François Mairesse


Update: I joined the Amazon Alexa team in 2013

[ Research | Publications | Theses | Online demos | Datasets and tools | Talks | Teaching ]
[ Google Scholar | Linkedin profile | Young Researchers' Roundtable on Spoken Dialogue Systems ]

Machine learning manager and individual contributor at Amazon, leading a team working on conversational AI for Alexa. I am hiring software engineers and applied scientists in San Francisco. Previously I was a research associate at the Cambridge University Machine Intelligence Lab, in the Dialogue Systems Group headed by Prof. Steve Young. I was part of the EU CLASSiC project (Computational Learning in Adaptive Systems for Spoken Conversation), which focuses on statistical methods for data-driven semantic parsing, dialogue management and natural language generation. I completed my Ph.D. thesis in 2008 under the supervision of Prof. Marilyn Walker, at the Computer Science Department of the University of Sheffield, United Kingdom. I obtained a Master of Engineering and Computer Science in 2004 from the Université Catholique de Louvain in Belgium.

I have been working on statistical methods for natural language understanding, natural language generation and opinion mining. These problems require learning structured prediction models from a large amount of annotated data. I have been especially interested in crowdsourcing for collecting data, in order to model the wide range of speaking styles found in natural language.

Research Interests:

  • Scalable and maintainable conversational AI
  • Domain expansion for spoken language understanding in the absence of data
  • Deep learning to generate complex utterance responses
  • Expressive language generation and text-to-speech synthesis
  • Learning to detect mood, emotion and personality for user modelling

Journal articles (Google Scholar):

Peer-reviewed publications at international conferences:

Online demos:
  • CamInfo: The Cambridge Tourist Information Dialogue System (requires a microphone)
    This Java applet is an interface to our group's live dialogue system, which provides information about most places in Cambridge, including pubs, restaurants, colleges, museums, etc. The system can also be called using the number +44 1223 852 453. The system implements the HIS framework, i.e. it relies on Partially-observable Markov Decision Processes to reason over multiple hypotheses about the user input, which are provided by the ATK speech recogniser. Some functionalities of Personage are used for language generation (e.g., syntactic aggregation, WordNet synonym selection). The speech synthesiser is an HTS voice trained on emphasis-dependent context features using the two-pass context clustering method.
  • Automatic personality recognition
    What does your language reveal about you? The personality recognition models can estimate your scores along the 5 main personality dimensions based on your input text. Models are detailed in this paper.

Datasets and software packages:

Here are various human-annotated datasets and freely available software. Feel free to use and modify them for non-commercial purposes.

  • BAGEL training and evaluation data
    This contains the 404 semantically aligned utterances used for training and evaluating the BAGEL statistical language generator, together with the naturalness and informativeness ratings of 1616 utterances generated using different learning configurations, i.e. using active learning and random sampling. More details in this paper.

  • Emphasis-annotated ARCTIC database for speaker AWB
    This corpus contains word-level emphasis annotations for the first 597 utterances (set A) of the ARCTIC speech database, i.e. the words or phrases perceived as the focus of speaker AWB's utterances.

  • Personage: Language Generation with Personality
    The Personage generator can produce personality-rich utterances for presenting information in the restaurant domain, by varying the target personality scores along the Big Five traits. Personage is based on supervised machine learning models predicting generation parameters from human personality ratings, detailed in this and this paper. The generator has been used in a wide range of publications by the UCSC NLDS Group. Download the Java stand-alone generator (65 MB, unsupported) and the Personage manual for more details.

  • Personage dataset: a personality-annotated corpus
    This dataset contains 580 utterances annotated with personality/stylistic ratings from human judges, for each Big Five trait. The data also includes the generation decisions made for each utterance, as well as the intermediary content plan tree, sentence plan tree and syntactic structures. Naturalness ratings are also included. This data was used for evaluating the Personage generator, as well as for training parameter estimation models (Mairesse & Walker, 2007, 2008). More details in the Personage dataset readme file.

  • EAR speaker features and personality ratings
    Prosodic, LIWC and MRC features extracted from each speaker of the EAR dataset, as used in the JAIR 2007 paper. One file per personality trait, for observed and self-reported ratings.

  • Personality Recognizer v1.02 (not supported anymore)
    This Java command-line application extracts psycholinguistic features from multiple text files and runs the included models to compute personality scores for all Big Five traits.

  • jMRC - MRC Psycholinguistic Database Java Interface v0.9
    This Java interface allows you to query the MRC Psycholinguistic Database from your Java programs, providing psycholinguistic features for over 150,000 words.