🧑‍🔬 Laurent Charlin

Laurent Charlin's papers here deal with dialogue systems, specifically open-domain conversational models that produce a reply from the preceding exchange alone. Two questions recur: how a neural generator can keep track of a whole conversation instead of only the next word, and how such systems should be judged. The methods range from recurrent encoder-decoder networks with variational training to empirical studies that compare automatic scores with human ratings. That second theme links the work to evaluation and benchmarks.

The most cited of these works is A Hierarchical Latent Variable Encoder-Decoder Model for Generating Dialogues. It introduced VHRED, which adds a continuous Gaussian latent variable, drawn once for each utterance, to a hierarchical recurrent dialogue model. Training optimises a bound on the likelihood, much as variational autoencoders do. The variable is meant to counter the dull, generic answers that come from purely token-level randomness. Close behind is How NOT To Evaluate Your Dialogue System. That study gathered human judgments of responses from several generative and retrieval models. It then checked how well word-overlap measures such as BLEU, METEOR and ROUGE, along with embedding-based measures, agree with those judgments. Both papers appeared in 2016 and are the natural starting point for a reader.

Since both papers date from the same year, the most recent work here is the same pair: one builds better conversational generators and the other questions the tools used to score them. Charlin is currently affiliated with HEC Montréal, Mila - Quebec Artificial Intelligence Institute and the Canadian Institute for Advanced Research. Earlier bylines list McGill University, Columbia University and the University of Toronto. Co-authors on these papers include Iulian Vlad Serban, Joelle Pineau and Ryan Lowe, as well as Aaron Courville, Alessandro Sordoni and Yoshua Bengio.

Laurent Charlin has 2 papers in our library, in 2016, most of them on Dialogue systems, Evaluation & benchmarks; often with Iulian Vlad Serban, Joelle Pineau, Ryan Lowe.