def process_phrasal(phrase, normalize=True):
"""
Tokenize and stem the phrase words and compute the frequency of each word in the query list
Arguments:
phrase string representing the phrasal query
Returns:
A tuple containing the tokenized sentence and a counter object containing the counts
for the terms in the query.
"""
query_list = []
sentences = nltk.sent_tokenize(phrase)
for sentence in sentences:
words = nltk.word_tokenize(sentence)
if normalize:
for word in words:
normalized = text_processing.normalize(word)
if normalized is not None:
query_list.append(normalized)
else:
query_list.extend(words)
# count the frequency of each term
query_count = Counter(query_list)
return (query_list, query_count)
def process_query(query_str):
"""
Tokenize and stem the query words and compute the frequency of each word in the query list
Arguments:
query_str string of query words
Returns:
query_count a dictionary with the stemmed words and the its frequency in the query
"""
query_list = []
sentences = nltk.sent_tokenize(query_str)
for sentence in sentences:
words = nltk.word_tokenize(sentence)
for word in words:
normalized = text_processing.normalize(word)
if normalized is not None:
query_list.append(normalized)
# count the frequency of each term
query_count = Counter(query_list)
# set the tf value for each term
query_weight = {}
for query_term, term_count in query_count.items():
query_weight[query_term] = 1 + math.log10(term_count)
return query_weight
def process_phrasal(phrase, normalize=True):
"""
Tokenize and stem the phrase words and compute the frequency of each word in the query list
Arguments:
phrase string representing the phrasal query
Returns:
A tuple containing the tokenized sentence and a counter object containing the counts
for the terms in the query.
"""
query_list = []
sentences = nltk.sent_tokenize(phrase)
for sentence in sentences:
words = nltk.word_tokenize(sentence)
if normalize:
for word in words:
normalized = text_processing.normalize(word)
if normalized is not None:
query_list.append(normalized)
else:
query_list.extend(words)
# count the frequency of each term
query_count = Counter(query_list)
return (query_list, query_count)
def processText(text):
words = text_processing.tokenize(text)
words = text_processing.normalize(words)
new_words = []
for word in words:
if word not in stopwords_:
new_words.append(word)
return new_words
def processDescription(desc):
"""
Process products' description
"""
global idx
idx += 1
if (idx % 10000 == 0):
print(idx)
if type(desc) == str:
text = desc
text = text_processing.tokenize(text)
text = text_processing.normalize(text)
return text
def __tokenize_string(self, the_string):
"""
Tokenize and stem a string and remove punctuation and stopwords
"""
word_list = []
sentences = nltk.sent_tokenize(the_string)
for sentence in sentences:
words = nltk.word_tokenize(sentence)
for word in words:
normalized = text_processing.normalize(word)
if normalized is not None:
word_list.append(normalized)
return word_list
def __tokenize_string(self, the_string):
"""
Tokenize and stem a string and remove punctuation and stopwords
"""
word_list = []
sentences = nltk.sent_tokenize(the_string)
for sentence in sentences:
words = nltk.word_tokenize(sentence)
for word in words:
normalized = text_processing.normalize(word)
if normalized is not None:
word_list.append(normalized)
return word_list
def __init__(self, text_path, mel_dir, lin_dir, data_in_memory=True):
self.data = []
self.data_in_memory = data_in_memory
if data_in_memory and os.path.exists(os.path.join(mel_dir, "all.npy")):
mels = np.load(os.path.join(mel_dir, "all.npy"), allow_pickle=True)
else:
mels = None
with io.open(text_path, "r", encoding="utf-8-sig") as f:
lines = f.readlines()
for i, line in enumerate(lines):
line = line.split("|", maxsplit=1)
text = line[1]
text = text_processing.normalize(text)
text = text + Config.vocab_end_of_text
# Skip if text is too long
if len(text) > Config.max_N:
print(
"Warning: Text with id '{}' is too long! Line will be skipped"
.format(line[0]))
continue
text = text_processing.vocab_lookup(text)
text = torch.tensor(text, dtype=torch.long)
if data_in_memory:
if mels is not None:
mel = mels[i]
t = mel.shape[0] # Needed for lin padding
mel = self._process_mel(mel)
else:
mel_path = os.path.join(mel_dir, line[0]) + ".npy"
mel = np.load(mel_path)
t = mel.shape[0] # Needed for lin padding
mel = self._process_mel(np.load(mel_path))
# Skip if mel is too long
if mel.shape[0] > Config.max_T:
print(
"Warning: Mel with id '{}' is too long! Line will be skipped"
.format(line[0]))
continue
self.data.append({
"name": line[0],
"text": text,
"mel": mel,
"t": t
})
else:
self.data.append({"name": line[0], "text": text})
self.mel_dir = mel_dir
self.lin_dir = lin_dir
def indexing(training_path, postings_file, dictionary_file, patent_info_file):
"""
Create an index of the corpus at training_path, placing the index in
dictionary_file, postings_file and patent_info_file.
dictionary_file will contain a list of the terms contained in the corpus.
On every line, the following information will be included (separated by spaces):
<term indexed> <document frequency> <postings pointer>
The postings pointer is a line number in the postings file.
Every line of the postings file is a postings list for the term that points to that
line. A line contains several entries. Each entry has the following form:
<patent ID> <log tf> <tf> <list of positions>
The list of positions indicates the word positions at which the indexed word can
be found in the given patent (a standard positional index). The list is always
<tf> elements long.
"""
pats = sorted(os.listdir(training_path))
pi = open(patent_info_file, 'w')
postings = dict()
for pat in pats:
tree = et.parse(os.path.join(training_path, pat))
pat_id = os.path.splitext(pat)[0]
root = tree.getroot()
title_content = ""
abstract_content = ""
year = ""
cites = "0"
ipc = ""
inventor = ""
content = ""
for child in root:
# extract patent content
if child.get('name') == 'Title':
content += child.text.encode('utf-8') + " "
if child.get('name') == 'Abstract':
content += child.text.encode('utf-8') + " "
# extract patent info (meta data)
if child.get('name') == 'Publication Year':
year = child.text.encode('utf-8').strip()
if child.get('name') == 'Cited By Count':
cites = child.text.encode('utf-8').strip()
if child.get('name') == 'IPC Primary':
ipc = child.text.encode('utf-8').strip()
if child.get('name') == '1st Inventor':
inventor = child.text.encode('utf-8').strip()
pi.write(pat_id + " | " + year + " | " + cites + " | " + ipc + " | " +
inventor + "\n")
stemmer = PorterStemmer()
# remove non utf-8 characters, http://stackoverflow.com/a/20078869
content = re.sub(r'[^\x00-\x7F]+', ' ', content)
sentences = nltk.sent_tokenize(content)
i = 0
occurences = dict()
for sentence in sentences:
# tokenize sentences in words
words = nltk.word_tokenize(sentence)
for word in words:
normalized = text_processing.normalize(word)
if normalized is None:
continue
if normalized in occurences:
occurences[normalized].append(i)
else:
occurences[normalized] = [i]
i += 1
for word, positions in occurences.iteritems():
if word in postings:
postings[word].append((pat_id, positions))
else:
postings[word] = [(pat_id, positions)]
pi.close()
write_dict_postings(postings, postings_file, dictionary_file,
training_path)
def test_normalize(self):
test_str = "This is an example."
pred = tp.normalize(test_str)
self.assertEqual(pred, "this is an example.")
test_str = " EXTRA SPACE "
pred = tp.normalize(test_str)
self.assertEqual(pred, "extra space")
test_str = "THIS IS ALL CAPS!!"
pred = tp.normalize(test_str)
self.assertEqual(pred, "this is all caps!!")
test_str = " "
pred = tp.normalize(test_str)
self.assertEqual(pred, "")
test_str = "this is all lower space..."
pred = tp.normalize(test_str)
self.assertEqual(pred, "this is all lower space...")
test_str = " H e L l O !"
pred = tp.normalize(test_str)
self.assertEqual(pred, "h e l l o !")
test_str = ""
pred = tp.normalize(test_str)
self.assertEqual(pred, "")
test_str = "........"
pred = tp.normalize(test_str)
self.assertEqual(pred, "........")
test_str = "EX A M P LE"
pred = tp.normalize(test_str)
self.assertEqual(pred, "ex a m p le")
test_str = "Test Text Normalization"
pred = tp.normalize(test_str)
self.assertEqual(pred, "test text normalization")
test_str = "AbCd EfGh IjKl MnOp"
pred = tp.normalize(test_str)
self.assertEqual(pred, "abcd efgh ijkl mnop")
def processDescription(desc):
if type(desc)==str:
text = desc
text = text_processing.tokenize(text)
text = text_processing.normalize(text)
return text
def fparse(kv):
return kv[0], compress(unidecode(normalize(bytes_to_text(kv[1]).lower())))
def process_raw_text(text):
return normalize(extras.sub('', pAt.sub('', replaceURLs('', text)))).strip()
def indexing(training_path, postings_file, dictionary_file, patent_info_file):
"""
Create an index of the corpus at training_path, placing the index in
dictionary_file, postings_file and patent_info_file.
dictionary_file will contain a list of the terms contained in the corpus.
On every line, the following information will be included (separated by spaces):
<term indexed> <document frequency> <postings pointer>
The postings pointer is a line number in the postings file.
Every line of the postings file is a postings list for the term that points to that
line. A line contains several entries. Each entry has the following form:
<patent ID> <log tf> <tf> <list of positions>
The list of positions indicates the word positions at which the indexed word can
be found in the given patent (a standard positional index). The list is always
<tf> elements long.
"""
pats = sorted(os.listdir(training_path))
pi = open(patent_info_file, 'w')
postings = dict()
for pat in pats:
tree = et.parse(os.path.join(training_path, pat))
pat_id = os.path.splitext(pat)[0]
root = tree.getroot()
title_content = ""
abstract_content = ""
year = ""
cites = "0"
ipc = ""
inventor = ""
content = ""
for child in root:
# extract patent content
if child.get('name') == 'Title':
content += child.text.encode('utf-8') + " "
if child.get('name') == 'Abstract':
content += child.text.encode('utf-8') + " "
# extract patent info (meta data)
if child.get('name') == 'Publication Year':
year = child.text.encode('utf-8').strip()
if child.get('name') == 'Cited By Count':
cites = child.text.encode('utf-8').strip()
if child.get('name') == 'IPC Primary':
ipc = child.text.encode('utf-8').strip()
if child.get('name') == '1st Inventor':
inventor = child.text.encode('utf-8').strip()
pi.write(pat_id + " | " + year + " | " + cites + " | " + ipc + " | " + inventor + "\n")
stemmer = PorterStemmer()
# remove non utf-8 characters, http://stackoverflow.com/a/20078869
content = re.sub(r'[^\x00-\x7F]+',' ', content)
sentences = nltk.sent_tokenize(content)
i = 0
occurences = dict()
for sentence in sentences:
# tokenize sentences in words
words = nltk.word_tokenize(sentence)
for word in words:
normalized = text_processing.normalize(word)
if normalized is None:
continue
if normalized in occurences:
occurences[normalized].append(i)
else:
occurences[normalized] = [i]
i += 1
for word, positions in occurences.iteritems():
if word in postings:
postings[word].append((pat_id, positions))
else:
postings[word] = [(pat_id, positions)]
pi.close()
write_dict_postings(postings, postings_file, dictionary_file, training_path)
