MT Summit Conference Recap
This year's MT Summit conference was held in Macau the first week of September, with five days of insights and innovations in the world of automated translation. Research from TransPerfect, with Celia Soler Uguet from the AI team as the lead author, was also presented at the event. Our paper discussed a method of mitigating gender bias in machine translation.
MT researchers have shown that bias is present in all MT systems, and MT systems contribute to this bias with systematically unfair translations. A classic example is the sentence "The doctor asked the nurse to help her with the procedure," which is translated by most MT systems as "El médico le pidió a la enferma que le ayudara con el procedmiento", using the masculine form of the Spanish word for doctor, despite the fact that the doctor in the sentence is clearly feminine. Our work extended and generalized a previously developed method for combating these biases by introducing the gender information at translation time, and proposed two ways to infer this information during translation. Our results showed significant improvement over the baseline, with a method based on semantic analysis performing better than using LLMs.
On Monday and Tuesday, participants engaged in a series of workshops covering diverse aspects of the field. On Monday, we had the pleasure of listening to Santosh Thottingal from the Wikimedia Foundation, who shed light on the invaluable role of machine translation in bridging knowledge gaps across different Wikipedia languages. He introduced Wikipedia's content translation tool, launched in 2015, which allows editors to leverage MT to fill knowledge gaps in different Wikipedia languages.
Another workshop that caught our attention was the Second Workshop on Corpus Generation and Corpus Augmentation for Machine Translation (CoCo4MT). The workshop featured interesting research on unsupervised morphological segmentation to enhance indigenous language translation, a task riddled with challenges such as a lack of standard orthography and polysynthetic structures.
Wednesday brought forth an array of enriching sessions. In an invited session on Machine Translation for Asian Languages, Dr. Tong Xiao from Niutrans shared their endeavors in translating from Chinese into 400+ languages. Dr. Mitesh Khapra introduced the Bharat dataset, an open-source parallel corpus between 22 Indic languages and English, and their MT system trained on this dataset, IndicTrans2, all available as open-source resources. Dr. Isao Goto from Japan's NHK broadcasting company delivered a fascinating talk on the art of news translation, highlighting the structural differences between English and Japanese news.
Thursday's keynote speech by Dr. Ondrej Bojar emphasized the need for improved quality metrics and training methodologies for machine translation. He cautioned against allowing MT to be overshadowed by large language models (LLMs) and stressed the importance of diversity in research approaches.
Recommended by LinkedIn
Different conference tracks explored such topics as quality estimation, MT in the classroom, MT for subtitling, transfer learning, training with auxiliary information, and new workflows with MT/LLMs. Notable research included leveraging multilingual knowledge graphs in GPT-based translation and creating neural machine translation systems capable of handling input from disfluent L2 speakers.
On Friday, Prof. Min Zhang delivered a thought-provoking keynote speech on the pros and cons of LLMs in translation and strategies to mitigate their shortcomings.
The MT Summit conference showcased the continued dynamism and diversity of the machine translation field. From groundbreaking research to practical applications and discussions about the future of translation, it was a gathering that left attendees enriched and inspired to continue pushing the boundaries of this transformative technology.
If you are curious to know more, you can read the full proceedings of the conference, or contact aiteam@transperfect.com.