Instroduction:
- Chip2018_task2 and Biendata competitions links:http://icrc.hitsz.edu.cn/chip2018/Task.html and https://biendata.com/competition/chip2018/
- Dataset Download Link:https://drive.google.com/drive/folders/1kDuR24xeIO97DSSxlPQU-F_GvZ6WAaUv?usp=sharing
Code instruction:
-
data_preprocess.py: Analysis data and generate features we need.
-
attention_model.py: Design model structure
-
train.py: Train model
-
predict.py: Load our pretrain_model and predict the result of testset.
-
label_transform.py : Transform the result that are predicted in predict.py into formal format, then you can submit.
-
requirements.txt: Related packages environment, before train model you should to execute command 'pip install -r requirements' to install dependency.
-
Trainseting: the number of 0: 10000, the number of 1: 10000, total number:20000
-
Testset: total number:10000
1.base on word:
-
The size of word_embedding.txt: 300
-
Total number of question pairs for training: 20000
-
Duplicate pairs: 50.0%
-
Total number of questions in the training data: 21787
-
Number of questions that appear multiple times: 14259
-
Word count: mean-train 7.07, std-train 3.44, min-train 1.00, max-train 43.00, Total word count:7611, mean-test 6.96, std-test 3.41, min-test 1.00, max-test 40.00
-
8features: q1_word_count, q2_word_count, word_count_diff, word_overlap, uni_BLEU, bi_BLEU,BLEU2, char_unigram_overlap
-
feature1: word_match, tfidf_word_match
-
feature2: q1_hash, q2_hash, q1_freq, q2_freq
-
magic_features: min_freq, common_neighbours, q_len1, q_len2 , diff_len, word_len1, word_len2, common_words, fuzzy_qratio, fuzzy_wratio, fuzzy_partial_ratio, fuzzy_partial_token_set_ratio, fuzzy_partial_token_sort_ratio,fuzzy_token_set_ratio,fuzzy_token_sort_ratio
-
questions_distance_features: cosine_distance,cityblock_distance,jaccard_distance,canberra_distance,euclidean_distance,minkowski_distance,braycurtis_distance,skew_q1vec,skew_q2vec,kur_q1vec
2.base on char:
-
The size of char_embedding.txt: 300
-
(Base on char) Total number of question pairs for training: 20000
-
(Base on char) Duplicate pairs: 50.0%
-
(Base on char) Total number of questions in the training data: 21786
-
(Base on char) Number of questions that appear multiple times: 14258
-
Char count: mean-train 13.73, std-train 5.34, min-train 2.00, max-train 54.00, Total char count:2085, mean-test 13.48, std-test 5.40, min-test 2.00, max-test 57.00
-
7features: q1_char_count, q2_char_count, char_count_diff, char_overlap, uni_BLEU, bi_BLEU,BLEU2
-
feature1: char_match, tfidf_char_match
-
feature2: q1_hash, q2_hash, q1_freq, q2_freq
-
magic_features: min_freq, common_neighbours, q_len1, q_len2 , diff_len, char_len1, char_len2, common_chars, fuzzy_qratio, fuzzy_wratio, fuzzy_partial_ratio, fuzzy_partial_token_set_ratio, fuzzy_partial_token_sort_ratio,fuzzy_token_set_ratio,fuzzy_token_sort_ratio
-
questions_distance_features: cosine_distance,cityblock_distance,jaccard_distance,canberra_distance,euclidean_distance,minkowski_distance,braycurtis_distance,skew_q1vec,skew_q2vec,kur_q1vec
Dataset Analysis
-
Dataset Analysis Base on word
Figure: q1_q2_intersect -
Dataset Analysis Base on char
Figure: q1_q2_intersect
