Document-Similarity-via-Hidden-Topics: A Python repository from Buratinator

This is the repository for paper "Document Similarity for Texts of Varying Lengths via Hidden Topics" to appear in ACL 2018.
This is a system implemented by Hongyu Gong in python for matching long documents and short documents (e.g. summary).

Structure:
src/ is the source code of the document matching system.
concept-project-mapping-dataset/ prepares the data for the task of concept-project matching.=


Required package:
a) gensim
b) sklearn
c) nltk
1. Input data format
raw_short_docs: a list of strings where each string is the abstract;
raw_long_docs: a list of strings where each string is a long document;

2. Run main.py
- Set hyperparameters properly (the default parameters are what we used in the paper);
- Specify the path of word embeddings.
- You need to write function to read data into specified format. You can refer to the function readData()in test.py.
Return: The function mapping() in main.py will return a score matrix, and the (i,j)-th entry in the matrix is the matching score of the i-th short document and the j-th long document. 
Note that the score is a real value between 0 and 1. The higher the score is, the better they match each other.

<b> Note that you can run test.py which returns matching scores for the project-concept pairs.</b>


Furthermore, we provide science-domain corpara extracted from the WikiCorpus, which are used to train science-domain embeddings:
(1) [science_corpus_d2](https://drive.google.com/open?id=1BIyKO9xMqpNOO7EordP9_kS0Zr2tIiNx): a refined corpus of 90 MB consists of documents within depth of 2 in science categories;
(2) [science_corpus_d3](https://drive.google.com/open?id=1xsJymoKZi2Fr2SCTDwSJzZid13Lp7PO4): a larger corpus of 2GB consists of documents within depth of 3 in science categories.

If you have any questions, please contact Hongyu Gong (hgong6@illinois.edu).

If you use our code/data, please cite our work:
Hongyu Gong, Tarek Sakakini, Suma Bhat and Jinjun Xiong, “Document Similarity for Texts of Varying Lengths via Hidden Topics”, accepted by the 56th Annual Meeting of the Association for Computational Linguistics (ACL), 2018.
@inproceedings{docsim,
  author    = {Hongyu Gong, Tarek Sakakini, Suma Bhat and Jinjun Xiong},
  title     = {Document Similarity for Texts of Varying Lengths via Hidden Topics},
  booktitle = {the 56th Annual Meeting of the Association for Computational Linguistics (ACL)},
  month     = {July},
  year      = {2018},
  publisher = {Association for Computational Linguistics}
}
Buratinator/Document-Similarity-via-Hidden-Topics
