winogrande.py 4.38 KB
Newer Older
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
"""
WinoGrande: An Adversarial Winograd Schema Challenge at Scale
https://arxiv.org/pdf/1907.10641.pdf

WinoGrande is a collection of 44k problems, inspired by Winograd Schema Challenge
(Levesque, Davis, and Morgenstern 2011), but adjusted to improve the scale and
robustness against the dataset-specific bias. Formulated as a fill-in-a-blank
task with binary options, the goal is to choose the right option for a given
sentence which requires commonsense reasoning.

NOTE: This evaluation of Winogrande uses partial evaluation as described by
Trinh & Le in Simple Method for Commonsense Reasoning (2018). 
See: https://arxiv.org/abs/1806.02847

Homepage: https://leaderboard.allenai.org/winogrande/submissions/public

@article{sakaguchi2019winogrande,
    title={WinoGrande: An Adversarial Winograd Schema Challenge at Scale},
    author={Sakaguchi, Keisuke and Bras, Ronan Le and Bhagavatula, Chandra and Choi, Yejin},
    journal={arXiv preprint arXiv:1907.10641},
    year={2019}
}
"""
Charles Foster's avatar
Charles Foster committed
24
import numpy as np
25
from . common import HFTask
&'s avatar
& committed
26
27
from lm_eval.base import rf
from ..metrics import mean
28

Charles Foster's avatar
Charles Foster committed
29
30

class Winogrande(HFTask):
Leo Gao's avatar
Leo Gao committed
31
    VERSION = 0
Charles Foster's avatar
Charles Foster committed
32
33
34
    DATASET_PATH = "winogrande"
    DATASET_NAME = "winogrande_xl"

35
36
    answer_to_num = {'1': 0, '2': 1}

Charles Foster's avatar
Charles Foster committed
37
38
39
40
41
42
43
    def has_training_docs(self):
        return True

    def has_validation_docs(self):
        return True

    def has_test_docs(self):
44
        return False
Charles Foster's avatar
Charles Foster committed
45

46
47
    def doc_to_text(self, doc):
        return self.partial_context(doc, doc["option" + doc["answer"]])
Charles Foster's avatar
Charles Foster committed
48

49
    @classmethod
50
51
    def partial_context(cls, doc, option):
        # Substitute the pronoun in the sentence with the specified option
52
53
        # and ignore everything after.
        pronoun_loc = doc["sentence"].index("_")
54
55
56
        return doc["sentence"][:pronoun_loc] + option

    def doc_to_target(self, doc):
Leo Gao's avatar
Leo Gao committed
57
        return self.partial_target(doc)
58
59
60
61
62

    @classmethod
    def partial_target(cls, doc):
        # The target is everything after the document specified pronoun.
        pronoun_loc = doc["sentence"].index("_") + 1
Leo Gao's avatar
Leo Gao committed
63
        return " " + doc["sentence"][pronoun_loc:].strip()
64

Leo Gao's avatar
Leo Gao committed
65
    def construct_requests(self, doc, ctx):
66
        """Uses RequestFactory to construct Requests and returns an iterable of
Leo Gao's avatar
Leo Gao committed
67
        Requests which will be sent to the LM.
68

Leo Gao's avatar
Leo Gao committed
69
70
71
        :param doc:
            The document as returned from training_docs, validation_docs, or test_docs.
        :param ctx: str
72
            The context string, generated by fewshot_context. This includes the natural
Leo Gao's avatar
Leo Gao committed
73
            language description, as well as the few shot examples, and the question
74
            part of the document for `doc`.
Leo Gao's avatar
Leo Gao committed
75
        """
76
        target = self.partial_target(doc)
77
        lls = []
78
79
80
        for option in [doc["option1"], doc["option2"]]:
            partial_ctx = self.partial_context(doc, option)
            full_ctx = self.append_context(ctx, partial_ctx)
81
82
            lls.append(rf.loglikelihood(full_ctx, target)[0])
        return lls
83
84
85

    @classmethod
    def append_context(cls, ctx, partial_ctx):
86
        ctx = ctx.split("\n\n")  # Each fewshot context is on its own new line.
87
88
        ctx.pop()  # Remove the correct context put in by `doc_to_text`.
        return "\n\n".join([*ctx, partial_ctx]) if ctx else partial_ctx
89

Leo Gao's avatar
Leo Gao committed
90
    def process_results(self, doc, results):
91
92
        """Take a single document and the LM results and evaluates, returning a
        dict where keys are the names of submetrics and values are the values of
Leo Gao's avatar
Leo Gao committed
93
94
95
96
97
98
99
        the metric for that one document

        :param doc:
            The document as returned from training_docs, validation_docs, or test_docs.
        :param results:
            The results of the requests created in construct_requests.
        """
100
        return {
101
            "acc": np.argmax(results) == self.answer_to_num[doc["answer"]]
102
        }
Leo Gao's avatar
Leo Gao committed
103
104
105
106

    def aggregation(self):
        """
        :returns: {str: [float] -> float}
107
            A dictionary where keys are the names of submetrics and values are
Leo Gao's avatar
Leo Gao committed
108
109
            functions that aggregate a list of metrics
        """
110
111
112
        return {
            "acc": mean
        }
Leo Gao's avatar
Leo Gao committed
113
114
115
116

    def higher_is_better(self):
        """
        :returns: {str: bool}
117
            A dictionary where keys are the names of submetrics and values are
Leo Gao's avatar
Leo Gao committed
118
119
            whether a higher value of the submetric is better
        """
120
121
122
        return {
            "acc": True
        }