Introduction

• This program classifies legal issues into a binary value for each National Subject Matter Index (NSMI). (https://nsmi.lsntap.org/browse-v2) "Category" means 20 indexes. "Class" means sub categories under the category.

Data

• Train Data: 2777 labeled articles from https://spot.suffolklitlab.org/data/2019-12-06_95p-confidence_binary.csv
• Test Data: 900,000 submissions from reddit.com/r/legaladvice using subreddit downloader(https://github.com/heeh/subreddit_downloader/blob/master/stream_downloader.py)

Preprocessing

• Each article has a binary value(0 or 1) that indicates if this article is related to a specific legal class.
• We ignore unlabeled entries when constructing a model.
• Among 100+ classes from NSMI-v2, we extracted 36 classes which has more than 10 positive submissions.
• After getting the classifier, we chose 16 classes that has reasonable recall.

Classifier

• First, the program converts each text into an tf-idf vector.
• After getting the vectors, we apply sklearn LogisticRegression model with L1 loss.
• We predict the model with 10-fold cross-validation.

Test

• We apply the classifier to the 900,000 submissions on reddit data.
• The classifier produces hard and soft labels.
• We get the final prevalence using freq-e (https://github.com/slanglab/freq-e)