def load_models(self):
for i, t in enumerate(self.types):
model = StanfordNERModel(self.basepath + "_" + t,
t,
subtypes=self.basemodel.subtypes)
model.load_tagger(self.baseport + i)
self.models[t] = model
def __init__(self, basepath, baseport=9191, **kwargs):
self.models = {}
self.basepath = basepath
self.corpus = kwargs.get("corpus")
submodels = []
self.baseport = baseport
self.types = []
if basepath.split("/")[-1].startswith("chemdner+ddi"):
self.types = self.DDI_TYPES + self.CHEMDNER_TYPES + [
"chemdner", "ddi"
]
elif basepath.split("/")[-1].startswith("ddi"):
self.types = self.DDI_TYPES + ["all"]
elif basepath.split("/")[-1].startswith("chemdner") or basepath.split(
"/")[-1].startswith("cemp"):
self.types = ["all"] + self.CHEMDNER_TYPES
elif basepath.split("/")[-1].startswith("gpro"):
self.types = self.GPRO_TYPES + ["all"]
self.basemodel = StanfordNERModel(self.basepath, "all")
def train_types(self):
"""
Train models for each subtype of entity, and a general model.
:param types: subtypes of entities to train individual models, as well as a general model
"""
self.basemodel.load_data(self.corpus,
feature_extractors.keys(),
subtype="all")
for t in self.types:
typepath = self.basepath + "_" + t
model = StanfordNERModel(typepath,
subtypes=self.basemodel.subtypes)
model.copy_data(self.basemodel, t)
logging.info("training subtype %s" % t)
model.train()
self.models[t] = model
def main():
start_time = time.time()
parser = argparse.ArgumentParser(description='')
parser.add_argument("actions",
default="classify",
help="Actions to be performed.",
choices=[
"load_corpus", "annotate", "classify",
"write_results", "write_goldstandard", "train",
"test", "train_multiple", "test_multiple",
"train_matcher", "test_matcher", "crossvalidation",
"train_relations", "test_relations"
])
parser.add_argument(
"--goldstd",
default="",
dest="goldstd",
nargs="+",
help="Gold standard to be used. Will override corpus, annotations",
choices=config.paths.keys())
parser.add_argument("--submodels",
default="",
nargs='+',
help="sub types of classifiers"),
parser.add_argument(
"-i",
"--input",
dest="input",
action="store",
default='''Administration of a higher dose of indinavir should be \\
considered when coadministering with megestrol acetate.''',
help="Text to classify.")
parser.add_argument(
"--corpus",
dest="corpus",
nargs=2,
default=[
"chemdner",
"CHEMDNER/CHEMDNER_SAMPLE_JUNE25/chemdner_sample_abstracts.txt"
],
help="format path")
parser.add_argument("--annotations", dest="annotations")
parser.add_argument("--tag",
dest="tag",
default="0",
help="Tag to identify the text.")
parser.add_argument("--models",
dest="models",
help="model destination path, without extension")
parser.add_argument("--entitytype",
dest="etype",
help="type of entities to be considered",
default="all")
parser.add_argument("--pairtype",
dest="ptype",
help="type of pairs to be considered",
default="all")
parser.add_argument("--doctype",
dest="doctype",
help="type of document to be considered",
default="all")
parser.add_argument("--annotated",
action="store_true",
default=False,
dest="annotated",
help="True if the input has <entity> tags.")
parser.add_argument(
"-o",
"--output",
"--format",
dest="output",
nargs=2,
help="format path; output formats: xml, html, tsv, text, chemdner.")
parser.add_argument("--crf",
dest="crf",
help="CRF implementation",
default="stanford",
choices=["stanford", "crfsuite"])
parser.add_argument("--log",
action="store",
dest="loglevel",
default="WARNING",
help="Log level")
parser.add_argument("--kernel",
action="store",
dest="kernel",
default="svmtk",
help="Kernel for relation extraction")
options = parser.parse_args()
# set logger
numeric_level = getattr(logging, options.loglevel.upper(), None)
if not isinstance(numeric_level, int):
raise ValueError('Invalid log level: %s' % options.loglevel)
while len(logging.root.handlers) > 0:
logging.root.removeHandler(logging.root.handlers[-1])
logging_format = '%(asctime)s %(levelname)s %(filename)s:%(lineno)s:%(funcName)s %(message)s'
logging.basicConfig(level=numeric_level,
format=logging_format,
filename="debug.log")
logging.getLogger().setLevel(numeric_level)
logging.getLogger("requests.packages").setLevel(30)
logging.info("Processing action {0} on {1}".format(options.actions,
options.goldstd))
# set configuration variables based on the goldstd option if the corpus has a gold standard,
# or on corpus and annotation options
# pre-processing options
if options.actions == "load_corpus":
if len(options.goldstd) > 1:
print("load only one corpus each time")
sys.exit()
options.goldstd = options.goldstd[0]
corpus_format = config.paths[options.goldstd]["format"]
corpus_path = config.paths[options.goldstd]["text"]
corpus_ann = config.paths[options.goldstd]["annotations"]
corenlp_client = StanfordCoreNLP('http://localhost:9000')
corpus = load_corpus(options.goldstd, corpus_path, corpus_format,
corenlp_client)
corpus.save(config.paths[options.goldstd]["corpus"])
if corpus_ann: #add annotation if it is not a test set
corpus.load_annotations(corpus_ann, options.etype, options.ptype)
corpus.save(config.paths[options.goldstd]["corpus"])
elif options.actions == "annotate": # rext-add annotation to corpus
if len(options.goldstd) > 1:
print("load only one corpus each time")
sys.exit()
options.goldstd = options.goldstd[0]
corpus_path = config.paths[options.goldstd]["corpus"]
corpus_ann = config.paths[options.goldstd]["annotations"]
logging.info("loading corpus %s" % corpus_path)
corpus = pickle.load(open(corpus_path, 'rb'))
logging.debug("loading annotations...")
corpus.clear_annotations(options.etype)
corpus.load_annotations(corpus_ann, options.etype, options.ptype)
# corpus.get_invalid_sentences()
corpus.save(config.paths[options.goldstd]["corpus"])
else:
corpus = Corpus("corpus/" + "&".join(options.goldstd))
for g in options.goldstd:
corpus_path = config.paths[g]["corpus"]
logging.info("loading corpus %s" % corpus_path)
this_corpus = pickle.load(open(corpus_path, 'rb'))
corpus.documents.update(this_corpus.documents)
if options.actions == "write_goldstandard":
model = BiasModel(options.output[1])
model.load_data(corpus, [])
results = model.test()
#results = ResultsNER(options.output[1])
#results.get_ner_results(corpus, model)
results.save(options.output[1] + ".pickle")
#logging.info("saved gold standard results to " + options.output[1] + ".txt")
# training
elif options.actions == "train":
if options.crf == "stanford":
model = StanfordNERModel(options.models, options.etype)
elif options.crf == "crfsuite":
model = CrfSuiteModel(options.models, options.etype)
model.load_data(corpus, feature_extractors.keys(), options.etype)
model.train()
elif options.actions == "train_matcher": # Train a simple classifier based on string matching
model = MatcherModel(options.models)
model.train(corpus)
# TODO: term list option
#model.train("TermList.txt")
elif options.actions == "train_multiple": # Train one classifier for each type of entity in this corpus
# logging.info(corpus.subtypes)
models = TaggerCollection(basepath=options.models,
corpus=corpus,
subtypes=corpus.subtypes)
models.train_types()
elif options.actions == "train_relations":
if options.kernel == "jsre":
model = JSREKernel(corpus, options.ptype)
elif options.kernel == "svmtk":
model = SVMTKernel(corpus, options.ptype)
elif options.kernel == "stanfordre":
model = StanfordRE(corpus, options.ptype)
elif options.kernel == "multir":
model = MultiR(corpus, options.ptype)
elif options.kernel == "scikit":
model = ScikitRE(corpus, options.ptype)
elif options.kernel == "crf":
model = CrfSuiteRE(corpus, options.ptype)
model.train()
# testing
elif options.actions == "test":
base_port = 9191
if len(options.submodels) > 1:
allresults = ResultSetNER(corpus, options.output[1])
for i, submodel in enumerate(options.submodels):
model = StanfordNERModel(options.models + "_" + submodel)
model.load_tagger(base_port + i)
# load data into the model format
model.load_data(corpus,
feature_extractors.keys(),
mode="test")
# run the classifier on the data
results = model.test(corpus, port=base_port + i)
allresults.add_results(results)
model.kill_process()
# save the results to an object that can be read again, and log files to debug
final_results = allresults.combine_results()
else:
if options.crf == "stanford":
model = StanfordNERModel(options.models, options.etype)
elif options.crf == "crfsuite":
model = CrfSuiteModel(options.models, options.etype)
model.load_tagger()
model.load_data(corpus, feature_extractors.keys(), mode="test")
final_results = model.test(corpus)
#with codecs.open(options.output[1] + ".txt", 'w', 'utf-8') as outfile:
# lines = final_results.corpus.write_chemdner_results(options.models, outfile)
#final_results.lines = lines
final_results.save(options.output[1] + ".pickle")
elif options.actions == "test_matcher":
if "mirna" in options.models:
model = MirnaMatcher(options.models)
else:
model = MatcherModel(options.models)
results = ResultsNER(options.models)
results.corpus, results.entities = model.test(corpus)
allentities = set()
for e in results.entities:
allentities.add(results.entities[e].text)
with codecs.open(options.output[1] + ".txt", 'w',
'utf-8') as outfile:
outfile.write('\n'.join(allentities))
results.save(options.output[1] + ".pickle")
elif options.actions == "test_multiple":
logging.info("testing with multiple classifiers... {}".format(
' '.join(options.submodels)))
allresults = ResultSetNER(corpus, options.output[1])
if len(options.submodels) < 2:
models = TaggerCollection(basepath=options.models)
models.load_models()
results = models.test_types(corpus)
final_results = results.combine_results()
else:
base_port = 9191
for submodel in options.submodels:
models = TaggerCollection(basepath=options.models + "_" +
submodel,
baseport=base_port)
models.load_models()
results = models.test_types(corpus)
logging.info("combining results...")
submodel_results = results.combine_results()
allresults.add_results(submodel_results)
base_port += len(models.models)
final_results = allresults.combine_results()
logging.info("saving results...")
final_results.save(options.output[1] + ".pickle")
elif options.actions == "test_relations":
if options.kernel == "jsre":
model = JSREKernel(corpus, options.ptype, train=False)
elif options.kernel == "svmtk":
model = SVMTKernel(corpus, options.ptype)
elif options.kernel == "rules":
model = RuleClassifier(corpus, options.ptype)
elif options.kernel == "stanfordre":
model = StanfordRE(corpus, options.ptype)
elif options.kernel == "scikit":
model = ScikitRE(corpus, options.ptype)
elif options.kernel == "crf":
model = CrfSuiteRE(corpus, options.ptype, test=True)
model.load_classifier()
model.test()
results = model.get_predictions(corpus)
results.save(options.output[1] + ".pickle")
total_time = time.time() - start_time
logging.info("Total time: %ss" % total_time)
def load_models(self):
# Run load_tagger method of all models
for i, a in enumerate(self.entity_annotators.keys()):
self.create_annotationset(a[0])
if a[1] == "stanfordner":
model = StanfordNERModel("annotators/{}/{}".format(a[2], a[0]), a[2])
model.load_tagger(self.baseport + i)
self.entity_annotators[a] = model
elif a[1] == "crfsuite":
model = CrfSuiteModel("annotators/{}/{}".format(a[2], a[0]), a[2])
model.load_tagger(self.baseport + i)
self.entity_annotators[a] = model
elif a[1] == "banner":
model = BANNERModel("annotators/{}/{}".format(a[2], a[0]), a[2])
# model.load_tagger(self.baseport + i)
self.entity_annotators[a] = model
for i, a in enumerate(self.relation_annotators.keys()):
self.create_annotationset(a[0])
if a[1] == "jsre":
model = JSREKernel(None, a[2], train=False, modelname="annotators/{}/{}.model".format(a[2], a[0]), ner="all")
model.load_classifier()
self.relation_annotators[a] = model
elif a[1] == "smil":
model = MILClassifier(None, a[2], relations=[], modelname="{}.model".format(a[0]),
ner="all", generate=False, test=True)
model.basedir = "annotators/{}".format(a[2])
model.load_kb("corpora/transmir/transmir_relations.txt")
model.load_classifier()
self.relation_annotators[a] = model
def run_crossvalidation(goldstd_list,
corpus,
model,
cv,
crf="stanford",
entity_type="all",
cvlog="cv.log"):
logfile = open(cvlog, 'w')
doclist = corpus.documents.keys()
random.shuffle(doclist)
size = int(len(doclist) / cv)
sublists = chunks(doclist, size)
logging.debug("Chunks:")
logging.debug(sublists)
p, r = [], []
all_results = ResultsNER(model)
all_results.path = model + "_results"
for nlist in range(cv):
testids, trainids = None, None
testids = sublists[nlist]
trainids = list(itertools.chain.from_iterable(sublists[:nlist]))
trainids += list(itertools.chain.from_iterable(sublists[nlist + 1:]))
train_corpus, test_corpus = None, None
print 'CV{} - test set: {}; train set: {}'.format(
nlist, len(testids), len(trainids))
train_corpus = Corpus(
corpus.path + "_train",
documents={did: corpus.documents[did]
for did in trainids})
test_corpus = Corpus(
corpus.path + "_test",
documents={did: corpus.documents[did]
for did in testids})
# logging.debug("train corpus docs: {}".format("\n".join(train_corpus.documents.keys())))
#test_entities = len(test_corpus.get_all_entities("goldstandard"))
#train_entities = len(train_corpus.get_all_entities("goldstandard"))
#logging.info("test set entities: {}; train set entities: {}".format(test_entities, train_entities))
basemodel = model + "_cv{}".format(nlist)
logging.debug('CV{} - test set: {}; train set: {}'.format(
nlist, len(test_corpus.documents), len(train_corpus.documents)))
'''for d in train_corpus.documents:
for s in train_corpus.documents[d].sentences:
print len([t.tags.get("goldstandard") for t in s.tokens if t.tags.get("goldstandard") != "other"])
sys.exit()'''
# train
logging.info('CV{} - TRAIN'.format(nlist))
# train_model = StanfordNERModel(basemodel)
train_model = None
if crf == "stanford":
train_model = StanfordNERModel(basemodel, entity_type)
elif crf == "crfsuite":
train_model = CrfSuiteModel(basemodel, entity_type)
train_model.load_data(train_corpus, feature_extractors.keys())
train_model.train()
# test
logging.info('CV{} - TEST'.format(nlist))
test_model = None
if crf == "stanford":
test_model = StanfordNERModel(basemodel, entity_type)
elif crf == "crfsuite":
test_model = CrfSuiteModel(basemodel, entity_type)
test_model.load_tagger(port=9191 + nlist)
test_model.load_data(test_corpus,
feature_extractors.keys(),
mode="test")
final_results = None
final_results = test_model.test(test_corpus, port=9191 + nlist)
if crf == "stanford":
test_model.kill_process()
final_results.basepath = basemodel + "_results"
final_results.path = basemodel
all_results.entities.update(final_results.entities)
all_results.corpus.documents.update(final_results.corpus.documents)
# validate
"""if config.use_chebi:
logging.info('CV{} - VALIDATE'.format(nlist))
final_results = add_chebi_mappings(final_results, basemodel)
final_results = add_ssm_score(final_results, basemodel)
final_results.combine_results(basemodel, basemodel)"""
# evaluate
logging.info('CV{} - EVALUATE'.format(nlist))
test_goldset = set()
for gs in goldstd_list:
goldset = get_gold_ann_set(config.paths[gs]["format"],
config.paths[gs]["annotations"],
entity_type, "pairtype",
config.paths[gs]["text"])
for g in goldset[0]:
if g[0] in testids:
test_goldset.add(g)
precision, recall = get_results(final_results, basemodel, test_goldset,
{}, [])
# evaluation = run_chemdner_evaluation(config.paths[goldstd]["cem"], basemodel + "_results.txt", "-t")
# values = evaluation.split("\n")[1].split('\t')
p.append(precision)
r.append(recall)
# logging.info("precision: {} recall:{}".format(str(values[13]), str(values[14])))
pavg = sum(p) / cv
ravg = sum(r) / cv
print "precision: average={} all={}".format(
str(pavg), '|'.join([str(pp) for pp in p]))
print "recall: average={} all={}".format(str(ravg),
'|'.join([str(rr) for rr in r]))
all_goldset = set()
for gs in goldstd_list:
goldset = get_gold_ann_set(config.paths[gs]["format"],
config.paths[gs]["annotations"],
entity_type, config.paths[gs]["text"])
for g in goldset:
all_goldset.add(g)
get_results(all_results, model, all_goldset, {}, [])
class TaggerCollection(object):
"""
Collection of tagger classifiers used to train and test specific subtype models
"""
CHEMDNER_TYPES = [
"IDENTIFIER", "MULTIPLE", "FAMILY", "FORMULA", "SYSTEMATIC",
"ABBREVIATION", "TRIVIAL"
]
GPRO_TYPES = ["NESTED", "IDENTIFIER", "FULL_NAME", "ABBREVIATION"]
DDI_TYPES = ["drug", "group", "brand", "drug_n"]
def __init__(self, basepath, baseport=9191, **kwargs):
self.models = {}
self.basepath = basepath
self.corpus = kwargs.get("corpus")
submodels = []
self.baseport = baseport
self.types = []
if basepath.split("/")[-1].startswith("chemdner+ddi"):
self.types = self.DDI_TYPES + self.CHEMDNER_TYPES + [
"chemdner", "ddi"
]
elif basepath.split("/")[-1].startswith("ddi"):
self.types = self.DDI_TYPES + ["all"]
elif basepath.split("/")[-1].startswith("chemdner") or basepath.split(
"/")[-1].startswith("cemp"):
self.types = ["all"] + self.CHEMDNER_TYPES
elif basepath.split("/")[-1].startswith("gpro"):
self.types = self.GPRO_TYPES + ["all"]
self.basemodel = StanfordNERModel(self.basepath, "all")
def train_types(self):
"""
Train models for each subtype of entity, and a general model.
:param types: subtypes of entities to train individual models, as well as a general model
"""
self.basemodel.load_data(self.corpus,
feature_extractors.keys(),
subtype="all")
for t in self.types:
typepath = self.basepath + "_" + t
model = StanfordNERModel(typepath,
subtypes=self.basemodel.subtypes)
model.copy_data(self.basemodel, t)
logging.info("training subtype %s" % t)
model.train()
self.models[t] = model
def load_models(self):
for i, t in enumerate(self.types):
model = StanfordNERModel(self.basepath + "_" + t,
t,
subtypes=self.basemodel.subtypes)
model.load_tagger(self.baseport + i)
self.models[t] = model
def process_type(self, modelst, t, corpus, basemodel, basepath, port):
# load data only for one model since this takes at least 5 minutes each time
logging.debug("{}: copying data...".format(t))
modelst.copy_data(basemodel)
#logging.debug("pre test %s" % model)
logging.debug("{}: testing...".format(t))
res = modelst.test(corpus, port)
logging.info("{}:done...".format(t))
return res
def test_types(self, corpus):
"""
Classify the corpus with multiple classifiers from different subtypes
:return ResultSetNER object with the results obtained for the models
"""
# TODO: parallelize
results = ResultSetNER(corpus, self.basepath)
self.basemodel.load_data(corpus, feature_extractors.keys())
all_results = []
tasks = [(self.models[t], t, corpus, self.basemodel, self.basepath,
self.baseport + i) for i, t in enumerate(self.types)]
all_results = []
for t in tasks:
r = self.process_type(*t)
all_results.append(r)
logging.info("adding results...")
for res, i in enumerate(all_results):
#logging.debug("adding these results: {}".format(self.types[i]))
results.add_results(res)
return results
def run_crossvalidation(goldstd_list, corpus, model, cv, crf="stanford", entity_type="all", cvlog="cv.log"):
logfile = open(cvlog, 'w')
doclist = corpus.documents.keys()
random.shuffle(doclist)
size = int(len(doclist)/cv)
sublists = chunks(doclist, size)
logging.debug("Chunks:")
logging.debug(sublists)
p, r = [], []
all_results = ResultsNER(model)
all_results.path = model + "_results"
for nlist in range(cv):
testids, trainids = None, None
testids = sublists[nlist]
trainids = list(itertools.chain.from_iterable(sublists[:nlist]))
trainids += list(itertools.chain.from_iterable(sublists[nlist+1:]))
train_corpus, test_corpus = None, None
print 'CV{} - test set: {}; train set: {}'.format(nlist, len(testids), len(trainids))
train_corpus = Corpus(corpus.path + "_train", documents={did: corpus.documents[did] for did in trainids})
test_corpus = Corpus(corpus.path + "_test", documents={did: corpus.documents[did] for did in testids})
# logging.debug("train corpus docs: {}".format("\n".join(train_corpus.documents.keys())))
#test_entities = len(test_corpus.get_all_entities("goldstandard"))
#train_entities = len(train_corpus.get_all_entities("goldstandard"))
#logging.info("test set entities: {}; train set entities: {}".format(test_entities, train_entities))
basemodel = model + "_cv{}".format(nlist)
logging.debug('CV{} - test set: {}; train set: {}'.format(nlist, len(test_corpus.documents), len(train_corpus.documents)))
'''for d in train_corpus.documents:
for s in train_corpus.documents[d].sentences:
print len([t.tags.get("goldstandard") for t in s.tokens if t.tags.get("goldstandard") != "other"])
sys.exit()'''
# train
logging.info('CV{} - TRAIN'.format(nlist))
# train_model = StanfordNERModel(basemodel)
train_model = None
if crf == "stanford":
train_model = StanfordNERModel(basemodel, entity_type)
elif crf == "crfsuite":
train_model = CrfSuiteModel(basemodel, entity_type)
train_model.load_data(train_corpus, feature_extractors.keys())
train_model.train()
# test
logging.info('CV{} - TEST'.format(nlist))
test_model = None
if crf == "stanford":
test_model = StanfordNERModel(basemodel, entity_type)
elif crf == "crfsuite":
test_model = CrfSuiteModel(basemodel, entity_type)
test_model.load_tagger(port=9191+nlist)
test_model.load_data(test_corpus, feature_extractors.keys(), mode="test")
final_results = None
final_results = test_model.test(test_corpus, port=9191+nlist)
if crf == "stanford":
test_model.kill_process()
final_results.basepath = basemodel + "_results"
final_results.path = basemodel
all_results.entities.update(final_results.entities)
all_results.corpus.documents.update(final_results.corpus.documents)
# validate
"""if config.use_chebi:
logging.info('CV{} - VALIDATE'.format(nlist))
final_results = add_chebi_mappings(final_results, basemodel)
final_results = add_ssm_score(final_results, basemodel)
final_results.combine_results(basemodel, basemodel)"""
# evaluate
logging.info('CV{} - EVALUATE'.format(nlist))
test_goldset = set()
for gs in goldstd_list:
goldset = get_gold_ann_set(config.corpus_paths.paths[gs]["format"], config.corpus_paths.paths[gs]["annotations"], entity_type,
"pairtype", config.corpus_paths.paths[gs]["text"])
for g in goldset[0]:
if g[0] in testids:
test_goldset.add(g)
precision, recall = get_results(final_results, basemodel, test_goldset, {}, [])
# evaluation = run_chemdner_evaluation(config.paths[goldstd]["cem"], basemodel + "_results.txt", "-t")
# values = evaluation.split("\n")[1].split('\t')
p.append(precision)
r.append(recall)
# logging.info("precision: {} recall:{}".format(str(values[13]), str(values[14])))
pavg = sum(p)/cv
ravg = sum(r)/cv
print "precision: average={} all={}".format(str(pavg), '|'.join([str(pp) for pp in p]))
print "recall: average={} all={}".format(str(ravg), '|'.join([str(rr) for rr in r]))
all_goldset = set()
for gs in goldstd_list:
goldset = get_gold_ann_set(config.corpus_paths.paths[gs]["format"], config.corpus_paths.paths[gs]["annotations"], entity_type, "",
config.corpus_paths.paths[gs]["text"])
for g in goldset[0]:
all_goldset.add(g)
get_results(all_results, model, all_goldset, {}, [])
