SHONGLAP: A Large Bengali Open-Domain Dialogue Corpus

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 5797–5804 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 5797 SHONGLAP: A Large Bengali Open-Domain Dialogue Corpus Syed Mostofa Monsur†, Sakib Chowdhury†∗, Md Shahrar Fatemi†∗, Shafayat Ahmed‡ †Celloscope Ltd.,‡Virginia Polytechnic Institute and State University mostofa.monsur@cellosco.pe, 1606044@eee.buet.ac.bd, 1605007@ugrad.cse.buet.ac.bd shafayatpiyal@vt.edu Abstract We introduce SHONGLAP, a large annotated open-domain dialogue corpus in Bengali language. Due to unavailability of high-quality dialogue datasets for low-resource languages like Bengali, existing neural open-domain dialogue systems suffer from data scarcity. We propose a framework to prepare large-scale open-domain dialogue datasets from publicly available multi-party discussion podcasts, talk-shows and label them based on weak-supervision techniques which is particularly suitable for low-resource settings. Using this framework, we prepared our corpus, the first reported Bengali open-domain dialogue corpus (7.7k+ fully annotated dialogues in total) which can serve as a strong baseline for future works. Experimental results show that our corpus improves performance of large language models (BanglaBERT) in case of downstream classification tasks during fine-tuning. Keywords: open-domain dialogues, weak supervision, low-resource dialogue 1. Introduction Due to recent advances in conversational systems re- search, a number of high quality dialogue datasets are being made publicly available (Rameshkumar and Bai- ley, 2020; Li et al., 2017b; McCowan et al., 2005; Janin et al., 2004; Canavan et al., 1997). These datasets have a wide range of applications including dialogue sum- marization, dialogue understanding, dialogue act clas- sification, disagreement modeling and dialogue state tracking. In general, dialogue systems can be classified into two major types: task-oriented and open-domain. Task- oriented dialogue systems are intended to assist peo- ple in achieving certain goals. Tasks are well-defined according to their use cases, and conversational sys- tems are customized to these domains (Yan, 2018). Be- cause of their attractive business value, these systems have been successfully deployed in a wide range of real-world applications such as online shopping (Yan et al., 2017), restaurant searching and reservation (Wen et al., 2017), customer care services at financial orga- nizations and technical assistance (Huang et al., 2020a; Peng et al., 2018; Li et al., 2017a). Additionally, an increasing demand for task-oriented dialogue agents is observed in healthcare services such as automatic di- agnosis,(Wei et al., 2018), mental state classification (Dosovitsky et al., 2020) and promoting health edu- cation (Liednikova et al., 2021; Brixey et al., 2017). Open-domain dialogue systems, on the other hand, are far more challenging to build due to their open-ended objective. Task-oriented dialogue systems have a pre- defined task-dependent workflow. In contrast with that, most of the currently available open-domain neural dialogue systems are not grounded in the real world which prevents these systems from seamlessly convers- ing about subjects that relate to the user’s environmen- *These authors contributed equally to this work. tal context (Huang et al., 2020b). Researchers have re- cently begun to investigate methods to ground open- domain conversation systems in real-world events and knowledge (Ghazvininejad et al., 2018; Mostafazadeh et al., 2017; Qin et al., 2019). …

Content truncated to 500 words