def makeVerb(topic, parents, num, w2v, jux=False):
key = topic + "".join(parents)
if key in cache:
final = cache[key]
else:
choices = set()
stripped_topic = h.strip_tag(topic)
for p in parents:
for r in ['trg']: #bga is terrible
addAll(
choices,
h.pos([
x[0] for x in dm.query(dm.related(r, p),
dm.topics(stripped_topic))
], 'v'))
for r in ['CapableOf', 'UsedFor', 'Desires']:
for phrase in [x[1] for x in cn.getOutgoing(p, r)]:
addAll(choices, h.pos(phrase.split(), 'v'))
final = list(choices)
if jux:
temp = set()
for w in final:
addAll(
temp,
h.pos([
x[0] for x in dm.query(dm.related('ant', w),
dm.topics(stripped_topic))
], 'v'))
for phrase in [x[1] for x in cn.getOutgoing(w, 'Antonym')]:
addAll(choices, h.pos(phrase.split(), 'v'))
final = list(temp)
#word2vec sort
parents = [x + '_NN' for x in parents]
relations = parents + [topic]
#word2vec threshold
final = [pen.conjugate(x) + '_VB' for x in final]
final = h.w2vWeightsListNew(final, relations, w2v)
if len(final) > 20:
final = final[:-len(final) / 5]
cache[key] = final
if len(final) < 1:
print "VERB FINAL IS EMPTY:", topic, parents, num, list(choices)
#just in case
if len(final) <= num:
return [x[0] for x in final]
return [pickOne(final) for x in range(num)]
def doit(topic, parent, pos, w2v=None):
choices = {}
for r in dmrs:
addAllDict(
choices,
h.pos([
x[0] for x in dm.query(dm.related(r, parent), dm.topics(topic))
], pos), 'dm:' + r)
for r in cnrs:
for phrase in [x[1] for x in cn.getOutgoing(parent, r)]:
addAllDict(choices, h.pos(phrase.split(), pos), 'cn:' + r)
#add w2v "custom relation" thing
if w2v is not None:
addAllDict(choices, h.pos(h.relation(parent, custom_rels, w2v), pos),
'custom')
print len(choices.keys())
scores = {}
for k in choices.keys():
s = ask(topic, parent, k)
s -= 5 #(-5 to 5 scale)
choices[k]['score'] += s
for src in choices[k]['sources']:
addScore(scores, src, s)
print "Best words", sorted(choices.keys(),
key=lambda x: choices[x]['score'],
reverse=True)
print "Best sources", scores
def genrec(node, parent, prev, force, w2v, fillin, w2vmax, w2vmin, verbgen):
i = node['index']
if not force and fillin[i]:
word = fillin[i]
if parent:
force = True #only force regen if its not the root (i.e. it has a parent)
else:
if parent['replace']:
parent = parent['replace']
nodep = node['pos']
startTag = '_' + parent['pos']
endTag = '_' + nodep
#maybe just have a set list to draw from for some restricted POS like IN, etc?
cacheK = (prev, parent['word'], node['word'])
if parent['dep'] == 'root' and cacheK in choiceCache:
choices = choiceCache[cacheK]
else:
choices = w2vChoices(prev, parent['word'], startTag, node['word'],
endTag, w2v, w2vmax, w2vmin)
if parent[
'dep'] == 'root': #only cache choices from root (more likely to be used, caching all is too much data for too little overlap)
choiceCache[cacheK] = choices
cacheK = (parent['word'], node['word'])
if cacheK not in relsCache:
relsCache[cacheK] = cn.getRels(parent['word'], node['word'])
cnRels = relsCache[cacheK]
for rel in cnRels:
choices += [cn.stripPre(t[0]) for t in cn.getOutgoing(prev, rel)]
final = []
for c in choices:
if c == node['word']: #try to find a different word
continue
p = h.getPOS(c)
if p == nodep:
final.append(c)
elif p in nodep or nodep in p or ('VB' in p and 'VB' in nodep):
newc = h.tryPOS(c, p, nodep)
if newc:
final.append(newc)
if not final:
word = wb.get(
nodep) #grab from wordbag insteadof using node['word']
else:
if verbgen:
print parent['word'], ":", node[
'word'], "::", prev, ":\n", final
print
word = random.choice(final) #Can this be smarter?
if not word:
word = node['word']
fillin[i] = word
if len(node['children']):
for child in node['children']:
genrec(child, node, word, force, w2v, fillin, w2vmax, w2vmin,
verbgen)
def get_words(topic, parents, cn_relations, w2v_relations, w2v, isIncoming=True):
res = []
stripped_topic = h.strip_tag(topic)
for p in parents:
stripped_parent = h.strip_tag(p)
for rel in cn_relations:
if isIncoming:
res += (x[1] for x in cn.getIncoming(stripped_parent, rel))
else:
res += (x[1] for x in cn.getOutgoing(stripped_parent, rel))
# TODO: find good datamuse relationships for outgoing edges.
if isIncoming:
res += (x[0] for x in dm.query(dm.related('trg',stripped_parent),dm.topics(stripped_topic), dm.metadata('p'))
if ('tags' in x[1] and x[1]['tags'][0] == 'n')) # get only nouns
w2v_words = h.get_nouns_from_verb(stripped_parent, w2v_relations, w2v)
res += w2v_words
#res = filterNoun(res)
res = oldFilterNoun(res)
removeMatch(res, topic, parents)
res = [x+'_NN' for x in res] # should hopefully all be nouns at this point...
return res
def comesBefore(word):
return cn.getIncoming(word, "Causes") + cn.getIncoming(
word, "HasSubevent") + cn.getOutgoing(word, "HasPrerequisite")