def expand_from_preext_sent_rule(self):
if not hasattr(self, 'cursor'):
self.cursor = fever_db.get_cursor()
if not hasattr(self, 'preext_sent_dict'):
d_list = read_jsonl(config.RESULT_PATH / \
"sent_retri_nn/2018_07_17_16-34-19_r/train_scale(0.1).jsonl")
self.preext_sent_dict = {item['id']: item for item in d_list}
item = self.item
# if len(item['prioritized_docids']) < 5:
new_pdocids = []
structured_docids_sent = {}
sent_ids = self.preext_sent_dict[item['id']]['scored_sentids']
for sent_id, score, probability in sent_ids:
docid, sent_ind = sent_id.split('<SENT_LINE>')
sent_ind = int(sent_ind)
id_list, sent_list, sent_links = \
fever_db.get_evidence(self.cursor,
docid,
sent_ind)
sent_links = json.loads(sent_links)
all_links = np.array(sent_links)
all_links = np.array(all_links)
all_links = all_links.reshape(-1, 2)[:, 1]
all_links = list(map(fever_db.reverse_convert_brc, all_links))
all_links = list(map(lambda x: x.replace(' ', '_'), all_links))
prio_docids = [(id_link, score) for id_link in all_links]
new_pdocids.extend(prio_docids)
structured_docids_sent.update({sent_id: prio_docids})
item['prioritized_docids_sent'] = new_pdocids
item['structured_docids_sent'] = structured_docids_sent
return self
def disambiguous_from_preext_sent_rule(self):
if not hasattr(self, 'cursor'):
self.cursor = fever_db.get_cursor()
if not hasattr(self, 'preext_sent_dict'):
d_list = read_jsonl(config.RESULT_PATH / \
"sent_retri_nn/2018_07_17_16-34-19_r/train_sent.jsonl")
self.preext_sent_dict = {item['id']: item for item in d_list}
item = self.item
if len(item['prioritized_docids']) > 60:
sent_ids = self.preext_sent_dict[item['id']]['']
return self
def tf_idf_rank(args, top_k=5):
dev_path = config.PRO_ROOT / \
'results_old/doc_retri/docretri.basic.nopageview/dev.jsonl'
cursor = get_cursor()
d_list = read_jsonl(dev_path)
d_list_test = d_list
for i, item in enumerate(spcl(d_list_test)):
all_sent = []
all_ids = [it[0] for it in item['prioritized_docids']]
try:
for doc_id in all_ids:
r_list, _ = get_all_sent_by_doc_id(cursor,
doc_id,
with_h_links=False)
all_sent.append(' '.join(r_list))
ranker = OnlineTfidfDocRanker(args, args.hash_size, args.ngram,
all_sent)
except Exception as e:
if i - 1 >= 0:
print(f'Early quit at {i-1} because of {e}')
save_path = config.RESULT_PATH / \
'doc_retri/docretri.tfidfrank/' \
f'dev_quit_dump_{uuid4()}.json'
DocRetrievalExperiment.dump_results(d_list_test[:i], save_path)
raise e
rank_ind, rank_score = \
ranker.closest_docs(' '.join(item['claim_tokens']), k=100)
id_score_dict = {docid: 0 for docid in all_ids}
id_score_dict.update({all_ids[ri]: rs \
for ri, rs in zip(rank_ind, rank_score)})
item['prioritized_docids'] = [(k, v) for k, v in id_score_dict.items()]
item['predicted_docids'] = \
list(set([k for k, v \
in sorted(item['prioritized_docids'],
key=lambda x: (-x[1], x[0]))][:top_k]))
save_path = config.RESULT_PATH / 'doc_retri/docretri.tfidfrank/dev.json'
DocRetrievalExperiment.dump_results(d_list_test, save_path)
def main():
doc_exp = DocRetrievalExperimentTwoStep()
d_list = read_jsonl(config.FEVER_DEV_JSONL)
# d_list = read_jsonl(config.FEVER_TRAIN_JSONL)
doc_exp.sample_answer_with_priority(d_list)
doc_exp.print_eval(d_list)
doc_exp.feed_sent_file(
"../../results/"
"sent_retri_nn/2018_07_17_16-34-19_r/dev_scale(0.1).jsonl")
doc_exp.find_sent_link_with_priority(d_list, predict=True)
doc_exp.print_eval(d_list)
# path = "../../results/doc_retri/docretri.spiral.aside/train.jsonl"
# doc_exp.dump_results(d_list, path)
from IPython import embed
embed()
import os
os._exit(1)
def main():
import os
from chaonan_src._config import old_result_path
from chaonan_src._utils.doc_utils import read_jsonl
from chaonan_src._utils.spcl import spcl
# pageview_path = os.path.join(config.RESULT_PATH,
# 'doc_retri/docretri.rawpageview/train.jsonl')
pageview_path = config.RESULT_PATH / \
'doc_retri/docretri.rawpageview/dev.jsonl'
# ori_path = os.path.join(old_result_path,
# 'doc_retri/docretri.pageview/dev.jsonl')
# d_list = read_jsonl(config.FEVER_DEV_JSONL)
# item_rb_exp = ItemRuleRawPageview()
# doc_exp = DocRetrievalExperiment(item_rb_exp)
# doc_exp.sample_answer_with_priority(d_list)
# doc_exp.dump_results(d_list, save_path)
# doc_exp.print_eval(d_list)
# d_list_ori = read_jsonl(ori_path)
d_list = read_jsonl(pageview_path)
# DocRetrievalExperiment.dump_results(d_list, pageview_path)
# item_rb = ItemRuleRawPageview()
top_k = 5
for item in spcl(d_list):
item['predicted_docids'] = \
list(set([k for k, v \
in sorted(item['docid_pageviews'],
key=lambda x: (-x[1], x[0]))][:top_k]))
# DocRetrievalExperiment.dump_results(d_list, save_path)
DocRetrievalExperiment.print_eval(d_list)
from IPython import embed
embed()
import os
os._exit(1)
def pageview_analysis():
from chaonan_src._doc_retrieval.item_rules import ItemRuleBuilder
from chaonan_src._utils.doc_utils import read_jsonl
from utils.fever_db import convert_brc
wiki_pv = WikiPageviews()
d_list = read_jsonl(
"../../../results/doc_retri/docretri.titlematch/dev.jsonl")
gt_evidences, pre_evidences = [], []
for item in d_list:
gt_evidences.extend(ItemRuleBuilder\
.get_all_docid_in_evidence(item['evidence']))
pre_evidences.extend([it[0] for it in item['prioritized_docids']])
gt_evidences = set(gt_evidences)
pre_evidences = set(pre_evidences)
gt_count = [wiki_pv[convert_brc(it)] for it in gt_evidences]
pre_count = [wiki_pv[convert_brc(it)] for it in pre_evidences]
from IPython import embed
embed()
import os
os._exit(1)
def feed_sent_score_result(self, path):
d_list = read_jsonl(path)
self.preext_sent_dict = {item['id']: item for item in d_list}
if predict:
porg = \
set([k for k, v \
in sorted(item['prioritized_docids'],
key=lambda x: (-x[1], x[0]))][:top_k])
paside = \
set([k for k, v \
in sorted(item['prioritized_docids_aside'],
key=lambda x: (-x[1], x[0]))][:top_k])
item['predicted_docids'] = list(porg | paside)
item['predicted_docids_origin'] = list(porg)
item['predicted_docids_aside'] = list(paside)
d_list[i] = item
ws_second.close()
if __name__ == "__main__":
from chaonan_src._utils.doc_utils import read_jsonl
import config
d_list = read_jsonl(config.FEVER_DEV_JSONL)
d_list_test = d_list[:20]
DocRetrievalClient.sample_answer_with_priority(d_list_test)
DocRetrievalClient.print_eval(d_list_test)
# DocRetrievalClient.feed_sent_file(
# "../../results"
# "/sent_retri_nn/2018_07_17_16-34-19_r/dev_scale(0.1).jsonl")
# DocRetrievalClient.find_sent_link_with_priority(d_list_test, predict=True)
DocRetrievalExperiment.print_eval(d_list_test)
def eval():
save_path = config.RESULT_PATH / 'doc_retri/docretri.tfidfrank/dev_toy.json'
d_list = read_jsonl(save_path)
DocRetrievalExperiment.print_eval(d_list)
def __init__(self, jl_path, train=False):
self.d_list = read_jsonl(jl_path)
self.train = train
self.initialized = False
self.batch_size = None