कॉल िकया था”/“I yesterday him call did­past”, where call is transliterated. Against this backdrop of diverse textual expres­ sion of relational facts, collecting high­quality la­ beled data for Indian language RE is a major chal­ lenge, particularly given data­hungry state­of­the­ art neural RE …

Content truncated to 500 words