Overprotective Training Environments Fall Short at Testing Time: Let Models Contribute to Their Own Training
Fiche du document
3 septembre 2021
- doi: 10.4000/books.aaccademia.8934
- eisbn: 9791280136336
https://www.openedition.org/12554 , info:eu-repo/semantics/openAccess
Mots-clés
Computational Linguistics Fine-grained sentiment analysis Distributional Semantics Quantitative Linguistic Investigations Gender Bias Depression from Social Media Online Hate Speech Automatic Sarcasm Detection TrAVaSI AriEmozione AEREST COVID-19 Linguistic Ostracism in Social Networks Multilingual NLU E3C Project DistilBERT Twitter during Pandemic LAN000000 Linguistics CBXSujets proches
Skills trainingCiter ce document
Alberto Testoni et al., « Overprotective Training Environments Fall Short at Testing Time: Let Models Contribute to Their Own Training », Accademia University Press, ID : 10.4000/books.aaccademia.8934
Métriques
Partage / Export
Résumé
Despite important progress, conversational systems often generate dialogues that sound unnatural to humans. We conjecture that the reason lies in their different training and testing conditions: agents are trained in a controlled “lab” setting but tested in the “wild”. During training, they learn to generate an utterance given the human dialogue history. On the other hand, during testing, they must interact with each other, and hence deal with noisy data. We propose to fill this gap by training the model with mixed batches containing both samples of human and machine-generated dialogues. We assess the validity of the proposed method on GuessWhat?!, a visual referential game.