nlpnet --- Natural Language Processing with neural networks

This is an obsolete version of the toolkit. A fully redesigned version is available as DeepNL at https://github.com/attardi/deepnl

nlpnet is a Python library for Natural Language Processing tasks based on neural networks.

It performs part-of-speech tagging, Named Entity tagging and semantic role labeling.

This branch is rewrite by Giuseppe Attardi (http://www.di.unipi.it/~attardi) to make it consistent with the architecture of SENNA:

The original version by Erick Fonseca is available at: https://github.com/erickrf/nlpnet.

Dependencies

nlpnet requires numpy.

NLTK is also needed to perform Portuguese tokenization and sentence splitting when used as a library (see Library usage).

The POS and NER command line taggers expect instead properly tokenized input.

Cython is required in development for generating C extensions that 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.

Basic usage

nlpnet can be used both as a Python library or through command line scripts. Both usages are explained below.

Library usage

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

>>> import nlpnet
>>> nlpnet.set_data_dir('/path/to/nlpnet-data/')
>>> tagger = nlpnet.POSTagger()
>>> 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 call to set_data_dir indicates where the data directory is located. This location must be given whenever nlpnet is imported.

Calling a tagger is pretty straightforward. The provided taggers are: POSTagger, NERTagger and SRLTagger, all having a method tag which receives strings with text to be tagged. 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 NERTagger is in IOB notation.

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'.

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. The scripts expect tokenized input, one token per line, with an empty line to separate sentences. When training, the token attributes are supplied in tsv format.

$ nlpnet-tag.py pos /path/to/nlpnet-data/
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

Or 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.

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

Benchmarks

The NER tagger replicates the performance of SENNA in the CoNLL 2003 benchmark. The CoNLL-2003 shared task data can be downloaded from http://www.cnts.ua.ac.be/conll2003/ner/.

The train and test data must be cleaned and converted to the more recent IOB2 notation, by calling:

sed '/-DOCSTART-/,+1d' train | bin/toIOB.py | cut -f 1,2,4 > train.iob
sed '/-DOCSTART-/,+1d' testa | bin/toIOB.py | cut -f 1,2,4 > testa.iob
sed '/-DOCSTART-/,+1d' testb | bin/toIOB.py | cut -f 1,2,4 > testb.iob
cat train.iob testa.iob > train+dev.iob

Assuming that the SENNA distribution is in directory senna, the embeddings from SENNA can be adapted for use with nlpnet with the command:

bin/nlpnet-load-embeddings.py --task ner senna senna/embeddings/embeddings.txt -v senna/hash/words.lst -o model

The gazetters from SENNA can be used to produce a single entity list as follows:

iconv -f ISO-8859-1 -t UTF-8 < senna/hash/ner.loc.lst | awk '{printf "LOC\t%s\n", $$0}'  > model/eng.list
iconv -f ISO-8859-1 -t UTF-8 < senna/hash/ner.misc.lst | awk '{printf "MISC\t%s\n", $$0}' >> model/eng.list
iconv -f ISO-8859-1 -t UTF-8 < senna/hash/ner.org.lst | awk '{printf "ORG\t%s\n", $$0}' >> model/eng.list
iconv -f ISO-8859-1 -t UTF-8 < senna/hash/ner.per.lst | awk '{printf "PER\t%s\n", $$0}' >> model/eng.list

You also need the list of suffixes:

cp -p senna/hash/suffix.lst model/suffixes.txt

The tagger can then be trained as follows:

bin/nlpnet-train.py ner --load_features --gazetteer \
     --data model --gold train+dev.iob -e 40 --variant senna \
     -l 0.0001 --lf 0.01 --lt 0.01 -w 5 -n 300 --caps --suffix -v

The benchmark can be run as:

bin/nlpnet-tag.py ner model < testb.iob > testb.out.iob

The results I achieved are:

processed 46435 tokens with 5648 phrases; found: 5640 phrases; correct: 5031. accuracy: 97.62%; precision: 89.20%; recall: 89.08%; FB1: 89.14 LOC: precision: 93.30%; recall: 91.01%; FB1: 92.14 MISC: precision: 78.24%; recall: 77.35%; FB1: 77.79 ORG: precision: 84.59%; recall: 87.24%; FB1: 85.89 PER: precision: 94.71%; recall: 94.06%; FB1: 94.38