Datasets:

Modalities:
Text
Formats:
json
ArXiv:
Libraries:
Datasets
pandas
License:
BLEnD / evaluation /exact_match.py
nayeon212's picture
ADD: Data & Code
605bdce verified
from evaluation_utils import *
import unicodedata as ud
# pip install konlpy
from konlpy.tag import Okt
# pip install hausastemmer
import hausastemmer
# git clone https://github.com/aznlp-disc/stemmer.git, cp word.txt & suffix.txt.
from stemmer.stemmer import Stemmer as AZStemmer
from string import punctuation
# pip install nlp-id
from nlp_id.lemmatizer import Lemmatizer as IDLemmatizer
# pip install hazm
from hazm import Lemmatizer as PRLemmatizer
# pip install qalsadi
from qalsadi.lemmatizer import Lemmatizer as ARLeammatizer
# pip install cltk
from cltk import NLP
# !pip install spark-nlp==5.3.3 pyspark==3.3.1
from sparknlp.base import *
from sparknlp.annotator import *
from sparknlp.pretrained import PretrainedPipeline
import sparknlp
from SUSTEM.SUSTEM_S import *
import spacy
# pip install jieba
import jieba
# git clone https://github.com/anoopkunchukuttan/indic_nlp_library.git & https://github.com/anoopkunchukuttan/indic_nlp_resources.git
# The path to the local git repo for Indic NLP library
INDIC_NLP_LIB_HOME=os.path.abspath("./indic_nlp_library")
# The path to the local git repo for Indic NLP Resources
INDIC_NLP_RESOURCES=os.path.abspath("./indic_nlp_resources")
sys.path.append(INDIC_NLP_LIB_HOME)
from indicnlp import common
from indicnlp import loader
from indicnlp.tokenize import indic_tokenize
def lemma_check(answer,llm_response,nlp_pipeline,language='Korean'):
if answer in llm_response or answer.replace('-',' ') in llm_response or answer.replace(' ','-') in llm_response:
return True
if language == 'Korean':
okt = Okt()
answer_tokens = okt.morphs(' '.join([w for w,p in okt.pos(answer) if p!='Josa']),stem=True)
llm_tokens = okt.morphs(' '.join([w for w,p in okt.pos(llm_response) if p!='Josa']),stem=True)
elif language == 'Hausa':
answer_tokens = [hausastemmer.stem(term.strip('-')) for term in answer.split()]
llm_tokens = [hausastemmer.stem(term.strip('-')) for term in llm_response.split()]
elif language == 'Amharic':
answer_tokens = [token.result if lemma.result.startswith('_') else lemma.result for token,lemma in zip(nlp_pipeline.fullAnnotate(answer)[0]['lemma'],nlp_pipeline.fullAnnotate(answer)[0]['token'])]
llm_tokens = [token.result if lemma.result.startswith('_') else lemma.result for token,lemma in zip(nlp_pipeline.fullAnnotate(llm_response)[0]['lemma'],nlp_pipeline.fullAnnotate(llm_response)[0]['token'])]
elif language == 'Azerbaijani':
# Instantiate Stemmer object
my_stemmer = AZStemmer()
def stem_words(my_text):
my_text=my_text.replace("İ", "I")
my_text=my_text.replace("“", "")
my_text=my_text.replace("”", "")
my_text=my_text.replace("'", "")
my_text=my_text.replace('"', "")
my_text=my_text.split()
my_words=[]
for word in my_text:
my_words.append(''.join(c for c in word if (c not in punctuation) or (c == '-')))
# Apply stemming to the list of words
my_words = my_stemmer.stem_words(my_words)
# Print words after stemming
return my_words
answer_tokens = stem_words(answer)
llm_tokens = stem_words(llm_response)
elif language == 'Indonesian':
lemmatizer = IDLemmatizer()
answer_tokens = lemmatizer.lemmatize(answer).split()
llm_tokens = lemmatizer.lemmatize(llm_response).split()
elif language == 'Persian':
lemmatizer = PRLemmatizer()
answer_tokens = [lemmatizer.lemmatize(term) for term in answer.split()]
llm_tokens = [lemmatizer.lemmatize(term) for term in llm_response.split()]
elif language == 'Arabic':
lemmatizer = ARLeammatizer()
answer_tokens = lemmatizer.lemmatize(answer)
llm_tokens = lemmatizer.lemmatize(llm_response)
elif language == 'Greek':
cltk_nlp = NLP(language="grc", suppress_banner=True)
answer_tokens = cltk_nlp.analyze(text=answer).lemmata
llm_tokens = cltk_nlp.analyze(text=llm_response).lemmata
elif language == 'Spanish':
answer_tokens = [lemma.result for lemma in nlp_pipeline.fullAnnotate(answer)[0]['lemma']]
llm_tokens = [lemma.result for lemma in nlp_pipeline.fullAnnotate(llm_response)[0]['lemma']]
elif language == 'Sundanese':
stemmer = EcsStemmer()
answer_tokens = [stemmer.stemmingProcess(word.replace('(','').replace(')','')) for word in answer.split()]
llm_tokens = [stemmer.stemmingProcess(word.replace('(','').replace(')','')) for word in llm_response.split()]
elif language == 'English':
answer_tokens = [token.lemma_ for token in nlp_pipeline(answer)]
llm_tokens = [token.lemma_ for token in nlp_pipeline(llm_response)]
elif language == 'Chinese':
answer_tokens = list(jieba.cut(answer))
llm_tokens = list(jieba.cut(llm_response))
elif language == 'Assamese':
common.set_resources_path(INDIC_NLP_RESOURCES)
loader.load()
answer_tokens = indic_tokenize.trivial_tokenize(answer)
llm_tokens = indic_tokenize.trivial_tokenize(llm_response)
d = {ord('\N{COMBINING ACUTE ACCENT}'):None}
answer_tokens = [ud.normalize('NFD',term).translate(d).lower() for term in answer_tokens if term not in punctuation and term != '']
llm_tokens = [ud.normalize('NFD',term).translate(d).lower() for term in llm_tokens if term not in punctuation and term != '']
for a in answer_tokens:
if a not in llm_tokens:
return False
return True
def hard_exact_match(annotation_dict,response_df,id_col,r_col,annotations_key='annotations'):
binary_score = 0
weight_score = 0
for qid,data in annotation_dict.items():
llm_response = get_llm_response_by_id(response_df,qid,id_col,r_col)
if llm_response and data[annotations_key]:
max_vote = max(list(data[annotations_key].values()))
for k,v in sorted(data[annotations_key].items(), key=lambda item: item[1],reverse=True):
if k == llm_response:
binary_score += 1
weight_score += v/max_vote
break
binary_score = binary_score / len(annotation_dict) * 100
weight_score = weight_score / len(annotation_dict) * 100
print(binary_score)
print(weight_score)
return binary_score, weight_score
def soft_exact_match(country,language,annotation_dict,response_df,id_col,r_col,annotations_key='aggregated_answers'):
binary_score = 0
weight_score = 0
valid_question_cnt = 0
if language == 'Spanish':
spark = sparknlp.start()
document_assembler = DocumentAssembler() \
.setInputCol("text") \
.setOutputCol("document")
tokenizer = Tokenizer() \
.setInputCols(["document"]) \
.setOutputCol("token")
lemmatizer = LemmatizerModel.pretrained("lemma", "es") \
.setInputCols(["token"]) \
.setOutputCol("lemma")
nlp_pipeline = Pipeline(stages=[document_assembler, tokenizer, lemmatizer])
nlpPipeline = LightPipeline(nlp_pipeline.fit(spark.createDataFrame([['']]).toDF('text')))
elif language == 'Amharic':
spark = sparknlp.start()
document_assembler = DocumentAssembler() \
.setInputCol("text") \
.setOutputCol("document")
tokenizer = Tokenizer() \
.setInputCols(["document"]) \
.setOutputCol("token")
lemmatizer = LemmatizerModel.pretrained("lemma", "am") \
.setInputCols(["token"]) \
.setOutputCol("lemma")
nlp_pipeline = Pipeline(stages=[document_assembler,tokenizer,lemmatizer])
nlpPipeline = LightPipeline(nlp_pipeline.fit(spark.createDataFrame([['']]).toDF('text')))
else:
nlpPipeline = None
en_lemmatizer = spacy.load("en_core_web_sm")
response_df['binary_score'] = [None]*response_df.shape[0]
response_df['weight_score'] = [None]*response_df.shape[0]
pb = tqdm(annotation_dict.items(),total=len(annotation_dict))
for qid,data in pb:
pb.set_description(qid)
if data['idks']['no-answer']+data['idks']['not-applicable'] >= 3 or data['idks']['idk']>=5 or len(data[annotations_key])==0:
continue
valid_question_cnt += 1
llm_response = get_llm_response_by_id(response_df,qid,id_col,r_col)
flag = False
if llm_response and data[annotations_key]:
max_vote = data[annotations_key][0]['count']
for agg_ans in data[annotations_key]:
if language != 'English':
for a in agg_ans['answers']:
if lemma_check(a,llm_response,nlpPipeline,language):
binary_score += 1
weight_score += agg_ans['count']/max_vote
flag = True
break
if not flag:
for a in agg_ans['en_answers']:
if lemma_check(a,llm_response,en_lemmatizer,'English'):
binary_score += 1
weight_score += agg_ans['count']/max_vote
flag = True
break
if flag:
break
if flag:
response_df.loc[response_df[id_col]==qid,'binary_score'] = 1
response_df.loc[response_df[id_col]==qid,'weight_score'] = agg_ans['count']/max_vote
print(response_df.loc[response_df[id_col]==qid])
else:
response_df.loc[response_df[id_col]==qid,'binary_score'] = 0
response_df.loc[response_df[id_col]==qid,'weight_score'] = 0
pb.set_postfix({'bs':binary_score/valid_question_cnt*100,'ws':weight_score/valid_question_cnt*100})
binary_score = binary_score / valid_question_cnt * 100
weight_score = weight_score / valid_question_cnt * 100
print(binary_score)
print(weight_score)
return binary_score, weight_score, response_df