Description du projet
Une nouvelle ère pour la traduction automatique
La traduction automatique du langage humain est objectif de longue date du domaine du traitement automatique du langage naturel (TALN). Cependant, les approches actuelles, telles que la traduction automatique statistique (TAS), négligent souvent les indices contextuels essentiels des traductions humaines. Les traductions manquent par conséquent d’informations pertinentes ou véhiculent des significations incorrectes, ce qui entrave la compréhension de la lecture et les rend bien souvent caduques. Dans ce contexte, le projet MultiMT, financé par le CER, adopte une approche innovante en exploitant des informations multimodales mondiales. Il développera des méthodes permettant d’intégrer des indices contextuels tels que des images, des documents connexes et des métadonnées dans les modèles de traduction. Les messages de X (ex-Twitter) et les évaluations de produits serviront d’ensembles de données de test. Cette initiative interdisciplinaire combine l’expertise du TALN, de la vision par ordinateur et de l’apprentissage automatique.
Objectif
Automatically translating human language has been a long sought-after goal in the field of Natural Language Processing (NLP). Machine Translation (MT) can significantly lower communication barriers, with enormous potential for positive social and economic impact. The dominant paradigm is Statistical Machine Translation (SMT), which learns to translate from human-translated examples.
Human translators have access to a number of contextual cues beyond the actual segment to translate when performing translation, for example images associated with the text and related documents. SMT systems, however, completely disregard any form of non-textual context and make little or no reference to wider surrounding textual content. This results in translations that miss relevant information or convey incorrect meaning. Such issues drastically affect reading comprehension and may make translations useless. This is especially critical for user-generated content such as social media posts -- which are often short and contain non-standard language -- but applies to a wide range of text types.
The novel and ambitious idea in this proposal is to devise methods and algorithms to exploit global multi-modal information for context modelling in SMT. This will require a significantly disruptive approach with new ways to acquire multilingual multi-modal representations, and new machine learning and inference algorithms that can process rich context models. The focus will be on three context types: global textual content from the document and related texts, visual cues from images and metadata including topic, date, author, source. As test beds, two challenging user-generated datasets will be used: Twitter posts and product reviews.
This highly interdisciplinary research proposal draws expertise from NLP, Computer Vision and Machine Learning and claims that appropriate modelling of multi-modal context is key to achieve a new breakthrough in SMT, regardless of language pair and text type.
Champ scientifique
- humanitieslanguages and literaturegeneral language studies
- natural sciencescomputer and information sciencesdata sciencenatural language processing
- natural sciencescomputer and information sciencesartificial intelligencecomputer vision
- natural sciencescomputer and information sciencesartificial intelligencemachine learning
Programme(s)
Thème(s)
Régime de financement
ERC-STG - Starting GrantInstitution d’accueil
SW7 2AZ LONDON
Royaume-Uni