Key Takeaways from the University of Surrey’s Convergence Conference

University of Surrey Conference on Language Services

In February 2023, the University of Surrey Center for Translation Studies (CTS) in the UK held the “Convergence – Human-Machine Integration in Translation and Interpreting” conference. 

Slator attended by invitation from Elena Davitti, Associate Professor in Translation Studies, MA Interpreting and MA Translation and Interpreting Program Leader, and Principal Investigator of the SMART project.

The conference brought together a diverse group of experts from the academic, technology, and business sectors. The main purpose of this event, the first of its kind at the now 40-year-old CTS, was “to promote responsible and smart integration of the capabilities of human agents and machines in the fields of translation and interpreting.”

“Embracing the complexity: hybrid practices for interlingual communication in real time” was the title given to a very packed and dynamic opening panel. Elena Davitti began with an introduction to the SMART project, which addresses interlingual respeaking, a technique Davitti first explained to Slator in August 2022. When respeaking, a person repeats live audio content into ASR software to improve subtitle accuracy.

Davitti was joined by several representatives of media conglomerates and language services providers, including Sky, Ai-Media, ONCALL, Speechmatics, Substream, and institutions such as the University of Leeds, the University of Vigo, the National Deaf Children’s Society and Disabled Students Commission, and members of the SMART project.

Collaborating for Instant Language Conversion

The first panel addressed the interaction between language professionals and AI-enabled technologies, including speech recognition and machine translation (MT), in real time. This interaction occurs typically in live events, including media broadcasts, and may involve multiple sources of language (speakers, interpreters, etc).

The presenters discussed how automated captions compare to human captions and the role and status of speech recognition technology in multiple settings. Luke Barrett, from Sky, started his presentation by stating that “demand for subtitles is very high.”

Hybridization

It is in those settings that hybridization takes a number of forms. Illustrating the hybridization of subtitling workflows in practice, which Davitti explained in the panel introduction, industry representatives offered as examples of their applications in broadcasting a mix of technologies that include automatic speech recognition (ASR) and MT, as well as human respeakers for both monolingual and interlingual subtitling.

Eva Dorrestein and Denise Kroesen, from Substream, addressed attendees with insights on the company’s experience with hybrid work models. A combination of interpreters, human captioners, and respeakers, aided at times by MT, collaborate for online streams, webinars, and conferences and post-production tasks for businesses and institutions.

There were multiple group discussions where presenters provided a few examples of human and automation combinations, and debated the challenges of quality output and low-resource languages. Some interactions also delved into what can be potentially expected as more automation enters captioning/subtitling services. “There will always be a niche market for clients who are willing to pay more for better results, so I don’t think the human influence will disappear completely,” was the observation Kroesen offered on the matter.

Augmented Translation and Interpreting; MT errors

A few of the sessions dealt with the concepts of “augmented translation” and “augmented interpreting.” To some, this idea of the machine augmenting the job of the linguist is indeed still conceptual. On the other hand, people from industry consider this a reality and back their conviction with the fact that tools already enhance the daily work of translators and interpreters. All in all, it could be said that it is a matter of perception.

The idea that was common to all presentations on the subject is that humans remain at the center of the linguistic task. Constantin Orasan, of the University of Surrey, put it this way: “the focus is on human centric technologies meant to support rather than replace translators,” echoing the conclusion that many in all sectors represented in the conference also reached.

Other situations addressed by more than one speaker were critical errors introduced by machine translation. Recurrent MT issues are the subject of numerous research projects, and include hallucination, negation, and other limitations that at times boil down to a lack of proper resources (i.e. datasets or training approaches), especially for low-resource languages. 

One interesting concept explored was that of emotion or sentiment as a function of context, and how machine translation can fail to capture and convert the emotional load of a given discourse, particularly social media texts.

There were some dramatic examples of the erroneous emotion/sentiment issue in the English <> Chinese and English <> Arabic combinations presented by NLP scientists, where meaning was not only lost in the target, but disconcertingly twisted by Google Translate. 

Results are at times nonsensical (“countless grass and mud horses floated in my hearts” → correct: “countless ‘expletive’ your mother appeared in my mind”) and at times inflammatory (“If you blew yourself up, God would forgive you” → correct: “he blew himself up, God will never forgive”), but all times incorrect in the examples provided.

Shaping Collaboration Beyond Tested Workflows

Collaboration in language services happens not only at events like the Convergence Conference. It also occurs fluidly in organizations that bring together industry professionals and individuals. 

Gökhan Firat, of the University of Surrey, referred to the practitioner-centric collaboration that happens organically using the term “transcooperation.” It basically expresses the concept of people who contribute to the evolving human-machine interaction through organizations like The Community Language Cooperative, Barcelona’s Linguistic Services Cooperative, and many more.

Crowd-supported language conversion is one of those contributions, and a topic discussed by several speakers. Crowd translation has its uses, such as in emergency/crisis situations that require multinational intervention. The question many people raised in different ways was how artificial intelligence is going to affect crowd translation workflows. And there is no one answer, but several possible approaches involving human-machine collaboration are on the horizon.

The Convergence Conference was an extraordinary series of multidisciplinary thematic panels, each bringing together experts with innovative and often eyebrow-raising perspectives, new empirical approaches, and a deep breakdown of the different technologies and real-world applications and their limitations. 

In the era of open collaboration, we can expect to see more of these events stressing the cooperative aspect of language conversion as it increasingly revolves more around technology than it does about how it can be best utilized to aid human applications.