Microsoft Research Sequential Question Answering (SQA) Dataset
Contact Persons: Scott Wen-tau Yih scottyih@microsoft.com Mohit Iyyer m.iyyer@gmail.com Ming-Wei Chang minchang@microsoft.com
The SQA dataset was created to explore the task of answering sequences of inter-related questions on HTML tables. A detailed description of the dataset, as well as some experimental studies, can be found in the following paper:
Mohit Iyyer, Wen-tau Yih, Ming-Wei Chang. "Answering Complicated Question Intents Expressed in Decomposed Question Sequences." arXiv preprint arXiv:1611.01242 https://arxiv.org/abs/1611.01242
Version 1.0: November 9, 2016
SUMMARY
Recent work in semantic parsing for question answering has focused on long and complicated questions, many of which would seem unnatural if asked in a normal conversation between two humans. In an effort to explore a conversational QA setting, we present a more realistic task: answering sequences of simple but inter-related questions.
We created SQA by asking crowdsourced workers to decompose 2,022 questions from WikiTableQuestions (WTQ)*, which contains highly-compositional questions about tables from Wikipedia. We had three workers decompose each WTQ question, resulting in a dataset of 6,066 sequences that contain 17,553 questions in total. Each question is also associated with answers in the form of cell locations in the tables.
- Panupong Pasupat, Percy Liang. "Compositional Semantic Parsing on Semi-Structured Tables" ACL-2015. http://www-nlp.stanford.edu/software/sempre/wikitable/
LIST OF FILES
train.tsv -- Training question sequences
test.tsv -- Testing question sequences
table_csv -- All the tables used in the questions (originally from WTQ)
random-split-* -- Five different 80-20 training/dev splits based on training.tsv.
The splits follow those provided in WTQ.
eval.py -- The evaluation script in Python
rndfake.tsv -- A fake output file for demonstrating the usage of eval.py
license.docx -- License
readme.txt -- This file
DATA FORMAT
train.tsv, test.tsv, random-split-* -- id: question sequence id (the id is consistent with those in WTQ) -- annotator: 0, 1, 2 (the 3 annotators who annotated the question intent) -- position: the position of the question in the sequence -- question: the question given by the annotator -- table_file: the associated table -- answer_coordinates: the table cell coordinates of the answers (0-based, where 0 is the first row after the table header) -- answer_text: the content of the answer cells Note that some text fields may contain Tab or LF characters and thus start with quotes. It is recommended to use a CSV parser like the Python CSV package to process the data.
table_csv (from WTQ; below is the original description in the WTQ release)
-- Comma-separated table (The first row is treated as the column header)
The escaped characters include:
double quote (" => \") and backslash (\ => \\).
Newlines are represented as quoted line breaks.
rndfake.tsv -- A fake output file for the test questions; fields are: id, annotator, position, answer_coordinates
EVALUATION
$ python eval.py test.tsv rndfake.tsv Sequence Accuracy = 14.83% (152/1025) Answer Accuracy = 50.33% (1516/3012)