Commit Graph

363 Commits

Author SHA1 Message Date
Matthew Honnibal
6b34a2f34b * Move morphological analysis into its own module, morphology.pyx 2014-12-09 21:16:17 +11:00
Matthew Honnibal
b962fe73d7 * Make suffixes file use full-power regex, so that we can handle periods properly 2014-12-09 19:04:27 +11:00
Matthew Honnibal
accdbe989b * Remove Tokens.extend method 2014-12-09 17:09:23 +11:00
Matthew Honnibal
495e1c7366 * Use fused type in Tokens.push_back, simplifying the use of the cache 2014-12-09 16:50:01 +11:00
Matthew Honnibal
302e09018b * Work on fixing special-cases, reading them in as JSON objects so that they can specify lemmas 2014-12-09 14:48:01 +11:00
Matthew Honnibal
99bbbb6feb * Work on morphological processing 2014-12-08 21:12:15 +11:00
Matthew Honnibal
7b68f911cf * Add WordNet lemmatizer 2014-12-08 01:39:13 +11:00
Matthew Honnibal
c20dd79748 * Fiddle with const correctness and comments 2014-12-08 00:03:55 +11:00
Matthew Honnibal
b031c7c430 * Remove language-general context module 2014-12-07 23:53:01 +11:00
Matthew Honnibal
ef4398b204 * Rearrange POS stuff, so that language-specific stuff can live in language-specific modules 2014-12-07 23:52:41 +11:00
Matthew Honnibal
327383e38a * Remove unused code in tagger.pyx 2014-12-07 22:16:17 +11:00
Matthew Honnibal
9f17467c2e * Fix EMPTY_TOKEN 2014-12-07 22:07:41 +11:00
Matthew Honnibal
3819a88e1b * Add support for tag dictionary, and fix error-code for predict method 2014-12-07 22:07:16 +11:00
Matthew Honnibal
f00afe12c4 * Load POS tagger in load() function if path exists 2014-12-07 22:05:57 +11:00
Matthew Honnibal
5fe5e6e66b * Move context functions to header, inlining them. 2014-12-07 21:59:04 +11:00
Matthew Honnibal
5caabec789 * Link in tagger, to work on integrating POS tagging 2014-12-07 15:29:41 +11:00
Matthew Honnibal
0c7aeb9de7 * Begin revising tagger, focussing on POS tagging 2014-12-07 15:29:04 +11:00
Matthew Honnibal
f5c4f2eb52 * Revise context, focussing on POS tagging for now 2014-12-07 15:28:22 +11:00
Matthew Honnibal
e27b912ef9 * Remove need for confusing _data pointer to be stored on Tokens 2014-12-05 16:31:30 +11:00
Matthew Honnibal
1c9253701d * Introduce a TokenC struct, to handle token indices, pos tags and sense tags 2014-12-05 15:56:14 +11:00
Matthew Honnibal
187372c7f3 * Allow the lexicon to create lexemes using an external memory pool, so that it can decide to make some lexemes temporary, rather than cached 2014-12-05 03:29:50 +11:00
Matthew Honnibal
75b8dfb348 * Remove upper_pc from lexeme.pyx 2014-12-04 22:14:34 +11:00
Matthew Honnibal
49f3780ff5 * Fiddle with lexeme attrs 2014-12-04 21:22:38 +11:00
Matthew Honnibal
564082e48e * Hack Token class to take lex.dense inplace of the old lex.norm. This needs to be fixed... 2014-12-04 20:51:29 +11:00
Matthew Honnibal
69bb022204 * Add as_array and count_by method 2014-12-04 20:46:55 +11:00
Matthew Honnibal
e1b1f45cc9 * Add STEM attribute to lexeme 2014-12-04 20:46:20 +11:00
Matthew Honnibal
d7952634ca * Make the string-store serve const pointers to Utf8Str 2014-12-03 16:01:47 +11:00
Matthew Honnibal
7e04c22f8f * const added to Lexicon interface. Seems to work. 2014-12-03 15:58:17 +11:00
Matthew Honnibal
d70d31aa45 * Introduce first attempt at const-ness 2014-12-03 15:44:25 +11:00
Matthew Honnibal
4560ada85b * Add typedef for attr_t. Change flag_t to flags_t 2014-12-03 11:06:31 +11:00
Matthew Honnibal
e600f7b327 * Move String struct stuff into the utf8string module, from spacy.lang 2014-12-03 11:06:00 +11:00
Matthew Honnibal
e170faf5b0 * Hack Tokens to work without tagger.pyx 2014-12-03 11:05:15 +11:00
Matthew Honnibal
b463a7eb86 * Make flag-setting a language-specific thing 2014-12-03 11:04:32 +11:00
Matthew Honnibal
71b009e323 * Fix bug in refactored StringStore.__getitem__ 2014-12-03 11:02:24 +11:00
Matthew Honnibal
14097311ae * Make StringStore.__getitem__ accept unicode-typed keys. 2014-12-03 01:33:20 +11:00
Matthew Honnibal
522bb0346e * Work on get_array method of Tokens 2014-12-02 23:48:05 +11:00
Matthew Honnibal
8c2938fe01 * Rename Lexicon._dict to Lexicon._map 2014-12-02 23:46:59 +11:00
Matthew Honnibal
33dfb4933c * Remove taggers from Language class. Work on doc strings 2014-11-26 19:53:55 +11:00
Matthew Honnibal
80baa2e3db * Work on beam parser 2014-11-20 19:49:33 +11:00
Matthew Honnibal
5c3016bac8 * Tmp commit of ner code 2014-11-14 18:27:47 +11:00
Matthew Honnibal
33c421bcf8 * More feature tweaks 2014-11-12 23:59:16 +11:00
Matthew Honnibal
41dedfb14e * Add label features for NER parsing 2014-11-12 23:55:10 +11:00
Matthew Honnibal
cf55b48ba6 * Switch to predict label on shift. Big increase in accuracy. 2014-11-12 23:50:12 +11:00
Matthew Honnibal
8f84e8a78b * Neaten oracle 2014-11-12 23:38:07 +11:00
Matthew Honnibal
7e0a9077dd * Add context files 2014-11-12 23:22:36 +11:00
Matthew Honnibal
3b0b902384 * IOB-style parsing working. Accuracy down from BILOU, form 87-88 to 85-86 2014-11-12 23:21:09 +11:00
Matthew Honnibal
e6bb8aa3a9 * Move moves to bilou_moves. Refactor context, returning to the simpler giant-enum style 2014-11-12 00:54:50 +11:00
Matthew Honnibal
c788633429 * Add tokens_from_list method to Language 2014-11-11 23:43:14 +11:00
Matthew Honnibal
95282d4993 * Use the dynamic oracle 'follow' strategy 2014-11-11 21:11:17 +11:00
Matthew Honnibal
5aaf7a024d * Move ner features to ner subdir 2014-11-11 21:09:03 +11:00
Matthew Honnibal
ff8989b63c * Use greedy NER parser 2014-11-11 21:08:35 +11:00
Matthew Honnibal
0d943ab358 * Fixed greedy NER parsing. With static oracle, replicates accuracy from tagger. 2014-11-11 17:17:54 +11:00
Matthew Honnibal
399239760b * Fix moves for new State struct 2014-11-10 22:16:05 +11:00
Matthew Honnibal
82247169f2 * Implement validation and oracle on pystate, for testing 2014-11-10 22:15:32 +11:00
Matthew Honnibal
3709ed9d6d * Add curr field to State, to handle entity being built 2014-11-10 22:14:36 +11:00
Matthew Honnibal
af9ed18cf1 * Bug fixes to NER 2014-11-10 17:39:23 +11:00
Matthew Honnibal
9f2587f5ec * Work on shift-reduce NER 2014-11-10 16:28:56 +11:00
Matthew Honnibal
f307eb2e36 * Refactor context extraction, and start breaking out gold standards into their own functions 2014-11-09 15:43:07 +11:00
Matthew Honnibal
602f993af9 * Moving tagger to accept multiple correct answers 2014-11-09 15:18:33 +11:00
Matthew Honnibal
f37d896a42 * Upd NER feats. With adadelta learner, getting 76.9 on NER 2014-11-07 04:43:54 +11:00
Matthew Honnibal
68d1cdad62 * When encoding POS/NER tags, accept '-' as a missing value 2014-11-07 04:42:31 +11:00
Matthew Honnibal
949a6245f9 * Increase default number of iterations from 5 to 10 2014-11-07 04:42:04 +11:00
Matthew Honnibal
3cab1d9a29 * Refine word_shape feature, by trimming the max sequence length 2014-11-07 04:41:29 +11:00
Matthew Honnibal
b4454cf036 * Add extra context tokens 2014-11-07 04:40:36 +11:00
Matthew Honnibal
50309e6e49 * Fix context vector, importing all features 2014-11-05 22:11:39 +11:00
Matthew Honnibal
07a23768de * Play with NER feats a bit. Up to 82.00 training on MUC7. 2014-11-05 21:47:17 +11:00
Matthew Honnibal
4ecbe8c893 * Complete refactor of Tagger features, to use a generic list of context names. 2014-11-05 20:45:29 +11:00
Matthew Honnibal
0a8c84625d * Moving feature context stuff to a generalized place 2014-11-05 19:55:10 +11:00
Matthew Honnibal
3733444101 * Generalize tagger code, in preparation for NER and supersense tagging. 2014-11-05 03:42:14 +11:00
Matthew Honnibal
abbe3e44b0 * Move spacy.pos tagger to spacy.tagger, and generalize it so that it can take on other tagging tasks, given a different set of feature templates. 2014-11-05 00:37:59 +11:00
Matthew Honnibal
954c970415 * Add __iter__ method to tokens 2014-11-04 01:07:08 +11:00
Matthew Honnibal
f07457a91f * Remove POS alignment stuff. Now use training data based on raw text, instead of clumsy detokenization stuff 2014-11-04 01:06:43 +11:00
Matthew Honnibal
ae52f9f38c * Remove vocab10k from tokens 2014-11-03 00:23:20 +11:00
Matthew Honnibal
32fb50dc35 * Remove non_sparse method --- features wanting this can do it easily enough. 2014-11-03 00:15:47 +11:00
Matthew Honnibal
b5ae1471db * Fiddle with POS tag features 2014-11-03 00:15:03 +11:00
Matthew Honnibal
70ea862703 * Remove vocab10k field, and add flags for gazetteers 2014-11-03 00:13:51 +11:00
Matthew Honnibal
711ed0f636 * Whitespace 2014-11-02 14:22:32 +11:00
Matthew Honnibal
fcd9490d56 * Add pos_tag method to Language 2014-11-02 14:21:43 +11:00
Matthew Honnibal
829bb2bdbe * Add mappings to Twitter POS tag corpus 2014-11-02 13:21:19 +11:00
Matthew Honnibal
437cd2217d * Fix strings i/o, removing use of ujson library in favour of plain text file. Allows better control of codecs. 2014-11-02 13:20:37 +11:00
Matthew Honnibal
3352e89e21 * Use LIKE_URL and LIKE_NUMBER flag features. Seems to improve accuracy on onto web 2014-11-02 13:19:54 +11:00
Matthew Honnibal
8335706321 * Add LIKE_URL and LIKE_NUMBER flag features 2014-11-02 13:19:23 +11:00
Matthew Honnibal
5484fbea69 * Implement is_number 2014-11-01 19:13:24 +11:00
Matthew Honnibal
f685218e21 * Add is_urlish function 2014-11-01 17:39:34 +11:00
Matthew Honnibal
09a3e54176 * Delete print statements from stringstore 2014-10-31 17:45:26 +11:00
Matthew Honnibal
b186a66bae * Rename Token.lex_pos to Token.postype, and Token.lex_supersense to Token.sensetype 2014-10-31 17:44:39 +11:00
Matthew Honnibal
a8ca078b24 * Restore lexemes field to lexicon 2014-10-31 17:43:25 +11:00
Matthew Honnibal
6c807aa45f * Restore id attribute to lexeme, and rename pos field to postype, to store clustered tag dictionaries 2014-10-31 17:43:00 +11:00
Matthew Honnibal
aaf6953fe0 * Add count_tags functionto pos.pyx, which should probably live in another file. Feature set achieves 97.9 on wsj19-21, 95.85 on onto web. 2014-10-31 17:42:15 +11:00
Matthew Honnibal
f67cb9a5a3 * Add count_tags functionto pos.pyx, which should probably live in another file. Feature set achieves 97.9 on wsj19-21, 95.85 on onto web. 2014-10-31 17:42:04 +11:00
Matthew Honnibal
ea8f1e7053 * Tighten interfaces 2014-10-30 18:14:42 +11:00
Matthew Honnibal
ea85bf3a0a * Tighten the interface to Language 2014-10-30 18:01:27 +11:00
Matthew Honnibal
c6fcd03692 * Small efficiency tweak to lexeme init 2014-10-30 17:56:11 +11:00
Matthew Honnibal
87c2418a89 * Fiddle with data types on Lexeme, to compress them to a much smaller size. 2014-10-30 15:42:15 +11:00
Matthew Honnibal
ac88893232 * Fix Token after lexeme changes 2014-10-30 15:30:52 +11:00
Matthew Honnibal
e6b87766fe * Remove lexemes vector from Lexicon, and the id and hash attributes from Lexeme 2014-10-30 15:21:38 +11:00
Matthew Honnibal
889b7b48b4 * Fix POS tagger, so that it loads correctly. Lexemes are being read in. 2014-10-30 13:38:55 +11:00
Matthew Honnibal
67c8c8019f * Update lexeme serialization, using a binary file format 2014-10-30 01:01:00 +11:00
Matthew Honnibal
13909a2e24 * Rewriting Lexeme serialization. 2014-10-29 23:19:38 +11:00
Matthew Honnibal
234d49bf4d * Seems to be working after refactor. Need to wire up more POS tag features, and wire up save/load of POS tags. 2014-10-24 02:23:42 +11:00
Matthew Honnibal
08ce602243 * Large refactor, particularly to Python API 2014-10-24 00:59:17 +11:00
Matthew Honnibal
7baef5b7ff * Fix padding on tokens 2014-10-23 04:01:17 +11:00
Matthew Honnibal
96b835a3d4 * Upd for refactored Tokens class. Now gets 95.74, 185ms training on swbd_wsj_ewtb, eval on onto_web, Google POS tags. 2014-10-23 03:20:02 +11:00
Matthew Honnibal
e5e951ae67 * Remove the feature array stuff from Tokens class, and replace vector with array-based implementation, with padding. 2014-10-23 01:57:59 +11:00
Matthew Honnibal
ea1d4a81eb * Refactoring get_atoms, improving tokens API 2014-10-22 13:10:56 +11:00
Matthew Honnibal
ad49e2482e * Tagger now gets 97pc on wsj, parsing 19-21 in 500ms. Gets 92.7 on web text. 2014-10-22 12:57:06 +11:00
Matthew Honnibal
0a0e41f6c8 * Add prefix and suffix features 2014-10-22 12:56:09 +11:00
Matthew Honnibal
7018b53d3a * Improve array features in tokens 2014-10-22 12:55:42 +11:00
Matthew Honnibal
43d5964e13 * Add function to read detokenization rules 2014-10-22 12:54:59 +11:00
Matthew Honnibal
224bdae996 * Add POS utilities 2014-10-22 10:17:57 +11:00
Matthew Honnibal
5ebe14f353 * Add greedy pos tagger 2014-10-22 10:17:26 +11:00
Matthew Honnibal
12742f4f83 * Add detokenize method and test 2014-10-18 18:07:29 +11:00
Matthew Honnibal
99f5e59286 * Have tokenizer emit tokens for whitespace other than single spaces 2014-10-14 20:25:57 +11:00
Matthew Honnibal
43743a5d63 * Work on efficiency 2014-10-14 18:22:41 +11:00
Matthew Honnibal
6fb42c4919 * Add offsets to Tokens class. Some changes to interfaces, and reorganization of spacy.Lang 2014-10-14 16:17:45 +11:00
Matthew Honnibal
2805068ca8 * Have tokens track tuples that record the start offset and pos tag as well as a lexeme pointer 2014-10-14 15:21:03 +11:00
Matthew Honnibal
65d3ead4fd * Rename LexStr_casefix to LexStr_norm and LexInt_i to LexInt_id 2014-10-14 15:19:07 +11:00
Matthew Honnibal
868e558037 * Preparations in place to handle hyphenation etc 2014-10-10 20:23:23 +11:00
Matthew Honnibal
ff79dbac2e * More slight cleaning for lang.pyx 2014-10-10 20:11:22 +11:00
Matthew Honnibal
3d82ed1e5e * More slight cleaning for lang.pyx 2014-10-10 19:50:07 +11:00
Matthew Honnibal
02e948e7d5 * Remove counts stuff from Language class 2014-10-10 19:25:01 +11:00
Matthew Honnibal
71ee921055 * Slight cleaning of tokenizer code 2014-10-10 19:17:22 +11:00
Matthew Honnibal
59b41a9fd3 * Switch to new data model, tests passing 2014-10-10 08:11:31 +11:00
Matthew Honnibal
1b0e01d3d8 * Revising data model of lexeme. Compiles. 2014-10-09 19:53:30 +11:00
Matthew Honnibal
e40caae51f * Update Lexicon class to expect a list of lexeme dict descriptions 2014-10-09 14:51:35 +11:00
Matthew Honnibal
51d75b244b * Add serialize/deserialize functions for lexeme, transport to/from python dict. 2014-10-09 14:10:46 +11:00
Matthew Honnibal
d73d89a2de * Add i attribute to lexeme, giving lexemes sequential IDs. 2014-10-09 13:50:05 +11:00
Matthew Honnibal
096ef2b199 * Rename external hashing lib, from trustyc to preshed 2014-09-26 18:40:03 +02:00
Matthew Honnibal
11a346fd5e * Remove hashing modules, which are now taken over by external lib 2014-09-26 18:39:40 +02:00
Matthew Honnibal
93505276ed * Add German tokenizer files 2014-09-25 18:29:13 +02:00
Matthew Honnibal
2e44fa7179 * Add util.py 2014-09-25 18:26:22 +02:00
Matthew Honnibal
b15619e170 * Use PointerHash instead of locally provided _hashing module 2014-09-25 18:23:35 +02:00
Matthew Honnibal
ed446c67ad * Add typedefs file 2014-09-17 23:10:32 +02:00
Matthew Honnibal
316a57c4be * Remove own memory classes, which have now been broken out into their own package 2014-09-17 23:10:07 +02:00
Matthew Honnibal
ac522e2553 * Switch from own memory class to cymem, in pip 2014-09-17 23:09:24 +02:00
Matthew Honnibal
6266cac593 * Switch to using a Python ref counted gateway to malloc/free, to prevent memory leaks 2014-09-17 20:02:26 +02:00
Matthew Honnibal
5a20dfc03e * Add memory management code 2014-09-17 20:02:06 +02:00
Matthew Honnibal
0152831c89 * Refactor tokenization, enable cache, and ensure we look up specials correctly even when there's confusing punctuation surrounding the token. 2014-09-16 18:01:46 +02:00
Matthew Honnibal
143e51ec73 * Refactor tokenization, splitting it into a clearer life-cycle. 2014-09-16 13:16:02 +02:00
Matthew Honnibal
c396581a0b * Fiddle with the way strings are interned in lexeme 2014-09-15 06:34:45 +02:00
Matthew Honnibal
0bb547ab98 * Fix memory error in cache, where entry wasn't being null-terminated. Various other changes, some good for performance 2014-09-15 06:34:10 +02:00
Matthew Honnibal
7959141d36 * Add a few abbreviations, to get tests to pass 2014-09-15 06:32:18 +02:00
Matthew Honnibal
d235299260 * Few nips and tucks to hash table 2014-09-15 05:03:44 +02:00
Matthew Honnibal
e68a431e5e * Pass only the tokens vector to _tokenize, instead of the whole python object. 2014-09-15 04:01:38 +02:00
Matthew Honnibal
08cef75ffd * Switch to using a heap-allocated vector in tokens 2014-09-15 03:46:14 +02:00
Matthew Honnibal
f77b7098c0 * Upd Tokens to use vector, with bounds checking. 2014-09-15 03:22:40 +02:00
Matthew Honnibal
0f6bf2a2ee * Fix niggling memory error, which was caused by bug in the way tokens resized their internal vector. 2014-09-15 02:08:39 +02:00
Matthew Honnibal
df24e3708c * Move EnglishTokens stuff to Tokens 2014-09-15 01:31:44 +02:00
Matthew Honnibal
bd08cb09a2 * Remove short-circuiting of initial_size argument for PointerHash 2014-09-15 01:30:49 +02:00
Matthew Honnibal
f3393cf57c * Improve interface for PointerHash 2014-09-13 17:29:58 +02:00