NL4XAI NTP5

19-21 April 2022

Nancy, France

Morning talks will be available online. Talks marked "Visio" are given remotely. The link for the visio will be anounced here closer to the date. If you intend to follow online, please register here so we can easily keep track of participation.

Morning talks will take place in A008

Tuesday 19 April: Natural Language Generation (NLG)

  • 10:00 - 10:45: A Brief Introduction to Retrieval-Augmented Neural NLG. Claire Gardent, CNRS/LORIA, Nancy, France. Slides
  • 10:45 - 11:30: Grounding revisited: Do pretrained multimodal models learn to ground linguistic phenomena? Albert Gatt, U. Utrecht, The Netherlands. Slides
  • Coffee break
  • 12:00 - 12:45: NLG Evaluation and Challenges for explainable NLG. Ehud Reiter (Visio), U. of Aberdeen/ARRIA, Scotland. Slides
  • Lunch
  • 14:00 - 14:15: A model-agnostic measure of understandability of machine learning pipelines. Ettore Mariotti, CiTIUS USC, Spain.
  • 14:15 - 14:30: Explainability in Interactive User Interfaces for Process Mining. Luca Nannini, Minsait by Indra, USC, Spain.
  • 14:30 - 17:00
    Group work on analysing, classifying and quantifying errors in NLG outputs. ESRs work in small groups including 1 NLG ESR. The output of various NLG models/tasks are provided including outputs from simplification, data-to-text and image-to-text models. The aim is to identify semantic adequacy errors (hallucinations, omissions) as these mostly affect the trustworthiness of an NLG model.
  • 18:00 - 19:00: Nancy Guided Tour

  • 19:00 - ... : Restaurant Le Grand Café Foy

Wednesday 20 April: Explainable NLG

  • 10:00 - 10:45: Explainable Summarisation: Planning with Learned Entity Prompts for Abstractive Summarization. Shashi Narayan (Visio), Google London, UK. Slides
  • 10:45 - 11:30: Dealing with hallucination and omission in neural NLG: A use case on meteorology. Jose Maria Alonso, CiTIUS-USC, Spain. Slides
  • Coffee break
  • 12:00 - 12:15: Probing Omissions in BART-based RDF-to-Text Models. Juliette Faille, Université de Lorraine, Nancy, France.
  • 12:15 - 12:30: Grounding Language in Vision for Understanding and Generation. Michele Cafagna, Malta/Utrecht University.
  • Lunch
  • 14:00 - 14:15: A Dialogue Game for Explanations in a Multi-agent Recommender System. Qurat-ul-ain Shaheen, IIIA-CSIC, Spain.
  • 14:15 - 14:30: Improving the fluency of logic-to-language generation. Eduardo Calò, U. Utrecht, The Netherlands.
  • 14:30 - 17:00
    Group work on identifying and when possible, applying ways of detecting errors. ESRs work in small groups to identify which methods presented in the morning could be used or adapted to detect/explain the types of errors they identified on the previous day. The aim is to get a better understanding of the challenges involved in error detection and explainable NLG methods.

Thursday 21 April: Industry and Academia

    Talks and demos by industrials (ORANGE Lannion)
  • 10:00 - 10:40: CoQAR: Question Rewriting on CoQA. Lina Rojas-Barahona (Joint work with Quentin Brabant and Gwenole Lecorve) Slides
  • 10:40 - 11:00: Hyperbolic Temporal Knowledge Graph Embeddings with Relational and Time Curvatures. S├ębastien Montella Slides
  • 11:00 - 11:30 Analysis of AMR multilanguage. Johannes Heinecke (Visio)
  • Coffee break
    Pre-recorded Demonstrators
  • 12:00 - 12:05: VisualCRM : Orange opinion analysis and surveys tool. Nathalie Charbonniaud, pre-recorded
  • 12:05 - 12:10: CS4GK: Conversational Search for General Knowledge. Lina Rojas, pre-recorded
  • 12:10 - 12:15: ClusterWiz : Clustering user intentions in dialogue. Constance Scherer, pre-recorded.
  • Lunch
  • 14:00 - 14:15: Cognitive Bias Mitigation During Web Interactions. Alisa Rieger, TU Delft, The Netherlands
  • 14:15 - 14:30: Personalized explanations by conversational agents. Sumit Srivastava, U. Twente, The Netherlands.
  • 14:30 - 15:30
    ESRs present the results of their group work (error types for each type of tasks; classification; quantitative analysis; error detection and possible explainable models that could help track the source of the error).

© Claire Gardent