Skip to content

mynlp/multiCCG

BranchesTags

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
ccg_supertagger
ccg_supertagger
 
 
data
data
 
 
parsing
parsing
 
 
plms/mt5-base
plms/mt5-base
 
 
treebanks
treebanks
 
 
.gitattributes
.gitattributes
 
 
README.md
README.md
 
 
base.py
base.py
 
 
ccg_rules.py
ccg_rules.py
 
 
ccg_unification.py
ccg_unification.py
 
 
data_loader.py
data_loader.py
 
 
requirements.txt
requirements.txt
 
 
tools.py
tools.py
 
 

Repository files navigation

Introduction

This is a multilingual CCG parser for both CCGBank and multilingual CCG treebanks based on python. Refer to https://github.com/KunhangL/CCG-HPSG-Parser for the older parser (designed specifically for CCGBank) and instructions.

Installing

  1. Create an environment named ccg: conda create -n ccg python=3.10.4
  2. Install packages using pip: pip install -r requirements.txt
  3. If you plan to use CCGBank, please save data (ccgbank-wsj_00.auto, ccgbank-wsj_02-21.auto and ccgbank-wsj_23.auto) under ./data. If you plan to use data of treebanks, please save all data to ./treebanks like below.
treebanks
├── UD_Afrikaans-AfriBooms
│   ├── af_afribooms-ud-dev.auto
│   ├── af_afribooms-ud-dev.lexicon
│   ├── af_afribooms-ud-test.auto
│   ├── af_afribooms-ud-test.lexicon
│   ├── af_afribooms-ud-train.auto
│   └── af_afribooms-ud-train.lexicon
├── UD_Ancient_Greek-Perseus
│   ├── grc_perseus-ud-dev.auto
│   ├── grc_perseus-ud-dev.lexicon
│   ├── grc_perseus-ud-test.auto
│   ├── grc_perseus-ud-test.lexicon
│   ├── grc_perseus-ud-train.auto
│   └── grc_perseus-ud-train.lexicon
├── ...
...
  1. Please download and save the corresponding PLM directory under ./plms like below. (You can only prepare mT5-base, which is the default PLM we use in codes.)
plms
├── bert-base-multilingual-cased
│   ├── config.json
│   ├── pytorch_model.bin
│   └── vocab.txt
├── bert-base-uncased
│   ├── config.json
│   ├── pytorch_model.bin
│   └── vocab.txt
└── mt5-base
    ├── config.json
    ├── generation_config.json
    ├── pytorch_model.bin
    ├── special_tokens_map.json
    ├── spiece.model
    └── tokenizer_config.json

NOTE: For the current treebanks, CCGBank data and PLMs used in development, please refer to https://drive.google.com/drive/folders/1W3oE9J0bRhNHizbv7VO7ahsCxY5EhYW8?usp=share_link

Supertagging

The supertagging model is a classifier built upon Transformers. We use the encoder of mT5-base by default.

Current Best Result

PLM Params Best Result Multitagging Results on Dev Data (with the best epoch)
mT5-base n_epochs = 20;
dropout = 0.5;
AdamW;
lr = 1e-5;
batch_size = 10;		 best_epoch = 20,
train_eval acc = 91.378,
dev_eval acc = 86.959		 beta = 0.0005, topk = 10 --> averaged number of lexical categories = 5.03, acc = 97.984;

beta = 0.0001, topk = 10 --> averaged number of lexical categories = 6.16, acc = 98.123;

beta = 0.00005, topk = 10 --> averaged number of lexical categories = 6.62, acc = 98.155;

beta = 0.00001, topk = 10 --> averaged number of lexical categories = 7.51, acc = 98.187;

Evaluation of dev data for each single treebank. Results shown in the order of decreasing training treebank size in terms of n_tokens

[UD_German-HDT] || Training size ~ n_sents: 142623, n_tokens: 2475889, ave_token_per_sent: 17.36 || Testing size ~ n_sents: 17192, n_tokens: 286449, ave_token_per_sent: 16.66 || ave_loss: 0.20399958353359685, acc: 94.54492950439453

[UD_Russian-SynTagRus] || Training size ~ n_sents: 64737, n_tokens: 1067352, ave_token_per_sent: 16.49 || Testing size ~ n_sents: 8297, n_tokens: 135217, ave_token_per_sent: 16.3 || ave_loss: 0.36904021011194194, acc: 90.33775329589844

[UD_Czech-PDT] || Training size ~ n_sents: 59481, n_tokens: 956509, ave_token_per_sent: 16.08 || Testing size ~ n_sents: 7979, n_tokens: 129201, ave_token_per_sent: 16.19 || ave_loss: 0.3787599020686589, acc: 89.93041229248047

[UD_Icelandic-IcePaHC] || Training size ~ n_sents: 33083, n_tokens: 665658, ave_token_per_sent: 20.12 || Testing size ~ n_sents: 4687, n_tokens: 127390, ave_token_per_sent: 27.18 || ave_loss: 0.7265830685271383, acc: 81.82823944091797

[UD_Romanian-Nonstandard] || Training size ~ n_sents: 22546, n_tokens: 481614, ave_token_per_sent: 21.36 || Testing size ~ n_sents: 1017, n_tokens: 17712, ave_token_per_sent: 17.42 || ave_loss: 0.6031297251728236, acc: 83.96002960205078

[UD_Spanish-AnCora] || Training size ~ n_sents: 13360, n_tokens: 405954, ave_token_per_sent: 30.39 || Testing size ~ n_sents: 1519, n_tokens: 46797, ave_token_per_sent: 30.81 || ave_loss: 0.3458706690488677, acc: 91.57424926757812

[UD_Czech-CAC] || Training size ~ n_sents: 20419, n_tokens: 388531, ave_token_per_sent: 19.03 || Testing size ~ n_sents: 551, n_tokens: 9689, ave_token_per_sent: 17.58 || ave_loss: 1.9561946247039097, acc: 78.97615051269531

[UD_Persian-PerDT] || Training size ~ n_sents: 22445, n_tokens: 358035, ave_token_per_sent: 15.95 || Testing size ~ n_sents: 1246, n_tokens: 19678, ave_token_per_sent: 15.79 || ave_loss: 0.3706875228881836, acc: 90.41569519042969

[UD_Catalan-AnCora] || Training size ~ n_sents: 11254, n_tokens: 343074, ave_token_per_sent: 30.48 || Testing size ~ n_sents: 1455, n_tokens: 45785, ave_token_per_sent: 31.47 || ave_loss: 0.31733518139752626, acc: 92.33592224121094

[UD_Spanish-GSD] || Training size ~ n_sents: 12969, n_tokens: 334792, ave_token_per_sent: 25.81 || Testing size ~ n_sents: 1268, n_tokens: 32327, ave_token_per_sent: 25.49 || ave_loss: 0.4872328211942057, acc: 87.35113525390625

[UD_French-GSD] || Training size ~ n_sents: 13601, n_tokens: 322274, ave_token_per_sent: 23.69 || Testing size ~ n_sents: 1388, n_tokens: 32775, ave_token_per_sent: 23.61 || ave_loss: 0.2557284160912466, acc: 92.91227722167969

[UD_Estonian-EDT] || Training size ~ n_sents: 23232, n_tokens: 312121, ave_token_per_sent: 13.43 || Testing size ~ n_sents: 2843, n_tokens: 39383, ave_token_per_sent: 13.85 || ave_loss: 0.572241225080532, acc: 86.12599182128906

[UD_Italian-ISDT] || Training size ~ n_sents: 12747, n_tokens: 258514, ave_token_per_sent: 20.28 || Testing size ~ n_sents: 542, n_tokens: 10990, ave_token_per_sent: 20.28 || ave_loss: 0.31557465249841865, acc: 91.24658966064453

[UD_Polish-PDB] || Training size ~ n_sents: 16167, n_tokens: 245243, ave_token_per_sent: 15.17 || Testing size ~ n_sents: 2013, n_tokens: 30271, ave_token_per_sent: 15.04 || ave_loss: 0.3909931227029993, acc: 88.55339813232422

[UD_Belarusian-HSE] || Training size ~ n_sents: 21011, n_tokens: 235969, ave_token_per_sent: 11.23 || Testing size ~ n_sents: 1175, n_tokens: 13056, ave_token_per_sent: 11.11 || ave_loss: 0.40820041889230074, acc: 89.82843017578125

[UD_Classical_Chinese-Kyoto] || Training size ~ n_sents: 47905, n_tokens: 231433, ave_token_per_sent: 4.83 || Testing size ~ n_sents: 5520, n_tokens: 27179, ave_token_per_sent: 4.92 || ave_loss: 0.7152479682758829, acc: 79.96614837646484

[UD_Hindi-HDTB] || Training size ~ n_sents: 11062, n_tokens: 219991, ave_token_per_sent: 19.89 || Testing size ~ n_sents: 1400, n_tokens: 28176, ave_token_per_sent: 20.13 || ave_loss: 0.41201430011008466, acc: 88.63216400146484

[UD_Norwegian-Bokmaal] || Training size ~ n_sents: 14449, n_tokens: 215263, ave_token_per_sent: 14.9 || Testing size ~ n_sents: 2229, n_tokens: 32216, ave_token_per_sent: 14.45 || ave_loss: 0.22726076677224427, acc: 94.0495376586914

[UD_Norwegian-Nynorsk] || Training size ~ n_sents: 12972, n_tokens: 214325, ave_token_per_sent: 16.52 || Testing size ~ n_sents: 1738, n_tokens: 27584, ave_token_per_sent: 15.87 || ave_loss: 0.30332234258838425, acc: 92.50652313232422

[UD_German-GSD] || Training size ~ n_sents: 11888, n_tokens: 212175, ave_token_per_sent: 17.85 || Testing size ~ n_sents: 648, n_tokens: 9292, ave_token_per_sent: 14.34 || ave_loss: 0.6013183150153893, acc: 86.24623107910156

[UD_Korean-Kaist] || Training size ~ n_sents: 17574, n_tokens: 211493, ave_token_per_sent: 12.03 || Testing size ~ n_sents: 1576, n_tokens: 17836, ave_token_per_sent: 11.32 || ave_loss: 0.6054604967372327, acc: 84.58735656738281

[UD_Portuguese-GSD] || Training size ~ n_sents: 8383, n_tokens: 208941, ave_token_per_sent: 24.92 || Testing size ~ n_sents: 1033, n_tokens: 26019, ave_token_per_sent: 25.19 || ave_loss: 0.26256496946399027, acc: 92.45935821533203

[UD_Latin-ITTB] || Training size ~ n_sents: 14400, n_tokens: 206338, ave_token_per_sent: 14.33 || Testing size ~ n_sents: 1380, n_tokens: 16089, ave_token_per_sent: 11.66 || ave_loss: 0.4428451117128134, acc: 88.94896697998047

[UD_Italian-VIT] || Training size ~ n_sents: 7568, n_tokens: 190665, ave_token_per_sent: 25.19 || Testing size ~ n_sents: 681, n_tokens: 24516, ave_token_per_sent: 36.0 || ave_loss: 0.3891491754979327, acc: 89.7740249633789

[UD_English-EWT] || Training size ~ n_sents: 11254, n_tokens: 173967, ave_token_per_sent: 15.46 || Testing size ~ n_sents: 1835, n_tokens: 21945, ave_token_per_sent: 11.96 || ave_loss: 0.4353062829164707, acc: 87.96270751953125

[UD_Latvian-LVTB] || Training size ~ n_sents: 10614, n_tokens: 170801, ave_token_per_sent: 16.09 || Testing size ~ n_sents: 1675, n_tokens: 24793, ave_token_per_sent: 14.8 || ave_loss: 0.5098282019164235, acc: 86.78659057617188

[UD_Arabic-PADT] || Training size ~ n_sents: 4975, n_tokens: 159423, ave_token_per_sent: 32.04 || Testing size ~ n_sents: 698, n_tokens: 21296, ave_token_per_sent: 30.51 || ave_loss: 0.7306396612099239, acc: 82.31123352050781

[UD_Japanese-GSD] || Training size ~ n_sents: 6794, n_tokens: 158851, ave_token_per_sent: 23.38 || Testing size ~ n_sents: 489, n_tokens: 11706, ave_token_per_sent: 23.94 || ave_loss: 0.38452432091747013, acc: 90.25286102294922

[UD_Romanian-RRT] || Training size ~ n_sents: 6809, n_tokens: 150342, ave_token_per_sent: 22.08 || Testing size ~ n_sents: 626, n_tokens: 13547, ave_token_per_sent: 21.64 || ave_loss: 0.48707828966398087, acc: 88.25569915771484

[UD_Turkish-Penn] || Training size ~ n_sents: 13525, n_tokens: 147539, ave_token_per_sent: 10.91 || Testing size ~ n_sents: 576, n_tokens: 6277, ave_token_per_sent: 10.9 || ave_loss: 0.7424036254142893, acc: 78.3335952758789

[UD_Russian-Taiga] || Training size ~ n_sents: 14449, n_tokens: 145755, ave_token_per_sent: 10.09 || Testing size ~ n_sents: 809, n_tokens: 7858, ave_token_per_sent: 9.71 || ave_loss: 1.0193609206212892, acc: 75.74446105957031

[UD_Dutch-Alpino] || Training size ~ n_sents: 10425, n_tokens: 143975, ave_token_per_sent: 13.81 || Testing size ~ n_sents: 630, n_tokens: 9457, ave_token_per_sent: 15.01 || ave_loss: 0.38703948566837915, acc: 89.4575424194336

[UD_Finnish-TDT] || Training size ~ n_sents: 10954, n_tokens: 139746, ave_token_per_sent: 12.76 || Testing size ~ n_sents: 1237, n_tokens: 15688, ave_token_per_sent: 12.68 || ave_loss: 0.530244565869291, acc: 85.82994079589844

[UD_Portuguese-Bosque] || Training size ~ n_sents: 6014, n_tokens: 138083, ave_token_per_sent: 22.96 || Testing size ~ n_sents: 1007, n_tokens: 22775, ave_token_per_sent: 22.62 || ave_loss: 0.3896965224406507, acc: 89.88803100585938

[UD_Turkish-Kenet] || Training size ~ n_sents: 14663, n_tokens: 134272, ave_token_per_sent: 9.16 || Testing size ~ n_sents: 1562, n_tokens: 16414, ave_token_per_sent: 10.51 || ave_loss: 1.0017769503745304, acc: 74.0770034790039

[UD_Croatian-SET] || Training size ~ n_sents: 6174, n_tokens: 130442, ave_token_per_sent: 21.13 || Testing size ~ n_sents: 849, n_tokens: 18858, ave_token_per_sent: 22.21 || ave_loss: 0.4803919949952294, acc: 88.18537902832031

[UD_Hebrew-HTB] || Training size ~ n_sents: 5021, n_tokens: 127476, ave_token_per_sent: 25.39 || Testing size ~ n_sents: 472, n_tokens: 10802, ave_token_per_sent: 22.89 || ave_loss: 0.4419611959407727, acc: 89.69635009765625

[UD_Japanese-GSDLUW] || Training size ~ n_sents: 6816, n_tokens: 123653, ave_token_per_sent: 18.14 || Testing size ~ n_sents: 489, n_tokens: 9100, ave_token_per_sent: 18.61 || ave_loss: 0.2910939544743421, acc: 92.56044006347656

[UD_Old_French-SRCMF] || Training size ~ n_sents: 11874, n_tokens: 118316, ave_token_per_sent: 9.96 || Testing size ~ n_sents: 1575, n_tokens: 15131, ave_token_per_sent: 9.61 || ave_loss: 0.5612638631387602, acc: 85.61893463134766

[UD_Bulgarian-BTB] || Training size ~ n_sents: 8615, n_tokens: 117513, ave_token_per_sent: 13.64 || Testing size ~ n_sents: 1086, n_tokens: 15275, ave_token_per_sent: 14.07 || ave_loss: 1.1381425332585606, acc: 81.23731994628906

[UD_Icelandic-Modern] || Training size ~ n_sents: 5130, n_tokens: 115833, ave_token_per_sent: 22.58 || Testing size ~ n_sents: 742, n_tokens: 15775, ave_token_per_sent: 21.26 || ave_loss: 0.5410747796297073, acc: 87.10617065429688

[UD_Finnish-FTB] || Training size ~ n_sents: 13778, n_tokens: 111627, ave_token_per_sent: 8.1 || Testing size ~ n_sents: 1716, n_tokens: 13500, ave_token_per_sent: 7.87 || ave_loss: 0.5635133101279999, acc: 84.7851791381836

[UD_Czech-FicTree] || Training size ~ n_sents: 9002, n_tokens: 106830, ave_token_per_sent: 11.87 || Testing size ~ n_sents: 1162, n_tokens: 13451, ave_token_per_sent: 11.58 || ave_loss: 0.3540819593283356, acc: 89.74053955078125

[UD_Persian-Seraji] || Training size ~ n_sents: 4462, n_tokens: 105970, ave_token_per_sent: 23.75 || Testing size ~ n_sents: 560, n_tokens: 14045, ave_token_per_sent: 25.08 || ave_loss: 0.6739260980061123, acc: 81.77999114990234

[UD_Polish-LFG] || Training size ~ n_sents: 13667, n_tokens: 103604, ave_token_per_sent: 7.58 || Testing size ~ n_sents: 1739, n_tokens: 13024, ave_token_per_sent: 7.49 || ave_loss: 0.37631598101051034, acc: 88.52118682861328

[UD_Latin-PROIEL] || Training size ~ n_sents: 11443, n_tokens: 94757, ave_token_per_sent: 8.28 || Testing size ~ n_sents: 847, n_tokens: 7335, ave_token_per_sent: 8.66 || ave_loss: 0.7682609861387926, acc: 79.91819763183594

[UD_Slovenian-SSJ] || Training size ~ n_sents: 5689, n_tokens: 94561, ave_token_per_sent: 16.62 || Testing size ~ n_sents: 617, n_tokens: 11400, ave_token_per_sent: 18.48 || ave_loss: 0.30855319168298473, acc: 92.41228485107422

[UD_Ancient_Greek-PROIEL] || Training size ~ n_sents: 9399, n_tokens: 91901, ave_token_per_sent: 9.78 || Testing size ~ n_sents: 612, n_tokens: 6579, ave_token_per_sent: 10.75 || ave_loss: 2.6672701104994743, acc: 65.45067596435547

[UD_Old_East_Slavic-TOROT] || Training size ~ n_sents: 11317, n_tokens: 90658, ave_token_per_sent: 8.01 || Testing size ~ n_sents: 1574, n_tokens: 12032, ave_token_per_sent: 7.64 || ave_loss: 0.8016182303051406, acc: 78.85637664794922

[UD_Chinese-GSD] || Training size ~ n_sents: 3728, n_tokens: 90199, ave_token_per_sent: 24.2 || Testing size ~ n_sents: 474, n_tokens: 11773, ave_token_per_sent: 24.84 || ave_loss: 0.6287098446240028, acc: 85.22042083740234

[UD_English-GUM] || Training size ~ n_sents: 5096, n_tokens: 89355, ave_token_per_sent: 17.53 || Testing size ~ n_sents: 784, n_tokens: 13891, ave_token_per_sent: 17.72 || ave_loss: 0.400859921510461, acc: 88.80570220947266

[UD_Turkish-BOUN] || Training size ~ n_sents: 7244, n_tokens: 86869, ave_token_per_sent: 11.99 || Testing size ~ n_sents: 909, n_tokens: 10733, ave_token_per_sent: 11.81 || ave_loss: 1.1756272463353126, acc: 69.05804443359375

[UD_Indonesian-GSD] || Training size ~ n_sents: 4138, n_tokens: 85423, ave_token_per_sent: 20.64 || Testing size ~ n_sents: 503, n_tokens: 10761, ave_token_per_sent: 21.39 || ave_loss: 0.7579179996368932, acc: 80.03903198242188

[UD_Italian-PoSTWITA] || Training size ~ n_sents: 4646, n_tokens: 84682, ave_token_per_sent: 18.23 || Testing size ~ n_sents: 563, n_tokens: 10087, ave_token_per_sent: 17.92 || ave_loss: 0.7505603343771216, acc: 81.51085662841797

[UD_Chinese-GSDSimp] || Training size ~ n_sents: 3544, n_tokens: 83567, ave_token_per_sent: 23.58 || Testing size ~ n_sents: 444, n_tokens: 10831, ave_token_per_sent: 24.39 || ave_loss: 0.5980112575822406, acc: 85.6984634399414

[UD_Romanian-SiMoNERo] || Training size ~ n_sents: 2908, n_tokens: 81363, ave_token_per_sent: 27.98 || Testing size ~ n_sents: 348, n_tokens: 10675, ave_token_per_sent: 30.68 || ave_loss: 1.1469638807432991, acc: 81.42388153076172

[UD_Ukrainian-IU] || Training size ~ n_sents: 4901, n_tokens: 76928, ave_token_per_sent: 15.7 || Testing size ~ n_sents: 590, n_tokens: 10234, ave_token_per_sent: 17.35 || ave_loss: 0.4767779902381412, acc: 87.42427062988281

[UD_Galician-CTG] || Training size ~ n_sents: 2210, n_tokens: 76573, ave_token_per_sent: 34.65 || Testing size ~ n_sents: 846, n_tokens: 29115, ave_token_per_sent: 34.41 || ave_loss: 1.1364294052124024, acc: 75.31856536865234

[UD_Urdu-UDTB] || Training size ~ n_sents: 3077, n_tokens: 74832, ave_token_per_sent: 24.32 || Testing size ~ n_sents: 424, n_tokens: 10160, ave_token_per_sent: 23.96 || ave_loss: 0.7032842001942701, acc: 81.21063232421875

[UD_Turkish-Tourism] || Training size ~ n_sents: 15277, n_tokens: 70112, ave_token_per_sent: 4.59 || Testing size ~ n_sents: 2162, n_tokens: 10277, ave_token_per_sent: 4.75 || ave_loss: 0.3322122536979461, acc: 89.81220245361328

[UD_Irish-IDT] || Training size ~ n_sents: 3032, n_tokens: 66316, ave_token_per_sent: 21.87 || Testing size ~ n_sents: 316, n_tokens: 6011, ave_token_per_sent: 19.02 || ave_loss: 0.973869888111949, acc: 78.30643463134766

[UD_Serbian-SET] || Training size ~ n_sents: 3083, n_tokens: 65593, ave_token_per_sent: 21.28 || Testing size ~ n_sents: 486, n_tokens: 10243, ave_token_per_sent: 21.08 || ave_loss: 0.3817359300292268, acc: 91.07683563232422

[UD_Dutch-LassySmall] || Training size ~ n_sents: 5320, n_tokens: 64301, ave_token_per_sent: 12.09 || Testing size ~ n_sents: 605, n_tokens: 9151, ave_token_per_sent: 15.13 || ave_loss: 0.47722327010706067, acc: 86.41677856445312

[UD_Slovak-SNK] || Training size ~ n_sents: 7008, n_tokens: 64251, ave_token_per_sent: 9.17 || Testing size ~ n_sents: 971, n_tokens: 11209, ave_token_per_sent: 11.54 || ave_loss: 0.34890312445825156, acc: 91.12320709228516

[UD_Western_Armenian-ArmTDP] || Training size ~ n_sents: 3627, n_tokens: 63114, ave_token_per_sent: 17.4 || Testing size ~ n_sents: 437, n_tokens: 7654, ave_token_per_sent: 17.51 || ave_loss: 0.7502026706933975, acc: 80.74209594726562

[UD_Swedish-Talbanken] || Training size ~ n_sents: 3984, n_tokens: 58673, ave_token_per_sent: 14.73 || Testing size ~ n_sents: 444, n_tokens: 8123, ave_token_per_sent: 18.3 || ave_loss: 0.45270336733924016, acc: 88.18170166015625

[UD_Danish-DDT] || Training size ~ n_sents: 3428, n_tokens: 53991, ave_token_per_sent: 15.75 || Testing size ~ n_sents: 441, n_tokens: 7005, ave_token_per_sent: 15.88 || ave_loss: 0.5173318127791087, acc: 87.59457397460938

[UD_Scottish_Gaelic-ARCOSG] || Training size ~ n_sents: 2824, n_tokens: 51062, ave_token_per_sent: 18.08 || Testing size ~ n_sents: 576, n_tokens: 8174, ave_token_per_sent: 14.19 || ave_loss: 0.9974150678207134, acc: 77.75875091552734

[UD_Swedish-LinES] || Training size ~ n_sents: 2994, n_tokens: 49936, ave_token_per_sent: 16.68 || Testing size ~ n_sents: 957, n_tokens: 16173, ave_token_per_sent: 16.9 || ave_loss: 0.5265347788420817, acc: 87.1205062866211

[UD_English-LinES] || Training size ~ n_sents: 2833, n_tokens: 49176, ave_token_per_sent: 17.36 || Testing size ~ n_sents: 898, n_tokens: 15662, ave_token_per_sent: 17.44 || ave_loss: 0.5170677006244659, acc: 86.34912109375

[UD_English-Atis] || Training size ~ n_sents: 4193, n_tokens: 47617, ave_token_per_sent: 11.36 || Testing size ~ n_sents: 554, n_tokens: 6414, ave_token_per_sent: 11.58 || ave_loss: 0.28618574192764107, acc: 92.89054870605469

[UD_French-Sequoia] || Training size ~ n_sents: 2153, n_tokens: 46955, ave_token_per_sent: 21.81 || Testing size ~ n_sents: 394, n_tokens: 9156, ave_token_per_sent: 23.24 || ave_loss: 0.35191274154931307, acc: 90.09392547607422

[UD_Latin-LLCT] || Training size ~ n_sents: 1802, n_tokens: 45118, ave_token_per_sent: 25.04 || Testing size ~ n_sents: 241, n_tokens: 6194, ave_token_per_sent: 25.7 || ave_loss: 0.36391018378490114, acc: 92.05683135986328

[UD_Italian-ParTUT] || Training size ~ n_sents: 1708, n_tokens: 44976, ave_token_per_sent: 26.33 || Testing size ~ n_sents: 151, n_tokens: 2795, ave_token_per_sent: 18.51 || ave_loss: 0.37600249610841274, acc: 90.76922607421875

[UD_Estonian-EWT] || Training size ~ n_sents: 3756, n_tokens: 42832, ave_token_per_sent: 11.4 || Testing size ~ n_sents: 747, n_tokens: 8593, ave_token_per_sent: 11.5 || ave_loss: 0.5895925744374593, acc: 85.12742614746094

[UD_Basque-BDT] || Training size ~ n_sents: 3677, n_tokens: 42361, ave_token_per_sent: 11.52 || Testing size ~ n_sents: 1223, n_tokens: 13663, ave_token_per_sent: 11.17 || ave_loss: 0.6827819407955418, acc: 82.95396423339844

[UD_Russian-GSD] || Training size ~ n_sents: 2497, n_tokens: 42223, ave_token_per_sent: 16.91 || Testing size ~ n_sents: 370, n_tokens: 6353, ave_token_per_sent: 17.17 || ave_loss: 0.6850266198854189, acc: 83.00016021728516

[UD_English-ParTUT] || Training size ~ n_sents: 1707, n_tokens: 40136, ave_token_per_sent: 23.51 || Testing size ~ n_sents: 151, n_tokens: 2580, ave_token_per_sent: 17.09 || ave_loss: 0.5216125813312829, acc: 87.24806213378906

[UD_Greek-GDT] || Training size ~ n_sents: 1543, n_tokens: 37999, ave_token_per_sent: 24.63 || Testing size ~ n_sents: 376, n_tokens: 9513, ave_token_per_sent: 25.3 || ave_loss: 0.7168417451413054, acc: 84.82077026367188

[UD_Korean-GSD] || Training size ~ n_sents: 3465, n_tokens: 36747, ave_token_per_sent: 10.61 || Testing size ~ n_sents: 750, n_tokens: 7773, ave_token_per_sent: 10.36 || ave_loss: 1.3697897024949393, acc: 67.83738708496094

[UD_Turkish-Atis] || Training size ~ n_sents: 4225, n_tokens: 35586, ave_token_per_sent: 8.42 || Testing size ~ n_sents: 567, n_tokens: 4793, ave_token_per_sent: 8.45 || ave_loss: 0.559838297531793, acc: 86.22991180419922

[UD_Lithuanian-ALKSNIS] || Training size ~ n_sents: 1890, n_tokens: 33866, ave_token_per_sent: 17.92 || Testing size ~ n_sents: 436, n_tokens: 7353, ave_token_per_sent: 16.86 || ave_loss: 0.7538191428916021, acc: 81.58574676513672

[UD_Armenian-ArmTDP] || Training size ~ n_sents: 1704, n_tokens: 33051, ave_token_per_sent: 19.4 || Testing size ~ n_sents: 218, n_tokens: 4248, ave_token_per_sent: 19.49 || ave_loss: 0.701201053505594, acc: 79.73163604736328

[UD_Turkish-IMST] || Training size ~ n_sents: 3404, n_tokens: 32776, ave_token_per_sent: 9.63 || Testing size ~ n_sents: 904, n_tokens: 8536, ave_token_per_sent: 9.44 || ave_loss: 1.3688426391109005, acc: 62.81631088256836

[UD_Afrikaans-AfriBooms] || Training size ~ n_sents: 1195, n_tokens: 28866, ave_token_per_sent: 24.16 || Testing size ~ n_sents: 185, n_tokens: 4820, ave_token_per_sent: 26.05 || ave_loss: 0.855294040943447, acc: 79.83402252197266

[UD_Old_Church_Slavonic-PROIEL] || Training size ~ n_sents: 3448, n_tokens: 28049, ave_token_per_sent: 8.13 || Testing size ~ n_sents: 897, n_tokens: 7382, ave_token_per_sent: 8.23 || ave_loss: 0.845256151093377, acc: 78.42047882080078

[UD_Gothic-PROIEL] || Training size ~ n_sents: 2819, n_tokens: 25256, ave_token_per_sent: 8.96 || Testing size ~ n_sents: 797, n_tokens: 7082, ave_token_per_sent: 8.89 || ave_loss: 1.1171187922358512, acc: 70.95453643798828

[UD_Ancient_Greek-Perseus] || Training size ~ n_sents: 2347, n_tokens: 22816, ave_token_per_sent: 9.72 || Testing size ~ n_sents: 206, n_tokens: 2880, ave_token_per_sent: 13.98 || ave_loss: 1.0867802983238584, acc: 72.11805725097656

[UD_Wolof-WTB] || Training size ~ n_sents: 1145, n_tokens: 22129, ave_token_per_sent: 19.33 || Testing size ~ n_sents: 426, n_tokens: 9315, ave_token_per_sent: 21.87 || ave_loss: 1.245098349659942, acc: 69.74771881103516

[UD_Faroese-FarPaHC] || Training size ~ n_sents: 1014, n_tokens: 22097, ave_token_per_sent: 21.79 || Testing size ~ n_sents: 296, n_tokens: 8310, ave_token_per_sent: 28.07 || ave_loss: 0.6107086320718129, acc: 85.10228729248047

[UD_Maltese-MUDT] || Training size ~ n_sents: 1072, n_tokens: 21002, ave_token_per_sent: 19.59 || Testing size ~ n_sents: 393, n_tokens: 8950, ave_token_per_sent: 22.77 || ave_loss: 2.7809042781591415, acc: 66.43575286865234

[UD_French-ParTUT] || Training size ~ n_sents: 717, n_tokens: 19722, ave_token_per_sent: 27.51 || Testing size ~ n_sents: 104, n_tokens: 1755, ave_token_per_sent: 16.88 || ave_loss: 0.43172089891000226, acc: 88.8888931274414

[UD_Italian-TWITTIRO] || Training size ~ n_sents: 951, n_tokens: 19321, ave_token_per_sent: 20.32 || Testing size ~ n_sents: 123, n_tokens: 2481, ave_token_per_sent: 20.17 || ave_loss: 0.7226600096775935, acc: 82.5070571899414

[UD_Vietnamese-VTB] || Training size ~ n_sents: 1336, n_tokens: 18680, ave_token_per_sent: 13.98 || Testing size ~ n_sents: 769, n_tokens: 10726, ave_token_per_sent: 13.95 || ave_loss: 1.447933057685951, acc: 66.12903594970703

[UD_Turkish-FrameNet] || Training size ~ n_sents: 2281, n_tokens: 16268, ave_token_per_sent: 7.13 || Testing size ~ n_sents: 204, n_tokens: 1406, ave_token_per_sent: 6.89 || ave_loss: 0.7962737267925626, acc: 77.24039459228516

[UD_Uyghur-UDT] || Training size ~ n_sents: 1449, n_tokens: 16019, ave_token_per_sent: 11.06 || Testing size ~ n_sents: 801, n_tokens: 8869, ave_token_per_sent: 11.07 || ave_loss: 1.2262895026324707, acc: 66.7042465209961

[UD_French-Rhapsodie] || Training size ~ n_sents: 1185, n_tokens: 16006, ave_token_per_sent: 13.51 || Testing size ~ n_sents: 1008, n_tokens: 11036, ave_token_per_sent: 10.95 || ave_loss: 0.5652795010569072, acc: 85.04893493652344

[UD_Hungarian-Szeged] || Training size ~ n_sents: 732, n_tokens: 14869, ave_token_per_sent: 20.31 || Testing size ~ n_sents: 320, n_tokens: 7412, ave_token_per_sent: 23.16 || ave_loss: 0.6865596557036042, acc: 83.59416961669922

[UD_Welsh-CCG] || Training size ~ n_sents: 731, n_tokens: 14493, ave_token_per_sent: 19.83 || Testing size ~ n_sents: 402, n_tokens: 8782, ave_token_per_sent: 21.85 || ave_loss: 0.7600899207882765, acc: 81.31404876708984

[UD_Latin-UDante] || Training size ~ n_sents: 480, n_tokens: 13552, ave_token_per_sent: 28.23 || Testing size ~ n_sents: 170, n_tokens: 4225, ave_token_per_sent: 24.85 || ave_loss: 1.0932055816930883, acc: 73.5621337890625

[UD_Norwegian-NynorskLIA] || Training size ~ n_sents: 1866, n_tokens: 10722, ave_token_per_sent: 5.75 || Testing size ~ n_sents: 388, n_tokens: 2051, ave_token_per_sent: 5.29 || ave_loss: 0.6516136208978983, acc: 83.61774444580078

[UD_Coptic-Scriptorium] || Training size ~ n_sents: 335, n_tokens: 9173, ave_token_per_sent: 27.38 || Testing size ~ n_sents: 173, n_tokens: 4771, ave_token_per_sent: 27.58 || ave_loss: 2.308559331629011, acc: 50.995601654052734

[UD_Turkish_German-SAGT] || Training size ~ n_sents: 503, n_tokens: 7973, ave_token_per_sent: 15.85 || Testing size ~ n_sents: 683, n_tokens: 10237, ave_token_per_sent: 14.99 || ave_loss: 1.4520702638487886, acc: 65.7614517211914

[UD_Czech-CLTT] || Training size ~ n_sents: 421, n_tokens: 6495, ave_token_per_sent: 15.43 || Testing size ~ n_sents: 49, n_tokens: 769, ave_token_per_sent: 15.69 || ave_loss: 0.5334645748138428, acc: 87.64629364013672

[UD_Tamil-TTB] || Training size ~ n_sents: 385, n_tokens: 6045, ave_token_per_sent: 15.7 || Testing size ~ n_sents: 80, n_tokens: 1263, ave_token_per_sent: 15.79 || ave_loss: 1.2758336886763573, acc: 65.71654510498047

[UD_Telugu-MTG] || Training size ~ n_sents: 1050, n_tokens: 5073, ave_token_per_sent: 4.83 || Testing size ~ n_sents: 131, n_tokens: 662, ave_token_per_sent: 5.05 || ave_loss: 0.6876610810203212, acc: 82.17522430419922

[UD_Marathi-UFAL] || Training size ~ n_sents: 357, n_tokens: 2795, ave_token_per_sent: 7.83 || Testing size ~ n_sents: 42, n_tokens: 366, ave_token_per_sent: 8.71 || ave_loss: 1.0074745655059814, acc: 73.22404479980469

[UD_Lithuanian-HSE] || Training size ~ n_sents: 108, n_tokens: 2103, ave_token_per_sent: 19.47 || Testing size ~ n_sents: 41, n_tokens: 796, ave_token_per_sent: 19.41 || ave_loss: 0.9869348645210266, acc: 73.24120330810547

[UD_Swedish_Sign_Language-SSLC] || Training size ~ n_sents: 63, n_tokens: 350, ave_token_per_sent: 5.56 || Testing size ~ n_sents: 48, n_tokens: 311, ave_token_per_sent: 6.48 || ave_loss: 2.6652204513549806, acc: 31.511253356933594

Train a Supertagging Model

cd ccg_supertagger
bash run_trainer.sh

NOTE 1: Using AutoTokenizer in trainer.py and supertagger.py, and MT5EncoderModel in models.py. Needed to switch to BertModel in models.py when using BERT-family models.
NOTE 2: Please remember to update --treebanks_train_data_paths, --treebanks_dev_data_paths and --treebanks_test_data_paths when updating the data of treebanks.
NOTE 3: Please specify different parameters in run_trainer.sh so as to use different functions.

Important Parameters

--lexical_category2idx_path: The relative path to the dictionary mapping each CCG category to its index. Please specify either ../data/lexical_category2idx_cutoff_ccgbank.json or ../data/lexical_category2idx_cutoff_treebanks.json.

  • If you want to build your new category2idx dictionary for new treebanks data, please go to ./multiCCG and run the build_category2idx function in tools.py, e.g., build_category2idx(folder_path='./treebanks', result_path='./data/treebanks_lexical_category2idx_cutoff.json').

--model_dir: The relative path to the directory of the pretrained language model. Default to ../plms/mt5-base.
--checkpoints_dir: The directory created under ccg_supertagger to save checkpoints.
--n_epochs: The total number of epochs, default to be 20.
--device: The device to run the experiment, default to be cuda:0.
--batch_size: Default to be 8.
--lr: Default to be 1e-5.
--dropout_p: The probability of dropout, default to be 0.5.
--mode: The mode to use the trainer. Choices include train, train_on and test, default to train. train is for training from scratch. train_on is for training from a specific checkpoint, in which case --checkpoint_epoch should be specified. test is for testing on one dataset using the model from a specific checkpoint, in which case --checkpoint_epoch and --test_mode should be specified.
--test_mode: Only specified when it is test mode. Choices include train_eval, dev_eval and test_eval, default to dev_eval.
--checkpoint_epoch: Only for train_on and test mode, the specific epoch of checkpoint to use.
--data_mode: This specifies which dataset to use. Choices include ccgbank and treebanks.
--load_mode: If running the first script on a certain dataset (ccgbank or treebanks), please specify first, in which case the codes will build all necessary datasets from raw data, and store them as .pkl files under ccg_supertagger. Note that it takes long (one hour~) to prepare data of treebanks for the first time. Afterwards, please always specify reuse so as to directly load these saved files.

Some Example Scripts

  • Training on CCGBank from scratch
#!/bin/bash

DATA="ccgbank"
python -u trainer.py \
 --batch_size 10 \
 --dropout_p 0.5 \
 --lexical_category2idx_path ../data/lexical_category2idx_cutoff_ccgbank.json \
 --model_dir ../plms/mt5-base \
 --checkpoints_dir ./checkpoints_$DATA \
 --mode train \
 --n_epochs 20 \
 --data_mode ccgbank \
 --load_mode first \
 2>&1 | tee -a supertagging_train_$DATA.log
  • Training on treebanks from scratch
    • NOTE: When training for the first time, if CUDA returns out of memory during the first epoch, it is better to set the batch size to be 1 and train for one epoch, so as to filter out ids of problematic data which will be printed out when running. E.g., ('../treebanks/UD_Belarusian-HSE/be_hse-ud-train.auto_ID=radyjosvaboda-6682 PARSER=GOLD NUMPARSE=1',). Remember to collect such data entries and store them in problemantic_ids.txt (Check this file for how the entries should be kept.)
#!/bin/bash

DATA="treebanks"
python -u trainer.py \
 --batch_size 10 \
 --dropout_p 0.5 \
 --lexical_category2idx_path ../data/lexical_category2idx_cutoff_treebanks.json \
 --model_dir ../plms/mt5-base \
 --checkpoints_dir ./checkpoints_$DATA \
 --mode train \
 --n_epochs 20 \
 --data_mode treebanks \
 --load_mode first \
 2>&1 | tee -a supertagging_train_$DATA.log
  • Continuing to train the supertagging model on CCGBank from checkpoint of epoch 3
#!/bin/bash

DATA="ccgbank"
python -u trainer.py \
 --batch_size 10 \
 --dropout_p 0.5 \
 --lexical_category2idx_path ../data/lexical_category2idx_cutoff_ccgbank.json \
 --model_dir ../plms/mt5-base \
 --checkpoints_dir ./checkpoints_$DATA \
 --mode train_on \
 --n_epochs 20 \
 --checkpoint_epoch 3 \
 --data_mode ccgbank \
 --load_mode reuse \
 2>&1 | tee -a supertagging_train_$DATA.log
  • Continuing to train the supertagging model on treebanks from checkpoint of epoch 3
#!/bin/bash

DATA="treebanks"
python -u trainer.py \
 --batch_size 10 \
 --dropout_p 0.5 \
 --lexical_category2idx_path ../data/lexical_category2idx_cutoff_treebanks.json \
 --model_dir ../plms/mt5-base \
 --checkpoints_dir ./checkpoints_$DATA \
 --mode train_on \
 --n_epochs 20 \
 --checkpoint_epoch 3 \
 --data_mode treebanks \
 --load_mode reuse \
 2>&1 | tee -a supertagging_train_$DATA.log
  • Testing the supertagging model on dev data of CCGBank using the checkpoint from epoch 8
#!/bin/bash

DATA="ccgbank"
python -u trainer.py \
 --batch_size 10 \
 --lexical_category2idx_path ../data/lexical_category2idx_cutoff_ccgbank.json \
 --model_dir ../plms/mt5-base \
 --checkpoints_dir ./checkpoints_$DATA \
 --mode test \
 --test_mode dev_eval \
 --checkpoint_epoch 8 \
 --data_mode ccgbank \
 --load_mode reuse \
 2>&1 | tee -a supertagging_test_$DATA.log
  • Testing the supertagging model on dev data of treebanks using the checkpoint from epoch 20 (different from CCGBank, loading checkpoints and testing on treebank data will return sorted evaluation results for each treebank)
#!/bin/bash

DATA="treebanks"
python -u trainer.py \
 --batch_size 10 \
 --lexical_category2idx_path ../data/lexical_category2idx_cutoff_treebanks.json \
 --model_dir ../plms/mt5-base \
 --checkpoints_dir ./checkpoints_$DATA \
 --mode test \
 --test_mode dev_eval \
 --checkpoint_epoch 20 \
 --data_mode treebanks \
 --load_mode reuse \
 2>&1 | tee -a supertagging_test_$DATA.log

Use the Supertagger

cd ccg_supertagger
bash run_supertagger.sh

NOTE 1: Please remember to update --treebanks_dev_data_paths and --treebanks_test_data_paths when updating the data of treebanks.
NOTE 2: Please specify different parameters in run_supertagger.sh so as to use different functions.

Important Parameters

--lexical_category2idx_path: The relative path to the dictionary mapping each CCG category to its index. Please specify either ../data/lexical_category2idx_cutoff_ccgbank.json or ../data/lexical_category2idx_cutoff_treebanks.json.
--model_dir: The relative path to the directory of the pretrained language model. Default to ../plms/mt5-base.
--checkpoint_path: The path to the designated checkpoint file.
--device: The device to run the experiment, default to be cuda:0.
--batch_size: Default to be 8.
--top_k: The maximum number of supertags allowed for one word, default to 10.
--beta: The coefficient used to prune predicted categories, default to 0.0005.
--mode: The mode of the supertagger, choices include predict and sanity_check. If predict, you can specify the .json file path where you put a list of pretokenized sentences using --pretokenized_sents_path (default to ../data/pretokenized_sents.json), and you should also specify the output file path with --batch_predicted_path (default to ./batch_predicted_supertags.json). If sanity_check, the supertagger will just run on the designated data and print the (multi)tagging accuracy and average number of categories per word. Default to sanity_check.
--sanity_check_mode: Used when --mode is specified sanity_check. Choices include single_treebank, all_treebanks_dev, all_treebanks_test, ccgbank_dev and ccgbank_test. The supertagger will first load the designated data from corresponding .auto files. When specifying single_treebank, please modify --sanity_check_data_path to choose the path to the tested treebank (default to ../treebanks/UD_English-Atis/en_atis-ud-dev.auto).

Some Example Scripts

  • Using the checkpoint file ./checkpoints_ccgbank/fc_mt5-base_drop0.5_epoch_8.pt in the mode sanity_check, testing dev data of CCGBank.
#!/bin/bash

DATA="ccgbank"
TOPK=10
BETA=0.0005

python -u supertagger.py \
 --lexical_category2idx_path ../data/lexical_category2idx_cutoff_ccgbank.json \
 --model_dir ../plms/mt5-base \
 --checkpoint_path ./checkpoints_ccgbank/fc_mt5-base_drop0.5_epoch_8.pt \
 --device cuda:0 \
 --batch_size 8 \
 --top_k ${TOPK} \
 --beta ${BETA} \
 --mode sanity_check \
 --sanity_check_mode ccgbank_dev \
 2>&1 | tee -a supertagger_${DATA}_${TOPK}_${BETA}.log
  • Using the checkpoint file ./checkpoints_treebanks/fc_mt5-base_drop0.5_epoch_20.pt in the mode sanity_check, testing dev data of UD_Chinese-GSD.
#!/bin/bash

DATA="single_treebank"
TOPK=10
BETA=0.0005

python -u supertagger.py \
 --lexical_category2idx_path ../data/lexical_category2idx_cutoff_treebanks.json \
 --model_dir ../plms/mt5-base \
 --checkpoint_path ./checkpoints_treebanks/fc_mt5-base_drop0.5_epoch_20.pt \
 --device cuda:0 \
 --batch_size 8 \
 --top_k ${TOPK} \
 --beta ${BETA} \
 --mode sanity_check \
 --sanity_check_mode single_treebank \
 --sanity_check_data_path ../treebanks/UD_Chinese-GSD/zh_gsd-ud-dev.auto \
 2>&1 | tee -a supertagger_${DATA}_${TOPK}_${BETA}.log
  • Using the checkpoint file ./checkpoints_treebanks/fc_mt5-base_drop0.5_epoch_20.pt in the mode sanity_check, testing dev data of all treebanks.
#!/bin/bash

DATA="treebanks"
TOPK=10
BETA=0.0005

python -u supertagger.py \
 --lexical_category2idx_path ../data/lexical_category2idx_cutoff_treebanks.json \
 --model_dir ../plms/mt5-base \
 --checkpoint_path ./checkpoints_treebanks/fc_mt5-base_drop0.5_epoch_20.pt \
 --device cuda:0 \
 --batch_size 8 \
 --top_k ${TOPK} \
 --beta ${BETA} \
 --mode sanity_check \
 --sanity_check_mode all_treebanks_dev \
 2>&1 | tee -a supertagger_${DATA}_${TOPK}_${BETA}.log
  • After using the checkpoint file ./checkpoints_treebanks/fc_mt5-base_drop0.5_epoch_20.pt in the mode predict, the supertagger will load the pretokenized sentences from --pretokenized_sents_path, return all possible supertags for each token, and save them to --batch_predicted_path.
#!/bin/bash

MODE="predict"
TOPK=10
BETA=0.0005

python -u supertagger.py \
 --lexical_category2idx_path ../data/lexical_category2idx_cutoff_treebanks.json \
 --model_dir ../plms/mt5-base \
 --checkpoint_path ./checkpoints_treebanks/fc_mt5-base_drop0.5_epoch_20.pt \
 --device cuda:0 \
 --batch_size 8 \
 --top_k ${TOPK} \
 --beta ${BETA} \
 --mode predict \
 2>&1 | tee -a supertagger_${MODE}_${TOPK}_${BETA}.log

Parsing

This parser is built upon A* supertag-factored algorithms, i.e., it uses the scores from supertagging to conduct A* parsing.

Use the parser

cd parsing
bash run_parser.sh

NOTE 1: Please remember to update --treebanks_dev_data_paths and --treebanks_test_data_paths when updating the data of treebanks.
NOTE 2: Please specify different parameters in run_parser.sh so as to use different functions.

Important Parameters

--lexical_category2idx_path: The relative path to the dictionary mapping each CCG category to its index. Please specify either ../data/lexical_category2idx_cutoff_ccgbank.json or ../data/lexical_category2idx_cutoff_treebanks.json. Default to ../data/lexical_category2idx_cutoff_treebanks.json.
--instantiated_unary_rules_path: The relative path to the .json file storing instantiated unary rules. Default to ../data/instantiated_unary_rules.json.
--unary_rules_n: The number of unary rules to use (the first n rules sorted according to frequency in the corpora), default to 20.
--instantiated_binary_rules_path: The relative path to the .json file storing instantiated binary rules, default to ../data/instantiated_binary_rules.json.

  • If you need to build new instantiated unary rules and binary rules from treebanks. Please first go to ./data/utils.py, and follow the script in the main function to generate instantiated_unary_rules_raw.json and instantiated_binary_rules_raw.json. Then go to multiCCG/tools.py, and run the collect_unary_rules and collect_binary_rules functions following the sample use in the main function. The resulting unary rules are sorted according to their frequency in the corpora.

--supertagging_model_dir: The relative path to the directory stoing the PLM of the supertagging model. Default to ../plms/mt5-base.
--supertagging_model_checkpoint_path: The relative path to the designated supertagging checkpoint file.
--predicted_auto_files_dir: The relative path the directory storing predicted .auto files. Default to ./evaluation.
--apply_supertagging_pruning / --no-apply_supertagging_pruning: To control whether to apply the supertagging pruning method, if True, please specify --beta, default to --apply_supertagging_pruning.
--top_k_supertags: The maximum number of supertags allowed for one word, default to 10.
--beta: The coefficient to prune predicted categories whose probabilities lie within $\beta$ of that of the best category, default to 0.0005.
--batch_size: The number of sentences to parse in a batch. Default to be 10.
--possible_roots: Allowed categories at the root of one parse, default to S|NP|S/NP|S\\NP for treebanks. Please specify "S[dcl]|NP|S[wq]|S[q]|S[qem]|S[b]\\NP" for CCGBank.
--device: The device to conduct parsing, default to cuda:0.
--mode: The mode of the parser, choices include sanity_check, predict_sent, batch_sanity_check and predict_batch. If sanity_check, the parser reads data from --sample_data_path (default to ../data/sample.auto) and returns the parsing result with its golden supertags. If predict_sent, the parser reads data from --sample_data_path and returns the parsing result after supertagging. If batch_sanity_check or predict_batch, the parser reads in data designated by batch_data_mode, saves the predicted .auto files using their golden or predicted supertags under ./parsing/evaluation, and the parseval scores under ./parsing/evaluation folder. No default value for --mode and it must be specified.
--batch_data_mode: Used when --mode is specified batch_sanity_check or predict_batch. Choices include treebanks_dev, treebanks_test, ccgbank_dev and ccgbank_test. No default value and it must be specified when used.

Some Example Scripts

  • Sanity checking ../data/sample.auto, with the supertagging model trained from CCGBank.
#!/bin/bash

DATA="ccgbank"
MODEL_NAME="mt5base"
TOPK=10
BETA=0.0005

python -u parser.py \
--sample_data_path ../data/sample.auto \
--lexical_category2idx_path ../data/lexical_category2idx_cutoff_ccgbank.json \
--instantiated_unary_rules_path ../data/instantiated_unary_rules_ccgbank.json \
--instantiated_binary_rules_path ../data/instantiated_binary_rules_ccgbank.json \
--unary_rules_n 20 \
--supertagging_model_dir ../plms/mt5-base \
--supertagging_model_checkpoint_path ../ccg_supertagger/checkpoints_ccgbank/fc_mt5-base_drop0.5_epoch_8.pt \
--predicted_auto_files_dir ./evaluation \
--beta ${BETA} \
--top_k_supertags ${TOPK} \
--batch_size 10 \
--possible_roots "S[dcl]|NP|S[wq]|S[q]|S[qem]|S[b]\\NP" \
--mode sanity_check \
2>&1 | tee -a AStarParsing_${DATA}_${MODEL_NAME}_${TOPK}_${BETA}.log
  • Predicting parses for ../data/sample.auto, with the supertagging model trained from treebanks.
#!/bin/bash

DATA="treebanks"
MODEL_NAME="mt5base"
TOPK=10
BETA=0.0005

python -u parser.py \
--sample_data_path ../data/sample.auto \
--lexical_category2idx_path ../data/lexical_category2idx_cutoff_treebanks.json \
--instantiated_unary_rules_path ../data/instantiated_unary_rules_treebanks.json \
--instantiated_binary_rules_path ../data/instantiated_binary_rules_treebanks.json \
--unary_rules_n 20 \
--supertagging_model_dir ../plms/mt5-base \
--supertagging_model_checkpoint_path ../ccg_supertagger/checkpoints_treebanks/fc_mt5-base_drop0.5_epoch_20.pt \
--predicted_auto_files_dir ./evaluation \
--beta ${BETA} \
--top_k_supertags ${TOPK} \
--batch_size 10 \
--possible_roots "S|NP|S/NP|S\\NP" \
--mode predict_sent \
2>&1 | tee -a AStarParsing_${DATA}_${MODEL_NAME}_${TOPK}_${BETA}.log
  • Batch sanity checking dev data of CCGBank.
#!/bin/bash

DATA="ccgbank"
MODEL_NAME="mt5base"
TOPK=10
BETA=0.0005

python -u parser.py \
 --lexical_category2idx_path ../data/lexical_category2idx_cutoff_ccgbank.json \
 --instantiated_unary_rules_path ../data/instantiated_unary_rules_ccgbank.json \
 --instantiated_binary_rules_path ../data/instantiated_binary_rules_ccgbank.json \
 --unary_rules_n 20 \
 --supertagging_model_dir ../plms/mt5-base \
 --supertagging_model_checkpoint_path ../ccg_supertagger/checkpoints_ccgbank/fc_mt5-base_drop0.5_epoch_8.pt \
 --predicted_auto_files_dir ./evaluation \
 --beta ${BETA} \
 --top_k_supertags ${TOPK} \
 --batch_size 10 \
 --possible_roots "S[dcl]|NP|S[wq]|S[q]|S[qem]|S[b]\\NP" \
 --mode batch_sanity_check \
 --batch_data_mode ccgbank_dev \
 2>&1 | tee -a AStarParsing_${DATA}_${MODEL_NAME}_${TOPK}_${BETA}.log
  • Batch predicting parses for dev data of treebanks.
#!/bin/bash

DATA="treebanks"
MODEL_NAME="mt5base"
TOPK=10
BETA=0.0005

python -u parser.py \
 --lexical_category2idx_path ../data/lexical_category2idx_cutoff_treebanks.json \
 --instantiated_unary_rules_path ../data/instantiated_unary_rules_treebanks.json \
 --instantiated_binary_rules_path ../data/instantiated_binary_rules_treebanks.json \
 --unary_rules_n 20 \
 --supertagging_model_dir ../plms/mt5-base \
 --supertagging_model_checkpoint_path ../ccg_supertagger/checkpoints_treebanks/fc_mt5-base_drop0.5_epoch_20.pt \
 --predicted_auto_files_dir ./evaluation \
 --beta ${BETA} \
 --top_k_supertags ${TOPK} \
 --batch_size 10 \
 --possible_roots "S|NP|S/NP|S\\NP" \
 --mode predict_batch \
 --batch_data_mode treebanks_dev \
 2>&1 | tee -a AStarParsing_${DATA}_${MODEL_NAME}_${TOPK}_${BETA}.log

About

A multilingual CCG Parser

Resources

Readme
Activity
Custom properties

Stars

0 stars

Watchers

2 watching

Forks

1 fork
Report repository

Releases

No releases published

Packages

No packages published

Languages