diaparser

Direct Attentive Dependency Parser


License
MIT
Install
pip install diaparser==1.0.0

Documentation

DiaParser

DiaParser

build docs release downloads LICENSE

DiaParser is a state-of-the-art dependency parser, that extends the architecture of the Biaffine Parser (Dozat and Manning, 2017) by exploiting both embeddings and attentions provided by transformers.

By exploiting the rich hidden linguistic information in contextual embeddings from transformers, DiaParser can avoid using intermediate annotations like POS, lemma and morphology, used in traditional parsers. Therefore the only stages in the parsing pipeline are tokenization and parsing.

The parser may also work directly on plain text. The parser automatically dowloads pretrained models as well as tokenizers and produces dependency parsing trees, as detailed in Usage.

Exploiting attentions from transformer heads provides improvements in accuracy, without resorting to fine tuning or training its own attention. Overall, this simplifies the architecture and lowers the cost of resources needed during training, especially memory, and allows the parser to improve as new versions of transformers become available. The parser uses the HuggingFace Transformers API and in particular the generic AutoClasses interface to access the transformer models avaiable.

We plan to track the improvements in transformer technology and to realease models of the parser that incorporate these models. Currently we provide pretrained models for 18 languages.

We encourage anyone to contribute trained models for other combinatiobns of language/transformer pairs, that we will publish. We will soon provide a web form where to upload new models.

You can also train your own models and contribute them to the repository, to share with others.

DiaParser uses pretrained contextual embeddings for representing input from models in transformers.

Pretrained tokenizers are provided by Stanza.

Alternatively to contextual embeddings, DiaParser also allows to utilize CharLSTM layers to produce character/subword-level features. Both BERT and CharLSTM avoid the need of generating POS tags.

DiaParser is derived from SuPar, which provides additional variants of dependency and constituency parsers.

Contents

Installation

DiaParser can be installed via pip:

$ pip install -U diaparser

Installing from sources is also possible:

$ git clone https://github.com/Unipisa/diaparser && cd diaparser
$ python setup.py install

The package has the following requirements:

Performance

DiaParser provides pretrained models for English, Chinese and other 21 languages from the Universal Dependencies treebanks v2.6. English models are trained on the Penn Treebank (PTB) with Stanford Dependencies, with 39,832 training sentences, while Chinese models are trained on Penn Chinese Treebank version 7 (CTB7) with 46,572 training sentences.

The accuracy and parsing speed of these models are listed in the following tables. The first table shows the result of parsing starting from gold tokenized text. Notably, punctuation is ignored in the evaluation metrics for PTB, but included in all the others. The numbers in bold represent state-of-the-art values.

Language Corpus Name UAS LAS Speed (Sents/s)
English PTB en_ptb.electra 96.03 94.37 352
Chinese CTB zh_ptb.hfl 92.14 85.74 319
Catalan AnCora ca_ancora.mbert 95.55 93.78 249
German HDT de_htd.dbmdz-bert-base 97.97 96.97 184
Japanese GSD ja_gsd.mbert 95.41 93.98 397
Latin ITTB, LLCT la_ittb_llct.mbert 94.03 91.70 139
Norwegian Nynorsk no_nynorsk.mbert 92.50 90.13 185
Romanian RRT ro_rrt.mbert 93.03 87.18 286
Spanish AnCora es_ancora.mbert 96.03 94.37 352
Turkish Boun tr_boun.electra-base 83.53 75.87 1198

Below are the results on the dataset of the IWPT 2020 Shared Task on Enhanced Dependencies, where the tokenization was done by the parser itself:

Language Corpus Name UAS LAS Speed (Sents/s)
Arabic PADT ar_padt.bert 87.75 83.25 99
Bulgarian BTB bg_btb.DeepPavlov 95.02 92.20 479
Czech PDT, CAC, FicTree cs_pdt_cac_fictree.DeepPavlov 94.02 92.06 403
English EWT en_ewt.electra 91.66 89.51 397
Estonian EDT, EWT et_edt_ewt.mbert 86.39 82.44 247
Finnish TDT fi_tdt.turkunlp 94.28 92.56 364
French sequoia fr_sequoia.camembert 92.81 89.55 200
German HDT de_hdt.dbmdz-bert-base 97.97 96.97 381
Italian ISDT it_isdt.dbmdz-electra-xxl 95.48 94.16 379
Latvian LVBT lv_lvtb.mbert 87.46 83.51 290
Lithuanian ALKSNIS lt_alksnis.mbert 80.09 75.14 290
Dutch Alpino, LassySmall nl_alpino_lassysmall.wietsedv 90.80 88.34 367
Polish PDB, LFG pl_pdb_lfg.dkleczek 94.38 91.70 563
Russian SynTagRus ru_syntagrus.DeepPavlov 94.97 93.72 445
Slovak SNK sk_snk.mbert 93.11 90.44 381
Swediskh Talbanken sv_talbanken.KB 90.79 88.08 491
Tamil TTB ta_ttb.mbert 74.20 66.49 175
Ukrainian IU uk_iu.TurkuNLP 90.39 87.61 301

These results were obtained on a server with Intel(R) Xeon(R) Gold 6132 CPU @ 2.60GHz and Nvidia T4 GPU.

Usage

DiaParser is simple to use: you can just download a pretrained model and run syntactic parsing over sentences with a few lines of code:

>>> from diaparser.parsers import Parser
>>> parser = Parser.load('en_ewt-electra')
>>> dataset = parser.predict([['She', 'enjoys', 'playing', 'tennis', '.']], prob=True)

The call to parser.predict will return an instance of diaparser.utils.Dataset containing the predicted syntactic trees. You can access each sentence within the dataset:

>>> print(dataset.sentences[0])
1       She     _       _       _       _       2       nsubj   _       _
2       enjoys  _       _       _       _       0       root    _       _
3       playing _       _       _       _       2       xcomp   _       _
4       tennis  _       _       _       _       3       dobj    _       _
5       .       _       _       _       _       2       punct   _       _

To parse plain text just requires specifying the language code:

>>> dataset = parser.predict('She enjoys playing tennis.', text='en')

An SVG picture illusrating the parse tree can be produced with:

>>> sent = dataset.sentences[0]
>>> displacy.render(sent.to_displacy(), style='dep', manual=True, options={'compact': True, 'distance': 120})

parse tree

The input can be provided in a file in CoNLL-U format.

Further examples of how to use the parser and experiment with it can be found in this notebook.

Training

To train a model from scratch, it is preferred to use the command-line option, which is more flexible and customizable. Here are some training examples:

# Biaffine Dependency Parser
# some common and default arguments are stored in config.ini
$ python -m diaparser.cmds.biaffine_dependency train -b -d 0  \
    -c config.ini  \
    -p exp/en_ptb.char/model  \
    -f char
# to use BERT, `-f` and `--bert` (default to bert-base-cased) should be specified
$ python -m diaparser.cmds.biaffine_dependency train -b -d 0  \
    -p exp/en_ptb.bert-base/model  \
    -f bert  \
    --bert bert-base-cased

Warning. There is currently a limit of 500 to the length of tokenized sentences, due to the maximum size of embeddings in most pretrained trnsformer models.

For further instructions on training, please type python -m diaparser.cmds.<parser> train -h.

Alternatively, DiaParser provides an equivalent command entry points registered in setup.py: diaparser.

$ diaparser train -b -d 0 -c config.ini -p exp/en_ptb.electra-base/model -f bert --bert google/electra-base-discriminator

For handling large models, distributed training is also supported:

$ python -m torch.distributed.launch --nproc_per_node=4 --master_port=10000  \
    -m parser.cmds.biaffine_dependency train -b -d 0,1,2,3  \
    -p exp/en_ptb.electra-base/model  \
    -f bert --bert google/electra-base-discriminator

You may consult the PyTorch documentation and tutorials for more details.

Evaluation

The evaluation process resembles prediction:

>>> parser = Parser.load('biaffine-dep-en')
>>> loss, metric = parser.evaluate('data/ptb/test.conllx')
2020-07-25 20:59:17 INFO Loading the data
2020-07-25 20:59:19 INFO
Dataset(n_sentences=2416, n_batches=11, n_buckets=8)
2020-07-25 20:59:19 INFO Evaluating the dataset
2020-07-25 20:59:20 INFO loss: 0.2326 - UCM: 61.34% LCM: 50.21% UAS: 96.03% LAS: 94.37%
2020-07-25 20:59:20 INFO 0:00:01.253601s elapsed, 1927.25 Sents/s

TODO

  • Provide a repository where to upload models, like HuggingFace.

References