La compréhension des verbatims doit permettre de mettre en œuvre plus facilement des actions appropriées.
“We may not necessarily notice it, but automatic processing of natural language is present almost everywhere to some extent: in spell checkers, in automatic translators, in search engines, in online chat agents…” Of Polish origin, Aleksandra Guerraz arrived in Paris in 2000 to take a degree in Language Sciences. She quickly became interested in the automatic processing of natural language (TALN), which combines linguistics and informatics, and allows you to “Transcend the domain of pure linguistics and widen it to that of applied linguistics”.
In 2004, at Orange, she embarked on a thesis about speech synthesis – which transforms a written text into speech. She was then recruited by the R & D Centre at Lannion, in the TALN team: she has recently joined the team for man-machine dialogue processing, and tends to specialise in “text mining”. “At the time”, she recalled, “the big issues were concerned with processing search engine requests, or data extraction, including the recognition of named entities, i.e. of informative words (the names of individuals, names of places, organisations, dates, etc.) within a document.”
A simple and intuitive interface
Nowadays, Aleksandra Guerraz works mainly on EVA, a simple semantic analysis tool. EVA is intended for non-expert colleagues, who need to analyse thousands of transcripts from polls, surveys or telephone calls. To fully exploit these transcripts, statistical classification and semantic analysis tools are available. However, these rely on advanced skills in linguistics, statistics and computer science.
“At present, our colleagues rely either on us or external providers to exploit the content of these transcripts betsti, Aleksandra explained. We wanted to facilitate the use of tools for statistical and sematic analysis via one simple and intuitive interface that masks the complexity, so they could create their own analyses.”
Basically, EVA allows the transcripts to be grouped into categories that contain similar words. “The advantage of this tool, she told us, is that it allows users to define their categories themselves … They are guided a little by suggestions for statistical groupings, but they can then refine them to adapt to the needs of their buisiness.” The tool also allows transcripts consisting of open questions, of the type “Your opinion is important to us” to be processed.
An oriented “UX” design
Once the transcripts have been classified, they have to be presented in a readable way, which allows a group of comments on a particular issue to be retained or dismissed . “The understanding of transcripts by operational bodies should allow them to implement appropriate actions more easily, by comparing them to those observed in their business, said the scientist. Currently, we are working with several entities (human resources, information management system…) to find out if the tool corresponds effectively to their requirements”.
The design is “UX” oriented, i.e. it takes the needs of the users into account. Except that in this case, the users are not Orange clients, but a group of their employees. “We first observed how they went about analysing their transcripts, and thinking about the way we could help them .”
One meeting conducted in collaboration with various players at Orange HQ: “EVA was using tools like Khiops, which had been developed by another team, who we have obviously collaborated with. We have also done a lot of work with an ergonomist, graphic designers, developers … “, ended Aleksandra.