Dowsing for Math Answers - the MathDowsers team's submission for ARQMath @ CLEF
MathDowsers is a team of researchers from the University of Waterloo who are interested in dowsing for answers to math questions:
Given a math question, look for potential answers to this math question from an existing math answer database.
ARQMath-1 & ARQMath-2 The team produces the best participant run of this Answer Retrieval task in the ARQMath (Answer Retrieval for Questions on Math) Lab in both year 2020 and 2021; and also the best automatic run of the Formula Retrieval task in the Lab in year 2021.
The math-aware search engine in use is Tangent-L. More details of the supporting research project can be found in the BrushSearch site.
ARQMath-3 With a new implemented engine mtextsearch, the team simplifies experimentation, improves retrieval effectiveness, and shows that traditional text retrieval systems can easily be enhanced to become effective math-aware search engines.
Click here for the Demo. (Last updated: 2022-10)
This demo displays the task runs, in particular the submitted runs from the MathDowsers for ARQMath Lab 2020, 2021, and 2022. The math answer database is the MathStackExchange corpus from year 2010 to 2018, and math questions are selected from the same corpus from year 2019 to 2020.
The aim of this demo is data exploration. To make an actual search to the math questions (with formulas and keywords), visit the user interface of the search engine (which supports a pen-based input).
The template for the document corpus is src_2021/template_minimal_v2.html, which is an HTML page that stores an answer together with its parent math question (that is, a question-answer pair) and other meta information.
To create the document corpus,
-
First, download the two folders from this repository
/data /src_2021
followed by the Lab-provided Math Stack Exchange (MSE) collection at the designated paths. (Check individual README in the
datafolder) -
Next, navigate to the source code folder
src_2021. -
Create all preprocessing files by running
./main_preprocessing.sh
The expected files to be created are documented at each individual python file with the
prepro-prefix . -
Then, demo documents can be created by running
python main_generate_document_corpus.py --style minimal
To create all documents (MSE question-answer pairs from year 2010 to year 2018)
python main_generate_document_corpus.py --style minimal --year all
The generated files will be stored at
/data/ARQMath/html_minimal_2021. See the filemain_generate_document_corpus.pyfor more available options.
The given ARQMath task topics is a collection of mathematical questions from MathStackExchange with title, question body, and tags. However, the search engine requires keywords and formulas as formal search queries for retrieval tasks. MathDowsers created the formal search queries with the following files:
query_prepro.py: with pre-processing functions to convert and clean the given task topic files into json files.query_filters.py: with helper functions that create and read list of stopwords and "mathy" words for keyword selection.query_model.py: with rule-based models to generate search queries.query_postpro.py: with post-processing functions to print or analyze the generated search queries.
See descriptions in each file for more information.
2022-10-29: Update demo with ARQMath-3 runs
2022-03-18: Freeze ARQMath-2 version (add git tag)
2022-02-21: Fix the followings: the running order in the preprocessing script; naming of the map_of_comments files; and the order of the duplicate/related-post titles in the generated htmls (thanks to Andrew Kane for pointing them out)
2021-12-22: Update resources for generating formal search queries from the given task topics.
2021-11: I finish my MMath thesis :)
2021-09-16: Finish most features of the demo. Host the full demo at cs.uwaterloo.ca instead, and keep only a limited version in GitHub.
2021-07-15: Initialize a demo page.
2021-07-12: Add instructions to create the document corpus for the Answer Retrieval task.
2021-07-04: Initialize the repository with basic configuration.
MMath Thesis:
Yin Ki Ng (2021). Dowsing for Math Answers: Exploring MathCQA with a Math-aware Search Engine. UWSpace. http://hdl.handle.net/10012/17696
Publications:
Andrew Kane, Yin Ki Ng, Frank Wm. Tompa. Dowsing for Answers to Math Questions: Doing Better with Less, in: CLEF 2022, volume 3180 of CEUR Workshop Proceddings, 2022
Yin Ki NG, Dallas J. Fraser, Besat Kassaie, Frank Wm. Tompa. Dowsing for Math Answers, in: Candan K.S. et al. (eds) Experimental IR Meets Multilinguality, Multimodality, and Interaction. CLEF 2021. Lecture Notes in Computer Science, vol 12880. Springer, Cham. https://doi.org/10.1007/978-3-030-85251-1_16
Yin Ki NG, Dallas J. Fraser, Besat Kassaie, Frank Wm. Tompa. Dowsing for Answers to Math Questions: Ongoing Viability of Traditional MathIR , in: CLEF 2021, volume 2936 of CEUR Workshop Proceddings, 2021
Yin Ki NG, Dallas J. Fraser, Besat Kassaie, George Labahn, Mirette S. Marzouk, Frank Wm Tompa, and Kevin Wang. Dowsing for Math Answers with Tangent-L, in: CLEF 2020, volume 2696 of CEUR Workshop Proceedings, 2020