Alignment-Augmented Consistent Translation for Multilingual Open Information Extraction

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 2502 - 2517 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Alignment-Augmented Consistent Translation for Multilingual Open Information Extraction Keshav Kolluru1*, Mohammed Muqeeth1*, Shubham Mittal1, Soumen Chakrabarti2, and Mausam1 1 Indian Institute of Technology Delhi 2 Indian Institute of Technology Bombay keshav.kolluru@gmail.com, muqeeth101@gmail.com, shubhamiitd18@gmail.com soumen@cse.iitb.ac.in, mausam@cse.iitd.ac.in Abstract Progress with supervised Open Information Extraction (OpenIE) has been primarily lim- ited to English due to the scarcity of train- ing data in other languages. In this pa- per, we explore techniques to automatically convert English text for training OpenIE sys- tems in other languages. We introduce the Alignment-Augmented Consistent Translation (AACTRANS) model to translate English sen- tences and their corresponding extractions con- sistently with each other — with no changes to vocabulary or semantic meaning which may result from independent translations. Using the data generated with AACTRANS, we train a novel two-stage generative OpenIE model, which we call GEN2OIE, that outputs for each sentence: 1) relations in the first stage and 2) all extractions containing the relation in the second stage. GEN2OIE increases rela- tion coverage using a training data transforma- tion technique that is generalizable to multiple languages, in contrast to existing models that use an English-specific training loss. Evalua- tions on 5 languages — Spanish, Portuguese, Chinese, Hindi and Telugu — show that the GEN2OIE with AACTRANS data outperforms prior systems by a margin of 6-25% F1.1 1 Introduction Open Information Extraction (OpenIE) is the task of converting unstructured text to semi-structured tuples of the format <subject; relation; object>, where these three components are textual phrases, broadly extracted from the original text (Etzioni et al., 2011). OpenIE tuples have shown util- ity in various downstream tasks (Mausam, 2016) like Question Answering (Fader et al., 2013; Khot et al., 2017), Machine Reading (Poon et al., * denotes equal contribution 1Code and models released at github.com:dair-iitd/moie 2010), Multi-Document Summarization (Chris- tensen et al., 2014; Fan et al., 2019), Schema Induc- tion (Balasubramanian et al., 2013), and Knowl- edge Base Construction (Gupta et al., 2019; Chan- drahas and Talukdar, 2021). With widespread adoption of Deep Learning in NLP, Open Information Extraction (OpenIE) sys- tems have gone through a paradigm shift from us- ing rule-based, statistical systems to supervised neural models. However, both types of systems have been limited to only a few languages – ear- lier systems required language-specific OpenIE in- sights, and current systems require annotated train- ing corpus that pose a barrier, particularly for low- resource languages. Related tasks such as Semantic Role Labeling face similar challenges in extending to multiple lan- guages. X-SRL (Daza and Frank, 2020) addresses this by automatic translation of English sentences to the target language followed by label projection to infer the semantic role labels in the translated sentence. However, translating the sentence alone may be insufficient for OpenIE because the gener- ated tuples (also referred to as extractions) can in- clude additional words absent in the sentence or …

Content truncated to 500 words