def vectorizer(tokens, w2v_db):
db_path = w2v_db
# POS TAGGING
tagger = POSTagger('tagger/english-left3words-distsim.tagger', 'tagger/stanford-postagger.jar')
tagged_tokens = tagger.tag(tokens)
unsorted_kw = OrderedDict()
for (w,t) in tagged_tokens:
if t in ['NNP', 'NNPS', 'FW']:
label = 1.5
elif t in ['NN', 'NNS']:
label = 1
else:
continue
w = w.lower()
try:
unsorted_kw[w] += label
except KeyError:
unsorted_kw[w] = label
# Get the vectors of words. Maintain order as in document.
token_vecs = OrderedDict()
conn = SQLCon(db_path)
words = (word.lower() for word in unsorted_kw)
for word in words:
try:
if token_vecs[word]: continue
except KeyError:
v = conn.read(word)
if not v is None:
token_vecs[word] = list(v)
print("kw_len: {0} vec_len: {1}".format(len(unsorted_kw), len(token_vecs))) #Output for debugging; total vs unique words.
conn.close()
return unsorted_kw, token_vecs
def search(self, search_str, search_dir, depth=2):
# Depth -> number of top nodes to search. Default 2 (arbitrary) has been sufficient so far.
results = Counter() # {url:score}
dist_neu_searchq = Counter(
) # {nid:dist} Contains the similarity of each neuron to the aggregated search query
neuron_lookup = self._neu_index
neuron_labels = [[
k for (k, n) in Counter(neuron.get_keylist()).most_common()[:10]
] for neuron in self._neurons]
glomat = self._glomat # Global matrix. Contains similarity of each search term in the query to all neurons. Something like a cache.
conn = SQLCon()
searchvecs = [(x, list(conn.read(x))) for x in search_str.split()
if not conn.read(x) is None
] # Obtain (word,vec) of search terms
search_len = len(searchvecs)
for (w, v) in searchvecs: # For colour coding the map
try:
for nid in glomat[w]:
if glomat[w][nid] > dist_neu_searchq[nid]:
dist_neu_searchq[nid] += glomat[w][nid] / search_len
except KeyError:
glomat[w] = {}
for nid, neuron in enumerate(self._neurons):
glomat[w][nid] = distance(
neuron.get_weights(), v
) # cosine similarity, hence 1 is best. 0 is bleh. -1 is opposite.
if glomat[w][nid] > dist_neu_searchq[nid]:
dist_neu_searchq[nid] += glomat[w][nid] / search_len
# Union of all doclists with minimum dist_from_neuron.
doclist = {}
for nid in dist_neu_searchq.most_common()[:depth]:
neuron = neuron_lookup[nid[0]]
doclist.update(neuron.get_top_docs(30))
files = (open(doc) for doc in doclist)
for json_file in files:
data = json.load(json_file)
centroids = data['centroids']
url = data['url']
json_file.close()
wc_sim = [distance(v, c) for c in centroids]
max_wc_sim = max(wc_sim)
results[url] += max_wc_sim / len(searchvecs)
results = OrderedDict(results.most_common(20))
htmlVars = {'query': search_str, 'results': results}
htmlCode = template.render(htmlVars)
result_path = os.path.join(search_dir, search_str + '.html')
map_path = os.path.join(search_dir, search_str + '_map.html')
with open(result_path, 'w') as f:
f.write(htmlCode)
self.draw_SOM(search_str, dist_neu_searchq, neuron_labels, map_path)
result_path = "file://{}".format(pathname2url(result_path))
map_path = "file://{}".format(pathname2url(map_path))
webbrowser.open(result_path)
def processor(name, url, tokens, db_path, json_dir, USE_TITLE_WORDS=False):
# POS TAGGING
tagger = POSTagger('tagger/english-left3words-distsim.tagger',
'tagger/stanford-postagger.jar')
tagged_tokens = tagger.tag(tokens)
unsorted_kw = OrderedDict()
for (w, t) in tagged_tokens:
if t in ['NNP', 'NNPS', 'FW']:
label = 1.5
elif t in ['NN', 'NNS']:
label = 1
else:
continue
w = w.lower()
try:
unsorted_kw[w] += label
except KeyError:
unsorted_kw[w] = label
# Get the vectors list
token_vecs = OrderedDict()
conn = SQLCon(db_path)
words = (word.lower() for word in unsorted_kw)
for word in words:
try:
if token_vecs[word]: continue
except KeyError:
v = conn.read(word)
if not v is None:
token_vecs[word] = list(v)
print("kw_len: {0} vec_len: {1}".format(len(unsorted_kw), len(token_vecs)))
conn.close()
#Compute cluster centers:
nk = round(len(token_vecs) / 4)
data = numpy.array(list(token_vecs.values()))
cent, _ = kmeans2(data, nk, iter=20, minit='points')
centroids = cent.tolist()
# Create the JSON object for this webpage.
if not os.path.exists(json_dir):
os.makedirs(json_dir)
json_path = os.path.join(json_dir, name + '.json')
file_dest = open(json_path, 'w')
json.dump(
{
'url': url,
'vectors': token_vecs,
'keyword_frequency': unsorted_kw,
'centroids': centroids
}, file_dest)
file_dest.close()
def search(self,search_str,search_dir,depth=2):
# Depth -> number of top nodes to search. Default 2 (arbitrary) has been sufficient so far.
results = Counter() # {url:score}
dist_neu_searchq=Counter() # {nid:dist} Contains the similarity of each neuron to the aggregated search query
neuron_lookup = self._neu_index
neuron_labels = [[k for (k,n) in Counter(neuron.get_keylist()).most_common()[:10]] for neuron in self._neurons]
glomat = self._glomat # Global matrix. Contains similarity of each search term in the query to all neurons. Something like a cache.
conn = SQLCon()
searchvecs = [(x,list(conn.read(x))) for x in search_str.split() if not conn.read(x) is None] # Obtain (word,vec) of search terms
search_len = len(searchvecs)
for (w,v) in searchvecs: # For colour coding the map
try:
for nid in glomat[w]:
if glomat[w][nid] > dist_neu_searchq[nid]:
dist_neu_searchq[nid] += glomat[w][nid]/search_len
except KeyError:
glomat[w]={}
for nid,neuron in enumerate(self._neurons):
glomat[w][nid] = distance(neuron.get_weights(),v) # cosine similarity, hence 1 is best. 0 is bleh. -1 is opposite.
if glomat[w][nid] > dist_neu_searchq[nid]:
dist_neu_searchq[nid] += glomat[w][nid]/search_len
# Union of all doclists with minimum dist_from_neuron.
doclist = {}
for nid in dist_neu_searchq.most_common()[:depth]:
neuron = neuron_lookup[nid[0]]
doclist.update(neuron.get_top_docs(30))
files = (open(doc) for doc in doclist)
for json_file in files:
data = json.load(json_file)
centroids = data['centroids']
url = data['url']
json_file.close()
wc_sim = [distance(v,c) for c in centroids]
max_wc_sim = max(wc_sim)
results[url] += max_wc_sim/len(searchvecs)
results = OrderedDict(results.most_common(20))
htmlVars = {'query': search_str, 'results':results}
htmlCode = template.render(htmlVars)
result_path = os.path.join(search_dir,search_str+'.html')
map_path = os.path.join(search_dir,search_str+'_map.html')
with open(result_path,'w') as f:
f.write(htmlCode)
self.draw_SOM(search_str,dist_neu_searchq,neuron_labels,map_path)
result_path = "file://{}".format(pathname2url(result_path))
map_path = "file://{}".format(pathname2url(map_path))
webbrowser.open(result_path)
def processor(name, url, tokens, db_path,json_dir, USE_TITLE_WORDS = False):
# POS TAGGING
tagger = POSTagger('tagger/english-left3words-distsim.tagger', 'tagger/stanford-postagger.jar')
tagged_tokens = tagger.tag(tokens)
unsorted_kw = OrderedDict()
for (w,t) in tagged_tokens:
if t in ['NNP', 'NNPS', 'FW']:
label = 1.5
elif t in ['NN', 'NNS']:
label = 1
else:
continue
w = w.lower()
try:
unsorted_kw[w] += label
except KeyError:
unsorted_kw[w] = label
# Get the vectors list
token_vecs = OrderedDict()
conn = SQLCon(db_path)
words = (word.lower() for word in unsorted_kw)
for word in words:
try:
if token_vecs[word]: continue
except KeyError:
v = conn.read(word)
if not v is None:
token_vecs[word] = list(v)
print("kw_len: {0} vec_len: {1}".format(len(unsorted_kw), len(token_vecs)))
conn.close()
#Compute cluster centers:
nk = round(len(token_vecs)/4)
data = numpy.array(list(token_vecs.values()))
cent, _ = kmeans2(data,nk,iter=20,minit='points')
centroids = cent.tolist()
# Create the JSON object for this webpage.
if not os.path.exists(json_dir):
os.makedirs(json_dir)
json_path = os.path.join(json_dir,name+'.json')
file_dest = open(json_path, 'w')
json.dump({'url': url, 'vectors' : token_vecs, 'keyword_frequency': unsorted_kw, 'centroids' : centroids}, file_dest)
file_dest.close()
