def semantics(reply, stmt=None, **kwargs):
global nlp
nlp = kwargs.get('nlp', nlp)
if kwargs is None or kwargs['nlp'] is None or not stmt:
return 0.0
cos_sim = nlp(reply).similarity(nlp(stmt))
return cos_sim
def generate_sentence(spec=SENTENCE_SPEC, sentence_id=None):
""" Generate random sentence using word probabilities specified in SENTENCE_SPEC
>>> spec = {
... "answers":[[{"HDL":0.95,"good_cholesterol":0.05}, {"150": 0.01,"145": 0.01,"unk": 0.98}],
... "sentences":["Patient LDL level is 100, ________ level is 50, and the total is ______ .",]
... }
>>> s = generate_sentence(spec=spec, sentence_id=0) # doctest: +ELLIPSIS, +NORMALIZE_WHITESPACE
>>> s
'Patient LDL level is 100, ... level is 50, and the total is ... .'
>>> s[26:42] in ('HDL level is 50,', 'good_cholesterol')
True
>>> s[60:63] in ('150', '145', 'unk')
True
"""
sentences = spec['sentences']
if sentence_id is None:
sentence_id = np.random.randint(0, len(sentences))
sentence = sentences[sentence_id]
answer = spec['answers'][sentence_id]
i_unk = 0
tokens = []
for i, tok in enumerate(nlp(sentence)):
if re.match(r'^(_+|unk|\[MASK\])$', tok.text):
possible_tokens, p = list(zip(*answer[i_unk].items()))
tokens.append(np.random.choice(a=possible_tokens, p=p))
i_unk += 1
def get_sentences(df, size_limit=50000, vector_dim=None):
vector_dim = len(
nlp('word').vector) if vector_dim is None else int(vector_dim)
sents = []
docvecs = np.zeros((len(df), vector_dim))
encodings = []
for file_id, row in tqdm(df.iterrows(), total=len(df)):
sentvecs = []
encodings.append('utf8')
if row['size'] <= size_limit and row['is_journal']:
try:
with open(row['path'], 'rb') as fin:
bintext = fin.read()
try:
text = bintext.decode()
except UnicodeDecodeError:
encodings[-1] = 'latin'
log.warning(f"LATIN?: {row['path']}")
text = bintext.decode('latin')
doc = nlp(text)
except UnicodeDecodeError:
log.error(f"UnicodeDecodeError: {row['path']}")
continue
docvecs[file_id, :] = np.array(list(doc.vector))
docsents = [
dict(sentence_pos=f'{file_id}-{j}',
file_id=file_id,
text=s.text) for j, s in enumerate(doc.sents)
]
log.info(f"Read {len(docsents)} sentences: {row['path']}")
# print(doc.vector)
sents.extend(docsents)
sentvecs.extend([s.vector for s in doc.sents])
else:
log.warn(f"skipped {row['path']}")
df['encoding'] = encodings
df = pd.concat([df, pd.DataFrame(np.array(docvecs))], axis=1)
df_sents = pd.DataFrame(sents, index=list(range(len(sents))))
df_sents = pd.concat([df_sents, pd.DataFrame(np.array(sentvecs))], axis=1)
df_sents.index.name = 'sentence_id'
return df, df_sents
def semantics(reply, stmt=None, **kwargs):
""" Compute word2vec docvec cosine similarity (fall back to character IOU)
>>> semantics('Hello world!', 'Goodbye big earth!') > .5
True
"""
global nlp
nlp = kwargs.get('nlp', nlp)
if kwargs is None or nlp is None or not stmt or not reply:
return 0.0
reply_doc, stmt_doc = nlp(reply), nlp(stmt)
if not reply_doc or not stmt_doc or not reply_doc.has_vector or not stmt_doc.has_vector:
# FIXME: levenshtien would be better or fuzzywuzzy
return iou(reply, stmt)
cos_sim = nlp(reply).similarity(nlp(stmt))
log.debug(f'cos_sim={cos_sim}')
return cos_sim
def term_vector_dict(terms, keys=None):
terms = [str(t) if t else '' for t in terms]
keys = terms if keys is None else list(keys)
vector_list = []
log.info(f'Computing doc vectors for {len(terms)} terms...')
for k, term in zip(keys, terms):
vec = nlp(term).vector # s can sometimes (rarely) be a float because of pd.read_csv (df_titles)
vec /= np.linalg.norm(vec) or 1.
# vec = vec.round(7)
mask_zeros = np.abs(vec) > 0
if mask_zeros.sum() < len(mask_zeros):
log.warning(f'BAD VEC: {term} [0]*{mask_zeros.sum()}')
vector_list.append((k, vec))
# columns = [f'x{i}' for i in range(300)''
# dtypes = {c: pd.np.float16 for c in columns}
# df_vectors
# dtypes.update(page_title=str)
# self.df_vectors = pd.read_csv(filepath, dtype=dtypes)
return dict(vector_list)
def search_csv(
csv_path='/midata/private/journal/files.csv',
query='Misima island port harbor derelict ship PNG Papua New Guinnea Australis harbor storm sailing cliffs anchor drag',
num_results=10,
num_dims=300):
df = pd.read_csv(csv_path, index_col=0)
index_path = os.path.join(os.path.dirname(csv_path), 'files_index.ann')
index = AnnoyIndex(f=num_dims)
index.load(index_path)
vec = nlp(query).vector
paths = []
for i in index.get_nns_by_vector(vec, num_results):
path = df.iloc[i]['path']
paths.append(path)
print(path)
with open(path, 'rb') as fin:
bintext = b''.join(fin.readlines()[:10])
try:
text = bintext.decode()
except UnicodeDecodeError:
text = bintext.decode('latin')
print(text)
print('-' * 120)
return
def scrape_articles(titles=TITLES, exclude_headings=EXCLUDE_HEADINGS,
see_also=True, max_articles=10000, max_depth=3):
""" Download text for an article and parse into sections and sentences
>>> nlp('hello') # to eager-load spacy model
hello
>>> df = scrape_articles(['ELIZA'], see_also=False)
>>> df.shape
(87, 3)
>>> df.columns
Index(['title', 'section', 'sentence'], dtype='object')
"""
titles = list([titles] if isinstance(titles, str) else titles)
exclude_headings = set([eh.lower().strip() for eh in (exclude_headings or [])])
depths = list([0] * len(titles))
title_depths = list(zip(titles, depths))
sentences = []
# FIXME: breadth-first search so you can do a tqdm progress bar for each depth
# FIXME: record title tree (see also) so that .2*title1+.3*title2+.5*title3 can be semantically appended to sentences
titles_scraped = set([''])
title, d = '', 0
wiki = Wikipedia()
for depth in range(max_depth):
for i in range(max_articles):
title = None
while not title or title in titles_scraped:
# log.warn(f"Skipping {title} (already scraped)")
try:
title, d = title_depths.pop()
except IndexError:
log.warn(f'Out of titles: {title_depths}')
break
title = title.strip()
if d > max_depth or not title:
log.info(f"{d} > {max_depth} or title ('{title}') is empty")
continue
titles_scraped.add(title)
page = wiki.article(title)
if not (len(page.text) + len(page.summary)):
log.error(f"Unable to retrieve {title}")
time.sleep(2.17)
continue
if see_also and d + 1 < max_depth:
# .full_text() includes the section heading ("See also"). .text does not
section = page.section_by_title('See also')
if not section:
continue
for t in section.text.split('\n')[1:]:
if t in page.links:
title_depths.append((t, d + 1))
log.debug(f'extended title_depths at depth {d}: {title_depths}')
for section in page.sections:
if section.title.lower().strip() in exclude_headings:
continue
# TODO: use pugnlp.to_ascii() or nlpia.to_ascii()
text = section.text.replace('’', "'") # spacy doesn't handle "latin" (extended ascii) apostrophes well.
# FIXME: need to rejoin short names before colons, like 'ELIZA:' 'Tell me...', and 'Human:' 'What...'
# FIXME: need to split on question marks without white space but where next word is capitalized: ...to be unhappy?Though designed strictly...
sentences.extend([
(d, title, section.title, s.text) for s in nlp(text).sents if (
len(s.text.strip().strip('"').strip("'").strip()) > 1)
])
log.debug(f'Parsed {len(sentences)} sentences.')
# retval = parse_sentences(
# title=title, sentences=sentences, title_depths=title_depths, see_also=see_also,
# exclude_headings=exclude_headings, d=d, depth=depth, max_depth=max_depth)
# if retval is None:
# continue
# else:
# sentences, title_depths = retval
log.info(str([depth, d, i, title]))
if d > depth:
log.info(f"{d} > {depth}")
break
return pd.DataFrame(sentences, columns='depth title section sentence'.split())
