Movie QA Benchmarking Dataset
For one particular application of YodaQA, we want to enhance and speed up its capability to answer "noisy" questions on a structured knowledge base in a narrow domain. To start prototyping, we have chosen the "movies" domain.
To get started, we extracted movie-related questions from WebQuestions (http://nlp.stanford.edu/software/sempre/ - Berant et al., 2013, CC-BY) using the machinery in https://github.com/brmson/dataset-factoid-webquestions (we use the same JSON structure and scripts in this repo). This is the moviesB dataset.
The moviesC dataset also includes "mfb" questions which stand for
"movie feedback", as reported by the YodaQA feedback tool when testing
the YodaQA Movies engine by internet users (mainly interns of the
eClub Prague foundation). The
GoogleDocs2json.py script extracts
the feedback data from a Google Docs spreadsheet.
We intend to follow up with even larger and better datasets, using next consecutive letters.
Using with YodaQA
YodaQA typically excepts datasets in a TSV format rather than JSON. (JSON collection reader in YodaQA is work-in-progress.) To get the data to TSV format, run
../dataset-factoid-webquestions/scripts/json2tsv.py moviesC train moviesC ../dataset-factoid-webquestions/scripts/json2tsv.py moviesC test moviesC
The dataset is called moviesA - the A letter represents our intention to develop it further. It is currently rather noisy, mixed with sports questions and not that large either.
moviesC is a dataset created by merging the t-movies dataset (here named moviesB for reference) from https://github.com/brmson/dataset-factoid-webquestions/t-movies and public feedback in our 2 spreadsheets (downloaded 17.8.2015):
moviesD is an update of moviesC on 2015-10-19.
moviesE is an update of moviesD on 2015-12-10 and inclusion of synthetic questions gen v0.
moviesF is an update of moviesE on 2016-01-04 with a variety of bugs related to the synthetic questions fixed.
Licence and Acknowledgements
This dataset may be distributed under the terms of the CC-BY 4.0 licence. Work on this project has been supported in part by the Medialab foundation.