Investigating Effectiveness of Multi-Encoder for Conversational Neural Machine Translation

Investigating Effectiveness of Multi-Encoder for Conversational Neural Machine Translation Baban Gain1, Ramakrishna Appicharla1, Soumya Chennabasavaraj2, Nikesh Garera2, Asif Ekbal1 and Muthusamy Chelliah2 1Department of Computer Science and Engineering, Indian Institute of Technology Patna, India {gainbaban, ramakrishnaappicharla, asif.ekbal}@gmail.com 2Flipkart, India {soumya.cb, nikesh.garera, muthusamy.c}@flipkart.com Abstract Multilingual chatbots are the need of the hour for modern business. There is increasing de- mand for such systems all over the world. A multilingual chatbot can help to connect distant parts of the world together, without sharing a common language. We participated in WMT22 Chat Translation Shared Task. In this paper, we report descriptions of methodologies used for participation. We submit outputs from multi- encoder based transformer model, where one encoder is for context and another for source utterance. We consider one previous utterance as context. We obtain COMET scores of 0.768 and 0.907 on English-to-German and German- to-English directions, respectively. We submit- ted outputs without using context at all, which generated worse results in English-to-German direction. While for German-to-English, the model achieved a lower COMET score but slightly higher chrF and BLEU scores. Further, to understand the effectiveness of the context encoder, we submitted a run after removing the context encoder during testing and we obtain similar results. 1 Introduction Translation of Dialogues is a crucial part of build- ing multilingual chatbots. With easier access to the internet than ever, we have the opportunity to connect with different people with different lan- guages. However, language remains a barrier to smooth communication. Using automated machine translation systems can alleviate such issues. How- ever, most of the general MT systems are not very suitable for conversations. This is due to additional challenges chat translation possesses that general domains do not have. This includes the presence of noisy utterances. Compared to other domains, chat is more prone to contain noisy sentences. This comes from multiple sources, as follows. a) Key- board typos: Spelling mistakes that occurred due to quick typing. In this case, often, some char- acters are replaced by nearby characters on the keyboard. Further, the insertion of extra charac- ters or the absence of some characters is also com- mon. b) Intentional shortening of Words: Users often use short forms of words by removing cer- tain characters (primarily vowels) while keeping the pronunciation similar to the correct word (For example, ‘hw’ instead of ‘how’). c) Grammatical Errors: Conversations usually occur in an informal setting, and grammar is mostly ignored as long as the meaning is understood correctly. However, this makes it difficult to translate. Further, there are other challenges, like context dependency. That is, the utterances can be ambiguous, and the cor- rect meaning of an utterance can not be understood without referring to its dialogue history. In this paper, we use a multi-encoder transformer to translate chat utterances. We use six encoder layers for source text and one encoder layer for context. For better comparison, we have submit- ted translations from two other models. To test the effectiveness of context, we did not provide context …

Content truncated to 500 words