A Python wrapper for the Java Stanford Core NLP tools
This is a Wordseer-specific fork of Dustin Smith's stanford-corenlp-python, a Python interface to Stanford CoreNLP. It can either use as python package, or run as a JSON-RPC server.
Edited
- Tested only with the current annotator configuration: not a general-purpose wrapper
- Update to Stanford CoreNLP v3.5.2
- Added multi-threaded load balancing
- Fix many bugs & improve performance
- Using jsonrpclib for stability and performance
- Can edit the constants as argument such as Stanford Core NLP directory
- Adjust parameters not to timeout in high load
- Fix a problem with long text input by Johannes Castner stanford-corenlp-python
- Packaging
Requirements
- pexpect
- unidecode
- jsonrpclib (optionally)
Download and Usage
To use this program you must download and unpack the zip file containing Stanford's CoreNLP package. By default, corenlp.py looks for the Stanford Core NLP folder as a subdirectory of where the script is being run.
Then, to launch a server:
python corenlp/corenlp.py
Optionally, you can specify a host or port:
python corenlp/corenlp.py -H 0.0.0.0 -p 3456
For additional concurrency, you can add a load-balancing layer on top:
python corenlp/corenlp.py --ports=8081,8082,8083,8084
That will run a public JSON-RPC server on port 3456. And you can specify Stanford CoreNLP directory:
python corenlp/corenlp.py -S stanford-corenlp-full-2013-06-20/
Assuming you are running on port 8080 and CoreNLP directory is stanford-corenlp-full-2013-06-20/ in current directory, the code in client.py shows an example parse:
import jsonrpclib
from simplejson import loads
server = jsonrpclib.Server("http://localhost:8080")
result = loads(server.parse("Hello world. It is so beautiful"))
print "Result", result
If you are using the load balancing component, then you can use the following approach:
import jsonrpclib
from simplejson import loads
server = jsonrpclib.Server("http://localhost:8080")
result = loads(server.send("Hello world. It is so beautiful"))
print "Result", server.getForKey(result['key'])
# asynchronous parsing and retrieval
sents = [ 'add in as many sentences as you want', 'your mileage may vary' ]
for sent in sents:
server.send(sent)
# this approach is non-blocking
print server.getCompleted()
# this approach waits for all in-progress parses to finish (i.e. blocks)
print server.getAll()
That returns a dictionary containing the keys sentences and (when applicable) corefs. The key sentences contains a list of dictionaries for each sentence, which contain parsetree, text, tuples containing the dependencies, and words, containing information about parts of speech, NER, etc:
{u'sentences': [{u'parsetree': u'(ROOT (S (VP (NP (INTJ (UH Hello)) (NP (NN world)))) (. !)))',
u'text': u'Hello world!',
u'tuples': [[u'dep', u'world', u'Hello'],
[u'root', u'ROOT', u'world']],
u'words': [[u'Hello',
{u'CharacterOffsetBegin': u'0',
u'CharacterOffsetEnd': u'5',
u'Lemma': u'hello',
u'NamedEntityTag': u'O',
u'PartOfSpeech': u'UH'}],
[u'world',
{u'CharacterOffsetBegin': u'6',
u'CharacterOffsetEnd': u'11',
u'Lemma': u'world',
u'NamedEntityTag': u'O',
u'PartOfSpeech': u'NN'}],
[u'!',
{u'CharacterOffsetBegin': u'11',
u'CharacterOffsetEnd': u'12',
u'Lemma': u'!',
u'NamedEntityTag': u'O',
u'PartOfSpeech': u'.'}]]},
{u'parsetree': u'(ROOT (S (NP (PRP It)) (VP (VBZ is) (ADJP (RB so) (JJ beautiful))) (. .)))',
u'text': u'It is so beautiful.',
u'tuples': [[u'nsubj', u'beautiful', u'It'],
[u'cop', u'beautiful', u'is'],
[u'advmod', u'beautiful', u'so'],
[u'root', u'ROOT', u'beautiful']],
u'words': [[u'It',
{u'CharacterOffsetBegin': u'14',
u'CharacterOffsetEnd': u'16',
u'Lemma': u'it',
u'NamedEntityTag': u'O',
u'PartOfSpeech': u'PRP'}],
[u'is',
{u'CharacterOffsetBegin': u'17',
u'CharacterOffsetEnd': u'19',
u'Lemma': u'be',
u'NamedEntityTag': u'O',
u'PartOfSpeech': u'VBZ'}],
[u'so',
{u'CharacterOffsetBegin': u'20',
u'CharacterOffsetEnd': u'22',
u'Lemma': u'so',
u'NamedEntityTag': u'O',
u'PartOfSpeech': u'RB'}],
[u'beautiful',
{u'CharacterOffsetBegin': u'23',
u'CharacterOffsetEnd': u'32',
u'Lemma': u'beautiful',
u'NamedEntityTag': u'O',
u'PartOfSpeech': u'JJ'}],
[u'.',
{u'CharacterOffsetBegin': u'32',
u'CharacterOffsetEnd': u'33',
u'Lemma': u'.',
u'NamedEntityTag': u'O',
u'PartOfSpeech': u'.'}]]}],
u'coref': [[[[u'It', 1, 0, 0, 1], [u'Hello world', 0, 1, 0, 2]]]]}
Not to use JSON-RPC, load the module instead:
from corenlp import StanfordCoreNLP
corenlp_dir = "stanford-corenlp-full-2013-06-20/"
corenlp = StanfordCoreNLP(corenlp_dir) # wait a few minutes...
corenlp.raw_parse("Parse it")
If you need to parse long texts (more than 30-50 sentences), you must use a batch_parse function. It reads text files from input directory and returns a generator object of dictionaries parsed each file results:
from corenlp import batch_parse
corenlp_dir = "stanford-corenlp-full-2013-06-20/"
raw_text_directory = "sample_raw_text/"
parsed = batch_parse(raw_text_directory, corenlp_dir) # It returns a generator object
print parsed #=> [{'coref': ..., 'sentences': ..., 'file_name': 'new_sample.txt'}]
The function uses XML output feature of Stanford CoreNLP, and you can take all information by raw_output option. If true, CoreNLP's XML is returned as a dictionary without converting the format.
parsed = batch_parse(raw_text_directory, corenlp_dir, raw_output=True)
(note: The function requires xmltodict now, you should install it by sudo pip install xmltodict)
Developers
- Hiroyoshi Komatsu [hiroyoshi.komat@gmail.com]
- Johannes Castner [jac2130@columbia.edu]
- Robert Elwell [robert@wikia-inc.com]
- Tristan Chong [tristan@wikia-inc.com]
- Aditi Muralidharan [aditi.shrikumar@gmail.com]
- Ian MacFarland [ianmacfarland@ischool.berkeley.edu]
