Name Entity Tagger

Using Max Entropy algorithm to predict the Name Entity Tag of the token in sentences

Author: Yuanxu Wu

Prerequisites

1. python 2.7.13
2. java 1.8.0

How to compile and run

1. get into the "code" folder in terminal (using cd code)
2. javac -cp ".:maxent-3.0.0.jar:trove.jar" MEtrain.java
3. javac -cp ".:maxent-3.0.0.jar:trove.jar" MEtag.java
4. python FeatureBuilder.py <output_filename> // get train file advanced feature
   for example, python FeatureBuilder.py input/CONLL_train.pos-chunk-name CONLL_train_ENHANCED_FEATURE.txt
5. python FeatureBuilder.py <output_filename> // get test file advanced feature
   for example, python FeatureBuilder.py input/CONLL_dev.pos-chunk CONLL_dev_test_ENHANCED_FEATURE.txt
6. java -cp ".:maxent-3.0.0.jar:trove.jar" MEtrain
   for example, java -cp ".:maxent-3.0.0.jar:trove.jar" MEtrain CONLL_train_ENHANCED_FEATURE.txt maxEntModel
7. java -cp ".:maxent-3.0.0.jar:trove.jar" MEtag
   for example, java -cp ".:maxent-3.0.0.jar:trove.jar" MEtag CONLL_dev_test_ENHANCED_FEATURE.txt maxEntModel response.name
8. python score.name.py

The outcome on development corpus

50240 out of 51578 tags correct
accuracy: 97.41
5917 groups in key
6162 groups in response
4957 correct groups
precision: 80.44
recall: 83.78
F1: 82.08

Local Features:

1. case feature
   The case feature contains initial case is upper or lower, all upper case or not, case sequence feature.
2. tag feature
   Tag feature contains the part-of-speech tag and BIO chunk tag.
3. Token Information
   These features are based on the token: such as contains digits, contains dollar sign, contains hyphen etc
   and the length of the token.
4. previous and next word feature
   the features of previous token and next token.
5. Zone feature
   this token is in Head Line (HL), Author Line (AU), Date Line (DL), or text (TXT).
6. Morphology feature
   prefix and suffix features
   and the maximum length of non-vowel characters in the token
   and the shape of the token, for example, "Cccds" means Capital Char + lower case char + lower case char + number + symbol.
7. Combined feature
   Combine some feature together:
   Combine the token is the first of a sentence or not, initial with capital or not, and Zone feature together.
   Combine the token has all capital letter or not and Zone feature together.
   Combine the token starts with a lower case letter and contains both upper and lower case letters, and Zone feature together.
8. Bigram trigram feature
   treat consecutive two words and consecutive three words as features.
9. Rare word feature
   The frequency of the token is under 3. Because this feature will take about 20mins to train and only get a very little improvement, so I removed it.
   10 Global features
   I find a name list from https://pypi.python.org/pypi/human-names
   and try to create a feature to check if the token is in this name set or not.
   Because this feature will take about 20mins to train and also only get a very little improvement, so I removed it.

Advanced Features: the final feature used in generate advanced feature txt

prev_init_case_F, curr_init_case_F, next_init_case_F, prev_tok_F, curr_tok_F, next_tok_F, prev_pos_F, curr_pos_F, next_pos_F, prev_all_upper_case_F, curr_all_upper_case_F, next_all_upper_case_F, curr_lower_case_F, curr_contain_hyphen_cap_F, curr_contain_upper_case_F, curr_contain_dot_F, curr_contain_quote_cap_F, curr_contain_number_F, curr_last_char_is_dot_F, curr_length, curr_max_consec_non_vowel_F, curr_first_init_cap_Zone_F, curr_all_caps_Zone_F, curr_mixed_caps_Zone_F, curr_sent_has_bracket_or_quotes_F, curr_suffix_3, curr_prefix_4, curr_bigram_F, trigram_F, cap_case_cap_F, Zone_F  

the other features I tried but didn't use in generating advanced feature are the features will generate negative improvement.

Tried Feature Set:

prev_init_case_F : the initial case of the previous token
curr_init_case_F : the initial case of current token
next_init_case_F : the initial case of next token
prev_tok_F : previous token
curr_tok_F : current token
next_tok_F : next token
prev_pos_F : previous part-of-speech tag
curr_pos_F : current part-of-speech tag
next_pos_F : next part of speech tag
prev_BIO_F : previous BIO chunk tag
curr_BIO_F : current BIO chunk tag
next_BIO_F : next BIO chunk tag
prev_all_upper_case_F : the previous token is all uppercase or not
curr_all_upper_case_F : current token is all uppercase or not
next_all_upper_case_F : next token is all uppercase or not
prev_contain_number_F : the previous token contain number or not
curr_contain_number_F : current token contain number or not
next_contain_number_F : next token contain number or not
curr_lower_case_F : transit current token to lower case
curr_contain_hyphen_cap_F : current token contain hyphen or not
curr_contain_upper_case_F : current token contain uppercase or not
curr_contain_dot_F : current token contain dot or not
curr_contain_quote_cap_F : current token contain quote and capital letter or not
curr_last_char_is_dot_F : last character is dot or not
curr_length_F : the length of the current token
curr_word_shape_F : the shape of the word, for example, "Cccds" means Capital Char + lower case char + lower case char + number + symbol
curr_first_init_cap_Zone_F : this is a combined feature : current token is the first of a sentence or not, initial with capital or not, and Zone feature
curr_all_caps_Zone_F : this is a combined feature : current token has all capital letter or not and Zone feature
curr_mixed_caps_Zone_F : this is a combined feature : the token starts with a lower case letter, and contains
both upper and lower case letters, and Zone feature
curr_sent_has_bracket_or_quotes_F : the sentence contained this token has bracket or quotes or not
Zone_F : this token is in Head Line (HL), Author Line (AU), Date Line (DL), or text (TXT)
curr_prefix_2 : first two character
curr_suffix_2 : last two character
curr_prefix_3 : first three character
curr_suffix_3 : last three character
curr_prefix_4 : first four character
curr_suffix_4 : last four character
consecutive_three_tok_init_case_F : the initial case of three consecutive token
curr_bigram_F : current word and next word
prev_bigram_F : previous word and next word
trigram_F : previous word and current word and next word
cap_case_cap_F : case sequence
curr_max_consec_non_vowel_F : the maximum length of non-vowel characters in the current token