🧑‍🔬 Iulian Vlad Serban

Iulian Vlad Serban works on open-domain dialogue systems, with papers here spanning 2015–2019. The work keeps returning to three questions: how conversational agents can learn directly from large collections of human exchanges, how their replies can be made less generic, and how anyone can tell whether a reply is good. The methods are mostly hierarchical recurrent neural networks and latent-variable generative models, applied to large scraped or crowdsourced corpora and checked against human ratings gathered on Mechanical Turk.

Among the most cited papers is Building End-To-End Dialogue Systems Using Generative Hierarchical Neural Network Models (2015). It recasts conversation as generative modelling by adapting the hierarchical recurrent encoder-decoder (HRED), originally built for query suggestion. In this setup one network reads each utterance, a second tracks the dialogue across turns, and a decoder writes the next reply. The paper also released the MovieTriples dataset. The Ubuntu Dialogue Corpus (2015) tackled the shortage of public conversational data by extracting multi-turn exchanges from Ubuntu support channels on IRC, separating interleaved threads with a heuristic. A Hierarchical Latent Variable Encoder-Decoder Model for Generating Dialogues (2016) introduced VHRED. This model extends HRED with a Gaussian latent variable, drawn once per utterance and trained through a variational bound, so that replies stay tied to the wider conversation.

The later papers turn toward evaluation. Towards an Automatic Turing Test (2017) frames the scoring of responses as a learning problem. Its model, ADEM, is trained on human appropriateness judgments of Twitter replies and compares context, reference and candidate to predict a human-like score, rather than counting shared words as BLEU does. ConvAI2 (2019) describes a NeurIPS 2018 competition built on Persona-Chat. It offered chit-chat research a shared benchmark that paired automatic metrics with human judging.

Serban's papers here were written at the Université de Montréal (2015–2020), with some later bylines also at Mila - Quebec Artificial Intelligence Institute. Frequent co-authors include Joelle Pineau, Ryan Lowe and Yoshua Bengio, along with Aaron Courville and Alessandro Sordoni. All of the work belongs to the topic of dialogue systems.

Iulian Vlad Serban has 6 papers in our library, in 2015–2019, most of them on Dialogue systems, Evaluation & benchmarks; often with Joelle Pineau, Ryan Lowe, Yoshua Bengio.