class SemEval2014Task4_Laptops(__SemEval2014Task4):
""" SemEval 2014 Task 4 Laptop Dataset for Aspect based Sentiment Analysis.
Download: http://alt.qcri.org/semeval2014/task4/index.php?id=data-and-tools
"""
# paths
TRAIN_FILE = FilePath(
"SemEval2014-Task4/Laptops_Train.xml",
"https://raw.githubusercontent.com/pedrobalage/SemevalAspectBasedSentimentAnalysis/master/semeval_data/Laptop_Train_v2.xml"
)
TEST_FILE = FilePath(
"SemEval2014-Task4/laptops-trial.xml",
"https://alt.qcri.org/semeval2014/task4/data/uploads/laptops-trial.xml"
)
n_train_items = lambda self: 3045
n_eval_items = lambda self: 100
def get_aspect_label_pairs(self, sentence):
# get aspect categories and terms
aspect_terms = sentence.find('aspectTerms')
# load aspect label pairs
aspect_label_pairs = []
if aspect_terms is not None:
aspect_label_pairs += [(aspect.attrib['term'],
aspect.attrib['polarity'])
for aspect in aspect_terms]
# return
return aspect_label_pairs
class SemEval2014Task4_Category(__SemEval2014Task4):
""" SemEval 2014 Task 4 Aspect-Category Dataset for Aspect based Sentiment Analysis.
Only provides examples for aspect-categories (not explicitly mentioned in the text).
Download: http://alt.qcri.org/semeval2014/task4/index.php?id=data-and-tools
"""
# paths
TRAIN_FILE = FilePath(
"SemEval2014-Task4/Restaurants_Train.xml",
"https://raw.githubusercontent.com/pedrobalage/SemevalAspectBasedSentimentAnalysis/master/semeval_data/Restaurants_Train_v2.xml"
)
TEST_FILE = FilePath(
"SemEval2014-Task4/restaurants-trial.xml",
"https://alt.qcri.org/semeval2014/task4/data/uploads/restaurants-trial.xml"
)
n_train_items = lambda self: 3041
n_eval_items = lambda self: 100
def get_aspect_label_pairs(self, sentence):
# only get categories
aspect_categories = sentence.find('aspectCategories')
# build aspect label pairs
aspect_label_pairs = []
if aspect_categories is not None:
aspect_label_pairs += [
(aspect.attrib['category'], aspect.attrib['polarity'])
for aspect in sentence.find('aspectCategories')
if aspect is not None
]
# return
return aspect_label_pairs
class SemEval2015Task12(AOEx_Dataset):
""" SemEval 2015 Task 12 dataset for Aspect-Opinion Extraction
Downlaod: https://github.com/happywwy/Coupled-Multi-layer-Attentions/tree/master/util/data_semEval
"""
n_train_items = lambda self: 1316
n_eval_items = lambda self: 686
def yield_items(self, aspect_fname:str, sent_opinion_fname:str):
# build full paths to files
aspect_fpath = self.data_base_dir / aspect_fname
sent_opinion_fpath = self.data_base_dir / sent_opinion_fname
# load file contents
with open(aspect_fpath, 'r', encoding='utf-8') as f:
all_aspects = f.read().replace('NULL', '').split('\n')
with open(sent_opinion_fpath, 'r', encoding='utf-8') as f:
all_sents_opinions = f.read().split('\n')
assert len(all_aspects) == len(all_sents_opinions)
# preprocess data
for sent_opinions, aspects in zip(all_sents_opinions, all_aspects):
# separate sentence from opinions
sent, opinions = sent_opinions.split('##') if '##' in sent_opinions else (sent_opinions, '')
# get aspects and opinions
opinions = [o.strip()[:-3] for o in opinions.split(',')] if len(opinions) > 0 else []
aspects = [a.strip() for a in aspects.split(',')] if len(aspects) > 0 else []
# build aspect and opinion spans
opinion_pos = [sent.find(o) for o in opinions]
opinion_spans = [(i, i + len(o)) for i, o in zip(opinion_pos, opinions)]
aspect_pos = [sent.find(a) for a in aspects]
aspect_spans = [(i, i + len(a)) for i, a in zip(aspect_pos, aspects)]
# yield dataset item
yield AOEx_DatasetItem(
sentence=sent,
aspect_spans=aspect_spans,
opinion_spans=opinion_spans
)
# yield train and test items
yield_train_items = lambda self: self.yield_items(
aspect_fname=FilePath(
"SemEval2015-Task12/aspectTerm_res15",
"https://raw.githubusercontent.com/happywwy/Coupled-Multi-layer-Attentions/master/util/data_semEval/aspectTerm_res15"
),
sent_opinion_fname=FilePath(
"SemEval2015-Task12/sentence_res15_op",
"https://raw.githubusercontent.com/happywwy/Coupled-Multi-layer-Attentions/master/util/data_semEval/sentence_res15_op"
)
)
yield_eval_items = lambda self: self.yield_items(
aspect_fname=FilePath(
"SemEval2015-Task12/aspectTerm_restest15",
"https://raw.githubusercontent.com/happywwy/Coupled-Multi-layer-Attentions/master/util/data_semEval/aspectTerm_restest15"
),
sent_opinion_fname=FilePath(
"SemEval2015-Task12/sentence_restest15_op",
"https://raw.githubusercontent.com/happywwy/Coupled-Multi-layer-Attentions/master/util/data_semEval/sentence_restest15_op"
)
)
class SemEval2015Task12_AspectPolarity(__SemEval2015Task12):
""" Dataset for the SemEval2014 Task4 data for Aspect-based Sentiment Analysis
Download: http://alt.qcri.org/semeval2015/task12/index.php?id=data-and-tools
"""
LABELS = ['positive', 'neutral', 'negative']
TRAIN_FILE = FilePath(
"SemEval2015-Task12/ABSA-15_Restaurants_Train_Final.xml",
"https://raw.githubusercontent.com/peace195/aspect-based-sentiment-analysis/master/data/ABSA_SemEval2015/Restaurants_Train_Final.xml"
)
EVAL_FILE = FilePath(
"SemEval2015-Task12/ABSA15_Restaurants_Test.xml",
"https://raw.githubusercontent.com/peace195/aspect-based-sentiment-analysis/master/data/ABSA_SemEval2015/Restaurants_Test.xml"
)
n_train_items = lambda self: 833
n_eval_items = lambda self: 402
def yield_items(self, fpath: str) -> iter:
# parse xml file
tree = ET.parse(fpath)
root = tree.getroot()
# parse all reviews
for review in root:
for sent in review[0].findall('sentence'):
# get sentence
text = sent.find('text').text
# find opinions
opinions = sent.find('Opinions')
if opinions is None:
continue
# get aspects and sentiments
aspects = [(int(o.attrib['from']), int(o.attrib['to']))
for o in opinions]
sentiments = [o.attrib['polarity'] for o in opinions]
# remove unvalids - no aspect target
sentiments = [
s for s, (b, e) in zip(sentiments, aspects) if b < e
]
aspects = [(b, e) for (b, e) in aspects if b < e]
# no aspects found
if len(aspects) == 0:
continue
# build dataset item
yield NEC_DatasetItem(
sentence=text,
entity_spans=aspects,
labels=[
SemEval2015Task12_AspectPolarity.LABELS.index(s)
for s in sentiments
])
class SemEval2015Task12_OpinionPolarity(__SemEval2015Task12):
""" Dataset for the SemEval2014 Task4 data for Opinion-based Sentiment Analysis
Downlaod: https://github.com/happywwy/Coupled-Multi-layer-Attentions/tree/master/util/data_semEval
"""
LABELS = ['positive', 'negative']
TRAIN_FILE = FilePath(
"SemEval2015-Task12/sentence_res15_op",
"https://raw.githubusercontent.com/happywwy/Coupled-Multi-layer-Attentions/master/util/data_semEval/sentence_res15_op"
)
EVAL_FILE = FilePath(
"SemEval2015-Task12/sentence_restest15_op",
"https://raw.githubusercontent.com/happywwy/Coupled-Multi-layer-Attentions/master/util/data_semEval/sentence_restest15_op"
)
n_train_items = lambda self: 760
n_eval_items = lambda self: 333
def yield_items(self, fpath: str) -> iter:
# load file content
with open(fpath, 'r', encoding='utf-8') as f:
all_sents_opinions = f.read().split('\n')
# preprocess data
for sent_opinions in all_sents_opinions:
# no opinions
if '##' not in sent_opinions:
continue
# separate sentence from opinions
sent, opinions = sent_opinions.split('##')
# get aspects and opinions
opinions = [o.strip() for o in opinions.split(',')
] if len(opinions) > 0 else []
opinions, sentiments = zip(*[(o[:-2].strip(), o[-2:])
for o in opinions])
# build opinion spans
opinion_pos = [sent.find(o) for o in opinions]
opinion_spans = [(i, i + len(o))
for i, o in zip(opinion_pos, opinions)]
# get sentiment labels
sentiments = [(-int(i) + 1) // 2 for i in sentiments]
# build dataset item
yield NEC_DatasetItem(sentence=sent,
entity_spans=opinion_spans,
labels=sentiments)
class SemEval2010Task8(RelExDataset):
""" SemEval2010 Task8 Dataset
Download: https://github.com/sahitya0000/Relation-Classification/blob/master/corpus/SemEval2010_task8_all_data.zip
"""
# training and testing files
TRAIN_FILE = FilePath(
"SemEval2010-Task8/SemEval2010_task8_training/TRAIN_FILE.TXT",
"https://raw.githubusercontent.com/sahitya0000/Relation-Classification/master/corpus/SemEval2010_task8_training/TRAIN_FILE.TXT"
)
EVAL_FILE = FilePath(
"SemEval2010-Task8/SemEval2010_task8_testing_keys/TEST_FILE_FULL.TXT",
"https://raw.githubusercontent.com/sahitya0000/Relation-Classification/master/corpus/SemEval2010_task8_testing_keys/TEST_FILE_FULL.TXT"
)
# set of valid labels
LABELS = [
"Other", "Component-Whole(e2,e1)", "Component-Whole(e1,e2)",
"Instrument-Agency(e2,e1)", "Instrument-Agency(e1,e2)",
"Member-Collection(e2,e1)", "Member-Collection(e1,e2)",
"Cause-Effect(e2,e1)", "Cause-Effect(e1,e2)",
"Entity-Destination(e2,e1)", "Entity-Destination(e1,e2)",
"Content-Container(e2,e1)", "Content-Container(e1,e2)",
"Message-Topic(e2,e1)", "Message-Topic(e1,e2)",
"Product-Producer(e2,e1)", "Product-Producer(e1,e2)",
"Entity-Origin(e2,e1)", "Entity-Origin(e1,e2)"
]
n_train_items = lambda self: 8000
n_eval_items = lambda self: 2717
# yield training and evaluation items
yield_train_items = lambda self: self.yield_item(
self.data_base_dir / SemEval2010Task8.TRAIN_FILE)
yield_eval_items = lambda self: self.yield_item(self.data_base_dir /
SemEval2010Task8.EVAL_FILE)
def yield_item(self, fpath: str) -> iter:
# load data
with open(fpath, 'r', encoding='utf-8') as f:
lines = f.read().strip().split('\n')
# read examples
for sent_line, relation_line in zip(lines[::4], lines[1::4]):
# get text
sent = sent_line.split('\t')[1].strip()
# clean up sentence
assert sent[0] == sent[-1] == '"'
sent = sent[1:-1]
# find entities in sentence
entity_A = re.search(r'<e1>(.*)</e1>', sent)
entity_B = re.search(r'<e2>(.*)</e2>', sent)
# get spans from matches with markers
entity_span_A = (entity_A.start(), entity_A.end() - 4 - 5)
entity_span_B = (entity_B.start(), entity_B.end() - 4 - 5)
if entity_span_A[0] < entity_span_B[0]:
entity_span_B = (entity_span_B[0] - 4 - 5,
entity_span_B[1] - 4 - 5)
else:
entity_span_A = (entity_span_A[0] - 4 - 5,
entity_span_A[1] - 4 - 5)
# remove markers from text
sent = re.sub(r'<(/?)e1>', '', sent)
sent = re.sub(r'<(/?)e2>', '', sent)
# get label
label = relation_line.strip()
# yield features
yield RelExDatasetItem(
sentence=sent,
source_entity_span=entity_span_A,
target_entity_span=entity_span_B,
relation_type=SemEval2010Task8.LABELS.index(label))