Tutorial
Linguistically Motivated Neural Machine Translation
27th June 09:00-12:30, Diamond LT2
Abstract: In this tutorial, we focus on a niche area of neural machine translation (NMT) that aims to incorporate linguistics into different stages in the NMT pipeline, from pre-processing to model training to evaluation. We first introduce the background of NMT and fundamental analysis tools, such as word segmenters, part-of-speech taggers, and dependency parsers. We then cover topics including 1) word/subword segmentation, and character decomposition during MT data pre-processing, 2) incorporating direct and indirect linguistic features into NMT models, and 3) fine-grained linguistic evaluation for MT systems. We reveal the impact of orthography, syntax, and semantics information on translation performance. This tutorial is mainly aimed at researchers interested in the intersection of linguistics and low-resource machine translation. We hope this tutorial inspires and encourages them to develop linguistically motivated high-quality MT systems and evaluation benchmarks.
Repository: The accompanying repository contains a comprehensive reading list and other useful information for the tutorial.
Presenters
Haiyue Song is a technical researcher at the Advanced Translation Technology Laboratory, National Institute of Information and Communications Technology (NICT), Japan. He obtained his Ph.D. at Kyoto University. His research interests include machine translation, large language models, subword segmentation, and decoding algorithms. He has MT and LLMs related publications in TALLIP, AACL, LREC, ACL, and EMNLP.
Hour Kaing is a researcher at the Advanced Translation Technology Laboratory, National Institute of Information and Communications Technology (NICT), Japan. He received his B.S. from Institute of Technology of Cambodia, Cambodia, his M.Sc from University of Grenoble 1, France, and his Ph.D. from NARA Institute of Science and Technology, Japan. He is interested in linguistic analysis, low-resource machine translation, language modeling, and speech processing. He has publications in TALLIP, EACL, PACLIC, LREC, and IWSLT.
Raj Dabre is a senior researcher at the Advanced Translation Technology Laboratory, National Institute of Information and Communications Technology (NICT), Japan and an Adjunct Faculty at IIT Madras, India. He received his Ph.D. from Kyoto University and Masters from IIT Bombay. His primary interests are in low-resource NLP, language modeling and efficiency. He has published in ACL, EMNLP, NAACL, TMLR, AAAI, AACL, IJCNLP and CSUR.
Programme
9:00-9:20: Introduction
9:20-10:20: Augmenting NMT Architectures with Linguistic Features
10:20-10:50: Coffee break
10:50-11:20: Linguistically Motivated Tokenization and Transfer Learning
11:20-11:40: Linguistically Aware Decoding
11:40-12:00: Linguistically Motivated Evaluation
12:00-12:15: Conclusions
12:15-12:30: QA