def __init__(self,
lucene_index_path: str,
min_df: int = 1,
verbose: bool = False):
self.min_df: int = min_df
self.verbose: bool = verbose
self.index_reader = index.IndexReader(lucene_index_path)
self.searcher = search.LuceneSearcher(lucene_index_path)
self.num_docs: int = self.searcher.num_docs
self.stats = self.index_reader.stats()
self.analyzer = Analyzer(get_lucene_analyzer())
# build vocabulary
self.vocabulary_ = set()
for term in self.index_reader.terms():
if term.df > self.min_df:
self.vocabulary_.add(term.term)
self.vocabulary_ = sorted(self.vocabulary_)
# build term to index mapping
self.term_to_index = {}
for i, term in enumerate(self.vocabulary_):
self.term_to_index[term] = i
self.vocabulary_size = len(self.vocabulary_)
if self.verbose:
print(
f'Found {self.vocabulary_size} terms with min_df={self.min_df}'
)
def __init__(self,
k1: float = 1.6,
b: float = 0.75,
index_path: str = None):
self.k1 = k1
self.b = b
self.use_corpus_estimator = False
self.analyzer = Analyzer(get_lucene_analyzer())
if index_path:
self.use_corpus_estimator = True
self.index_utils = IndexReader(index_path)
def get_term_query(term, field="contents", analyzer=get_lucene_analyzer()):
"""Searches the collection.
Parameters
----------
term : str
The query term string.
field : str
Field to search.
analyzer : Analyzer
Analyzer to use for tokenizing the query term.
Returns
-------
JTermQuery
"""
analyzer = Analyzer(analyzer)
return JTermQuery(JTerm(field, analyzer.analyze(term)[0]))
def batch_process(batch):
if (os.getcwd().endswith('ltr_msmarco')):
stopwords = read_stopwords('stopwords.txt', lower_case=True)
else:
stopwords = read_stopwords('./scripts/ltr_msmarco/stopwords.txt',
lower_case=True)
nlp = SpacyTextParser('en_core_web_sm',
stopwords,
keep_only_alpha_num=True,
lower_case=True)
analyzer = Analyzer(get_lucene_analyzer())
#nlp_ent = spacy.load("en_core_web_sm")
bert_tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
def process(line):
if not line:
return None
line = line[:maxDocSize] # cut documents that are too long!
fields = line.split('\t')
if len(fields) != 2:
return None
pid, body = fields
text, text_unlemm = nlp.proc_text(body)
#doc = nlp_ent(body)
#entity = {}
#for i in range(len(doc.ents)):
#entity[doc.ents[i].text] = doc.ents[i].label_
#entity = json.dumps(entity)
analyzed = analyzer.analyze(body)
for token in analyzed:
assert ' ' not in token
contents = ' '.join(analyzed)
doc = {
"id": pid,
"text": text,
"text_unlemm": text_unlemm,
'contents': contents,
"raw": body
}
doc["text_bert_tok"] = get_retokenized(bert_tokenizer, body.lower())
return doc
res = []
start = time.time()
for line in batch:
res.append(process(line))
if len(res) % 1000 == 0:
end = time.time()
print(f'finish {len(res)} using {end-start}')
start = end
return res
class Vectorizer:
"""Base class for vectorizer implemented on top of Pyserini.
Parameters
----------
lucene_index_path : str
Path to lucene index folder
min_df : int
Minimum acceptable document frequency
verbose : bool
Whether to print out debugging information
"""
def __init__(self,
lucene_index_path: str,
min_df: int = 1,
verbose: bool = False):
self.min_df: int = min_df
self.verbose: bool = verbose
self.index_reader = index.IndexReader(lucene_index_path)
self.searcher = search.LuceneSearcher(lucene_index_path)
self.num_docs: int = self.searcher.num_docs
self.stats = self.index_reader.stats()
self.analyzer = Analyzer(get_lucene_analyzer())
# build vocabulary
self.vocabulary_ = set()
for term in self.index_reader.terms():
if term.df > self.min_df:
self.vocabulary_.add(term.term)
self.vocabulary_ = sorted(self.vocabulary_)
# build term to index mapping
self.term_to_index = {}
for i, term in enumerate(self.vocabulary_):
self.term_to_index[term] = i
self.vocabulary_size = len(self.vocabulary_)
if self.verbose:
print(
f'Found {self.vocabulary_size} terms with min_df={self.min_df}'
)
def get_query_vector(self, query: str):
matrix_row, matrix_col, matrix_data = [], [], []
tokens = self.analyzer.analyze(query)
for term in tokens:
if term in self.vocabulary_:
matrix_row.append(0)
matrix_col.append(self.term_to_index[term])
matrix_data.append(1)
vectors = csr_matrix((matrix_data, (matrix_row, matrix_col)),
shape=(1, self.vocabulary_size))
return vectors
def batch_process(batch):
#assume call the script from the root dir
stopwords = read_stopwords('./scripts/ltr_msmarco/stopwords.txt', lower_case=True)
nlp = SpacyTextParser('en_core_web_sm', stopwords, keep_only_alpha_num=True, lower_case=True)
analyzer = Analyzer(get_lucene_analyzer())
bert_tokenizer =AutoTokenizer.from_pretrained("bert-base-uncased")
def process(line):
if not line:
return None
json_line = json.loads(line)
pid = json_line['id']
body = json_line['contents']
#url = json_line['url']
#title = json_line['title']
text, text_unlemm = nlp.proc_text(body)
#_,title_unlemm = nlp.proc_text(title)
analyzed = analyzer.analyze(body)
for token in analyzed:
assert ' ' not in token
contents = ' '.join(analyzed)
doc = {"id": pid,
"text": text,
"text_unlemm": text_unlemm,
'contents': contents,
#"title_unlemm": title_unlemm,
#"url": url,
"raw": body}
if (len(body)>512):
doc["text_bert_tok"] = get_retokenized(bert_tokenizer, body.lower()[:512])
else:
doc["text_bert_tok"] = get_retokenized(bert_tokenizer, body.lower())
return doc
res = []
start = time.time()
for line in batch:
res.append(process(line))
if len(res) % 10000 == 0:
end = time.time()
print(f'finish {len(res)} using {end-start}')
start = end
return res
class Bm25Reranker(Reranker):
def __init__(self,
k1: float = 1.6,
b: float = 0.75,
index_path: str = None):
self.k1 = k1
self.b = b
self.use_corpus_estimator = False
self.analyzer = Analyzer(get_lucene_analyzer())
if index_path:
self.use_corpus_estimator = True
self.index_utils = IndexReader(index_path)
def rerank(self, query: Query, texts: List[Text]) -> List[Text]:
query_words = self.analyzer.analyze(query.text)
sentences = list(map(self.analyzer.analyze, (t.text for t in texts)))
query_words_set = set(query_words)
sentence_sets = list(map(set, sentences))
if not self.use_corpus_estimator:
idfs = {
w: math.log(
len(sentence_sets) /
(1 + sum(int(w in sent) for sent in sentence_sets)))
for w in query_words_set
}
mean_len = np.mean(list(map(len, sentences)))
d_len = len(sentences)
texts = deepcopy(texts)
for sent_words, text in zip(sentences, texts):
tf = Counter(filter(query_words.__contains__, sent_words))
if self.use_corpus_estimator:
idfs = {
w: self.index_utils.compute_bm25_term_weight(
text.metadata['docid'], w)
for w in tf
}
score = sum(idfs[w] * tf[w] * (self.k1 + 1) /
(tf[w] + self.k1 * (1 - self.b + self.b *
(d_len / mean_len))) for w in tf)
if np.isnan(score):
score = 0
text.score = score
return texts
def test_lucene_analyzer_fr_book_examples(self):
analyzer = Analyzer(get_lucene_analyzer(name='french'))
tokens = analyzer.analyze(
'marche parler vélo randonnée rouler défilement')
self.assertEqual(['march', 'parl', 'vélo', 'randon', 'roul', 'defil'],
tokens)
tokens = analyzer.analyze('défilement roulant')
self.assertEqual(['defil', 'roulant'], tokens)
tokens = analyzer.analyze('biostatistique')
self.assertEqual(['biostatist'], tokens)
tokens = analyzer.analyze('antagoniste')
self.assertEqual(['antagonist'], tokens)
def test_lucene_analyzer_en_book_examples(self):
analyzer = Analyzer(get_lucene_analyzer())
tokens = analyzer.analyze(
'walking talking balking biking hiking rolling scrolling')
self.assertEqual(
['walk', 'talk', 'balk', 'bike', 'hike', 'roll', 'scroll'], tokens)
tokens = analyzer.analyze('rolling scrolling')
self.assertEqual(['roll', 'scroll'], tokens)
tokens = analyzer.analyze('biostatistics')
self.assertEqual(['biostatist'], tokens)
tokens = analyzer.analyze('adversarial')
self.assertEqual(['adversari'], tokens)
def test_lucene_analyzer_zh_book_examples(self):
analyzer = Analyzer(get_lucene_analyzer(name='cjk'))
tokens = analyzer.analyze('走路说话骑自行车远足滚动滚动')
self.assertEqual([
'走路', '路说', '说话', '话骑', '骑自', '自行', '行车', '车远', '远足', '足滚', '滚动',
'动滚', '滚动'
], tokens)
tokens = analyzer.analyze('滚动滚动')
self.assertEqual(['滚动', '动滚', '滚动'], tokens)
tokens = analyzer.analyze('生物统计学')
self.assertEqual(['生物', '物统', '统计', '计学'], tokens)
tokens = analyzer.analyze('对抗的')
self.assertEqual(['对抗', '抗的'], tokens)
parser.add_argument('--collection_path',
required=True,
help='MS MARCO .tsv collection file')
parser.add_argument('--predictions',
required=True,
help='File containing predicted queries.')
parser.add_argument('--output_folder', required=True, help='output folder')
parser.add_argument('--max_docs_per_file',
default=1000000,
type=int,
help='maximum number of documents in each jsonl file.')
args = parser.parse_args()
if not os.path.exists(args.output_folder):
os.makedirs(args.output_folder)
analyzer = Analyzer(get_lucene_analyzer())
print('Converting collection...')
file_index = 0
new_words = 0
total_words = 0
with open(args.collection_path) as f_corpus, open(
args.predictions) as f_pred:
for i, (line_doc, line_pred) in enumerate(zip(f_corpus, f_pred)):
# Write to a new file when the current one reaches maximum capacity.
if i % args.max_docs_per_file == 0:
if i > 0:
output_jsonl_file.close()
output_path = os.path.join(args.output_folder,
f'docs{file_index:02d}.json')
arg_vars = vars(args)
inpFile = open(args.input)
outFile = open(args.output, 'w')
minQueryTokQty = args.min_query_token_qty
if os.getcwd().endswith('ltr_msmarco'):
stopwords = read_stopwords('stopwords.txt', lower_case=True)
else:
stopwords = read_stopwords('./scripts/ltr_msmarco/stopwords.txt',
lower_case=True)
print(stopwords)
nlp = SpacyTextParser('en_core_web_sm',
stopwords,
keep_only_alpha_num=True,
lower_case=True)
analyzer = Analyzer(get_lucene_analyzer())
nlp_ent = spacy.load("en_core_web_sm")
bert_tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
# Input file is a TSV file
ln = 0
for line in tqdm(inpFile):
ln += 1
line = line.strip()
if not line:
continue
fields = line.split('\t')
if len(fields) != 2:
print('Misformated line %d ignoring:' % ln)
print(line.replace('\t', '<field delimiter>'))
continue
def analyze_query(query):
analyzer = Analyzer(get_lucene_analyzer())
query = analyzer.analyze(query)
return query
def test_analysis(self):
# Default is Porter stemmer
analyzer = Analyzer(get_lucene_analyzer())
self.assertTrue(isinstance(analyzer, Analyzer))
tokens = analyzer.analyze('City buses are running on time.')
self.assertEqual(tokens, ['citi', 'buse', 'run', 'time'])
# Specify Porter stemmer explicitly
analyzer = Analyzer(get_lucene_analyzer(stemmer='porter'))
self.assertTrue(isinstance(analyzer, Analyzer))
tokens = analyzer.analyze('City buses are running on time.')
self.assertEqual(tokens, ['citi', 'buse', 'run', 'time'])
# Specify Krovetz stemmer explicitly
analyzer = Analyzer(get_lucene_analyzer(stemmer='krovetz'))
self.assertTrue(isinstance(analyzer, Analyzer))
tokens = analyzer.analyze('City buses are running on time.')
self.assertEqual(tokens, ['city', 'bus', 'running', 'time'])
# No stemming
analyzer = Analyzer(get_lucene_analyzer(stemming=False))
self.assertTrue(isinstance(analyzer, Analyzer))
tokens = analyzer.analyze('City buses are running on time.')
self.assertEqual(tokens, ['city', 'buses', 'running', 'time'])
# No stopword filter, no stemming
analyzer = Analyzer(
get_lucene_analyzer(stemming=False, stopwords=False))
self.assertTrue(isinstance(analyzer, Analyzer))
tokens = analyzer.analyze('City buses are running on time.')
self.assertEqual(tokens,
['city', 'buses', 'are', 'running', 'on', 'time'])
# No stopword filter, with stemming
analyzer = Analyzer(get_lucene_analyzer(stemming=True,
stopwords=False))
self.assertTrue(isinstance(analyzer, Analyzer))
tokens = analyzer.analyze('City buses are running on time.')
self.assertEqual(tokens, ['citi', 'buse', 'ar', 'run', 'on', 'time'])
def test_invalid_analysis(self):
# Invalid configuration, make sure we get an exception.
with self.assertRaises(ValueError):
Analyzer(get_lucene_analyzer('blah'))
def test_invalid_analyzer_wrapper(self):
# Invalid JAnalyzer, make sure we get an exception.
with self.assertRaises(TypeError):
Analyzer('str')