LOQUATICS NEWS READER


It has been nearly two years since ChatGPT burst onto the scene in November 2022, taking AI mainstream and top of mind for pretty much all players in the language industry.

The way Mark Lawyer sees it, 2023 was the year of POCs for AI; 2024, the year of production — that is, implementing, putting into production, and adopting that technology. 

Lawyer, President of Regulated Industries & Linguistic AI at Super Agency RWS, joined RWS President of Enterprise Services Vasagi Kothandapani to discuss the company’s approach to AI. 

The pressure on the language industry has intensified throughout 2024, Lawyer said, and the results of a June 2024 RWS global research study of 200 C-suite executives shows that the appetite for AI is often client-led.

While 87% of respondents feel pressure themselves to implement GenAI solutions, and 76% are excited about the potential benefits, more than one-third (36%) have concerns, such as worries that resources would be better spent elsewhere.

The C-suite, in turn, is feeling the heat from their end clients. Content can be a make-or-break experience for the customer experience, adding up to a tangible business impact for global companies. Since 2016, Lawyer explained, the amount of data has seen an “explosive increase,” accelerating demand for personalized content across channels and formats.

Competitors who can localize faster and evolve to meet clients’ changing needs and preferences will always have a competitive advantage, he added.

RWS’ proposed solution, Genuine Intelligence, aims to refine the advancement of large language models and GenAI with human-in-the-loop expertise.

Evolve to Adapt

In particular, humans are absolutely critical for regulated industries; once hooked up to Evolve, RWS’ AI-powered translation solution, they can contribute to major gains for clients.

RWS’ TMS is Trados Enterprise, the interface through which expert humans-in-the-loop “interact” with content.

The content is first fed into adaptive MT, and then pushed through to a proprietary quality estimation LLM, built and calibrated on human decisions. The quality estimation LLM scores the content and sends it back into the workflow for automated post-editing in a private, secure LLM for improvement. The content is then rerouted back to quality estimation. 

This process continues until the content is deemed acceptable for its end use. Since all the edits go back into the model, it self-improves, resulting in overall higher-quality output over time.

According to Lawyer, users have seen 20% faster time to market and up to 65% improved efficiencies across the translation supply chain, depending on the language combination and content type.

Solving the Data Challenge

Data may be the backbone of building any AI, but that does not mean that it is easy. 

“Most of our TrainAI clients face significant challenges when it comes to preparing the data needed to train or fine-tune AI,” Kothandapani shared.

Even “insiders” can struggle with data, she added, noting that almost 80% of data scientists’ time is spent preparing data in order to train a machine learning model. That task can be daunting for clients without language expertise, whose data might be scattered, within the organization and/or in the public domain.

The granular jobs may seem unrelated — from mapping and collecting information to prepping data models, from weeding out bias to ensuring data privacy — but underpinning all of them is a major pain point: the quality and accuracy of the data. Simply put, can the data be trusted?

Only a human in the feedback loop can currently answer that question, via response rating, verification of content, content moderation, and fact extraction, among other work. 

Humans will be the linchpin in these emerging workflows and new technological systems for the foreseeable future, especially as multimodal models — those that combine text, speech, audio, and video, all embedded with language — become more prevalent.

“There’s a synergy that comes in between the technology as well as the human element,” Kothandapani concluded. “So let’s embrace ‘Genuine Intelligence,’ which is really where our transformative technology could head, but obviously with the help of our human intelligence.”



Source link

News provided by