Create coqa.py
Browse files
coqa.py
ADDED
@@ -0,0 +1,178 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
CoQA: A Conversational Question Answering Challenge
|
3 |
+
https://arxiv.org/pdf/1808.07042.pdf
|
4 |
+
|
5 |
+
CoQA is a large-scale dataset for building Conversational Question Answering
|
6 |
+
systems. The goal of the CoQA challenge is to measure the ability of machines to
|
7 |
+
understand a text passage and answer a series of interconnected questions that
|
8 |
+
appear in a conversation.
|
9 |
+
|
10 |
+
Homepage: https://stanfordnlp.github.io/coqa/
|
11 |
+
"""
|
12 |
+
import inspect
|
13 |
+
import transformers.data.metrics.squad_metrics as squad_metrics
|
14 |
+
import lm_eval.datasets.coqa.coqa
|
15 |
+
from lm_eval.base import Task, rf, mean
|
16 |
+
from itertools import zip_longest
|
17 |
+
|
18 |
+
|
19 |
+
_CITATION = """
|
20 |
+
@misc{reddy2018coqa,
|
21 |
+
title={CoQA: A Conversational Question Answering Challenge},
|
22 |
+
author={Siva Reddy and Danqi Chen and Christopher D. Manning},
|
23 |
+
year={2018},
|
24 |
+
eprint={1808.07042},
|
25 |
+
archivePrefix={arXiv},
|
26 |
+
primaryClass={cs.CL}
|
27 |
+
}
|
28 |
+
"""
|
29 |
+
|
30 |
+
|
31 |
+
class CoQA(Task):
|
32 |
+
VERSION = 1
|
33 |
+
DATASET_PATH = inspect.getfile(lm_eval.datasets.coqa.coqa)
|
34 |
+
DATASET_NAME = None
|
35 |
+
|
36 |
+
def has_training_docs(self):
|
37 |
+
return True
|
38 |
+
|
39 |
+
def has_validation_docs(self):
|
40 |
+
return True
|
41 |
+
|
42 |
+
def has_test_docs(self):
|
43 |
+
return False
|
44 |
+
|
45 |
+
def training_docs(self):
|
46 |
+
return self.dataset["train"]
|
47 |
+
|
48 |
+
def validation_docs(self):
|
49 |
+
return self.dataset["validation"]
|
50 |
+
|
51 |
+
def test_docs(self):
|
52 |
+
pass
|
53 |
+
|
54 |
+
def doc_to_text(self, doc):
|
55 |
+
# Given a passage p, the conversation history {q1, a1, . . . qi−1, ai−1}
|
56 |
+
# and a question qi, the task is to predict the answer ai
|
57 |
+
doc_text = doc["story"] + "\n\n"
|
58 |
+
for (q, a) in zip_longest(
|
59 |
+
doc["questions"]["input_text"], doc["answers"]["input_text"][:-1]
|
60 |
+
): # omit target answer ai
|
61 |
+
question = f"Q: {q}\n\n"
|
62 |
+
answer = f"A: {a}\n\n" if a is not None else "A:"
|
63 |
+
doc_text += question + answer
|
64 |
+
return doc_text
|
65 |
+
|
66 |
+
def should_decontaminate(self):
|
67 |
+
return True
|
68 |
+
|
69 |
+
def doc_to_decontamination_query(self, doc):
|
70 |
+
return doc["story"] + " " + "\n".join(doc["questions"]["input_text"])
|
71 |
+
|
72 |
+
@classmethod
|
73 |
+
def get_answers(cls, doc, turn_id):
|
74 |
+
# Returns unique answers and valid alternatives (Some questions in CoQA have multiple valid answers).
|
75 |
+
answers = []
|
76 |
+
answer_forturn = doc["answers"]["input_text"][turn_id - 1]
|
77 |
+
answers.append(answer_forturn)
|
78 |
+
|
79 |
+
additional_answers = doc.get("additional_answers")
|
80 |
+
if additional_answers:
|
81 |
+
for key in additional_answers:
|
82 |
+
additional_answer_for_turn = additional_answers[key]["input_text"][
|
83 |
+
turn_id - 1
|
84 |
+
]
|
85 |
+
if additional_answer_for_turn.lower() not in map(str.lower, answers):
|
86 |
+
answers.append(additional_answer_for_turn)
|
87 |
+
return answers
|
88 |
+
|
89 |
+
@classmethod
|
90 |
+
def get_answer_choice(self, raw_text):
|
91 |
+
# Function maps answers to CoQA answer categories
|
92 |
+
# ~ 1/5 of the CoQA answers are Yes/No
|
93 |
+
# ~ 2/3 of the CoQA answers are span-based
|
94 |
+
# (answers overlap with the passage ignoring punctuation and case mismatch)
|
95 |
+
if raw_text == "unknown":
|
96 |
+
return "0"
|
97 |
+
if squad_metrics.normalize_answer(raw_text) == "yes":
|
98 |
+
return "1"
|
99 |
+
if squad_metrics.normalize_answer(raw_text) == "no":
|
100 |
+
return "2"
|
101 |
+
return "3" # Not a yes/no question
|
102 |
+
|
103 |
+
@staticmethod
|
104 |
+
def compute_scores(gold_list, pred):
|
105 |
+
# tests for exact match and on the normalised answer (compute_exact)
|
106 |
+
# test for overlap (compute_f1)
|
107 |
+
f1_sum = 0.0
|
108 |
+
em_sum = 0.0
|
109 |
+
if len(gold_list) > 1:
|
110 |
+
for i in range(len(gold_list)):
|
111 |
+
gold_answers = gold_list[0:i] + gold_list[i + 1 :]
|
112 |
+
# predictions compared against (n) golds and take maximum
|
113 |
+
em_sum += max(
|
114 |
+
squad_metrics.compute_exact(a, pred) for a in gold_answers
|
115 |
+
)
|
116 |
+
f1_sum += max(squad_metrics.compute_f1(a, pred) for a in gold_answers)
|
117 |
+
else:
|
118 |
+
em_sum += max(squad_metrics.compute_exact(a, pred) for a in gold_list)
|
119 |
+
f1_sum += max(squad_metrics.compute_f1(a, pred) for a in gold_list)
|
120 |
+
|
121 |
+
return {
|
122 |
+
"em": em_sum / max(1, len(gold_list)),
|
123 |
+
"f1": f1_sum / max(1, len(gold_list)),
|
124 |
+
}
|
125 |
+
|
126 |
+
def doc_to_target(self, doc, turnid=None):
|
127 |
+
# Default to prediction of last turn.
|
128 |
+
if turnid is None:
|
129 |
+
turnid = len(doc["questions"]["input_text"])
|
130 |
+
raw_text = doc["answers"]["input_text"][turnid - 1]
|
131 |
+
return " " + raw_text
|
132 |
+
|
133 |
+
def construct_requests(self, doc, ctx):
|
134 |
+
"""Uses RequestFactory to construct Requests and returns an iterable of
|
135 |
+
Requests which will be sent to the LM.
|
136 |
+
|
137 |
+
:param doc:
|
138 |
+
The document as returned from training_docs, validation_docs, or test_docs.
|
139 |
+
:param ctx: str
|
140 |
+
The context string, generated by fewshot_context. This includes the natural
|
141 |
+
language description, as well as the few shot examples, and the question
|
142 |
+
part of the document for `doc`.
|
143 |
+
"""
|
144 |
+
cont_request = rf.greedy_until(ctx, {"until": ["\nQ:"]})
|
145 |
+
return cont_request
|
146 |
+
|
147 |
+
def process_results(self, doc, results):
|
148 |
+
"""Take a single document and the LM results and evaluates, returning a
|
149 |
+
dict where keys are the names of submetrics and values are the values of
|
150 |
+
the metric for that one document
|
151 |
+
|
152 |
+
:param doc:
|
153 |
+
The document as returned from training_docs, validation_docs, or test_docs.
|
154 |
+
:param results:
|
155 |
+
The results of the requests created in construct_requests.
|
156 |
+
"""
|
157 |
+
turn_id = len(doc["questions"]["input_text"])
|
158 |
+
gold_list = self.get_answers(doc, turn_id)
|
159 |
+
pred = results[0].strip().split("\n")[0]
|
160 |
+
|
161 |
+
scores = self.compute_scores(gold_list, pred)
|
162 |
+
|
163 |
+
return {
|
164 |
+
"f1": scores["f1"],
|
165 |
+
"em": scores["em"],
|
166 |
+
}
|
167 |
+
|
168 |
+
def higher_is_better(self):
|
169 |
+
return {
|
170 |
+
"f1": True,
|
171 |
+
"em": True,
|
172 |
+
}
|
173 |
+
|
174 |
+
def aggregation(self):
|
175 |
+
return {
|
176 |
+
"f1": mean,
|
177 |
+
"em": mean,
|
178 |
+
}
|