File size: 7,703 Bytes
f6f97d8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
"""
General utilities.
"""
import json
import os
from typing import List, Union, Dict
from functools import cmp_to_key
import math
from collections.abc import Iterable

from datasets import load_dataset

ROOT_DIR = os.path.join(os.path.dirname(__file__), "../")

def _load_table(table_path) -> dict:
    """
    attention: the table_path must be the .tsv path.
    Load the WikiTableQuestion from csv file. Result in a dict format like:
    {"header": [header1, header2,...], "rows": [[row11, row12, ...], [row21,...]... [...rownm]]}
    """

    def __extract_content(_line: str):
        _vals = [_.replace("\n", " ").strip() for _ in _line.strip("\n").split("\t")]
        return _vals

    with open(table_path, "r") as f:
        lines = f.readlines()

        rows = []
        for i, line in enumerate(lines):
            line = line.strip('\n')
            if i == 0:
                header = line.split("\t")
            else:
                rows.append(__extract_content(line))

    table_item = {"header": header, "rows": rows}

    # Defense assertion
    for i in range(len(rows) - 1):
        if not len(rows[i]) == len(rows[i - 1]):
            raise ValueError('some rows have diff cols.')

    return table_item


def majority_vote(
        nsqls: List,
        pred_answer_list: List,
        allow_none_and_empty_answer: bool = False,
        allow_error_answer: bool = False,
        answer_placeholder: Union[str, int] = '<error|empty>',
        vote_method: str = 'prob',
        answer_biased: Union[str, int] = None,
        answer_biased_weight: float = None,
):
    """
    Determine the final nsql execution answer by majority vote.
    """

    def _compare_answer_vote_simple(a, b):
        """
        First compare occur times. If equal, then compare max nsql logprob.
        """
        if a[1]['count'] > b[1]['count']:
            return 1
        elif a[1]['count'] < b[1]['count']:
            return -1
        else:
            if a[1]['nsqls'][0][1] > b[1]['nsqls'][0][1]:
                return 1
            elif a[1]['nsqls'][0][1] == b[1]['nsqls'][0][1]:
                return 0
            else:
                return -1

    def _compare_answer_vote_with_prob(a, b):
        """
        Compare prob sum.
        """
        return 1 if sum([math.exp(nsql[1]) for nsql in a[1]['nsqls']]) > sum(
            [math.exp(nsql[1]) for nsql in b[1]['nsqls']]) else -1

    # Vote answers
    candi_answer_dict = dict()
    for (nsql, logprob), pred_answer in zip(nsqls, pred_answer_list):
        if allow_none_and_empty_answer:
            if pred_answer == [None] or pred_answer == []:
                pred_answer = [answer_placeholder]
        if allow_error_answer:
            if pred_answer == '<error>':
                pred_answer = [answer_placeholder]

        # Invalid execution results
        if pred_answer == '<error>' or pred_answer == [None] or pred_answer == []:
            continue
        if candi_answer_dict.get(tuple(pred_answer), None) is None:
            candi_answer_dict[tuple(pred_answer)] = {
                'count': 0,
                'nsqls': []
            }
        answer_info = candi_answer_dict.get(tuple(pred_answer), None)
        answer_info['count'] += 1
        answer_info['nsqls'].append([nsql, logprob])

    # All candidates execution errors
    if len(candi_answer_dict) == 0:
        return answer_placeholder, [(nsqls[0][0], nsqls[0][-1])]

    # Sort
    if vote_method == 'simple':
        sorted_candi_answer_list = sorted(list(candi_answer_dict.items()),
                                          key=cmp_to_key(_compare_answer_vote_simple), reverse=True)
    elif vote_method == 'prob':
        sorted_candi_answer_list = sorted(list(candi_answer_dict.items()),
                                          key=cmp_to_key(_compare_answer_vote_with_prob), reverse=True)
    elif vote_method == 'answer_biased':
        # Specifically for Tabfact entailed answer, i.e., `1`.
        # If there exists nsql that produces `1`, we consider it more significant because `0` is very common.
        assert answer_biased_weight is not None and answer_biased_weight > 0
        for answer, answer_dict in candi_answer_dict.items():
            if answer == (answer_biased,):
                answer_dict['count'] *= answer_biased_weight
        sorted_candi_answer_list = sorted(list(candi_answer_dict.items()),
                                          key=cmp_to_key(_compare_answer_vote_simple), reverse=True)
    elif vote_method == 'lf_biased':
        # Assign weights to different types of logic forms (lf) to control interpretability and coverage
        for answer, answer_dict in candi_answer_dict.items():
            count = 0
            for nsql, _ in answer_dict['nsqls']:
                if 'map@' in nsql:
                    count += 10
                elif 'ans@' in nsql:
                    count += 10
                else:
                    count += 1
            answer_dict['count'] = count
        sorted_candi_answer_list = sorted(list(candi_answer_dict.items()),
                                          key=cmp_to_key(_compare_answer_vote_simple), reverse=True)
    else:
        raise ValueError(f"Vote method {vote_method} is not supported.")

    pred_answer_info = sorted_candi_answer_list[0]
    pred_answer, pred_answer_nsqls = list(pred_answer_info[0]), pred_answer_info[1]['nsqls']
    return pred_answer, pred_answer_nsqls


def load_data_split(dataset_to_load, split, data_dir=os.path.join(ROOT_DIR, 'datasets/')):
    dataset_split_loaded = load_dataset(
        path=os.path.join(data_dir, "{}.py".format(dataset_to_load)),
        cache_dir=os.path.join(data_dir, "data"))[split]

    # unify names of keys
    if dataset_to_load in ['wikitq', 'has_squall', 'missing_squall',
                           'wikitq', 'wikitq_sql_solvable', 'wikitq_sql_unsolvable',
                           'wikitq_sql_unsolvable_but_in_squall',
                           'wikitq_scalability_ori',
                           'wikitq_scalability_100rows',
                           'wikitq_scalability_200rows',
                           'wikitq_scalability_500rows',
                           'wikitq_robustness'
                           ]:
        pass
    elif dataset_to_load == 'tab_fact':
        new_dataset_split_loaded = []
        for data_item in dataset_split_loaded:
            data_item['question'] = data_item['statement']
            data_item['answer_text'] = data_item['label']
            data_item['table']['page_title'] = data_item['table']['caption']
            new_dataset_split_loaded.append(data_item)
        dataset_split_loaded = new_dataset_split_loaded
    elif dataset_to_load == 'hybridqa':
        new_dataset_split_loaded = []
        for data_item in dataset_split_loaded:
            data_item['table']['page_title'] = data_item['context'].split(' | ')[0]
            new_dataset_split_loaded.append(data_item)
        dataset_split_loaded = new_dataset_split_loaded
    elif dataset_to_load == 'mmqa':
        new_dataset_split_loaded = []
        for data_item in dataset_split_loaded:
            data_item['table']['page_title'] = data_item['table']['title']
            new_dataset_split_loaded.append(data_item)
        dataset_split_loaded = new_dataset_split_loaded
    else:
        raise ValueError(f'{dataset_to_load} dataset is not supported now.')
    return dataset_split_loaded


def pprint_dict(dic):
    print(json.dumps(dic, indent=2))


def flatten(nested_list):
    for x in nested_list:
        if isinstance(x, Iterable) and not isinstance(x, (str, bytes)):
            yield from flatten(x)
        else:
            yield x