We build a local search engine, then use the question as query to get the top results. Then we calculate the word count of the answers in these top results, the answers that get the highest word count is choosed as the prediction.

We use the wikipedia page of the ck-12 keywords to build index. The ck-12 keywords are scraped from the https://www.ck12.org using the scrape.py script. Then we get the wikipedia page content of these keywords using get_wikipedia_data.py.

We use the lucene to index and search. The index is built using IndexFiles.java(java org.apache.lucene.demo.IndexFiles -docs data/wikipedia_content_based_on_ck_12_keyword_one_file_per_keyword/), and search is done by Get_Top_Documents_Based_on_Lucene.java, which is just a modification of the SearchFiles.java.

For each answer, we calculate the word count of its words(excluding stop words) on the top N search results. Then the answer which get the highest word count is chosen as the prediction.

We can get a score of 0.36, please refer to https://www.kaggle.com/c/the-allen-ai-science-challenge/leaderboard for the current leaderboard.