wanjinchang / kaggle_allen Goto Github PK

Two simple solutions for Allen Institute Kaggle competiotion.

1) Glove based solution.
The already computed wiki glove word2vec is used co calculate cosine similaruty of questions and answers.
Scores are around 0.31875
Run:
- get glove data - fetch_glove_data.sh
- get prediciton - glove_predict.py

2)IR based part:
Retrivieng ck12 topics, then get wiki pages for each page and then ranking documents for question and answer pair.
Scores are around 0.35375.
Run:
To get wiki pages and prediction - python ck12_wiki_predict.py --get_data 1