def exe_extract_feature(argv):
window_path, doc_path, topic_path, judge_path, word_stat_path, out_path = argv;
judge_file = QRelFile(judge_path);
topics = StandardFormat().read(topic_path);
global window_db, doc_db, word_stat, model_factory;
window_db = bsddb.hashopen(window_path);
doc_db = bsddb.hashopen(doc_path);
word_stat = load_word_stat(word_stat_path);
model_factory = DocumentModelFactory(word_stat);
writer = open(out_path, 'w');
global topic_chain, window_chain, doc_chain;
topic_chain = TextChain([TextTokenizer(word_tokenize), TextStopRemover(stop_path), TextStemmer(stemmer), TextModeler(model_factory)]);
window_chain = topic_chain;
doc_chain = TextChain([TextTokenizer(word_tokenize),TextStopRemover(stop_path), TextStemmer(stemmer), TextModeler(model_factory), WindowWorker(window_chain), DocumentTitleWorker(topic_chain)])
global topic_id;
topic_ids = judge_file.keys();
for topic_id in topic_ids:
if not topics.has_key(topic_id):
continue;
topic_str = topics[topic_id];
print topic_id;
global topic;
topic = TextPiece(topic_str);
topic_chain.work(topic);
p = Pool(task_num);
lines_group = p.map(multithread_extract_feature, judge_file[topic_id].items());
for lines in lines_group:
for line in lines:
writer.write(line);
writer.write('\n');
writer.close();
def exe_extract_windows(argv):
topic_path, judge_path, text_db_path, windows_db_path = argv;
text_db = bsddb.hashopen(text_db_path);
window_db = bsddb.hashopen(windows_db_path, 'w');
judge_file = QRelFile(judge_path);
topics = StandardFormat().read(topic_path);
topic_chain = TextChain([TextTokenizer(word_tokenize), TextStopRemover('data/stoplist.dft'), TextStemmer(EnglishStemmer()), TextTokenNormalizer()]);
sentence_chain = TextChain([TextTokenizer(word_tokenize), TextStemmer(EnglishStemmer()), TextTokenNormalizer()]);
for topic_id, topic_str in topics.items():
print topic_id;
sys.stdout.flush();
topic = TextPiece(topic_str);
topic_chain.work(topic);
if not judge_file.has_key(topic_id):
continue;
docnos = judge_file[topic_id].keys();
for docno in docnos:
if not is_cluewebB(docno):
continue;
doc_text = text_db[docno];
window_candidates = match_window(topic, doc_text, sentence_chain);
sentences = map(lambda text_piece: text_piece.text, window_candidates);
text = '\n'.join(sentences);
window_db[docno] = text.encode('utf8');
window_db.close();
def exe_build_train(argv):
#1. create the workers;
judge_path, topic_path, word_stat_path, doc_path, window_path, out_path = argv;
global judge_file, topics, doc_db, window_db, word_stat, ranker;
judge_file = QRelFile(judge_path);
topics = StandardFormat().read(topic_path);
doc_db = bsddb.hashopen(doc_path);
window_db = bsddb.hashopen(window_path);
word_stat = load_word_stat(word_stat_path);
# aggregators = map(lambda k: Aggregator(k), K_options);
# ranker = DistanceWindowRanker(CosTextScorer(), DocumentModelFactory(word_stat),aggregators);
ranker = RetrievalWindowRanker(CosTextScorer(), DocumentModelFactory(word_stat));
#2. build the training data;
# p = Pool(4);
topic_ids = judge_file.keys();
# docs_groups = p.map(build_train, topic_ids);
docs_groups = map(build_train, topic_ids);
assert len(docs_groups) == len(topic_ids);
#3. write out the training data
writer = open(out_path, 'w');
for i in xrange(len(topic_ids)):
topic_id = topic_ids[i];
docs = docs_groups[i];
for doc in docs:
docno = doc.docno;
judge = judge_file[topic_id][docno];
for scores, sentence_id in doc.score_windows:
score_str = ','.join(map(str, scores));
writer.write('%s %s %s %d %s\n' % (topic_id, docno, judge, sentence_id, score_str));
writer.close();
def test_extract_text(judge_path, index_path):
judge_file = QRelFile(judge_path);
docnos = judge_file.key2s();
print 'doc number:', len(docnos);
for docno in filter(is_cluewebB, docnos)[:3]:
text = extract_text(docno, index_path);
print text
print '-' * 20
def test_extract_text(judge_path, index_path, collection_type):
judge_file = QRelFile(judge_path);
docnos = judge_file.key2s();
print 'doc number:', len(docnos);
for docno in docnos[:1]:
text = extract_text(docno, index_path, collection_type);
print text
print '-' * 20
def exe_extract_text(judge_path, index_path, out_path, collection_type = 'html'):
'''
extract texts of docs in qrel from an index, and store them in out_path in standard trec format
'''
import Corpus
judge_file = QRelFile(judge_path);
docnos = judge_file.key2s();
print 'doc number:', len(docnos);
writer = Corpus.TRECWriter(out_path);
for docno in docnos:
text = extract_text(docno, index_path, collection_type)
writer.write(Corpus.Document(docno, text))
def exe_extract_text(judge_path, index_path, text_db_path):
judge_file = QRelFile(judge_path);
docnos = judge_file.key2s();
docnos = filter(is_cluewebB, docnos);
#docnos = docnos[:1000];
print 'doc number:', len(docnos);
db = bsddb.hashopen(text_db_path, 'w');
count = 0;
texts = fastmap.fastmap(lambda docno: extract_text(docno, index_path), 30, docnos);
assert len(docnos) == len(texts);
for i in xrange(len(docnos)):
db[docnos[i]] = texts[i];
db.close();
def exe_stat_window(qrel_path, window_db_path):
window_db = bsddb.hashopen(window_db_path);
qrel = QRelFile(qrel_path);
sentence_nums = [];
sentence_lens = [];
for q in qrel.keys():
for d in qrel.get(q).keys():
if window_db.has_key(d):
window = window_db[d];
sentences = window.split('\n');
sentence_nums.append(len(sentences));
sentence_lens += map(lambda sentence: len(sentence.split()), sentences);
print np.mean(sentence_nums), np.median(sentence_nums), np.mean(sentence_lens), np.median(sentence_lens);
def exe_example(snippet_judge_path, doc_judge_path):
#def exe_example(snippet_judge_path, doc_judge_path, bing_path, sum_path, dsm_path):
from JudgeFile import QRelFile;
snippet_judge = load_snippet_judge(snippet_judge_path);
doc_judge = QRelFile(doc_judge_path);
sources = snippet_judge.keys();
for topic_id in snippet_judge[sources[0]].keys():
for docno in snippet_judge[sources[0]][topic_id]:
in_rel = int(doc_judge.get_value(topic_id, docno));
if in_rel <= 0:
in_rel = 0;
elif in_rel > 0:
in_rel = 1;
if snippet_judge['bing'].has_key(topic_id) and snippet_judge['pablo.short'].has_key(topic_id) and snippet_judge['windowshop.oq'].has_key(topic_id) :
bing_per_rel = int(snippet_judge['bing'][topic_id][docno]);
sum_per_rel = int(snippet_judge['pablo.short'][topic_id][docno]);
oq_per_rel = int(snippet_judge['windowshop.oq'][topic_id][docno]);
if (in_rel <> bing_per_rel or in_rel <> sum_per_rel) and in_rel == oq_per_rel == 1:
print topic_id, docno, in_rel, bing_per_rel, sum_per_rel, oq_per_rel;
import sys
from JudgeFile import QRelFile
from TRECTopics import StandardFormat
from TrainGenerator import *;
topics = StandardFormat().read(sys.argv[1]);
judge_file = QRelFile(sys.argv[2])
lemmas = set()
topic_ids = judge_file.keys()
for topic_id in topic_ids:
if not topics.has_key(topic_id):
continue
topic_str = topics[topic_id]
topic = TextPiece(topic_str)
lemmas.update(topic.lemmas)
for lemma in lemmas:
print lemma
