DATE

October 24, 2016

TIME

3pm

LOGISTICS

• hangouts

AGENDA

- report progress for all tasks

PARTICIPANTS

Jean-Pierre Lorre, Sammy BenHamiche, Tom Jorquera, Maxence Bunel, Zied Sellami (LINAGORA); Polykarpos Meladianos, Antoine Tixier (LIX)

MINUTES

• Linagora is working on the speech-to-text model for French. They are experimenting with corpus enrichment strategies such as noise/distortion and evaluating how it impacts performance. So far the gains are marginal. A new PhD student starts his thesis on speech-to-text for French.
• LIX has started investigating Natural Language Generation as a way to improve the quality of the final summaries (for the offline system). Indeed, the keywords extracted by the system are good but since it uses an extractive approach in the end the readability of the summaries is not optimal. Two 3rd year students from Polytechnique have started working on this topic using the following paper as a starting point: http://anthology.aclweb.org/W/W13/W13-21.pdf#page=152
• Zied Sellami recommends to consider the following papers as well: https://www.irit.fr/publis/LILAC/BVA-VERB01.pdf, http://thesesups.ups-tlse.fr/2741/1/2015TOU30023.pdf
• for communication between speech-to-text module and keyword extraction & recommendation module, a very simple restAPI will be developed by LIX enabling a start session request - then LIX needs to receive a stream of text (WebSocket could be used since Hubl.in is written in node.js) - LIX needs to receive chunks of text every 60 seconds, with time stamp. Interval duration could also be a user parameter. For the offline system a simple restAPI will be enough
• Linagora will prepare and send a proposal for restAPI specifications both for the real-time and offline system
• LIX is going to explore the implementation of automatic language detection in both systems to enable their use in other languages than English