def save_to_csv():
gold = get_claim_perspective_id_dict()
def routine(claims, out_path):
payloads = predict_by_elastic_search(claims, 50)
head = ['sentence1', 'sentence2', 'gold_label', 'cid', 'pid']
rows = []
for cid, data_list in payloads:
gold_pids = gold[cid]
all_pid_set = set(flatten(gold_pids))
for p_entry in data_list:
c_text = p_entry['claim_text']
p_text = p_entry['perspective_text']
y = 1 if p_entry['pid'] in all_pid_set else 0
row = [c_text, p_text, y, cid, p_entry['pid']]
rows.append(row)
f_out = csv.writer(open(out_path, "w", encoding="utf-8"),
dialect='excel-tab')
f_out.writerows([head] + rows)
claims, val = train_split()
routine(claims, get_file_path('train'))
d_ids: List[int] = list(load_dev_claim_ids())
claims = get_claims_from_ids(d_ids)
routine(claims, get_file_path('dev'))
d_ids: List[int] = list(load_test_claim_ids())
claims = get_claims_from_ids(d_ids)
routine(claims, get_file_path('test'))
def run_reweight():
top_k = 7
claims, val = train_split()
param = {'k1': 1}
target = claims[:50]
pred = predict_by_reweighter(get_bm25_module(), target, top_k, param)
print(param)
print(evaluate(pred))
def run_bm25_2():
claims, val = train_split()
top_k = 1000
candidate_dict: List[Tuple[int, List[int]]] = get_eval_candidates_w_q_text(
claim_as_query(claims), top_k)
pred = predict_by_bm25_from_candidate(get_bm25_module(), claims,
candidate_dict, top_k)
print(evaluate_recall(pred, True))
def run_a_relevant_lm():
claims, val = train_split()
top_k = 50
print("Building lms")
claim_lms = get_train_passage_a_lms()
print("Predicting")
pred = predict_by_lm(claim_lms, claims, top_k)
print(evaluate_map(pred))
def run_gold_lm_ap():
claims, val = train_split()
top_k = 50
print("Building lms")
claim_lms = build_gold_claim_lm_train()
print("Predicting")
pred = predict_by_lm(claim_lms, claims, top_k)
print(evaluate_map(pred))
def run_bert_baseline():
claims, val = train_split()
top_k = 50
target = filter_avail(val)
print("targets", len(target))
pc_score_d = load_from_pickle("pc_bert_baseline_score_d_train")
pred = predict_from_dict(pc_score_d, target, top_k)
print(evaluate(pred))
def run_gold_lm():
claims, val = train_split()
top_k = 5
print("Building lms")
claim_lms: List[ClaimLM] = build_gold_claim_lm_train()
print("Predicting")
pred = predict_by_lm(claim_lms, claims, top_k)
print(evaluate(pred))
def run_rel_scorer():
claims, val = train_split()
top_k = 6
target = filter_avail(val)
print("targets", len(target))
pc_score_d = load_from_pickle("pc_rel_based_score_train")
pred = predict_from_dict(pc_score_d, target, top_k)
print(evaluate(pred))
def run_bm25_ex():
claims, val = train_split()
top_k = 100
candidate_dict = get_eval_candidates_l(
get_expanded_query_text(claims, "train"))
pred = predict_by_bm25_from_candidate(get_bm25_module(), claims,
candidate_dict, top_k)
print(evaluate_recall(pred, True))
def run_baseline_lm():
claims, val = train_split()
claims = val
top_k = 50
print("Building lms")
claim_lms = build_baseline_lms(claims)
print("Predicting")
pred = predict_by_lm(claim_lms, claims, top_k)
print(evaluate_map(pred))
def generate_classification_payload():
claims, val = train_split()
top_k = 50
pred = predict_by_elastic_search(claims, top_k)
save_to_pickle(pred, "perspective_cls_train_X")
d_ids: List[int] = list(load_dev_claim_ids())
claims = get_claims_from_ids(d_ids)
top_k = 50
pred = predict_by_elastic_search(claims, top_k)
save_to_pickle(pred, "perspective_cls_dev_X")
def run_para_scorer():
claims, val = train_split()
top_k = 6
target = filter_avail(val)
print("targets", len(target))
score_pred_file: FileName = FileName("pc_para_D_pred")
cpid_resolute_file: FileName = FileName("resolute_dict_580_606")
pred = predict_by_para_scorer(score_pred_file, cpid_resolute_file, target,
top_k)
print(evaluate(pred))
def build_df():
claims, val = train_split()
gold = get_claim_perspective_id_dict()
tokenizer = PCTokenizer()
df = Counter()
dl_list = []
for claim in claims:
cid = claim["cId"]
gold_pids = flatten(gold[cid])
p_text_list: List[str] = lmap(perspective_getter, gold_pids)
tokens_list = lmap(tokenizer.tokenize_stem, p_text_list)
dl_list.extend(lmap(len, tokens_list))
for t in set(flatten(tokens_list)):
df[t] += 1
print(dl_list)
print("Avdl", average(dl_list))
print(len(claims))
print(df.most_common(30))
save_to_pickle(df, "pc_df")
def build_gold_claim_lm_train() -> List[ClaimLM]:
# load claims and perspectives
# Calculate term frequency for each terms.
claims, val = train_split()
return build_gold_lms(claims)
def run_bm25():
claims, val = train_split()
top_k = 20
pred = predict_by_bm25(get_bm25_module(), claims, top_k)
inspect(pred)
def run_bm25():
claims, val = train_split()
top_k = 20
pred = predict_by_bm25(get_bm25_module(), claims, top_k)
print(evaluate(pred))
def run_oracle_on_candiate_map():
claims, val = train_split()
top_k = 50
pred = predict_by_oracle_on_candidate(claims, top_k)
print(evaluate_map(pred))
def run_baseline():
claims, val = train_split()
top_k = 50
pred = predict_by_elastic_search(claims, top_k)
print(evaluate(pred))
