Investigate Question-Answering models working on tables

[FrancescoCasalegno]

Context

• Traditional transformers-based models for extractive question-answering tasks operate on contexts that are units of texts in natural language, e.g. a sentence or a paragraph.
• However, in many cases the values of parameters of interest for our neuroscientific applications are contained into tables of articles rather than in the text.
• For instance, the Wikipedia article on Michaelis constant (here) contains several values for this parameter of interest for us, but they are all in a table and no value is mentioned in the text. In fact this is not an isolated case: it's really hard to find Michaelis constant values in the text of any scientific article!
  
• There seem to be some models for question-answering that can operate on tabular or text/tabular mixed contexts, like TAPAS.

Actions

• How should the tables be represented for TAPAS (or another model) to be able to take it in input (html? csv? ...) ?
  Is this format compatible with what we can get out our parsing pipeline for the various formats (arXiv, medRxiv, bioRxiv, PMC, PubMed, ...) when the article contains a table?
• Can TAPAS take mixed inputs, i.e. contexts containing both text and tables?
• How does TableQuestionAnsweringPipeline differ from QuestionAnsweringPipeline in 🤗 transformers?
• Are there any other models a part from TAPAS that support question-answering on tabular contexts?
• Test TAPAS (or another model) on a sample related to neruoscience to see if it could potentially work on our use case.