class GeniaA2WriterTest(object):
'''
classdocs
'''
source = "E:/corpus/bionlp2011/project_data/"
def __init__(self):
'''
Constructor
'''
'''
Constructor
'''
out_path = "E:/corpus/bionlp2011/project_test/result/model1"
WD = WordDictionary(self.source)
WD.load("train")
TD = TriggerDictionary(self.source)
TD.load("train")
self.builder = DocumentBuilder(self.source, WD, TD)
self.a2writter = GeniaA2Writer(out_path)
def test1(self):
doc_id = 'PMC-2222968-04-Results-03'
o_doc = self.builder.build(doc_id)
self.a2writter.write(o_doc)
def test2(self):
dir_name_eval = "test-model-002-cause"
doc_ids = ['PMC-2222968-04-Results-03']
dict_type = 'train'
prediction = Prediction(self.source, dir_name_eval, dict_type)
prediction.predict(doc_ids)
o_doc = prediction.docs[doc_ids[0]]
for sen in o_doc.sen:
sen.test()
self.a2writter.write(o_doc)
def test3(self):
dir_name_eval = "test-model-013"
doc_ids = ['PMID-1763325']
dict_type = 'train'
prediction = Prediction(self.source, dir_name_eval, dict_type)
prediction.predict(doc_ids, write_result=False)
o_doc = prediction.docs[doc_ids[0]]
for sen in o_doc.sen:
sen.test()
self.a2writter.write(o_doc)
class GeniaA2WriterTest(object):
'''
classdocs
'''
source = "E:/corpus/bionlp2011/project_data/"
def __init__(self):
'''
Constructor
'''
'''
Constructor
'''
out_path = "E:/corpus/bionlp2011/project_test/result/model1"
WD = WordDictionary(self.source)
WD.load("train")
TD = TriggerDictionary(self.source)
TD.load("train")
self.builder = DocumentBuilder(self.source, WD, TD)
self.a2writter = GeniaA2Writer(out_path)
def test1(self):
doc_id = 'PMC-2222968-04-Results-03'
o_doc = self.builder.build(doc_id)
self.a2writter.write(o_doc)
def test2(self):
dir_name_eval = "test-model-002-cause"
doc_ids = ['PMC-2222968-04-Results-03']
dict_type = 'train'
prediction = Prediction(self.source, dir_name_eval, dict_type)
prediction.predict(doc_ids)
o_doc = prediction.docs[doc_ids[0]]
for sen in o_doc.sen:
sen.test()
self.a2writter.write(o_doc)
def test3(self):
dir_name_eval = "test-model-013"
doc_ids = ['PMID-1763325']
dict_type = 'train'
prediction = Prediction(self.source, dir_name_eval, dict_type)
prediction.predict(doc_ids, write_result=False)
o_doc = prediction.docs[doc_ids[0]]
for sen in o_doc.sen:
sen.test()
self.a2writter.write(o_doc)
def build_dict(self, corpus_dir):
"""
add trigger word to dictionary, including multi-word trigger
ex.
triggers: "Negative Regulator", "Expression"
store in dict:
"Negative Regulator" <== bigram version
"Negative" <== head of bigram
"Expression" <== standard unigram
"""
if corpus_dir not in self.CORPUS_DIR:
raise ValueError("wrong value. choose 'dev', 'train', or 'mix'")
# init document builder
doc_builder = DocumentBuilder(self.src)
# init default dict with counter
td = defaultdict(Counter)
# get list of document
doc_ids = self.get_doc_ids(corpus_dir)
for doc_id in doc_ids:
o_doc = doc_builder.build(doc_id)
for i in range(0, len(o_doc.sen)):
o_sen = o_doc.sen[i]
for twn in o_sen.trigger:
w = o_sen.words[twn]
ttype = w['type']
string = w['string'].lower()
stem = w['stem'].lower()
# skip short word
if len(string) < 4: continue
# adding unigram text and stem to dictionary
td[string][ttype] += 1
td[stem][ttype] += 1
# check full text
# add to dict if it's multi-word
full_string = o_sen.trigger_text[twn]
if ' ' in full_string:
td[full_string][ttype] += 1
print "the dictionary contains:", len(td), "trigger words"
return dict(td)
class DocumentBuilderTest(object):
'''
classdocs
'''
def __init__(self):
'''
Constructor
'''
self.source = "E:/corpus/bionlp2011/project_data/"
WD = WordDictionary(self.source)
WD.load("train")
TD = TriggerDictionary(self.source)
TD.load("train")
self.builder = DocumentBuilder(self.source, WD, TD)
def run(self):
self.test1()
self.test2()
self.test3()
def test1(self):
doc_id = "PMID-2160380"
o_doc = self.builder.build(doc_id, is_test = False)
print "Test 1: document from test corpus\n================================================="
self.print_info(o_doc)
def test2(self):
doc_id = "PMID-2083253"
o_doc = self.builder.build(doc_id, is_test = True)
print "\n\nTest 2: document from train corpus set is_test=True\n================================================="
self.print_info(o_doc)
def test3(self):
doc_id = "PMC-2222968-04-Results-03"
o_doc = self.builder.build(doc_id)
print "\n\nTest 3: document from dev corpus\n================================================="
self.print_info(o_doc)
def test4(self):
# document builder without dictionary
doc_id = 'PMID-8978306'
o_doc = self.builder.build(doc_id)
print "\n\nTest 4: document with multi word trigger\n================================================="
self.print_info(o_doc)
def print_info(self, o_doc):
print "doc id:", o_doc.doc_id
print "is test:", o_doc.is_test
for i in range(0, len(o_doc.sen)):
o_sen = o_doc.sen[i]
print "sen:", i
print "-------------------------------"
for j in range(0,o_sen.nwords):
w = o_sen.words[j]
print j, w['start'], w['string'], w['pos_tag'], w['type'], w['score']
# entity maps
print "entity maps"
print o_sen.entity_map
# list of word number which is marked as trigger candidate
print "trigger candidate:"
print o_sen.trigger_candidate
# list of protein word number
print "protein:"
print o_sen.protein
# list of trigger word number
print "trigger:"
print o_sen.trigger
print o_sen.trigger_text
# dependency
print "dependency"
print o_sen.dep.root
print o_sen.dep.graph
print o_sen.dep.pair
# chunk
print "chunk"
print o_sen.chunk.chunk_map
print o_sen.chunk.chunk_type
# tree
# relation representation
print "relation:"
if o_sen.rel != None:
print o_sen.rel.data
class Prediction(object):
'''
classdocs
'''
# suffix and extension of id file
DOCID_SUFFIX_EXT = "_doc_ids.json"
# directory for saving svm model
MODEL_DIR = "/model"
# directory for saving output a2 file
OUT_DIR = "/result"
# list of event name
EVENT_NAME = ["None",
"Gene_expression",
"Transcription",
"Protein_catabolism",
"Phosphorylation",
"Localization",
"Binding",
"Regulation",
"Positive_regulation",
"Negative_regulation"]
def __init__(self, source, dir_name, dict_type):
'''
Constructor
'''
self.src = source
self._model_path = ''
self._out_path = ''
self.set_path(source, dir_name)
self.dict_type = dict_type
self.wdict = None
self.tdict = None
self.doc_builder = None
self.extraction = None
self.docs = {}
self._set(dict_type)
def _set(self, dict_type):
"""
initialize dictionary type to be used in feature extraction process
initialize document builder
initialize feature extraction
"""
self.wdict = WordDictionary(self.src)
self.wdict.load(dict_type)
self.tdict = TriggerDictionary(self.src)
self.tdict.load(dict_type)
self.doc_builder = DocumentBuilder(self.src, self.wdict, self.tdict)
self.extraction = FeatureExtraction(self.src, self.wdict, self.tdict)
self.a2 = A2Writter(self._out_path)
def set_path(self, source, dir_name):
"""
check whether given dir_name is exist
raise error if it does not exist
return full _model_path of dir_name
"""
# model path
path = source + self.MODEL_DIR + '/' + dir_name
if not os.path.exists(path):
raise ValueError(path + "does not exist!!, chose another dir_name for prediction")
self._model_path = path
# output path
path = source + self.OUT_DIR + '/' + dir_name
if not os.path.exists(path):
os.makedirs(path)
self._out_path = path
def get_feature(self, step):
"""
extract feature and return X, Y for a given step
step are either one of these:
'tp' => trigger-protein relation
'tt' => trigger-trigger relation to predict regulation event with trigger argument
'tc' => trigger-theme-cause relation to predict regulation event with theme and cause (binary)
't2' => trigger-theme1-theme2 relation to predict theme2 in binding (binary)
"""
if step not in ['tt','tp','tc','t2']:
raise ValueError("only support step for tt, tp, tc and t2")
X = []
Y = []
info = []
dt_start = dt.now()
# reset statistic of extraction
self.extraction.reset_statistic()
# init feature
print "now extracting", len(self.docs), "docs"
for doc_id in self.docs.keys():
o_doc = self.docs[doc_id]
if step == 'tp':
samples = self.extraction.extract_tp(o_doc)
elif step == 'tt':
samples = self.extraction.extract_tt(o_doc)
elif step == 'tc':
samples = self.extraction.extract_tc(o_doc)
elif step == 't2':
samples = self.extraction.extract_t2(o_doc)
for sample in samples:
X.append(sample[2])
Y.append(sample[1])
info.append(sample[0])
print "time to extract feature", dt.now() - dt_start
return X,Y, info
def set_prediction_docs(self,docid_list_fname, is_test = True):
"""
build a document to be predicted
"""
dt_start = dt.now()
self.docs = {}
# get list of file
doc_ids = self.get_docid_list(docid_list_fname)
print "now building", len(doc_ids), "docs"
for doc_id in doc_ids:
self.docs[doc_id] = self.doc_builder.build(doc_id, is_test)
print "finish built docs in:", dt.now() - dt_start
def update_doc_info(self, list_info, list_target, arg_name, arg_type):
"""
update trigger and relation of document
"""
for i in range(0,len(list_info)):
target = list_target[i]
if target < 1: continue
info = list_info[i]
doc_id = info["doc"]
self.docs[doc_id].update(info['sen'], info['t'], self.EVENT_NAME[target], info['a'], arg_name, arg_type)
def update_doc_relation(self, rel_type, list_info, list_target):
"""
update only relation of document
"""
for i in range(0,len(list_info)):
target = list_target[i]
if target == 1:
info = list_info[i]
doc_id = info["doc"]
if rel_type == 'cause':
arg = info['c']
else:
arg = info['a2']
self.docs[doc_id].update_relation(rel_type, info['sen'], info['t'], arg)
def get_docid_list(self, docid_list_fname):
"""
return list of file
"""
if not isinstance(docid_list_fname, list):
# get list of doc ids from file
path = self.src + '/' + docid_list_fname + self.DOCID_SUFFIX_EXT
if not os.path.exists(path):
raise ValueError(path + " is not exist")
with open(path, 'r') as f:
doc_ids = json.loads(f.read())
else:
doc_ids = docid_list_fname
return doc_ids
def predict_tp(self, grid_search = True):
"""
return prediction of given docid_list
"""
if self.docs == {}:
raise ValueError("docs have not been created. call set_prediction_docs first!")
# get list of file
#doc_ids = self.get_docid_list(docid_list_fname)
# get features and target
X, Y, info = self.get_feature('tp')
# init svm classifier
svm = SVM(self._model_path, "trig-prot", "linear", grid_search = grid_search, class_weight = 'auto')
svm.load()
return svm.predict(X), Y, info
def predict_tt(self, grid_search = True):
"""
return prediction of given docid_list
"""
if self.docs == {}:
raise ValueError("docs have not been created. call set_prediction_docs first!")
# get list of file
#doc_ids = self.get_docid_list(docid_list_fname)
# get features and target
X, Y, info = self.get_feature('tt')
# init svm classifier
svm = SVM(self._model_path, "trig-trig", "linear", grid_search = grid_search, class_weight = 'auto')
svm.load()
return svm.predict(X), Y, info
def predict_tc(self, grid_search = True):
if self.docs == {}:
raise ValueError("docs have not been created. call set_prediction_docs first!")
# get list of file
#doc_ids = self.get_docid_list(docid_list_fname)
# get features and target
X, Y, info = self.get_feature('tc')
# init svm classifier
svm = SVM(self._model_path, "trig-theme-cause", "linear", grid_search = grid_search, class_weight = 'auto')
svm.load()
return svm.predict(X), Y, info
def predict_t2(self, grid_search = True):
if self.docs == {}:
raise ValueError("docs have not been created. call set_prediction_docs first!")
# get features and target
X, Y, info = self.get_feature('t2')
# init svm classifier
svm = SVM(self._model_path, "trig-theme1-2", "linear", grid_search = grid_search, class_weight = 'auto')
svm.load()
return svm.predict(X), Y, info
def predict(self, docid_list_fname, write_result = True):
# create document object for prediction
self.set_prediction_docs(docid_list_fname)
# predict trigger-protein relation
Ypred, _, info = self.predict_tp(grid_search = True)
# update document
self.update_doc_info(info, Ypred, "Theme", "P")
# predict trigger-trigger relation
for _ in range(0,2):
Ypred, _, info = self.predict_tt(grid_search = True)
self.update_doc_info(info, Ypred, "Theme", "E")
# predict trigger-theme-cause relation
Ypred, _, info = self.predict_tc(grid_search = True)
self.update_doc_relation('cause', info, Ypred)
# predict theme2 relation
Ypred, _, info = self.predict_t2(grid_search = True)
self.update_doc_relation('theme2', info, Ypred)
# write a2
if write_result:
self.write_result()
def write_result(self):
print "now writing", len(self.docs), "docs result to", self._out_path
for doc in self.docs.itervalues():
self.a2.write(doc)
class Learning(object):
'''
Learning steps:
1. define docs for learning
2. extract features
3. build input data for classifier
4. build a model and save it
'''
# suffix and extension of id file
DOCID_SUFFIX_EXT = "_doc_ids.json"
# directory for saving svm model
MODEL_DIR = "/model"
def __init__(self, source, dir_name, dict_type):
'''
Constructor
'''
self.src = source
self.path = self.set_path(source, dir_name)
self.dict_type = dict_type
self.wdict = None
self.tdict = None
self.doc_builder = None
self.extraction = None
self._set(dict_type)
def set_path(self, source, dir_name):
"""
check whether given dir_name is exist
raise error if exist, otherwise create new one
return full path of dir_name
"""
path = source + self.MODEL_DIR + '/' + dir_name
if os.path.exists(path):
raise ValueError(path +
"exist!!, chose anoher dir_name for learning")
else:
# create dir_name
os.makedirs(path)
return path
def _set(self, dict_type):
"""
initialize dictionary type to be used in learning process
initialize document builder
initialize feature extraction
"""
self.wdict = WordDictionary(self.src)
self.wdict.load(dict_type)
self.tdict = TriggerDictionary(self.src)
self.tdict.load(dict_type)
self.doc_builder = DocumentBuilder(self.src, self.wdict, self.tdict)
self.extraction = FeatureExtraction(self.src, self.wdict, self.tdict)
def get_docid_list(self, docid_list_fname):
"""
return list of file
"""
if not isinstance(docid_list_fname, list):
# get list of doc ids from file
path = self.src + '/' + docid_list_fname + self.DOCID_SUFFIX_EXT
if not os.path.exists(path):
raise ValueError(path + " is not exist")
with open(path, 'r') as f:
doc_ids = json.loads(f.read())
else:
doc_ids = docid_list_fname
return doc_ids
def get_feature(self, doc_ids, step):
"""
extract feature and return X, Y for a given step
step are either one of these:
'tp' => trigger-protein relation
'tt' => trigger-trigger relation to predict regulation event with trigger argument
'tc' => trigger-theme-cause relation to predict regulation event with theme and cause (binary)
't2' => trigger-theme1-theme2 relation to predict theme2 in binding (binary)
"""
if step not in ['tt', 'tp', 'tc', 't2']:
raise ValueError("only support step for tt, tp, tc and t2")
X = []
Y = []
dt_start = dt.now()
# reset statistic of extraction
self.extraction.reset_statistic()
# init feature
print "now extracting", len(doc_ids), "docs"
for doc_id in doc_ids:
o_doc = self.doc_builder.build(doc_id)
if step == 'tp':
samples = self.extraction.extract_tp(o_doc)
elif step == 'tt':
samples = self.extraction.extract_tt(o_doc)
elif step == 'tc':
samples = self.extraction.extract_tc(o_doc)
elif step == 't2':
samples = self.extraction.extract_t2(o_doc)
for sample in samples:
X.append(sample[2])
Y.append(sample[1])
# print statistic
pos = self.extraction.sample_pos
neg = self.extraction.sample_neg
stat = (pos, neg, pos + neg)
print stat
print "percentege of positif data:", pos * 100.0 / (pos + neg)
print "time to extract feature", dt.now() - dt_start
return X, Y
def learn_tp(self, docid_list_fname, grid_search):
# get list of file
doc_ids = self.get_docid_list(docid_list_fname)
# get features and target
X, Y = self.get_feature(doc_ids, 'tp')
# init svm classifier
svm = SVM(self.path,
'trig-prot',
'linear',
grid_search=grid_search,
class_weight='auto')
svm.create()
# fit training data
svm.learn(X, Y)
def learn_tt(self, docid_list_fname, grid_search):
# get list of file
doc_ids = self.get_docid_list(docid_list_fname)
# get features and target
X, Y = self.get_feature(doc_ids, 'tt')
# init svm classifier
svm = SVM(self.path,
'trig-trig',
'linear',
grid_search=grid_search,
class_weight='auto')
svm.create()
# fit training data
svm.learn(X, Y)
def learn_tc(self, docid_list_fname, grid_search):
# get list of file
doc_ids = self.get_docid_list(docid_list_fname)
# get features and target
X, Y = self.get_feature(doc_ids, 'tc')
# init svm classifier
svm = SVM(self.path,
'trig-theme-cause',
'linear',
grid_search=grid_search,
class_weight='auto')
svm.create()
# fit training data
svm.learn(X, Y)
def learn_t2(self, docid_list_fname, grid_search):
# get list of file
doc_ids = self.get_docid_list(docid_list_fname)
# get features and target
X, Y = self.get_feature(doc_ids, 't2')
# init svm classifier
svm = SVM(self.path,
'trig-theme1-2',
'linear',
grid_search=grid_search,
class_weight='auto')
svm.create()
# fit training data
svm.learn(X, Y)
class DocumentBuilderTest(object):
'''
classdocs
'''
def __init__(self):
'''
Constructor
'''
self.source = "E:/corpus/bionlp2011/project_data/"
WD = WordDictionary(self.source)
WD.load("train")
TD = TriggerDictionary(self.source)
TD.load("train")
self.builder = DocumentBuilder(self.source, WD, TD)
def run(self):
self.test1()
self.test2()
self.test3()
def test1(self):
doc_id = "PMID-2160380"
o_doc = self.builder.build(doc_id, is_test=False)
print "Test 1: document from test corpus\n================================================="
self.print_info(o_doc)
def test2(self):
doc_id = "PMID-2083253"
o_doc = self.builder.build(doc_id, is_test=True)
print "\n\nTest 2: document from train corpus set is_test=True\n================================================="
self.print_info(o_doc)
def test3(self):
doc_id = "PMC-2222968-04-Results-03"
o_doc = self.builder.build(doc_id)
print "\n\nTest 3: document from dev corpus\n================================================="
self.print_info(o_doc)
def test4(self):
# document builder without dictionary
doc_id = 'PMID-8978306'
o_doc = self.builder.build(doc_id)
print "\n\nTest 4: document with multi word trigger\n================================================="
self.print_info(o_doc)
def print_info(self, o_doc):
print "doc id:", o_doc.doc_id
print "is test:", o_doc.is_test
for i in range(0, len(o_doc.sen)):
o_sen = o_doc.sen[i]
print "sen:", i
print "-------------------------------"
for j in range(0, o_sen.nwords):
w = o_sen.words[j]
print j, w['start'], w['string'], w['pos_tag'], w['type'], w[
'score']
# entity maps
print "entity maps"
print o_sen.entity_map
# list of word number which is marked as trigger candidate
print "trigger candidate:"
print o_sen.trigger_candidate
# list of protein word number
print "protein:"
print o_sen.protein
# list of trigger word number
print "trigger:"
print o_sen.trigger
print o_sen.trigger_text
# dependency
print "dependency"
print o_sen.dep.root
print o_sen.dep.graph
print o_sen.dep.pair
# chunk
print "chunk"
print o_sen.chunk.chunk_map
print o_sen.chunk.chunk_type
# tree
# relation representation
print "relation:"
if o_sen.rel != None:
print o_sen.rel.data