def print_statistics(self, methods):
single_queries = Query(self.collection_path).get_queries_of_length(1)
queries = {ele['num']:ele['title'] for ele in single_queries}
cs = CollectionStats(self.collection_path)
performance = Performances(self.collection_path)
res = performance.gen_optimal_performances_queries(methods, queries.keys())
avdl = cs.get_avdl()
total_terms = cs.get_total_terms()
collection_freq = []
for qid in queries:
idx = 0
ctf = cs.get_term_collection_occur(queries[qid])
idf = cs.get_term_logidf1(queries[qid])
collection_freq.append( ctf*1.0/total_terms )
print avdl
print np.mean(collection_freq)
for ele in res:
label = ele[0]
p = ele[1]
para = float(ele[2].split(':')[1])
print label
if 'okapi' in label:
print 'b:', para, 'beta:', 1.2*para/avdl, 'c2:', 1.2*(1-para)
if 'pivoted' in label:
print 's:', para, 'beta:', para/avdl, 'c2:', 1-para
def gen_ranking_list(self, method, _callback, paras):
"""
We get the statistics from /collection_path/detailed_doc_stats/
so that we can get everything for the top 10,000 documents for
each query generated by Dirichlet language model method.
"""
single_queries = Query(self.collection_path).get_queries_of_length(1)
queries = {ele['num']:ele['title'] for ele in single_queries}
doc_details = GenSqaDocDetails(self.collection_path)
cs = CollectionStats(self.collection_path)
avdl = cs.get_avdl()
total_terms = cs.get_total_terms()
res = {}
for qid in queries:
print queries[qid]
res[qid] = []
idx = 0
ctf = cs.get_term_collection_occur(queries[qid])
idf = cs.get_term_logidf1(queries[qid])
#for row in cs.get_qid_details(qid):
for row in doc_details.get_qid_details(qid):
docid = row['docid']
total_tf = float(row['total_tf'])
doc_len = float(row['doc_len'])
localpara = copy.deepcopy(paras)
localpara.extend([total_tf, doc_len, avdl, ctf, total_terms, idf])
score = _callback(localpara)
res[qid].append((docid, score))
idx += 1
if idx >= 1000:
break
self.output_results(res, method)
self.eval(method)
def process(self, qid, method_name, method_paras, output_fn):
cs = CollectionStats(self.collection_path)
single_queries = Query(self.collection_path).get_queries_of_length(1)
queries = {ele['num']:ele['title'] for ele in single_queries}
#print qids
self.rel_docs = Judgment(self.collection_path).get_relevant_docs_of_some_queries([qid], 1, 'dict')
# idfs = [(qid, math.log(cs.get_term_IDF1(queries[qid]))) for qid in self.rel_docs]
# idfs.sort(key=itemgetter(1))
avdl = cs.get_avdl()
total_terms = cs.get_total_terms()
data = {True: [], False: []} # False: non-relevant True: relevant
ctf = cs.get_term_collection_occur(queries[qid])
collection_para = {
'avdl': avdl,
'total_terms': total_terms,
'ctf': ctf
}
for row in cs.get_qid_details(qid):
docid = row['docid']
total_tf = float(row['total_tf'])
doc_len = float(row['doc_len'])
rel_score = int(row['rel_score'])
rel = (rel_score>=1)
data[rel].append( {
'docid': docid,
'tf': total_tf,
'ln': doc_len
} )
method_para_dict = {ele.split(':')[0]:ele.split(':')[1] for ele in method_paras.split(',')}
max_map, max_para = self.learn(qid, data, collection_para, method_name, method_para_dict)
with open(output_fn, 'wb') as f:
json.dump({'map':max_map, 'para':max_para, 'eta':method_para_dict['eta']}, f, indent=2)
