ChatRAG-Bench / evaluation /get_scores.py
zihanliu's picture
Upload 7 files
f2ad62a verified
raw
history blame
13.2 kB
from evaluation_utils import quac_correct_retrieved_instance_idx_list
from evaluation_utils import unanswerable_keyphrases
import json
from metrics import F1Metric
import copy
import re
def compute_f1_score(predicted_answers, groundtruth_answer, exp_name="default"):
"""Evaluating F1 Score"""
print(len(predicted_answers), len(groundtruth_answer))
if len(predicted_answers) != len(groundtruth_answer):
groundtruth_answer = groundtruth_answer[:len(predicted_answers)]
guess_list = []
for guess in predicted_answers:
guess = guess.strip()
if "</s>" in guess:
guess = guess.replace("</s>", "")
guess_list.append(guess)
answer_list = []
for answer in groundtruth_answer:
answer_list.append(answer)
assert len(guess_list) == len(answer_list), \
"lengths of guess and answer are different!"
precision, recall, f1 = F1Metric.compute_all_pairs(guess_list, answer_list)
print('Method: %s; Precision: %.4f; recall: %.4f; f1: %.4f' % (\
exp_name, precision, recall, f1))
def load_groundtruth_file(data_file):
with open(data_file, "r") as f:
examples = json.load(f)
data = []
for instance in examples:
if "answers" in instance:
answers = instance["answers"]
elif "answer" in instance:
if type(instance["answer"]) is str:
answers = [instance["answer"]]
elif type(instance["answer"]) is list:
answers = instance["answer"]
else:
answers = [str(instance["answer"])]
else:
raise ValueError("need to have answer or answers")
data.append(answers)
return data
def load_prediction(data_file):
data = []
with open(data_file, "r") as f:
for line in f.readlines():
data.append(line.strip())
return data
def evaluate_f1(ground_truth_file, prediction_file):
groundtruth_answers = load_groundtruth_file(ground_truth_file)
if "inscit" in ground_truth_file:
groundtruth_answers_update = []
for answers in groundtruth_answers:
answers_update = []
for ans in answers:
## this answer is additionally added to the answer_list for inscit dataset, needs to remove
if ans != "Sorry. I cannot find the answer based on the context.":
answers_update.append(ans)
assert len(answers_update) > 0
groundtruth_answers_update.append(copy.deepcopy(answers_update))
groundtruth_answers = groundtruth_answers_update
predicted_answers = load_prediction(prediction_file)
if "quac" in prediction_file or "doqa" in prediction_file:
predicted_answers_new = []
for pred in predicted_answers:
pred = pred.lower()
for keyphrase in unanswerable_keyphrases:
if keyphrase in pred:
pred = "Sorry. I cannot find the answer based on the context."
break
predicted_answers_new.append(pred)
predicted_answers = predicted_answers_new
compute_f1_score(predicted_answers, groundtruth_answers)
def separate_cannot_answer(ground_truth_file, prediction_file):
# load ground truth
with open(ground_truth_file, "r") as f:
groundtruth_answers = json.load(f)
# load prediction
predicted_answers = load_prediction(prediction_file)
print(len(predicted_answers), len(groundtruth_answers))
if len(predicted_answers) != len(groundtruth_answers):
groundtruth_answers = groundtruth_answers[:len(predicted_answers)]
if "quac" in prediction_file:
"""
For answerable cases, we want to make sure the retrieved context list contains the gold chunk.
For QuAC dataset, we use top-5 retrieved contexts as inputs, quac_correct_retrieved_instance_idx_list
is the index list where the top-5 retrieved context contains the gold answer
"""
answerable_instance_idx_list = quac_correct_retrieved_instance_idx_list
else:
answerable_instance_idx_list = None
predicted_answers_new = []
for pred in predicted_answers:
pred = pred.lower()
for keyphrase in unanswerable_keyphrases:
if keyphrase in pred:
pred = "Sorry. I cannot find the answer based on the context."
break
predicted_answers_new.append(pred)
predicted_answers = predicted_answers_new
cannot_answer_idx_list = []
answerable_idx_list = []
if answerable_instance_idx_list:
count_idx = 0
for idx, item in enumerate(groundtruth_answers):
if 'answers' in item:
answer = item["answers"][0]
else:
answer = item['answer']
noanswer_response = "Sorry. I cannot find the answer based on the context."
if answer == noanswer_response:
cannot_answer_idx_list.append(idx)
continue
if answerable_instance_idx_list:
if count_idx in answerable_instance_idx_list:
answerable_idx_list.append(idx)
count_idx += 1
else:
answerable_idx_list.append(idx)
print("number of cannot answer cases: %d (out of %d)" % (len(cannot_answer_idx_list), len(groundtruth_answers)))
print("number of answerable cases: %d (out of %d)" % (len(answerable_idx_list), len(groundtruth_answers)))
return predicted_answers, cannot_answer_idx_list, answerable_idx_list
def get_cannot_answer_and_answerable_acc(predicted_answers, cannot_answer_idx_list, answerable_idx_list):
# cannot answer
noanswer_count = 0
for idx in cannot_answer_idx_list:
prediction = predicted_answers[idx]
prediction = prediction.lower()
# print(prediction)
if "sorry" in prediction and "cannot find the answer" in prediction:
# print(prediction)
noanswer_count += 1
cannot_answer_acc = noanswer_count / len(cannot_answer_idx_list)
print("accuracy of cannot answer cases: %.4f" % cannot_answer_acc)
# answerable
answerable_count = 0
for idx in answerable_idx_list:
prediction = predicted_answers[idx]
prediction = prediction.lower()
if "sorry" in prediction and "cannot find the answer" in prediction:
# print(prediction)
continue
answerable_count += 1
answerable_acc = answerable_count / len(answerable_idx_list)
print("accuracy of answerable cases: %.4f" % answerable_acc)
def evaluate_cannot_answer_acc(ground_truth_file, prediction_file):
predicted_answers, cannot_answer_idx_list, answerable_idx_list = \
separate_cannot_answer(ground_truth_file, prediction_file)
get_cannot_answer_and_answerable_acc(predicted_answers, cannot_answer_idx_list, answerable_idx_list)
def evaluate_convfinqa(ground_truth_file, prediction_file):
"""
Since the model will give a long answer output, while the gold answer for ConvFinQA are either
a arithmetic formula or a final executed number.
We consider the output containing either the executed number or the arithmetic formula as correct.
This script is to measure the proportion of the outputs containing these elements.
"""
def _is_float(string):
try:
float(string)
return True
except ValueError:
return False
with open(ground_truth_file, "r") as f:
gold_list = json.load(f)
groundtruth_answers = [item['exe_answer'] for item in gold_list]
groundtruth_answers_formula = [item['answers'][0] for item in gold_list]
## last turn question_list
question_list = [item['messages'][-1]['content'] for item in gold_list]
predicted_answers = load_prediction(prediction_file)
print(len(predicted_answers), len(groundtruth_answers))
if len(predicted_answers) != len(groundtruth_answers):
groundtruth_answers = groundtruth_answers[:len(predicted_answers)]
count_exact_match = 0
for question, pred, gold, gold_formula in zip(question_list, predicted_answers, groundtruth_answers, groundtruth_answers_formula):
original_pred = pred
## convert 1,000,000 into 1000000
original_pred = original_pred.replace(",", "")
## convert $10 million + $20 million into 10 + 20
original_pred = original_pred.replace("$", "").replace("million", "").replace("billion", "")
## convert 10 (2017) + 20 (2018) into 10 + 20
pattern = r'\((\b\w+\b)\)'
original_pred = re.sub(pattern, '', original_pred)
## make sure it each token only has one space in between
original_pred = " ".join(original_pred.split())
if str(gold) in original_pred:
count_exact_match += 1
elif str(gold_formula) in original_pred:
count_exact_match += 1
elif _is_float(gold) and (str(round(float(gold), 3)) in original_pred or str(round(float(gold), 2)) in original_pred):
count_exact_match += 1
elif "percent" in question and (str(float(gold)*100) in original_pred or str(round(float(gold)*100, 1)) in original_pred or str(round(float(gold)*100, 2)) in original_pred):
count_exact_match += 1
elif str(gold).endswith(".0") and str(int(gold)) in original_pred:
## gold is a integer like 80.0 then convert it into 80
count_exact_match += 1
elif "decrease" in original_pred and _is_float(gold) and gold < 0 and (str(-1 * gold) in original_pred):
## for the case where model generates something like a decrese of 10 million, while gold is -10.
count_exact_match += 1
print("accuracy of exact match: %.4f" % (count_exact_match/len(predicted_answers)))
def main():
## doc2dial
prediction_file = "PATH_TO_THE_GENERATED_OUTPUT" # e.g., outputs/doc2idal_output.txt
ground_truth_file = "PATH_TO_THE_TEST_DATA" # e.g., data/doc2dial/test.json
print("-"*80)
print(prediction_file)
print(ground_truth_file)
evaluate_f1(ground_truth_file, prediction_file)
## quac
prediction_file = "PATH_TO_THE_GENERATED_OUTPUT"
ground_truth_file = "PATH_TO_THE_TEST_DATA"
print("-"*80)
print(prediction_file)
print(ground_truth_file)
evaluate_f1(ground_truth_file, prediction_file)
evaluate_cannot_answer_acc(ground_truth_file, prediction_file)
## qrecc
prediction_file = "PATH_TO_THE_GENERATED_OUTPUT"
ground_truth_file = "PATH_TO_THE_TEST_DATA"
print("-"*80)
print(prediction_file)
print(ground_truth_file)
evaluate_f1(ground_truth_file, prediction_file)
## topiocqa
prediction_file = "PATH_TO_THE_GENERATED_OUTPUT"
ground_truth_file = "PATH_TO_THE_TEST_DATA"
print("-"*80)
print(prediction_file)
print(ground_truth_file)
evaluate_f1(ground_truth_file, prediction_file)
## inscit
prediction_file = "PATH_TO_THE_GENERATED_OUTPUT"
ground_truth_file = "PATH_TO_THE_TEST_DATA"
print("-"*80)
print(prediction_file)
print(ground_truth_file)
evaluate_f1(ground_truth_file, prediction_file)
## coqa
prediction_file = "PATH_TO_THE_GENERATED_OUTPUT"
ground_truth_file = "PATH_TO_THE_TEST_DATA"
print("-"*80)
print(prediction_file)
print(ground_truth_file)
evaluate_f1(ground_truth_file, prediction_file)
## hybridial
prediction_file = "PATH_TO_THE_GENERATED_OUTPUT"
ground_truth_file = "PATH_TO_THE_TEST_DATA"
print("-"*80)
print(prediction_file)
print(ground_truth_file)
evaluate_f1(ground_truth_file, prediction_file)
## sqa
prediction_file = "PATH_TO_THE_GENERATED_OUTPUT"
ground_truth_file = "PATH_TO_THE_TEST_DATA"
print("-"*80)
print(prediction_file)
print(ground_truth_file)
evaluate_f1(ground_truth_file, prediction_file)
## doqa_cooking
prediction_file = "PATH_TO_THE_GENERATED_OUTPUT"
ground_truth_file = "PATH_TO_THE_TEST_DATA"
print("-"*80)
print(prediction_file)
print(ground_truth_file)
evaluate_f1(ground_truth_file, prediction_file)
evaluate_cannot_answer_acc(ground_truth_file, prediction_file)
## doqa_travel
prediction_file = "PATH_TO_THE_GENERATED_OUTPUT"
ground_truth_file = "PATH_TO_THE_TEST_DATA"
print("-"*80)
print(prediction_file)
print(ground_truth_file)
evaluate_f1(ground_truth_file, prediction_file)
evaluate_cannot_answer_acc(ground_truth_file, prediction_file)
## doqa_movies
prediction_file = "PATH_TO_THE_GENERATED_OUTPUT"
ground_truth_file = "PATH_TO_THE_TEST_DATA"
print("-"*80)
print(prediction_file)
print(ground_truth_file)
evaluate_f1(ground_truth_file, prediction_file)
evaluate_cannot_answer_acc(ground_truth_file, prediction_file)
## convfinqa
prediction_file = "PATH_TO_THE_GENERATED_OUTPUT"
ground_truth_file = "PATH_TO_THE_TEST_DATA"
print("-"*80)
print(prediction_file)
print(ground_truth_file)
evaluate_convfinqa(ground_truth_file, prediction_file)
if __name__ == "__main__":
main()