qa_match

Deep learning model for match medical QA

IDEA 1 :

give more match count example (but not relevant answer) more weight to loss

step 1: 抽樣10000筆 從這些資料中觀察大概重疊字的分布 結巴斷詞 2.8276 2 gram and 3 gram(不含1gram) 2.4723 因為沒有包含1gram所以反而比斷詞重疊還要少

tfidf的hit rate(邱老師) hit rate @1 = 0.632 hit rate @2 = 0.763 hit rate @3 = 0.816 hit rate @4 = 0.895 hit rate @5 = 1.000

台灣e院 hit rate @1 = 0.767 hit rate @2 = 0.874 hit rate @3 = 0.920 hit rate @4 = 0.946 hit rate @5 = 0.970

尋醫問藥 hit rate @1 = 0.535 hit rate @2 = 0.638 hit rate @3 = 0.690 hit rate @4 = 0.723 hit rate @5 = 0.747

step 2: 從step1的結果中 設定超參數(給比較多match的負樣本的權重)