code for Modeling Dynamic Pairwise Attention for Crime Classification over Legal Articles, SIGIR 2018 [pdf]

@inproceedings{Wang:2018:MDP:3209978.3210057,
 author = {Wang, Pengfei and Yang, Ze and Niu, Shuzi and Zhang, Yongfeng and Zhang, Lei and Niu, ShaoZhang},
 title = {Modeling Dynamic Pairwise Attention for Crime Classification over Legal Articles},
 booktitle = {The 41st International ACM SIGIR Conference on Research \&\#38; Development in Information Retrieval},
 series = {SIGIR '18},
 year = {2018},
 isbn = {978-1-4503-5657-2},
 location = {Ann Arbor, MI, USA},
 pages = {485--494},
 numpages = {10},
 url = {http://doi.acm.org/10.1145/3209978.3210057},
 doi = {10.1145/3209978.3210057},
 acmid = {3210057},
 publisher = {ACM},
 address = {New York, NY, USA},
 keywords = {dynamic threshold predictor, multi-label classification, pairwise attention model},
} 

In juridical field, judges usually need to consult several relevant cases to determine the specific articles that the evidence violated, which is a task that is time consuming and needs extensive professional knowledge. we treat the evidences as documents, and articles as labels, thus the conviction process can be cast as a multi-label classification problem.

DPAM model is used to predict articles by the evidences. This model adopt the multi-task learning method to learn the multi-label classifier and the threshold predictor jointly.

1. clone the code
2. download the dataset
3. mv the dataset file into code dir
4. run the script: run.sh

To work with the DPAM, you can simply run the script: run.sh. and you can also modify the parameters in the DPAM_train.py, then run the script: python DPAM_train.py

The dataset contains 4 files: one_hot_vocab_70.txt src_rule_data.txt src_train_dev_data.txt stop_words.txt

the download url: https://drive.google.com/open?id=1TCcTzte2cQ2wxWw-kXsZdUApCsNejdn8

this file contains the mapping of the top 70 (labels)articles to one hot representation.

the file has two columns, the first column is article number, the second column is labels(articles) description.

this file contains 17160 samples, which is the train and dev source data, the file has two columns, the first column is evidence description, the second column is labels(articles).

this file contains some stop words, which is ignored in our train and dev samples.