41360


Taille cible

 
Michel GROSBOST
Directeur Général
IMA
DO TANK DATA SCIENCE
1er sujet  -  Long Form Question Answering  - Angela FAN – FAIR (Facebook Artificial Intelligence Research)
abstract: We will discuss long-form question answering, a task requiring elaborate and in-depth answers to open-ended questions. The dataset comprises 270K threads from the Reddit forum ``Explain Like I'm Five'' (ELI5) where an online community provides answers to questions which are comprehensible by five year olds. Compared to existing datasets, ELI5 comprises diverse questions requiring multi-sentence answers. We provide a large set of web documents to help answer the question. Automatic and human evaluations show that an abstractive model trained with a multi-task objective outperforms conventional Seq2Seq, language modeling, as well as a strong extractive baseline. However, our best model is still far from human performance since raters prefer gold responses in over 86% of cases, leaving ample opportunity for future improvement. In subsequent work, we propose constructing a local graph structured knowledge base for each query, which compresses the web search information and reduces redundancy. We show that by linearizing the graph into a structured input sequence, models can encode the graph representations within a standard Sequence-to-Sequence setting. We apply this approach to long form question answering. By feeding graph representations as input, we can achieve better performance than using retrieved text portions.

2 ème sujetautoML : Feedback on added value for Data Scientists & presentation of AIKIT, an open source AutoML solution developed by the Data Scientists team at Société Générale CIB

Les Do Tank DS sont co-organisés avec l’association deeplearning.ai

 

1er sujet  -  Long Form Question Answering  - Angela FAN – FAIR (Facebook Artificial Intelligence Research)
abstract: We will discuss long-form question answering, a task requiring elaborate and in-depth answers to open-ended questions. The dataset comprises 270K threads from the Reddit forum ``Explain Like I'm Five'' (ELI5) where an online community provides answers to questions which are comprehensible by five year olds. Compared to existing datasets, ELI5 comprises diverse questions requiring multi-sentence answers. We provide a large set of web documents to help answer the question. Automatic and human evaluations show that an abstractive model trained with a multi-task objective outperforms conventional Seq2Seq, language modeling, as well as a strong extractive baseline. However, our best model is still far from human performance since raters prefer gold responses in over 86% of cases, leaving ample opportunity for future improvement. In subsequent work, we propose constructing a local graph structured knowledge base for each query, which compresses the web search information and reduces redundancy. We show that by linearizing the graph into a structured input sequence, models can encode the graph representations within a standard Sequence-to-Sequence setting. We apply this approach to long form question answering. By feeding graph representations as input, we can achieve better performance than using retrieved text portions.

2 ème sujetautoML : Feedback on added value for Data Scientists & presentation of AIKIT, an open source AutoML solution developed by the Data Scientists team at Société Générale CIB

Les Do Tank DS sont co-organisés avec l’association deeplearning.ai


#datascience#intelligenceartificielle

Programme en cours d'élaboration
S’inscrire S’inscrire : Les inscriptions sont closes
Les inscriptions à l'événement sont désormais closes.
Vous pouvez néanmoins suivre l'événement en direct et en totalité à partir de Lundi matin (27/05) 9h via notre livestream directement sur le site de l'IMA.

Il vous suffit pour cela de cliquer sur le lien suivant : Live Stream IMAgine Day 'Journée de la Data Science'

Si vous souhaitez en savoir plus sur nos événements et accéder au prochain IMAgine Day, ou si vous souhaitez faire une demande spécifique pour participer à cet événement exceptionnel, vous pouvez contacter : michel.grosbost@ima-dt.org

N'hésitez pas à nous contacter si vous avez la moindre question, l'IMA est à votre écoute !
L'équipe IMA