def get_all_semsim():
accident_semsim = event2semsim("accident")
natdis_semsim = event2semsim("earthquake")
social_semsim = event2semsim("protest")
terror_semsim = event2semsim("shooting")
return {
"accident": accident_semsim,
"earthquake": natdis_semsim,
"storm": natdis_semsim,
"impact event": natdis_semsim,
"shooting": terror_semsim,
"hostage": terror_semsim,
"conflict": terror_semsim,
"bombing": terror_semsim,
"protest": social_semsim,
"riot": social_semsim,
}
def get_all_semsim():
accident_semsim = event2semsim("accident")
natdis_semsim = event2semsim("earthquake")
social_semsim = event2semsim("protest")
terror_semsim = event2semsim("shooting")
return {
"accident": accident_semsim,
"earthquake": natdis_semsim,
"storm": natdis_semsim,
"impact event": natdis_semsim,
"shooting": terror_semsim,
"hostage": terror_semsim,
"conflict": terror_semsim,
"bombing": terror_semsim,
"protest": social_semsim,
"riot": social_semsim,
}
all_results = []
data = []
with open("apsal.tsv", "w") as o:
for event in cuttsum.events.get_events():
if event.query_num < 26: continue
istream = get_input_stream(event, False)
with open("clusters-2015/{}.tsv".format(event.query_id), "r") as f:
df = pd.read_csv(f, sep="\t", converters={"stems": eval, "nuggets": eval})
#thresh = lm2thr[event2lm_name(event)]
thresh = .65
cache = None
semsim = event2semsim(event)
results = []
for ts, batch in df.groupby("timestamp"):
X = semsim.transform(batch["stems"].apply(lambda x: ' '.join(x)).tolist())
for i, (_, row) in enumerate(batch.iterrows()):
if cache is None:
cache = X[i]
results.append(row.to_dict())
all_results.append(row.to_dict())
else:
K = cosine_similarity(cache, X[i])
if (K < thresh).all():
cache = np.vstack([cache, X[i]])
results.append(row.to_dict())
all_results.append(row.to_dict())
for result in results:
def do_job_unit(self, event, corpus, unit, **kwargs):
if unit != 0:
raise Exception("Job unit {} out of range".format(unit))
service_configs = kwargs.get("service-configs", {})
cnlp_configs = service_configs.get("corenlp", {})
cnlp_port = int(cnlp_configs.get("port", 9999))
domain_lm_config = service_configs[event2lm_name(event)]
domain_lm_port = int(domain_lm_config["port"])
domain_lm_order = int(domain_lm_config.get("order", 3))
gw_lm_config = service_configs["gigaword-lm"]
gw_lm_port = int(gw_lm_config["port"])
gw_lm_order = int(gw_lm_config.get("order", 3))
thresh = kwargs.get("dedupe-sim-threshold", .8)
extractor = kwargs.get("extractor", "goose")
res = DedupedArticlesResource()
dfiter = res.dataframe_iter(event,
corpus,
extractor,
include_matches=None,
threshold=thresh)
domain_lm = cuttsum.srilm.Client(domain_lm_port, domain_lm_order, True)
gw_lm = cuttsum.srilm.Client(gw_lm_port, gw_lm_order, True)
cnlp_client = cnlp.client.CoreNLPClient(port=cnlp_port)
def make_query_synsets():
synonyms = []
hypernyms = []
hyponyms = []
print event.type.split(' ')[0]
for synset in wn.synsets(event.type.split(' ')[0]):
synonyms.extend([
lemma.name().lower().replace(u'_', u' ').encode(u'utf-8')
for lemma in synset.lemmas()
])
hypernyms.extend([
lemma.name().lower().replace(u'_', u' ').encode(u'utf-8')
for synset in synset.hypernyms()
for lemma in synset.lemmas()
])
hyponyms.extend([
lemma.name().lower().replace(u'_', u' ').encode(u'utf-8')
for synset in synset.hyponyms()
for lemma in synset.lemmas()
])
print hypernyms
print hyponyms
print synonyms
return set(synonyms), set(hypernyms), set(hyponyms)
def heal_text(sent_text):
sent_text = re.sub(
ur"[A-Z ]+, [A-Z][a-z ]+\( [A-Z]+ \) [-\u2014_]+ ", r"",
sent_text)
sent_text = re.sub(ur"^.*?[A-Z ]+, [A-Z][a-z]+ [-\u2014_]+ ", r"",
sent_text)
sent_text = re.sub(ur"^.*?[A-Z ]+\([^\)]+\) [-\u2014_]+ ", r"",
sent_text)
sent_text = re.sub(ur"^.*?[A-Z]+ +[-\u2014_]+ ", r"", sent_text)
sent_text = re.sub(r"\([^)]+\)", r" ", sent_text)
sent_text = re.sub(ur"^ *[-\u2014_]+", r"", sent_text)
sent_text = re.sub(u" ([,.;?!]+)([\"\u201c\u201d'])", r"\1\2",
sent_text)
sent_text = re.sub(r" ([:-]) ", r"\1", sent_text)
sent_text = re.sub(r"([^\d]\d{1,3}) , (\d\d\d)([^\d]|$)",
r"\1,\2\3", sent_text)
sent_text = re.sub(r"^(\d{1,3}) , (\d\d\d)([^\d]|$)", r"\1,\2\3",
sent_text)
sent_text = re.sub(ur" ('|\u2019) ([a-z]|ll|ve|re)( |$)", r"\1\2 ",
sent_text)
sent_text = re.sub(r" ([',.;?!]+) ", r"\1 ", sent_text)
sent_text = re.sub(r" ([',.;?!]+)$", r"\1", sent_text)
sent_text = re.sub(r"(\d\.) (\d)", r"\1\2", sent_text)
sent_text = re.sub(r"(a|p)\. m\.", r"\1.m.", sent_text)
sent_text = re.sub(r"U\. (S|N)\.", r"U.\1.", sent_text)
sent_text = re.sub(ur"\u201c ([^\s])", ur"\u201c\1", sent_text)
sent_text = re.sub(ur"([^\s]) \u201d", ur"\1\u201d", sent_text)
sent_text = re.sub(ur"\u2018 ([^\s])", ur"\u2018\1", sent_text)
sent_text = re.sub(ur"([^\s]) \u2019", ur"\1\u2019", sent_text)
sent_text = re.sub(ur"\u00e2", ur"'", sent_text)
sent_text = re.sub(r"^Photo:Reuters|^Photo:AP", r"", sent_text)
sent_text = sent_text.replace("\n", " ")
return sent_text.encode("utf-8")
def get_number_feats(sent):
feats = []
for tok in sent:
if tok.ne == "NUMBER" and tok.nne is not None:
for chain in get_dep_chain(tok, sent, 0):
feat = [tok.nne] + [elem[1].lem for elem in chain]
feats.append(feat)
return feats
def get_dep_chain(tok, sent, depth):
chains = []
if depth > 2:
return chains
for p in sent.dep2govs[tok]:
if p[1].is_noun():
for chain in get_dep_chain(p[1], sent, depth + 1):
chains.append([p] + chain)
elif p[1]:
chains.append([p])
return chains
import unicodedata as u
P = ''.join(
unichr(i) for i in range(65536) if u.category(unichr(i))[0] == 'P')
P = re.escape(P)
punc_patt = re.compile("[" + P + "]")
from collections import defaultdict
stopwords = english_stopwords()
mention_counts = defaultdict(int)
total_mentions = 0
from nltk.stem.porter import PorterStemmer
stemmer = PorterStemmer()
synonyms, hypernyms, hyponyms = make_query_synsets()
path = self.get_path(event, corpus, extractor, thresh)
dirname = os.path.dirname(path)
if not os.path.exists(dirname): os.makedirs(dirname)
meta_cols = [
"update id", "stream id", "sent id", "timestamp", "pretty text",
"tokens", "lemmas", "stems", "pos", "ne", "tokens stopped",
"lemmas stopped"
]
basic_cols = [
"BASIC length", "BASIC char length", "BASIC doc position",
"BASIC all caps ratio", "BASIC upper ratio", "BASIC lower ratio",
"BASIC punc ratio", "BASIC person ratio", "BASIC location ratio",
"BASIC organization ratio", "BASIC date ratio", "BASIC time ratio",
"BASIC duration ratio", "BASIC number ratio",
"BASIC ordinal ratio", "BASIC percent ratio", "BASIC money ratio",
"BASIC set ratio", "BASIC misc ratio"
]
lm_cols = [
"LM domain lp", "LM domain avg lp", "LM gw lp", "LM gw avg lp"
]
query_cols = [
"Q_query_sent_cov",
"Q_sent_query_cov",
"Q_syn_sent_cov",
"Q_sent_syn_cov",
"Q_hyper_sent_cov",
"Q_sent_hyper_cov",
"Q_hypo_sent_cov",
"Q_sent_hypo_cov",
]
sum_cols = [
"SUM_sbasic_sum",
"SUM_sbasic_amean",
"SUM_sbasic_max",
"SUM_novelty_gmean",
"SUM_novelty_amean",
"SUM_novelty_max",
"SUM_centrality",
"SUM_pagerank",
"SUM_sem_novelty_gmean",
"SUM_sem_novelty_amean",
"SUM_sem_novelty_max",
"SUM_sem_centrality",
"SUM_sem_pagerank",
]
stream_cols = [
"STREAM_sbasic_sum",
"STREAM_sbasic_amean",
"STREAM_sbasic_max",
"STREAM_per_prob_sum",
"STREAM_per_prob_max",
"STREAM_per_prob_amean",
"STREAM_loc_prob_sum",
"STREAM_loc_prob_max",
"STREAM_loc_prob_amean",
"STREAM_org_prob_sum",
"STREAM_org_prob_max",
"STREAM_org_prob_amean",
"STREAM_nt_prob_sum",
"STREAM_nt_prob_max",
"STREAM_nt_prob_amean",
]
semsim = event2semsim(event)
all_cols = meta_cols + basic_cols + query_cols + lm_cols + sum_cols + stream_cols
stream_uni_counts = defaultdict(int)
stream_per_counts = defaultdict(int)
stream_loc_counts = defaultdict(int)
stream_org_counts = defaultdict(int)
stream_nt_counts = defaultdict(int)
with gzip.open(path, "w") as f:
f.write("\t".join(all_cols) + "\n")
for df in dfiter:
if len(df) == 1: continue
df = df.head(20)
#df["lm"] = df["sent text"].apply(lambda x: lm.sentence_log_prob(x.encode("utf-8"))[1])
df["pretty text"] = df["sent text"].apply(heal_text)
df = df[df["pretty text"].apply(lambda x: len(x.strip())) > 0]
df = df[
df["pretty text"].apply(lambda x: len(x.split(" "))) < 200]
df = df.reset_index(drop=True)
if len(df) == 0:
print "skipping"
continue
doc_text = "\n".join(df["pretty text"].tolist())
doc = cnlp_client.annotate(doc_text)
df["tokens"] = map(lambda sent: [str(tok) for tok in sent],
doc)
df["lemmas"] = map(
lambda sent: [tok.lem.encode("utf-8") for tok in sent],
doc)
df["stems"] = map(
lambda sent:
[stemmer.stem(unicode(tok).lower()) for tok in sent], doc)
df["pos"] = map(lambda sent: [tok.pos for tok in sent], doc)
df["ne"] = map(lambda sent: [tok.ne for tok in sent], doc)
df["tokens stopped"] = map(
lambda sent: [str(tok) for tok in sent
if unicode(tok).lower() not in stopwords \
and len(unicode(tok)) < 50],
doc)
df["lemmas stopped"] = map(
lambda sent: [tok.lem.lower().encode("utf-8") for tok in sent
if unicode(tok).lower() not in stopwords \
and len(unicode(tok)) < 50],
doc)
df["num tuples"] = [get_number_feats(sent) for sent in doc]
### Basic Features ###
df["BASIC length"] = df["lemmas stopped"].apply(len)
df["BASIC doc position"] = df.index.values + 1
df = df[df["BASIC length"] > 0]
df = df.reset_index(drop=True)
df["BASIC char length"] = df["pretty text"].apply(
lambda x: len(x.replace(" ", "")))
df["BASIC upper ratio"] = df["pretty text"].apply(
lambda x: len(re.findall("[A-Z]", x))) \
/ df["BASIC char length"].apply(lambda x: float(max(x, 1)))
df[ "BASIC lower ratio"] = df["pretty text"].apply(
lambda x: len(re.findall("[a-z]", x))) \
/ df["BASIC char length"].apply(lambda x: float(max(x, 1)))
df["BASIC punc ratio"] = df["pretty text"].apply(
lambda x: len(re.findall(punc_patt, x))) \
/ df["BASIC char length"].apply(lambda x: float(max(x, 1)))
df["BASIC all caps ratio"] = df["tokens stopped"].apply(
lambda x: np.sum([1 if re.match("^[A-Z]+$", xi) else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC person ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "PERSON" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC location ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "LOCATION" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC organization ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "ORGANIZATION" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC date ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "DATE" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC time ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "TIME" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC duration ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "DURATION" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC number ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "NUMBER" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC ordinal ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "ORDINAL" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC percent ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "PERCENT" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC money ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "MONEY" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC set ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "SET" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC misc ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "MISC" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
### Language Model Features ###
dm_probs = df["lemmas"].apply(
lambda x: domain_lm.sentence_log_prob(" ".join([
xi.decode("utf-8").lower().encode("utf-8") for xi in x
if len(xi) < 50
])))
dm_log_probs = [lp for lp, avg_lp in dm_probs.tolist()]
dm_avg_log_probs = [avg_lp for lp, avg_lp in dm_probs.tolist()]
df["LM domain lp"] = dm_log_probs
df["LM domain avg lp"] = dm_avg_log_probs
gw_probs = df["lemmas"].apply(
lambda x: gw_lm.sentence_log_prob(" ".join([
xi.decode("utf-8").lower().encode("utf-8") for xi in x
if len(xi) < 50
])))
gw_log_probs = [lp for lp, avg_lp in gw_probs.tolist()]
gw_avg_log_probs = [avg_lp for lp, avg_lp in gw_probs.tolist()]
df["LM gw lp"] = gw_log_probs
df["LM gw avg lp"] = gw_avg_log_probs
### Query Features ###
self.compute_query_features(
df, set([q.lower() for q in event.query]), synonyms,
hypernyms, hyponyms)
### Single Doc Summarization Features ###
counts = []
doc_counts = defaultdict(int)
for lemmas in df["lemmas stopped"].tolist():
counts_i = {}
for lem in lemmas:
counts_i[lem.lower()] = counts_i.get(lem.lower(),
0) + 1
doc_counts[lem.lower()] += 1
doc_counts["__TOTAL__"] += len(lemmas)
counts.append(counts_i)
doc_counts["__TOTAL__"] *= 1.
doc_uni = {
key: val / doc_counts["__TOTAL__"]
for key, val in doc_counts.items() if key != "__TOTAL__"
}
sum_probs = []
amean_probs = []
max_probs = []
for lemmas in df["lemmas stopped"].tolist():
probs = [doc_uni[lem.lower()] for lem in lemmas]
sum_probs.append(np.sum(probs))
amean_probs.append(np.mean(probs))
max_probs.append(np.max(probs))
df["SUM_sbasic_sum"] = sum_probs
df["SUM_sbasic_amean"] = amean_probs
df["SUM_sbasic_max"] = max_probs
tfidfer = TfidfTransformer()
vec = DictVectorizer()
X = vec.fit_transform(counts)
X = tfidfer.fit_transform(X)
ctrd = X.mean(axis=0)
K = cosine_similarity(ctrd, X).ravel()
I = K.argsort()[::-1]
R = np.array([[i, r + 1] for r, i in enumerate(I)])
R = R[R[:, 0].argsort()]
df["SUM_centrality"] = R[:, 1]
L = semsim.transform(
df["stems"].apply(lambda x: ' '.join(x)).tolist())
ctrd_l = L.mean(axis=0)
K_L = cosine_similarity(ctrd_l, L).ravel()
I_L = K_L.argsort()[::-1]
R_L = np.array([[i, r + 1] for r, i in enumerate(I_L)])
R_L = R_L[R_L[:, 0].argsort()]
df["SUM_sem_centrality"] = R_L[:, 1]
K = cosine_similarity(X)
M = np.zeros_like(K)
M[np.diag_indices(K.shape[0])] = 1
Km = np.ma.masked_array(K, M)
D = 1 - Km
novelty_amean = D.mean(axis=1)
novelty_max = D.max(axis=1)
novelty_gmean = gmean(D, axis=1)
df["SUM_novelty_amean"] = novelty_amean
df["SUM_novelty_max"] = novelty_max
df["SUM_novelty_gmean"] = novelty_gmean
K_L = cosine_similarity(L)
M_L = np.zeros_like(K)
M_L[np.diag_indices(K_L.shape[0])] = 1
K_Lm = np.ma.masked_array(K_L, M_L)
D_L = 1 - K_Lm
sem_novelty_amean = D_L.mean(axis=1)
sem_novelty_max = D_L.max(axis=1)
sem_novelty_gmean = gmean(D_L, axis=1)
df["SUM_sem_novelty_amean"] = sem_novelty_amean
df["SUM_sem_novelty_max"] = sem_novelty_max
df["SUM_sem_novelty_gmean"] = sem_novelty_gmean
K = (K > 0).astype("int32")
degrees = K.sum(axis=1) - 1
edges_x_2 = K.sum() - K.shape[0]
if edges_x_2 == 0: edges_x_2 = 1
pr = 1. - degrees / float(edges_x_2)
df["SUM_pagerank"] = pr
K_L = (K_L > .2).astype("int32")
degrees_L = K_L.sum(axis=1) - 1
edges_x_2_L = K_L.sum() - K_L.shape[0]
if edges_x_2_L == 0: edges_x_2_L = 1
pr_L = 1. - degrees_L / float(edges_x_2_L)
df["SUM_sem_pagerank"] = pr_L
print df["pretty text"]
# print df[["SUM_sbasic_sum", "SUM_sbasic_amean", "SUM_sbasic_max"]]
# print df[
# ["SUM_pagerank", "SUM_centrality", "SUM_novelty_gmean",
# "SUM_novelty_amean", "SUM_novelty_max"]]
### Stream Features ###
for key, val in doc_counts.items():
stream_uni_counts[key] += val
denom = stream_uni_counts["__TOTAL__"]
sum_probs = []
amean_probs = []
max_probs = []
for lemmas in df["lemmas stopped"].tolist():
probs = [
stream_uni_counts[lem.lower()] / denom
for lem in lemmas
]
sum_probs.append(np.sum(probs))
amean_probs.append(np.mean(probs))
max_probs.append(np.max(probs))
df["STREAM_sbasic_sum"] = sum_probs
df["STREAM_sbasic_amean"] = amean_probs
df["STREAM_sbasic_max"] = max_probs
for lemmas, nes in izip(df["lemmas"].tolist(),
df["ne"].tolist()):
for lem, ne in izip(lemmas, nes):
if ne == "PERSON":
stream_per_counts[lem.lower()] += 1
stream_per_counts["__TOTAL__"] += 1.
if ne == "LOCATION":
stream_loc_counts[lem.lower()] += 1
stream_loc_counts["__TOTAL__"] += 1.
if ne == "ORGANIZATION":
stream_org_counts[lem.lower()] += 1
stream_org_counts["__TOTAL__"] += 1.
for tuples in df["num tuples"].tolist():
for nt in tuples:
for item in nt:
stream_nt_counts[item.lower()] += 1
stream_nt_counts["__TOTAL__"] += 1.
pdenom = stream_per_counts["__TOTAL__"]
ldenom = stream_loc_counts["__TOTAL__"]
odenom = stream_org_counts["__TOTAL__"]
ntdenom = stream_nt_counts["__TOTAL__"]
sum_per_probs = []
amean_per_probs = []
max_per_probs = []
sum_loc_probs = []
amean_loc_probs = []
max_loc_probs = []
sum_org_probs = []
amean_org_probs = []
max_org_probs = []
sum_nt_probs = []
amean_nt_probs = []
max_nt_probs = []
for tuples in df["num tuples"].tolist():
if ntdenom > 0:
nt_probs = [
stream_nt_counts[item.lower()] / ntdenom
for nt in tuples for item in nt
]
else:
nt_probs = []
if len(nt_probs) > 0:
sum_nt_probs.append(np.sum(nt_probs))
amean_nt_probs.append(np.mean(nt_probs))
max_nt_probs.append(np.max(nt_probs))
else:
sum_nt_probs.append(0)
amean_nt_probs.append(0)
max_nt_probs.append(0)
for lemmas, nes in izip(df["lemmas"].tolist(),
df["ne"].tolist()):
if pdenom > 0:
per_probs = [
stream_per_counts[lem.lower()] / pdenom
for lem, ne in izip(lemmas, nes) if ne == "PERSON"
]
else:
per_probs = []
if len(per_probs) > 0:
sum_per_probs.append(np.sum(per_probs))
amean_per_probs.append(np.mean(per_probs))
max_per_probs.append(np.max(per_probs))
else:
sum_per_probs.append(0)
amean_per_probs.append(0)
max_per_probs.append(0)
if ldenom > 0:
loc_probs = [
stream_loc_counts[lem.lower()] / ldenom
for lem, ne in izip(lemmas, nes)
if ne == "LOCATION"
]
else:
loc_probs = []
if len(loc_probs) > 0:
sum_loc_probs.append(np.sum(loc_probs))
amean_loc_probs.append(np.mean(loc_probs))
max_loc_probs.append(np.max(loc_probs))
else:
sum_loc_probs.append(0)
amean_loc_probs.append(0)
max_loc_probs.append(0)
if odenom > 0:
org_probs = [
stream_org_counts[lem.lower()] / odenom
for lem, ne in izip(lemmas, nes)
if ne == "ORGANIZATION"
]
else:
org_probs = []
if len(org_probs) > 0:
sum_org_probs.append(np.sum(org_probs))
amean_org_probs.append(np.mean(org_probs))
max_org_probs.append(np.max(org_probs))
else:
sum_org_probs.append(0)
amean_org_probs.append(0)
max_org_probs.append(0)
df["STREAM_per_prob_sum"] = sum_per_probs
df["STREAM_per_prob_max"] = max_per_probs
df["STREAM_per_prob_amean"] = amean_per_probs
df["STREAM_loc_prob_sum"] = sum_loc_probs
df["STREAM_loc_prob_max"] = max_loc_probs
df["STREAM_loc_prob_amean"] = amean_loc_probs
df["STREAM_org_prob_sum"] = sum_org_probs
df["STREAM_org_prob_max"] = max_org_probs
df["STREAM_org_prob_amean"] = amean_org_probs
df["STREAM_nt_prob_sum"] = sum_nt_probs
df["STREAM_nt_prob_max"] = max_nt_probs
df["STREAM_nt_prob_amean"] = amean_nt_probs
#print df[["STREAM_sbasic_sum", "STREAM_sbasic_amean", "STREAM_sbasic_max"]]
#print df[["STREAM_per_prob_sum", "STREAM_per_prob_amean", "STREAM_per_prob_max"]]
#print df[["STREAM_loc_prob_sum", "STREAM_loc_prob_amean", "STREAM_loc_prob_max"]]
#print df[["STREAM_nt_prob_sum", "STREAM_nt_prob_amean", "STREAM_nt_prob_max"]]
### Write dataframe to file ###
df[all_cols].to_csv(f, index=False, header=False, sep="\t")
def main(output_dir, sim_threshold, bucket_size):
if not os.path.exists(output_dir):
os.makedirs(output_dir)
dev_qids = set([19, 23, 27, 34, 35,] + [7,24])
summary_data = []
K_data = []
for event in cuttsum.events.get_events():
if event.query_num in dev_qids: continue
print event
semsim = event2semsim(event)
istream = get_input_stream(event, False, extractor="goose",
thresh=.8, delay=None, topk=20)
prev_time = 0
cache = None
clusters = []
max_h = len(event.list_event_hours()) - 1
for h, hour in enumerate(event.list_event_hours()):
if h % bucket_size != 0 and h != max_h:
continue
current_time = epoch(hour)
input_sents = istream[
(istream["timestamp"] < current_time) & \
(istream["timestamp"] >= prev_time)]
len_select = input_sents["lemmas stopped"].apply(len) > 10
input_sents = input_sents[len_select]
if len(input_sents) <= 1: continue
stems = input_sents["stems"].apply(lambda x: ' '.join(x)).tolist()
X = semsim.transform(stems)
K = -(1 - cosine_similarity(X))
K_ma = np.ma.masked_array(K, np.eye(K.shape[0]))
Kmin = np.ma.min(K_ma)
Kmax = np.ma.max(K_ma)
median = np.ma.median(K_ma)[0]
print "SYS TIME:", hour, "# SENTS:", K.shape[0],
print "min/median/max pref: {}/{}/{}".format(
Kmin, median, Kmax)
#
ap = AffinityPropagation(affinity="precomputed",
verbose=True, max_iter=1000)
ap.fit(K)
labels = ap.labels_
if ap.cluster_centers_indices_ != None:
for c in ap.cluster_centers_indices_:
if cache == None:
cache = X[c]
updates_df = \
input_sents.reset_index(drop=True).iloc[c]
updates_df["query id"] = event.query_num
updates_df["system timestamp"] = current_time
summary_data.append(
updates_df[
["query id", "stream id", "sent id",
"system timestamp", "sent text"]
].to_frame().T
)
else:
Ksum = cosine_similarity(cache, X[c])
if Ksum.max() < sim_threshold:
cache = np.vstack([cache, X[c]])
updates_df = \
input_sents.reset_index(drop=True).iloc[c]
updates_df["query id"] = event.query_num
updates_df["system timestamp"] = current_time
summary_data.append(
updates_df[
["query id", "stream id", "sent id",
"system timestamp", "sent text"]
].to_frame().T
)
prev_time = current_time
df = pd.DataFrame(K_data, columns=["min", "max", "median"])
print df
print df.mean()
print df.std()
print df.max()
df = pd.concat(summary_data)
df["conf"] = .5
df["team id"] = "AP"
df["run id"] = "sim{}_bs{}".format(
sim_threshold, bucket_size)
print df
of = os.path.join(output_dir, "ap." + "sim{}_bs{}.tsv".format(
sim_threshold, bucket_size))
cols = ["query id", "team id", "run id", "stream id", "sent id",
"system timestamp", "conf"]
df[cols].to_csv(of, sep="\t", header=False, index=False)
def main(output_dir, sim_threshold, bucket_size, pref_offset):
if not os.path.exists(output_dir):
os.makedirs(output_dir)
dev_qids = set([19, 23, 27, 34, 35] + [7, 24])
summary_data = []
K_data = []
for event in cuttsum.events.get_events():
if event.query_num in dev_qids: continue
print event
semsim = event2semsim(event)
istream = get_input_stream(event, False, extractor="goose",
thresh=.8, delay=None, topk=20)
prev_time = 0
cache = None
clusters = []
max_h = len(event.list_event_hours()) - 1
for h, hour in enumerate(event.list_event_hours()):
if h % bucket_size != 0 and h != max_h:
continue
current_time = epoch(hour)
input_sents = istream[
(istream["timestamp"] < current_time) & \
(istream["timestamp"] >= prev_time)]
len_select = input_sents["lemmas stopped"].apply(len) > 10
input_sents = input_sents[len_select]
if len(input_sents) <= 1: continue
stems = input_sents["stems"].apply(lambda x: ' '.join(x)).tolist()
X = semsim.transform(stems)
probs = input_sents["probs"]
p = probs.values
K = -(1 - cosine_similarity(X))
K_ma = np.ma.masked_array(K, np.eye(K.shape[0]))
Kmin = np.ma.min(K_ma)
Kmax = np.ma.max(K_ma)
median = np.ma.median(K_ma)[0]
pref = np.minimum(p + median, -.05)
print "SYS TIME:", hour, "# SENTS:", K.shape[0],
print "min/median/max pref: {}/{}/{}".format(
pref.min(), np.median(pref), pref.max())
#K_data.append({"min": Kmin, "max": Kmax, "median": median})
K_data.append({"min": (pref).min(), "max": (pref).max(),
"median": np.median((pref))})
#print K
# continue
#
ap = AffinityPropagation(preference=pref-pref_offset, affinity="precomputed",
verbose=True, max_iter=50000)
ap.fit(K)
# ##print input_sents["pretty text"]
#
labels = ap.labels_
if ap.cluster_centers_indices_ != None:
for c in ap.cluster_centers_indices_:
if cache == None:
cache = X[c]
updates_df = input_sents.reset_index(
drop=True).iloc[c]
updates_df["query id"] = event.query_num
updates_df["system timestamp"] = current_time
summary_data.append(
updates_df[
["query id", "stream id", "sent id",
"system timestamp", "sent text"]
].to_frame().T)
else:
Ksum = cosine_similarity(cache, X[c])
#print "MAX SIM", Ksum.max()
#print input_sents.reset_index(drop=True).iloc[c]["sent text"]
if Ksum.max() < sim_threshold:
cache = np.vstack([cache, X[c]])
updates_df = input_sents.reset_index(
drop=True).iloc[c]
updates_df["query id"] = event.query_num
updates_df["system timestamp"] = current_time
summary_data.append(
updates_df[
["query id", "stream id", "sent id",
"system timestamp", "sent text"]
].to_frame().T)
#
# for l, i in enumerate(af.cluster_centers_indices_):
# support = np.sum(labels == l)
# center = input_sents.iloc[i][["update id", "sent text", "pretty text", "stems", "nuggets"]]
# center = center.to_dict()
# center["support"] = support
# center["timestamp"] = current_time
# clusters.append(center)
#
prev_time = current_time
# df = pd.DataFrame(clusters, columns=["update id", "timestamp", "support", "sent text", "pretty text", "stems", "nuggets"])
#
# import os
# dirname = "clusters"
# if not os.path.exists(dirname):
# os.makedirs(dirname)
#
# with open(os.path.join(dirname, "{}.tsv".format(event.query_id)), "w") as f:
# df.to_csv(f, sep="\t", index=False)
#
df = pd.DataFrame(K_data, columns=["min", "max", "median"])
print df
print df.mean()
print df.std()
print df.max()
df = pd.concat(summary_data)
df["conf"] = .5
df["team id"] = "APSAL"
df["run id"] = "sim{}_bs{}_off{}".format(
sim_threshold, bucket_size, pref_offset)
print df
of = os.path.join(output_dir, "apsal" + "sim{}_bs{}_off{}.tsv".format(
sim_threshold, bucket_size, pref_offset))
cols = ["query id", "team id", "run id", "stream id", "sent id",
"system timestamp", "conf"]
df[cols].to_csv(of, sep="\t", header=False, index=False)
def main(output_dir, sim_threshold, bucket_size):
if not os.path.exists(output_dir):
os.makedirs(output_dir)
dev_qids = set([19, 23, 27, 34, 35])
summary_data = []
K_data = []
for event in cuttsum.events.get_events():
if event.query_num not in dev_qids: continue
print event
semsim = event2semsim(event)
istream = get_input_stream(event,
False,
extractor="goose",
thresh=.8,
delay=None,
topk=20)
prev_time = 0
cache = None
clusters = []
max_h = len(event.list_event_hours()) - 1
for h, hour in enumerate(event.list_event_hours()):
if h % bucket_size != 0 and h != max_h:
continue
current_time = epoch(hour)
input_sents = istream[
(istream["timestamp"] < current_time) & \
(istream["timestamp"] >= prev_time)]
len_select = input_sents["lemmas stopped"].apply(len) > 10
input_sents = input_sents[len_select]
if len(input_sents) <= 1: continue
stems = input_sents["stems"].apply(lambda x: ' '.join(x)).tolist()
X = semsim.transform(stems)
K = -(1 - cosine_similarity(X))
K_ma = np.ma.masked_array(K, np.eye(K.shape[0]))
Kmin = np.ma.min(K_ma)
Kmax = np.ma.max(K_ma)
median = np.ma.median(K_ma)[0]
print "SYS TIME:", hour, "# SENTS:", K.shape[0],
print "min/median/max pref: {}/{}/{}".format(Kmin, median, Kmax)
#
ap = AffinityPropagation(affinity="precomputed",
verbose=True,
max_iter=1000)
ap.fit(K)
labels = ap.labels_
if ap.cluster_centers_indices_ != None:
for c in ap.cluster_centers_indices_:
if cache == None:
cache = X[c]
updates_df = \
input_sents.reset_index(drop=True).iloc[c]
updates_df["query id"] = event.query_num
updates_df["system timestamp"] = current_time
summary_data.append(updates_df[[
"query id", "stream id", "sent id",
"system timestamp", "sent text"
]].to_frame().T)
else:
Ksum = cosine_similarity(cache, X[c])
if Ksum.max() < sim_threshold:
cache = np.vstack([cache, X[c]])
updates_df = \
input_sents.reset_index(drop=True).iloc[c]
updates_df["query id"] = event.query_num
updates_df["system timestamp"] = current_time
summary_data.append(updates_df[[
"query id", "stream id", "sent id",
"system timestamp", "sent text"
]].to_frame().T)
prev_time = current_time
df = pd.DataFrame(K_data, columns=["min", "max", "median"])
print df
print df.mean()
print df.std()
print df.max()
df = pd.concat(summary_data)
df["conf"] = .5
df["team id"] = "AP"
df["run id"] = "sim{}_bs{}".format(sim_threshold, bucket_size)
print df
of = os.path.join(
output_dir,
"ap." + "sim{}_bs{}.tsv".format(sim_threshold, bucket_size))
cols = [
"query id", "team id", "run id", "stream id", "sent id",
"system timestamp", "conf"
]
df[cols].to_csv(of, sep="\t", header=False, index=False)
def main(output_dir, sim_threshold, bucket_size, pref_offset):
if not os.path.exists(output_dir):
os.makedirs(output_dir)
dev_qids = set([19, 23, 27, 34, 35])
summary_data = []
K_data = []
for event in cuttsum.events.get_events():
if event.query_num not in dev_qids: continue
print event
semsim = event2semsim(event)
istream = get_input_stream(event, False, extractor="goose",
thresh=.8, delay=None, topk=20)
prev_time = 0
cache = None
clusters = []
max_h = len(event.list_event_hours()) - 1
for h, hour in enumerate(event.list_event_hours()):
if h % bucket_size != 0 and h != max_h:
continue
current_time = epoch(hour)
input_sents = istream[
(istream["timestamp"] < current_time) & \
(istream["timestamp"] >= prev_time)]
len_select = input_sents["lemmas stopped"].apply(len) > 10
input_sents = input_sents[len_select]
if len(input_sents) <= 1: continue
stems = input_sents["stems"].apply(lambda x: ' '.join(x)).tolist()
X = semsim.transform(stems)
probs = input_sents["probs"]
p = probs.values
K = -(1 - cosine_similarity(X))
K_ma = np.ma.masked_array(K, np.eye(K.shape[0]))
Kmin = np.ma.min(K_ma)
Kmax = np.ma.max(K_ma)
median = np.ma.median(K_ma)[0]
pref = np.minimum(p + median, -.05)
print "SYS TIME:", hour, "# SENTS:", K.shape[0],
print "min/median/max pref: {}/{}/{}".format(
pref.min(), np.median(pref), pref.max())
#K_data.append({"min": Kmin, "max": Kmax, "median": median})
K_data.append({"min": (pref).min(), "max": (pref).max(),
"median": np.median((pref))})
#print K
# continue
#
ap = AffinityPropagation(
preference=pref-pref_offset, affinity="precomputed",
verbose=True, max_iter=1000)
ap.fit(K)
# ##print input_sents["pretty text"]
#
labels = ap.labels_
if ap.cluster_centers_indices_ != None:
for c in ap.cluster_centers_indices_:
if cache == None:
cache = X[c]
updates_df = \
input_sents.reset_index(drop=True).iloc[c]
updates_df["query id"] = event.query_num
updates_df["system timestamp"] = current_time
summary_data.append(
updates_df[
["query id", "stream id", "sent id",
"system timestamp", "sent text"]
].to_frame().T
)
else:
Ksum = cosine_similarity(cache, X[c])
#print "MAX SIM", Ksum.max()
#print input_sents.reset_index(drop=True).iloc[c]["sent text"]
if Ksum.max() < sim_threshold:
cache = np.vstack([cache, X[c]])
updates_df = \
input_sents.reset_index(drop=True).iloc[c]
updates_df["query id"] = event.query_num
updates_df["system timestamp"] = current_time
summary_data.append(
updates_df[
["query id", "stream id", "sent id",
"system timestamp", "sent text"]
].to_frame().T
)
#
# for l, i in enumerate(af.cluster_centers_indices_):
# support = np.sum(labels == l)
# center = input_sents.iloc[i][["update id", "sent text", "pretty text", "stems", "nuggets"]]
# center = center.to_dict()
# center["support"] = support
# center["timestamp"] = current_time
# clusters.append(center)
#
prev_time = current_time
# df = pd.DataFrame(clusters, columns=["update id", "timestamp", "support", "sent text", "pretty text", "stems", "nuggets"])
#
# import os
# dirname = "clusters"
# if not os.path.exists(dirname):
# os.makedirs(dirname)
#
# with open(os.path.join(dirname, "{}.tsv".format(event.query_id)), "w") as f:
# df.to_csv(f, sep="\t", index=False)
#
df = pd.DataFrame(K_data, columns=["min", "max", "median"])
print df
print df.mean()
print df.std()
print df.max()
df = pd.concat(summary_data)
df["conf"] = .5
df["team id"] = "APSAL"
df["run id"] = "sim{}_bs{}_off{}".format(
sim_threshold, bucket_size, pref_offset)
print df
of = os.path.join(output_dir, "apsal" + "sim{}_bs{}_off{}.tsv".format(
sim_threshold, bucket_size, pref_offset))
cols = ["query id", "team id", "run id", "stream id", "sent id",
"system timestamp", "conf"]
df[cols].to_csv(of, sep="\t", header=False, index=False)
def get_classifier(self, event):
path = os.path.join(self.dir_, event.fs_name())
classifiers = []
semsim = event2semsim(event)
from nltk.stem.porter import PorterStemmer
stemmer = PorterStemmer()
if not os.path.exists(path):
return None
for nugget_id in os.listdir(path):
model = self.get_best_model(event, nugget_id)
if model is not None:
classifiers.append((nugget_id, model[0], model[1], model[2], model[3]))
# def remove_punctuation(text):
# return re.sub(
# ur"\p{P}+", "",
# text).lower().encode("utf-8")
def classify_nuggets(df):
sents = [" ".join(lemmas).lower() for lemmas in df["lemmas stopped"].tolist()]
all_stems = [' '.join(stems) for stems in df["stems"].tolist()]
sets = [set(sent.split(" ")) for sent in sents]
#sents = map(remove_punctuation, sents)
nuggets = [set() for sent in sents]
nugget_probs = [dict() for sent in sents]
all_probs = []
for nugget_id, vec, clf, nugget_lems, nugget_stems in classifiers:
if vec is not None and clf is not None:
X = vec.transform(sents).todense()
x_cov = [len(nugget_lems.intersection(set(lems))) / float(len(nugget_lems))
for lems in sets]
x_cov = np.array(x_cov)[:, np.newaxis]
K = cosine_similarity(semsim.transform(all_stems),
semsim.transform(nugget_stems))
X = np.hstack([X, x_cov, K, x_cov * K])
P = clf.predict_proba(X)
y = np.zeros(X.shape[0], dtype="int32")
#y[(P[:,1] > .95) | ((x_cov[:,0] > .75) & (len(nugget_lems) > 1))] = 1
for i in np.where(y == 1)[0]:
#if X[i].sum() < 3:
#if len(sents[i].split(" ")) < 6:
# continue
nuggets[i].add(nugget_id.encode("utf-8"))
for i in xrange(len(nugget_probs)):
#if P[i,1] > .5:
nugget_probs[i][nugget_id.encode("utf-8")] = P[i,1]
all_probs.append(P[:,1,np.newaxis])
#else:
#x_cov = [len(nugget_lems.intersection(set(lems))) / float(len(nugget_lems))
# for lems in sets]
#x_cov = np.array(x_cov)[:, np.newaxis]
#y = np.zeros(x_cov.shape[0], dtype="int32")
#y[((x_cov[:,0] > .75) & (len(nugget_lems) > 1))] = 1
#for i in np.where(y == 1)[0]:
# nuggets[i].add(nugget_id.encode("utf-8"))
#for i in xrange(len(nugget_probs)):
# if x_cov[i,0] > .5:
# nugget_probs[i][nugget_id.encode("utf-8")] = x_cov[i,0]
Pp = np.hstack(all_probs)
conf = np.max(Pp, axis=1)
return nuggets, conf, nugget_probs
#for sent in sents:
# sent
# return
return classify_nuggets
def do_job_unit(self, event, corpus, unit, **kwargs):
stopwords = english_stopwords()
### Preprocessing here. ###
df = cuttsum.judgements.get_merged_dataframe()
matches = df[df["query id"] == event.query_id]
matching_update_ids = set(matches["update id"].tolist())
#nuggets = matches.groupby("nugget id")
#thrsh_nuggets = all_nuggets.filter(lambda x: len(x) > 10)
nugget_ids = list(set(matches["nugget id"].tolist()))
#nugget_ids = list(set(all_nuggets["nugget id"].tolist()))
nugget_ids.sort()
nugget_id = nugget_ids[unit]
with corenlp.Server(port=9876 + event.query_num * 100 + unit, mem="20G", threads=4, max_message_len=524288,
annotators=["tokenize", "ssplit", "pos", "lemma"], #, "ner"],
corenlp_props={
"pos.maxlen": 50, "ssplit.eolonly": "true"}) as pipeline:
if event.query_id.startswith("TS13"):
updates = cuttsum.judgements.get_2013_updates()
elif event.query_id.startswith("TS14"):
updates = cuttsum.judgements.get_2014_sampled_updates()
elif event.query_id.startswith("TS15"):
updates = cuttsum.judgements.get_2015_sampled_updates()
updates = updates[updates["query id"] == event.query_id]
non_matching_updates = updates[updates["update id"].apply(
lambda x: x not in matching_update_ids)]
matching_updates = matches[matches["nugget id"] == nugget_id]
# if len(matching_updates) == 0:
# return
#matching_updates = df[df["nugget id"] == nugget_id]
nugget_text = matching_updates.iloc[0]["nugget text"]
n_matching = len(matching_updates)
n_nonmatching = min(n_matching, len(non_matching_updates))
n_instances = n_matching + n_nonmatching
semsim = event2semsim(event)
from nltk.stem.porter import PorterStemmer
stemmer = PorterStemmer()
nugget_doc = pipeline.annotate(nugget_text)
nugget_lems = []
nugget_stems = []
for sent in nugget_doc:
for tok in sent:
if unicode(tok).lower() not in stopwords and len(unicode(tok)) < 50:
nugget_lems.append(tok.lem.lower())
stem = stemmer.stem(unicode(tok).lower())
if len(stem) < 50:
nugget_stems.append(stem)
nugget_stems = [u" ".join(nugget_stems)]
if n_matching <= 10:
model_dir = self.get_model_dir(event, nugget_id)
if not os.path.exists(model_dir):
os.makedirs(model_dir)
joblib.dump([None, set(nugget_lems), nugget_stems], self.get_vectorizer_path(event, nugget_id), compress=9)
joblib.dump([], self.get_model_path(event, nugget_id, "gbc"), compress=9)
return
non_matching_updates = non_matching_updates.iloc[
np.random.permutation(len(non_matching_updates))]
non_matching_updates = non_matching_updates.iloc[
np.arange(n_nonmatching)]
#non_matching_updates["text"] = \
# non_matching_updates["text"].apply(lambda x: x.lower())
y = np.zeros(n_instances, dtype="int32")
y[:n_matching] = 1
X_string = matching_updates["update text"].tolist()
X_string += non_matching_updates.head(n_nonmatching)["text"].tolist()
assert len(X_string) == n_instances
p = np.random.permutation(n_instances)
y = y[p]
X_string = [X_string[i] for i in p]
print "pipeline start"
docs = pipeline.annotate_mp(X_string, n_procs=4)
print "pipeline done"
lemmas = []
all_stems = []
for doc in docs:
lems = []
stems = []
for sent in doc:
for tok in sent:
if unicode(tok).lower() not in stopwords and len(unicode(tok)) < 50:
lems.append(tok.lem.lower())
stem = stemmer.stem(unicode(tok).lower())
if len(stem) < 50:
stems.append(stem)
#print lems
lemmas.append(lems)
all_stems.append(u" ".join(stems))
# map(
# lambda doc: [str(tok)
# for doc in docs
# for sent in doc
# for tok in sent
K = cosine_similarity(
semsim.transform(all_stems),
semsim.transform(nugget_stems))
X_string = [u" ".join(lem) for lem in lemmas]
vec = TfidfVectorizer(
input=u"content", stop_words="english", ngram_range=(1,5))
vec.fit([u" ".join(nugget_lems)] + X_string)
X = vec.transform(X_string).todense()
nugget_lems = set(nugget_lems)
x_cov = [len(nugget_lems.intersection(set(lems))) / float(len(nugget_lems))
for lems in lemmas]
x_cov = np.array(x_cov)[:, np.newaxis]
X = np.hstack([X, x_cov, K, K * x_cov])
gbc = GradientBoostingClassifier(
n_estimators=500, learning_rate=.1,
max_depth=8, random_state=0, max_features="log2")
gbc.fit(X, y)
print "SCORE", gbc.score(X, y)
model_dir = self.get_model_dir(event, nugget_id)
if not os.path.exists(model_dir):
os.makedirs(model_dir)
joblib.dump([vec, nugget_lems, nugget_stems],
self.get_vectorizer_path(event, nugget_id), compress=9)
joblib.dump(
gbc, self.get_model_path(event, nugget_id, "gbc"), compress=9)
def do_job_unit(self, event, corpus, unit, **kwargs):
if unit != 0:
raise Exception("Job unit {} out of range".format(unit))
service_configs = kwargs.get("service-configs", {})
cnlp_configs = service_configs.get("corenlp", {})
cnlp_port = int(cnlp_configs.get("port", 9999))
domain_lm_config = service_configs[event2lm_name(event)]
domain_lm_port = int(domain_lm_config["port"])
domain_lm_order = int(domain_lm_config.get("order", 3))
gw_lm_config = service_configs["gigaword-lm"]
gw_lm_port = int(gw_lm_config["port"])
gw_lm_order = int(gw_lm_config.get("order", 3))
thresh = kwargs.get("dedupe-sim-threshold", .8)
extractor = kwargs.get("extractor", "goose")
res = DedupedArticlesResource()
dfiter = res.dataframe_iter(
event, corpus, extractor, include_matches=None,
threshold=thresh)
domain_lm = cuttsum.srilm.Client(domain_lm_port, domain_lm_order, True)
gw_lm = cuttsum.srilm.Client(gw_lm_port, gw_lm_order, True)
cnlp_client = cnlp.client.CoreNLPClient(port=cnlp_port)
def make_query_synsets():
synonyms = []
hypernyms = []
hyponyms = []
print event.type.split(' ')[0]
for synset in wn.synsets(event.type.split(' ')[0]):
synonyms.extend(
[lemma.name().lower().replace(u'_', u' ').encode(u'utf-8')
for lemma in synset.lemmas()])
hypernyms.extend(
[lemma.name().lower().replace(u'_', u' ').encode(u'utf-8')
for synset in synset.hypernyms()
for lemma in synset.lemmas()])
hyponyms.extend(
[lemma.name().lower().replace(u'_', u' ').encode(u'utf-8')
for synset in synset.hyponyms()
for lemma in synset.lemmas()])
print hypernyms
print hyponyms
print synonyms
return set(synonyms), set(hypernyms), set(hyponyms)
def heal_text(sent_text):
sent_text = re.sub(
ur"[A-Z ]+, [A-Z][a-z ]+\( [A-Z]+ \) [-\u2014_]+ ",
r"", sent_text)
sent_text = re.sub(
ur"^.*?[A-Z ]+, [A-Z][a-z]+ [-\u2014_]+ ",
r"", sent_text)
sent_text = re.sub(
ur"^.*?[A-Z ]+\([^\)]+\) [-\u2014_]+ ",
r"", sent_text)
sent_text = re.sub(
ur"^.*?[A-Z]+ +[-\u2014_]+ ",
r"", sent_text)
sent_text = re.sub(r"\([^)]+\)", r" ", sent_text)
sent_text = re.sub(ur"^ *[-\u2014_]+", r"", sent_text)
sent_text = re.sub(u" ([,.;?!]+)([\"\u201c\u201d'])", r"\1\2", sent_text)
sent_text = re.sub(r" ([:-]) ", r"\1", sent_text)
sent_text = re.sub(r"([^\d]\d{1,3}) , (\d\d\d)([^\d]|$)", r"\1,\2\3", sent_text)
sent_text = re.sub(r"^(\d{1,3}) , (\d\d\d)([^\d]|$)", r"\1,\2\3", sent_text)
sent_text = re.sub(ur" ('|\u2019) ([a-z]|ll|ve|re)( |$)", r"\1\2 ", sent_text)
sent_text = re.sub(r" ([',.;?!]+) ", r"\1 ", sent_text)
sent_text = re.sub(r" ([',.;?!]+)$", r"\1", sent_text)
sent_text = re.sub(r"(\d\.) (\d)", r"\1\2", sent_text)
sent_text = re.sub(r"(a|p)\. m\.", r"\1.m.", sent_text)
sent_text = re.sub(r"U\. (S|N)\.", r"U.\1.", sent_text)
sent_text = re.sub(
ur"\u201c ([^\s])",
ur"\u201c\1", sent_text)
sent_text = re.sub(
ur"([^\s]) \u201d",
ur"\1\u201d", sent_text)
sent_text = re.sub(
ur"\u2018 ([^\s])",
ur"\u2018\1", sent_text)
sent_text = re.sub(
ur"([^\s]) \u2019",
ur"\1\u2019", sent_text)
sent_text = re.sub(
ur"\u00e2",
ur"'", sent_text)
sent_text = re.sub(
r"^Photo:Reuters|^Photo:AP",
r"", sent_text)
sent_text = sent_text.replace("\n", " ")
return sent_text.encode("utf-8")
def get_number_feats(sent):
feats = []
for tok in sent:
if tok.ne == "NUMBER" and tok.nne is not None:
for chain in get_dep_chain(tok, sent, 0):
feat = [tok.nne] + [elem[1].lem for elem in chain]
feats.append(feat)
return feats
def get_dep_chain(tok, sent, depth):
chains = []
if depth > 2:
return chains
for p in sent.dep2govs[tok]:
if p[1].is_noun():
for chain in get_dep_chain(p[1], sent, depth + 1):
chains.append([p] + chain)
elif p[1]:
chains.append([p])
return chains
import unicodedata as u
P=''.join(unichr(i) for i in range(65536) if u.category(unichr(i))[0]=='P')
P = re.escape(P)
punc_patt = re.compile("[" + P + "]")
from collections import defaultdict
stopwords = english_stopwords()
mention_counts = defaultdict(int)
total_mentions = 0
from nltk.stem.porter import PorterStemmer
stemmer = PorterStemmer()
synonyms, hypernyms, hyponyms = make_query_synsets()
path = self.get_path(
event, corpus, extractor, thresh)
dirname = os.path.dirname(path)
if not os.path.exists(dirname): os.makedirs(dirname)
meta_cols = ["update id", "stream id", "sent id", "timestamp",
"pretty text", "tokens", "lemmas", "stems", "pos", "ne",
"tokens stopped", "lemmas stopped"]
basic_cols = ["BASIC length", "BASIC char length",
"BASIC doc position", "BASIC all caps ratio",
"BASIC upper ratio", "BASIC lower ratio",
"BASIC punc ratio", "BASIC person ratio",
"BASIC location ratio",
"BASIC organization ratio", "BASIC date ratio",
"BASIC time ratio", "BASIC duration ratio",
"BASIC number ratio", "BASIC ordinal ratio",
"BASIC percent ratio", "BASIC money ratio",
"BASIC set ratio", "BASIC misc ratio"]
lm_cols = ["LM domain lp", "LM domain avg lp",
"LM gw lp", "LM gw avg lp"]
query_cols = [
"Q_query_sent_cov",
"Q_sent_query_cov",
"Q_syn_sent_cov",
"Q_sent_syn_cov",
"Q_hyper_sent_cov",
"Q_sent_hyper_cov",
"Q_hypo_sent_cov",
"Q_sent_hypo_cov",
]
sum_cols = [
"SUM_sbasic_sum",
"SUM_sbasic_amean",
"SUM_sbasic_max",
"SUM_novelty_gmean",
"SUM_novelty_amean",
"SUM_novelty_max",
"SUM_centrality",
"SUM_pagerank",
"SUM_sem_novelty_gmean",
"SUM_sem_novelty_amean",
"SUM_sem_novelty_max",
"SUM_sem_centrality",
"SUM_sem_pagerank",
]
stream_cols = [
"STREAM_sbasic_sum",
"STREAM_sbasic_amean",
"STREAM_sbasic_max",
"STREAM_per_prob_sum",
"STREAM_per_prob_max",
"STREAM_per_prob_amean",
"STREAM_loc_prob_sum",
"STREAM_loc_prob_max",
"STREAM_loc_prob_amean",
"STREAM_org_prob_sum",
"STREAM_org_prob_max",
"STREAM_org_prob_amean",
"STREAM_nt_prob_sum",
"STREAM_nt_prob_max",
"STREAM_nt_prob_amean",
]
semsim = event2semsim(event)
all_cols = meta_cols + basic_cols + query_cols + lm_cols + sum_cols + stream_cols
stream_uni_counts = defaultdict(int)
stream_per_counts = defaultdict(int)
stream_loc_counts = defaultdict(int)
stream_org_counts = defaultdict(int)
stream_nt_counts = defaultdict(int)
with gzip.open(path, "w") as f:
f.write("\t".join(all_cols) + "\n")
for df in dfiter:
if len(df) == 1: continue
df = df.head(20)
#df["lm"] = df["sent text"].apply(lambda x: lm.sentence_log_prob(x.encode("utf-8"))[1])
df["pretty text"] = df["sent text"].apply(heal_text)
df = df[df["pretty text"].apply(lambda x: len(x.strip())) > 0]
df = df[df["pretty text"].apply(lambda x: len(x.split(" "))) < 200]
df = df.reset_index(drop=True)
if len(df) == 0:
print "skipping"
continue
doc_text = "\n".join(df["pretty text"].tolist())
doc = cnlp_client.annotate(doc_text)
df["tokens"] = map(lambda sent: [str(tok) for tok in sent],
doc)
df["lemmas"] = map(lambda sent: [tok.lem.encode("utf-8")
for tok in sent],
doc)
df["stems"] = map(lambda sent:
[stemmer.stem(unicode(tok).lower()) for tok in sent], doc)
df["pos"] = map(lambda sent: [tok.pos for tok in sent],
doc)
df["ne"] = map(lambda sent: [tok.ne for tok in sent],
doc)
df["tokens stopped"] = map(
lambda sent: [str(tok) for tok in sent
if unicode(tok).lower() not in stopwords \
and len(unicode(tok)) < 50],
doc)
df["lemmas stopped"] = map(
lambda sent: [tok.lem.lower().encode("utf-8") for tok in sent
if unicode(tok).lower() not in stopwords \
and len(unicode(tok)) < 50],
doc)
df["num tuples"] = [get_number_feats(sent) for sent in doc]
### Basic Features ###
df["BASIC length"] = df["lemmas stopped"].apply(len)
df["BASIC doc position"] = df.index.values + 1
df = df[df["BASIC length"] > 0]
df = df.reset_index(drop=True)
df["BASIC char length"] = df["pretty text"].apply(
lambda x: len(x.replace(" ", "")))
df["BASIC upper ratio"] = df["pretty text"].apply(
lambda x: len(re.findall("[A-Z]", x))) \
/ df["BASIC char length"].apply(lambda x: float(max(x, 1)))
df[ "BASIC lower ratio"] = df["pretty text"].apply(
lambda x: len(re.findall("[a-z]", x))) \
/ df["BASIC char length"].apply(lambda x: float(max(x, 1)))
df["BASIC punc ratio"] = df["pretty text"].apply(
lambda x: len(re.findall(punc_patt, x))) \
/ df["BASIC char length"].apply(lambda x: float(max(x, 1)))
df["BASIC all caps ratio"] = df["tokens stopped"].apply(
lambda x: np.sum([1 if re.match("^[A-Z]+$", xi) else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC person ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "PERSON" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC location ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "LOCATION" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC organization ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "ORGANIZATION" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC date ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "DATE" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC time ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "TIME" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC duration ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "DURATION" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC number ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "NUMBER" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC ordinal ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "ORDINAL" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC percent ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "PERCENT" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC money ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "MONEY" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC set ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "SET" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
df["BASIC misc ratio"] = df["ne"].apply(
lambda x: np.sum([1 if xi == "MISC" else 0
for xi in x])) \
/ df["BASIC length"].apply(float)
### Language Model Features ###
dm_probs = df["lemmas"].apply(
lambda x: domain_lm.sentence_log_prob(
" ".join([xi.decode("utf-8").lower().encode("utf-8")
for xi in x if len(xi) < 50])))
dm_log_probs = [lp for lp, avg_lp in dm_probs.tolist()]
dm_avg_log_probs = [avg_lp for lp, avg_lp in dm_probs.tolist()]
df["LM domain lp"] = dm_log_probs
df["LM domain avg lp"] = dm_avg_log_probs
gw_probs = df["lemmas"].apply(
lambda x: gw_lm.sentence_log_prob(
" ".join([xi.decode("utf-8").lower().encode("utf-8")
for xi in x if len(xi) < 50])))
gw_log_probs = [lp for lp, avg_lp in gw_probs.tolist()]
gw_avg_log_probs = [avg_lp for lp, avg_lp in gw_probs.tolist()]
df["LM gw lp"] = gw_log_probs
df["LM gw avg lp"] = gw_avg_log_probs
### Query Features ###
self.compute_query_features(df,
set([q.lower() for q in event.query]),
synonyms, hypernyms, hyponyms)
### Single Doc Summarization Features ###
counts = []
doc_counts = defaultdict(int)
for lemmas in df["lemmas stopped"].tolist():
counts_i = {}
for lem in lemmas:
counts_i[lem.lower()] = counts_i.get(lem.lower(), 0) + 1
doc_counts[lem.lower()] += 1
doc_counts["__TOTAL__"] += len(lemmas)
counts.append(counts_i)
doc_counts["__TOTAL__"] *= 1.
doc_uni = {key: val / doc_counts["__TOTAL__"]
for key, val in doc_counts.items()
if key != "__TOTAL__"}
sum_probs = []
amean_probs = []
max_probs = []
for lemmas in df["lemmas stopped"].tolist():
probs = [doc_uni[lem.lower()] for lem in lemmas]
sum_probs.append(np.sum(probs))
amean_probs.append(np.mean(probs))
max_probs.append(np.max(probs))
df["SUM_sbasic_sum"] = sum_probs
df["SUM_sbasic_amean"] = amean_probs
df["SUM_sbasic_max"] = max_probs
tfidfer = TfidfTransformer()
vec = DictVectorizer()
X = vec.fit_transform(counts)
X = tfidfer.fit_transform(X)
ctrd = X.mean(axis=0)
K = cosine_similarity(ctrd, X).ravel()
I = K.argsort()[::-1]
R = np.array([[i, r + 1] for r, i in enumerate(I)])
R = R[R[:,0].argsort()]
df["SUM_centrality"] = R[:,1]
L = semsim.transform(df["stems"].apply(lambda x: ' '.join(x)).tolist())
ctrd_l = L.mean(axis=0)
K_L = cosine_similarity(ctrd_l, L).ravel()
I_L = K_L.argsort()[::-1]
R_L = np.array([[i, r + 1] for r, i in enumerate(I_L)])
R_L = R_L[R_L[:, 0].argsort()]
df["SUM_sem_centrality"] = R_L[:,1]
K = cosine_similarity(X)
M = np.zeros_like(K)
M[np.diag_indices(K.shape[0])] = 1
Km = np.ma.masked_array(K, M)
D = 1 - Km
novelty_amean = D.mean(axis=1)
novelty_max = D.max(axis=1)
novelty_gmean = gmean(D, axis=1)
df["SUM_novelty_amean"] = novelty_amean
df["SUM_novelty_max"] = novelty_max
df["SUM_novelty_gmean"] = novelty_gmean
K_L = cosine_similarity(L)
M_L = np.zeros_like(K)
M_L[np.diag_indices(K_L.shape[0])] = 1
K_Lm = np.ma.masked_array(K_L, M_L)
D_L = 1 - K_Lm
sem_novelty_amean = D_L.mean(axis=1)
sem_novelty_max = D_L.max(axis=1)
sem_novelty_gmean = gmean(D_L, axis=1)
df["SUM_sem_novelty_amean"] = sem_novelty_amean
df["SUM_sem_novelty_max"] = sem_novelty_max
df["SUM_sem_novelty_gmean"] = sem_novelty_gmean
K = (K > 0).astype("int32")
degrees = K.sum(axis=1) - 1
edges_x_2 = K.sum() - K.shape[0]
if edges_x_2 == 0: edges_x_2 = 1
pr = 1. - degrees / float(edges_x_2)
df["SUM_pagerank"] = pr
K_L = (K_L > .2).astype("int32")
degrees_L = K_L.sum(axis=1) - 1
edges_x_2_L = K_L.sum() - K_L.shape[0]
if edges_x_2_L == 0: edges_x_2_L = 1
pr_L = 1. - degrees_L / float(edges_x_2_L)
df["SUM_sem_pagerank"] = pr_L
print df["pretty text"]
# print df[["SUM_sbasic_sum", "SUM_sbasic_amean", "SUM_sbasic_max"]]
# print df[
# ["SUM_pagerank", "SUM_centrality", "SUM_novelty_gmean",
# "SUM_novelty_amean", "SUM_novelty_max"]]
### Stream Features ###
for key, val in doc_counts.items():
stream_uni_counts[key] += val
denom = stream_uni_counts["__TOTAL__"]
sum_probs = []
amean_probs = []
max_probs = []
for lemmas in df["lemmas stopped"].tolist():
probs = [stream_uni_counts[lem.lower()] / denom for lem in lemmas]
sum_probs.append(np.sum(probs))
amean_probs.append(np.mean(probs))
max_probs.append(np.max(probs))
df["STREAM_sbasic_sum"] = sum_probs
df["STREAM_sbasic_amean"] = amean_probs
df["STREAM_sbasic_max"] = max_probs
for lemmas, nes in izip(df["lemmas"].tolist(), df["ne"].tolist()):
for lem, ne in izip(lemmas, nes):
if ne == "PERSON":
stream_per_counts[lem.lower()] += 1
stream_per_counts["__TOTAL__"] += 1.
if ne == "LOCATION":
stream_loc_counts[lem.lower()] += 1
stream_loc_counts["__TOTAL__"] += 1.
if ne == "ORGANIZATION":
stream_org_counts[lem.lower()] += 1
stream_org_counts["__TOTAL__"] += 1.
for tuples in df["num tuples"].tolist():
for nt in tuples:
for item in nt:
stream_nt_counts[item.lower()] += 1
stream_nt_counts["__TOTAL__"] += 1.
pdenom = stream_per_counts["__TOTAL__"]
ldenom = stream_loc_counts["__TOTAL__"]
odenom = stream_org_counts["__TOTAL__"]
ntdenom = stream_nt_counts["__TOTAL__"]
sum_per_probs = []
amean_per_probs = []
max_per_probs = []
sum_loc_probs = []
amean_loc_probs = []
max_loc_probs = []
sum_org_probs = []
amean_org_probs = []
max_org_probs = []
sum_nt_probs = []
amean_nt_probs = []
max_nt_probs = []
for tuples in df["num tuples"].tolist():
if ntdenom > 0:
nt_probs = [stream_nt_counts[item.lower()] / ntdenom
for nt in tuples
for item in nt]
else:
nt_probs = []
if len(nt_probs) > 0:
sum_nt_probs.append(np.sum(nt_probs))
amean_nt_probs.append(np.mean(nt_probs))
max_nt_probs.append(np.max(nt_probs))
else:
sum_nt_probs.append(0)
amean_nt_probs.append(0)
max_nt_probs.append(0)
for lemmas, nes in izip(df["lemmas"].tolist(), df["ne"].tolist()):
if pdenom > 0:
per_probs = [stream_per_counts[lem.lower()] / pdenom
for lem, ne in izip(lemmas, nes)
if ne == "PERSON"]
else:
per_probs = []
if len(per_probs) > 0:
sum_per_probs.append(np.sum(per_probs))
amean_per_probs.append(np.mean(per_probs))
max_per_probs.append(np.max(per_probs))
else:
sum_per_probs.append(0)
amean_per_probs.append(0)
max_per_probs.append(0)
if ldenom > 0:
loc_probs = [stream_loc_counts[lem.lower()] / ldenom
for lem, ne in izip(lemmas, nes)
if ne == "LOCATION"]
else:
loc_probs = []
if len(loc_probs) > 0 :
sum_loc_probs.append(np.sum(loc_probs))
amean_loc_probs.append(np.mean(loc_probs))
max_loc_probs.append(np.max(loc_probs))
else:
sum_loc_probs.append(0)
amean_loc_probs.append(0)
max_loc_probs.append(0)
if odenom > 0:
org_probs = [stream_org_counts[lem.lower()] / odenom
for lem, ne in izip(lemmas, nes)
if ne == "ORGANIZATION"]
else:
org_probs = []
if len(org_probs) > 0 :
sum_org_probs.append(np.sum(org_probs))
amean_org_probs.append(np.mean(org_probs))
max_org_probs.append(np.max(org_probs))
else:
sum_org_probs.append(0)
amean_org_probs.append(0)
max_org_probs.append(0)
df["STREAM_per_prob_sum"] = sum_per_probs
df["STREAM_per_prob_max"] = max_per_probs
df["STREAM_per_prob_amean"] = amean_per_probs
df["STREAM_loc_prob_sum"] = sum_loc_probs
df["STREAM_loc_prob_max"] = max_loc_probs
df["STREAM_loc_prob_amean"] = amean_loc_probs
df["STREAM_org_prob_sum"] = sum_org_probs
df["STREAM_org_prob_max"] = max_org_probs
df["STREAM_org_prob_amean"] = amean_org_probs
df["STREAM_nt_prob_sum"] = sum_nt_probs
df["STREAM_nt_prob_max"] = max_nt_probs
df["STREAM_nt_prob_amean"] = amean_nt_probs
#print df[["STREAM_sbasic_sum", "STREAM_sbasic_amean", "STREAM_sbasic_max"]]
#print df[["STREAM_per_prob_sum", "STREAM_per_prob_amean", "STREAM_per_prob_max"]]
#print df[["STREAM_loc_prob_sum", "STREAM_loc_prob_amean", "STREAM_loc_prob_max"]]
#print df[["STREAM_nt_prob_sum", "STREAM_nt_prob_amean", "STREAM_nt_prob_max"]]
### Write dataframe to file ###
df[all_cols].to_csv(f, index=False, header=False, sep="\t")