erickrf / nlpnet

A neural network architecture for NLP tasks, using cython for fast performance. Currently, it can perform POS tagging, SRL and dependency parsing.

Gitter is chat room for developers.

nlpnet --- Natural Language Processing with neural networks

nlpnet is a Python library for Natural Language Processing tasks based on neural networks. Currently, it performs part-of-speech tagging, semantic role labeling and dependency parsing. Most of the architecture is language independent, but some functions were specially tailored for working with Portuguese. This system was inspired by SENNA.

Important: in order to use the trained models for Portuguese NLP, you will need to download the data from http://nilc.icmc.usp.br/nlpnet/models.html.

Dependencies

nlpnet requires NLTK and numpy. Additionally, it needs to download some data from NLTK. After installing it, call

>>> nltk.download()

go to the Models tab and select the Punkt tokenizer. It is used in order to split the text into sentences.

Cython is used to generate C extensions and run faster. You probably won't need it, since the generated .c file is already provided with nlpnet, but you will need a C compiler. On Linux and Mac systems this shouldn't be a problem, but may be on Windows, because setuptools requires the Microsoft C Compiler by default. If you don't have it already, it is usually easier to install MinGW instead and follow the instructions here.

Basic usage

nlpnet can be used both as a Python library or by its standalone scripts. Both usages are explained below.

Library usage

You can use nlpnet as a library in Python code as follows:

>>> import nlpnet
>>> tagger = nlpnet.POSTagger('/path/to/pos-model/', language='pt')
>>> tagger.tag('O rato roeu a roupa do rei de Roma.')
[[(u'O', u'ART'), (u'rato', u'N'), (u'roeu', u'V'), (u'a', u'ART'), (u'roupa', u'N'), (u'do', u'PREP+ART'), (u'rei', u'N'), (u'de', u'PREP'), (u'Roma', u'NPROP'), (u'.', 'PU')]]

In the example above, the POSTagger constructor receives as the first argument the directory where its trained model is located. The second argument is the two letter language code (currently, onle pt and en are supported). This only has impact in tokenization.

Calling an annotation tool is pretty straightforward. The provided ones are POSTagger, SRLTagger and DependencyParser, all of them having a method tag which receives strings with text to be tagged (in DependencyParser, there is an alias to the method parse, which sounds more adequate). The tagger splits the text into sentences and then tokenizes each one (hence the return of the POSTagger is a list of lists).

The output of the SRLTagger is slightly more complicated:

>>> tagger = nlpnet.SRLTagger()
>>> tagger.tag(u'O rato roeu a roupa do rei de Roma.')
[<nlpnet.taggers.SRLAnnotatedSentence at 0x84020f0>]

Instead of a list of tuples, sentences are represented by instances of SRLAnnotatedSentence. This class serves basically as a data holder, and has two attributes:

>>> sent = tagger.tag(u'O rato roeu a roupa do rei de Roma.')[0]
>>> sent.tokens
[u'O', u'rato', u'roeu', u'a', u'roupa', u'do', u'rei', u'de', u'Roma', u'.']
>>> sent.arg_structures
[(u'roeu',
  {u'A0': [u'O', u'rato'],
   u'A1': [u'a', u'roupa', u'do', u'rei', u'de', u'Roma'],
   u'V': [u'roeu']})]

The arg_structures is a list containing all predicate-argument structures in the sentence. The only one in this example is for the verb roeu. It is represented by a tuple with the predicate and a dictionary mapping semantic role labels to the tokens that constitute the argument.

Note that the verb appears as the first member of the tuple and also as the content of label 'V' (which stands for verb). This is because some predicates are multiwords. In these cases, the "main" predicate word (usually the verb itself) appears in arg_structures[0], and all the words appear under the key 'V'.

Here's an example with the DependencyParser:

>>> parser = nlpnet.DependencyParser('dependency', language='en')
>>> parsed_text = parser.parse('The book is on the table.')
>>> parsed_text
[<nlpnet.taggers.ParsedSentence at 0x10e067f0>]
>>> sent = parsed_text[0]
>>> print(sent.to_conll())
1       The     _       DT      DT      _       2       NMOD
2       book    _       NN      NN      _       3       SBJ
3       is      _       VBZ     VBZ     _       0       ROOT
4       on      _       IN      IN      _       3       LOC-PRD
5       the     _       DT      DT      _       6       NMOD
6       table   _       NN      NN      _       4       PMOD
7       .       _       .       .       _       3       P

The to_conll() method of ParsedSentence objects prints them in the CoNLL notation. The tokens, labels and head indices are accessible through member variables:

>>> sent.tokens
[u'The', u'book', u'is', u'on', u'the', u'table', u'.']
>>> sent.heads
array([ 1,  2, -1,  2,  5,  3,  2])
>>> sent.labels
[u'NMOD', u'SBJ', u'ROOT', u'LOC-PRD', u'NMOD', u'PMOD', u'P']

The heads member variable is a numpy array. The i-th position in the array contains the index of the head of the i-th token, except for the root token, which has a head of -1. Notice that these indices are 0-based, while the ones shown in the to_conll() function are 1-based.

Standalone scripts

nlpnet also provides scripts for tagging text, training new models and testing them. They are copied to the scripts subdirectory of your Python installation, which can be included in the system PATH variable. You can call them from command line and give some text input.

$ nlpnet-tag.py pos --data /path/to/nlpnet-data/ --lang pt
O rato roeu a roupa do rei de Roma.
O_ART rato_N roeu_V a_ART roupa_N do_PREP+ART rei_N de_PREP Roma_NPROP ._PU

If --data is not given, the script will search for the trained models in the current directory. --lang defaults to en. If you have text already tokenized, you may use the -t option; it assumes tokens are separated by whitespaces.

With semantic role labeling:

$ nlpnet-tag.py srl /path/to/nlpnet-data/
O rato roeu a roupa do rei de Roma.
O rato roeu a roupa do rei de Roma .
roeu
    A1: a roupa do rei de Roma
    A0: O rato
    V: roeu

The first line was typed by the user, and the second one is the result of tokenization.

And dependency parsing:

$ nlpnet-tag.py dependency --data dependency --lang en
The book is on the table.
1       The     _       DT      DT      _       2       NMOD
2       book    _       NN      NN      _       3       SBJ
3       is      _       VBZ     VBZ     _       0       ROOT
4       on      _       IN      IN      _       3       LOC-PRD
5       the     _       DT      DT      _       6       NMOD
6       table   _       NN      NN      _       4       PMOD
7       .       _       .       .       _       3       P

To learn more about training and testing new models, and other functionalities, refer to the documentation at http://nilc.icmc.usp.br/nlpnet

References

The following references describe the design of nlpnet, as well as experiments carried out. Some improvements to the code have been implemented since their publication.

• Fonseca, Erick and Aluísio, Sandra M. A Deep Architecture for Non-Projective Dependency Parsing. Proceedings of the NAACL-HLT Workshop on Vector Space Modeling for NLP. 2015
• Fonseca, Erick and Rosa, João Luís G. A Two-Step Convolutional Neural Network Approach for Semantic Role Labeling. Proceedings of the International Joint Conference on Neural Networks. 2013.
• Fonseca, Erick, Rosa, João Luís G. and Aluísio, Sandra M. Evaluating word embeddings and a revised corpus for part-of-speech tagging in Portuguese. Journal of The Brazilian Computer Society. 2015.