nltkor 1.2.0__cp39-cp39-macosx_10_9_x86_64.whl

nltkor/metrics/classical.py

@@ -0,0 +1,814 @@
+import os
+import sys
+import argparse
+import numpy as np
+from typing import Callable, Iterable, List, Tuple, Union
+from copy import deepcopy
+import itertools
+from nltk.translate.bleu_score import *
+from nltk.metrics import confusionmatrix
+from collections import defaultdict
+from nltk.util import ngrams, skipgrams
+#from nltkor.tokenize.ko_tokenize import word_tokenize,sent_tokenize,syllable_tokenize
+from nltkor.make_requirement import make_requirement
+from nltkor.tokenize import Ko_tokenize
+from nltkor.cider.cider import Cider
+from nltkor.tag import EspressoTagger
+from nltkor.sejong import ssem
+
+try:
+    import torch
+except ImportError:
+	file_path = make_requirement(['torch'])
+	raise Exception(f"""
+    Need to install Libraries, please pip install below libraries
+    \t pip install torch
+	Or, use pip install requirement.txt
+    \t  pip install -r {file_path}
+    """)
+
+
+class DefaultMetric:
+
+	def __init__(self, lang="ko"):
+		self.lang = lang
+		#if lang not in ["korean", "english"]:
+		if lang not in ["ko", "en"]:
+			raise Exception("Only \"korean\" or \"english\" in lang parameter")
+		self.tokenize=lambda ref: Ko_tokenize.word(ref,self.lang)
+		self.skip_pos = ['EE']
+
+	def accuracy_score(self, true, pred):
+
+		mat=confusionmatrix.ConfusionMatrix(true,pred)
+
+		conf=mat._confusion
+		total=0
+		tp=0
+
+		for r, tmp in enumerate(conf):
+			for v, n in enumerate(tmp):
+				if r==v:
+					tp+=n
+				total+=n
+
+		return float(tp/total)
+
+	def recall_score(self, true, pred, avg='micro'):
+
+		mat=confusionmatrix.ConfusionMatrix(true,pred)
+
+		conf=mat._confusion
+		indices=mat._indices
+		values=mat._values
+		total=0
+
+
+		if len(values)==2:
+			tp=0
+			fn=0
+			for r,i in enumerate(conf):
+				for r2,v in enumerate(i):
+					if r==0:
+						continue
+					elif r2==0:
+						fn=v
+					elif r==1:
+						tp=v
+
+			return float(tp/(tp+fn))
+
+
+		c_tp=[]
+		c_fn=[]
+		recall_cls=[]
+
+		for r, tmp in enumerate(conf):
+			temp=0
+			for v, n in enumerate(tmp):
+				if r==v:
+					c_tp.append(n)
+				else:
+					temp+=n
+			c_fn.append(temp)
+
+		if avg=='macro':
+
+			for tmp in range(len(values)):
+				try:
+					recall_cls.append(float(c_tp[tmp]/(c_tp[tmp]+c_fn[tmp])))
+				except:
+					recall_cls.append(0)
+
+			temp=0
+
+			for tmp in recall_cls:
+				temp+=tmp
+
+			return float(temp/len(recall_cls))
+
+		if avg=='micro':
+			ja=0
+			mo=0
+
+			for tmp in range(len(values)):
+				ja+=c_tp[tmp]
+				mo+=c_tp[tmp]+c_fn[tmp]
+
+			return float(ja/mo)
+
+		else:
+			return "avg expect micro/macro"
+
+
+
+	def precision_score(self, true, pred,avg='micro'):
+
+
+		mat=confusionmatrix.ConfusionMatrix(true,pred)
+
+		conf=mat._confusion
+		values=mat._values
+
+		total=0
+
+		if len(values)==2:
+			tp=0
+			fp=0
+			for r,i in enumerate(conf):
+				for r2,v in enumerate(i):
+					if r2==0:
+						continue
+					elif r==0:
+						fp=v
+					elif r==1:
+						tp=v
+
+			return float(tp/(tp+fp))
+
+		c_tp=list()
+		c_fp=[0 for _ in range(len(values))]
+		recall_cls=[]
+
+		for r, tmp in enumerate(conf):
+			for v, n in enumerate(tmp):
+				if r==v:#tp
+					c_tp.append(n)
+				else:
+					c_fp[v]+=n
+
+		if avg=='macro':
+			for tmp in range(len(values)):
+				try:
+					recall_cls.append(float(c_tp[tmp]/(c_tp[tmp]+c_fp[tmp])))
+				except:
+					recall_cls.append(0)
+
+			temp=0
+
+			for tmp in recall_cls:
+				temp+=tmp
+
+			return float(temp/len(recall_cls))
+
+
+		elif avg=='micro':
+			ja=0
+			mo=0
+
+			for tmp in range(len(values)):
+				ja+=c_tp[tmp]
+				mo+=c_tp[tmp]+c_fp[tmp]
+
+			return float(ja/mo)
+
+		else:
+			return "avg expect micro/macro"
+
+
+	def f1_score(self, true, pred, avg='micro'):
+
+		if avg =='micro' or avg =='macro':
+
+			precision=self.precision_score(true,pred,avg)
+			recall=self.recall_score(true,pred,avg)
+		else:
+			return "avg expect micro/macro"
+
+		return (((precision*recall)/(precision+recall))*2)
+
+
+
+
+	def pos_eval(self, fin):
+
+		#temp=os.getcwd()+'/'+fin
+		file=open(fin,'r').read()
+		sents=file.split("\n\n")
+
+		acc = defaultdict(float)
+		t_avg = defaultdict(float)
+
+		for sent in sents:
+			lines=sent.split('\n')
+			for line in lines:
+				tot=line.split('\t')
+
+				if line=='':continue
+
+				wd=tot[0]
+				gold=tot[1]
+				pred=tot[2]
+
+				acc['all']+=1
+				gold_list=gold.split('+')
+				pred_list=pred.split('+')
+
+				t_avg["pr_all"]+=len(pred_list)
+				t_avg["rc_all"]+=len(gold_list)
+
+				if gold==pred:
+					acc["true"]+=1
+					t_avg['pr']+=len(pred_list)
+					t_avg['rc']+=len(gold_list)
+					continue
+				else :
+					intersect=0
+					for g in gold_list:
+						if not g in pred_list: continue
+						intersect+=1
+					t_avg['pr']+=intersect
+					t_avg['rc']+=intersect
+
+
+		t_avg['pr_result'] = t_avg['pr'] / t_avg['pr_all']
+		t_avg['rc_result'] = t_avg['rc'] / t_avg['rc_all']
+
+		return float(acc['true']/acc['all']) ,t_avg['pr_result'],t_avg['rc_result'], f1(t_avg['pr_result'], t_avg['rc_result'])
+
+
+	def f1(self, p, r):
+		return 2 * p * r / (p + r) if p + r else 0
+
+
+	def precision_at_k(self, true: List[int], pred: List[int], k: int) -> float:
+		"""
+		avg = ['micro', 'macro']
+		"""
+
+		relevant = 0
+
+		if k > len(pred):
+			raise ValueError("`k` is bigger than pred's length")
+
+		pred = pred[:k]
+
+		for t in true:
+			if t in pred:
+				relevant += 1
+
+
+		return float(relevant/len(pred))
+
+	def recall_at_k(self, true: List[int], pred: List[int], k: int) -> float:
+
+		relevant = 0
+
+		if k > len(pred):
+			raise ValueError("`k` is bigger than pred's length")
+
+		pred = pred[:k]
+
+		for t in true:
+			if t in pred:
+				relevant += 1
+
+
+		return float(relevant/len(true))
+
+	def hit_rate_at_k(self, user: List[List[int]], pred: List[List[int]], k: int) -> float:
+		hit = 0
+
+		for u_list, p_list in zip(user, pred):
+			try:
+				p_list = p_list[:k]
+			except:
+				raise ValueError("`k` is bigger than pred's length ")
+			for u in u_list:
+				if u in p_list:
+					hit += 1
+					break
+
+		return float(hit/len(user))
+
+	def mean_absolute_error(self, true: Union[torch.Tensor, np.ndarray], pred: Union[torch.Tensor, np.ndarray]) -> float:
+		pass
+
+	def root_mean_square_error(self, true: Union[torch.Tensor, np.ndarray], pred: Union[torch.Tensor, np.ndarray]) -> float:
+		pass
+
+	def _W_CER(self, r, h):
+
+		costs = [[0 for inner in range(len(h)+1)] for outer in range(len(r)+1)]
+
+		DEL_PENALTY=1 # Tact
+		INS_PENALTY=1 # Tact
+		SUB_PENALTY=1 # Tact
+
+		for i in range(1, len(r)+1):
+			costs[i][0] = DEL_PENALTY*i
+
+		for j in range(1, len(h) + 1):
+			costs[0][j] = INS_PENALTY*j
+
+		# computation
+		for i in range(1, len(r)+1):
+			for j in range(1, len(h)+1):
+				if r[i-1] == h[j-1]:
+					costs[i][j] = costs[i-1][j-1]
+				else:
+					substitutionCost = costs[i-1][j-1] + SUB_PENALTY # penalty is always 1
+					insertionCost	= costs[i][j-1] + INS_PENALTY   # penalty is always 1
+					deletionCost	 = costs[i-1][j] + DEL_PENALTY   # penalty is always 1
+
+					costs[i][j] = min(substitutionCost, insertionCost, deletionCost)
+
+		mo = len(r)
+		i = len(r)
+		j = len(h)
+
+		result=(costs[i][j])/mo
+
+		if result>1.0:
+			return 1.0
+		else:
+			return result
+
+
+	def wer(self, reference, candidate):
+		r = Ko_tokenize.word(reference)
+		h = Ko_tokenize.word(candidate)
+
+		return self._W_CER(r,h)
+
+
+	def cer(self, reference,candidate):
+		r = Ko_tokenize.syllable(reference)
+		h = Ko_tokenize.syllable(candidate)
+
+		return self._W_CER(r,h)
+
+
+	def bleu(self, reference, candidate,weights=(0.25,0.25,0.25,0.25), smoothing_function=None):
+
+		if type(candidate)!=list or type(reference)!=list:
+			print("parameter expect list type")
+			return
+
+		reference=list(map(self.tokenize,reference))
+		candidate=Ko_tokenize.word(candidate)
+
+		return sentence_bleu(reference,candidate,weights,smoothing_function=smoothing_function)
+
+
+	def bleu_n(self, reference,candiate,n=1, smoothing_function=None):
+
+		if n==1:
+			return self.bleu(reference,candiate,(1,0,0,0), smoothing_function=smoothing_function)
+		elif n==2:
+			return self.bleu(reference,candiate,(0,1,0,0), smoothing_function=smoothing_function)
+		elif n==3:
+			return self.bleu(reference,candiate,(0,0,1,0), smoothing_function=smoothing_function)
+		elif n==4:
+			return self.bleu(reference,candiate,(0,0,0,1), smoothing_function=smoothing_function)
+
+
+
+
+	def _hyp_sent_split_remove(self, can):
+
+		can_sent=[[tmp.rstrip('.'or'?'or'!'or','or'\n')] for tmp in Ko_tokenize.sentence(can)]
+		return can_sent
+
+	def _ref_sent_split_remove(self, ref):
+
+		ref_sent=[Ko_tokenize.sentence(tmp) for tmp in ref]
+		ref_sent_c=[]
+		for tmp in ref_sent:
+			ref_sent_in=[]
+			for tmp2 in tmp:
+				ref_sent_in.append(Ko_tokenize.word(tmp2.rstrip('.'or'?'or'!'or','or'\n')))
+			ref_sent_c.append(ref_sent_in)
+
+		return ref_sent_c
+
+	def _token(self, ref_stoken, can, n):
+
+		numer=[]
+		ref_len=0
+
+		can=list(ngrams(can,n))
+
+		for tmp in ref_stoken:
+
+			if n==1:
+				ref=list(ngrams(tmp,1))
+			elif n==2:
+				ref=list(ngrams(tmp,2))
+			elif n==3:
+				ref=list(ngrams(tmp,3))
+			else:return 0
+
+			intersect = [t for t in ref if t in can ]
+			numer.append(len(intersect))
+			ref_len+=len(ref)
+
+		try:
+			rec= sum(numer)/ref_len
+		except:
+			rec=0
+
+		return rec
+
+
+
+	def rouge_n(self, ref, can, n=1):
+
+		beta=1
+		rec,prec=0,0
+
+		can_sent=self._hyp_sent_split_remove(can)
+		can_word=list(itertools.chain(*[Ko_tokenize.word(tmp,self.lang) for tmp in can_sent]))
+		ref=self._ref_sent_split_remove(ref)
+
+		r_list=[]
+
+		for tmp in ref:
+			if n==1:
+				r_list.append(self._token(tmp, can_word, 1))
+			elif n==2:
+				r_list.append(self._token(tmp, can_word, 2))
+			elif n==3:
+				r_list.append(self._token(tmp, can_word, 3))
+
+		return max(r_list)
+
+
+
+	def rouge_l(self, ref, can):
+
+		beta=1
+		#check=0
+
+		can= self._hyp_sent_split_remove(can)
+		can=[Ko_tokenize.word(tmp,self.lang) for tmp in can]
+		refs=self._ref_sent_split_remove(ref)
+
+		can_word=list(itertools.chain(*can))
+
+		result_list=[]
+
+		for ref in refs:
+			lcs_list=[]
+			for ri in ref:
+				ri_C=[]
+				for ci in can:
+					temp=self._lcs(ci,ri)
+					ri_C.append(temp)
+
+				ri_C=list(itertools.chain(*ri_C))
+				ri_C=set(ri_C)
+				lcs_list.append(len(ri_C))
+
+			ref_word=list(itertools.chain(*ref))
+
+			R_lcs=sum(lcs_list)/len(ref_word)
+			P_lcs=sum(lcs_list)/len(can_word)
+
+			try:
+				F_lcs= (2*R_lcs*P_lcs)/(R_lcs+P_lcs)
+			except:
+				F_lcs=0
+			result_list.append(F_lcs)
+
+		return max(result_list)
+
+
+
+	def _lcs(self, can, ref):
+
+
+		s1=can
+		s2=ref
+		check=0
+
+		if len(s1)<=len(s2):
+			temp=s1
+			s1=s2
+			s2=temp
+			check=1
+
+		m = [[0] * (1 + len(s2)) for i in range(1 + len(s1))]
+
+		for x in range(1, 1 + len(s1)):
+			for y in range(1, 1 + len(s2)):
+				if s1[x - 1] == s2[y - 1]:
+					m[x][y] = m[x - 1][y - 1] +1
+				else:
+					m[x][y]=max(m[x][y-1],m[x-1][y])
+		f_x=len(s2)+1
+		lcs=m[len(s1)][len(s2)]
+		temp=[]
+
+
+		i=len(s1)
+		j=len(s2)
+
+		while m[i][j]!=0:
+			if(m[i][j]==m[i][j-1]):
+				j-=1
+			elif (m[i][j]==m[i-1][j]):
+				i-=1
+			else:
+				if check==0:
+					temp.append(s1[i-1])
+				if check==1:
+					temp.append(s2[j-1])
+				i-=1
+				j-=1
+
+		return temp
+		'''
+		for y in reversed(range(1,1+len(s1))):
+			for x in reversed(range(1,1+len(s2))):
+				if (m[y][x]-m[y-1][x-1]==1) and (m[y][x]-m[y-1][x]==1) and (m[y][x]-m[y][x-1]==1):
+					if (y==len(s1)+1) and (x==len(s2)):
+						temp.append(x)
+					else:
+						temp.append(x-1)
+
+		print('the police 만  나와줘야',temp)
+		if check==0:
+			word=s1
+		elif check==1:
+			word=s2
+
+		ret_list=[]
+
+		for tmp in range(len(temp)):
+			ret_list.append(word[temp[tmp]])
+
+		return ret_list
+		'''
+
+
+	def _skip_bigrams(self, ref_stoken, can_sent, can, n=1):
+
+		beta=1
+		numer=[]
+		ref_len=0
+
+		candidate=list(skipgrams(can,2,n))
+		can_sent=[Ko_tokenize.word(tmp,self.lang) for tmp in can_sent]
+		can_sk_len=0
+
+		for tmp in ref_stoken:
+			ref=list(skipgrams(tmp,2,n))
+			intersect=[t for t in ref if t in candidate]
+			numer.append(len(intersect))
+			ref_len+=len(ref)
+
+		for tmp in can_sent:
+			can_sk_len+=len(list(skipgrams(tmp,2,n)))
+
+		prec=sum(numer)/can_sk_len
+		rec=sum(numer)/ref_len
+
+		if(prec!=0 and rec!=0):
+			score = ((1 + beta**2)*prec*rec)/float(rec + beta**2*prec)
+		else:
+			score = 0.0
+		return score
+
+
+	def rouge_s(self, ref, can, n):
+
+		can_sent= self._hyp_sent_split_remove(can)
+		can_word=list(itertools.chain(*[Ko_tokenize.word(tmp,self.lang) for tmp in can_sent]))
+		ref= self._ref_sent_split_remove(ref)
+
+
+		r_list=[]
+
+		for tmp in ref:
+			#tmp=list(itertools.chain(*tmp))
+			r_list.append(self._skip_bigrams(tmp,can_sent,can_word,n))
+
+		return max(r_list)
+
+
+	def cider(self, ref, hyp):
+
+		ref_dict=dict()
+		hyp_dict=dict()
+
+		ref_dict[0]=ref
+		hyp_dict[0]=hyp
+
+		cider_score=Cider()
+		score=cider_score.compute_score(ref_dict,hyp_dict)
+
+		return float(score)
+
+	def _process_espresso_output_format(self, result_list):
+		temp_list = []
+		for k in result_list:
+			#k = k.split('_')
+			k = list(k)
+			if k[1] == 'SP' or k[1] == 'SY':
+				continue
+			temp_list.append(k)
+		return temp_list
+
+	def _generate_enum(self, ref, hyp):
+		result_hyp = []
+		result_ref = []
+		for h in hyp:
+			enum_hyp_list = list(enumerate(h))
+			result_hyp.append(enum_hyp_list)
+		for r in ref:
+			enum_ref_list = list(enumerate(r))
+			result_ref.append(enum_ref_list)
+		return result_hyp, result_ref
+
+	def _tag_pos_meteor(self, sent_list):
+		result_list = list()
+		for sent in sent_list:
+			tagged_sent = EspressoTagger(task='pos').tag(sent)
+			tagged_sent = self._process_espresso_output_format(tagged_sent)
+			result_list.append(tagged_sent)
+		return result_list
+
+	def _match_enums(self,
+		enum_hypothesis_list: List[Tuple[int, str]],
+		enum_reference_list: List[Tuple[int, str]],
+	) -> Tuple[List[Tuple[int, int]], List[Tuple[int, str]], List[Tuple[int, str]]]:
+		"""
+		matches exact words in hypothesis and reference and returns
+		a word mapping between enum_hypothesis_list and enum_reference_list
+		based on the enumerated word id.
+
+		:param enum_hypothesis_list: enumerated hypothesis list
+		:param enum_reference_list: enumerated reference list
+		:return: enumerated matched tuples, enumerated unmatched hypothesis tuples,
+				enumerated unmatched reference tuples
+		"""
+		word_match = []
+		# print("test 213" , enum_hypothesis_list)
+		# print("test 124" , enum_reference_list)
+		for i in range(len(enum_hypothesis_list))[::-1]:
+			for j in range(len(enum_reference_list))[::-1]:
+				# print(f"\n \t {enum_hypothesis_list[i][1]} \t {enum_reference_list[j][1]}")
+				if enum_hypothesis_list[i][1] == enum_reference_list[j][1]:
+
+					# print("Check!!")
+					word_match.append(
+						(enum_hypothesis_list[i][0], enum_reference_list[j][0])
+					)
+					enum_hypothesis_list.pop(i)
+					enum_reference_list.pop(j)
+					break
+		return word_match, enum_hypothesis_list, enum_reference_list
+
+
+	def _count_chunks(self, matches: List[Tuple[int, int]]) -> int:
+		"""
+		Counts the fewest possible number of chunks such that matched unigrams
+		of each chunk are adjacent to each other. This is used to calculate the
+		fragmentation part of the metric.
+
+		:param matches: list containing a mapping of matched words (output of align_words)
+		:return: Number of chunks a sentence is divided into post alignment
+		"""
+		i = 0
+		chunks = 1
+		while i < len(matches) - 1:
+			if (matches[i + 1][0] == matches[i][0] + 1) and (
+				matches[i + 1][1] == matches[i][1] + 1
+			):
+				i += 1
+				continue
+			i += 1
+			chunks += 1
+		return chunks
+
+	def _match_syn_with_sejong(self, hyp_list, ref_list):
+		print("2")
+		syn_match = []
+		for i in range(len(hyp_list))[::-1]:
+			temp_syn_list = []
+			#print("test 344434: ", hyp_list[i])
+			if hyp_list[i][1][1] not in self.skip_pos:
+				entrys = ssem.entrys(hyp_list[i][1][0])
+				print("1")
+				print(entrys)
+				for entry in entrys:
+					print(entry)
+					for sense in entry.senses():
+						if sense.syn():
+							temp_syn_list.append(sense.syn())
+			if temp_syn_list:
+				hyp_list[i][1].append(deepcopy(temp_syn_list))
+
+			for j in range(len(ref_list))[::-1]:
+				is_break = False
+				if len(hyp_list[i][1]) == 3:
+					for syn in hyp_list[i][1][2]:
+
+						if syn[0] == ref_list[j][1][0]:
+							syn_match.append(
+								(hyp_list[i][0], ref_list[j][0])
+							)
+							is_break = True
+							hyp_list.pop(i)
+							ref_list.pop(j)
+							break
+				else:
+					if hyp_list[i][1] == ref_list[1][1]:
+						syn_match.append(
+							(hyp_list[i][0], ref_list[j][0])
+						)
+						is_break = True
+						hyp_list.pop(i)
+						ref_list.pop(j)
+				if is_break:
+					break
+
+
+
+			# print("test 231232 ", hyp_list[i])
+
+
+		return syn_match, hyp_list, ref_list
+
+	def meteor(self, ref, hyp):
+		ref_tag = self._tag_pos_meteor(ref)
+		hyp_tag = self._tag_pos_meteor(hyp)
+		meteors = []
+		alpha = 0.9
+		beta = 3.0
+		gamma = 0.5
+		enum_hyp, enum_ref = self._generate_enum(ref_tag, hyp_tag)
+		print("$")
+		# print("test 13333 ", enum_hyp)
+		for reference in enum_ref:
+			hyp_len = len(enum_hyp[0])
+			ref_len = len(reference)
+
+			# 단어/어간 매칭
+			word_match, enum_hyp_list, enum_ref_list = self._match_enums(deepcopy(enum_hyp[0]), reference)
+			syn_match, enum_hyp_list, enum_ref_list = self._match_syn_with_sejong(enum_hyp_list, enum_ref_list)
+			# print("test 123344 " ,enum_ref_list) ## [(0, ['오늘', 'NN']), (6, ['이', 'VB']), (7, ['었다', 'EE'])]
+
+			final_match = sorted(word_match + syn_match)
+
+			#최종 결과 계산
+			final_match_count = len(final_match)
+
+
+			precision = float(final_match_count) / hyp_len
+			recall = float(final_match_count) / ref_len
+			fmean = (precision * recall) / (alpha * precision + (1 - alpha) * recall)
+			chunk_count = float(self._count_chunks(final_match))
+			frag = 0.0
+			if final_match_count != 0:
+				frag = chunk_count / final_match_count
+			else:
+				frag = 0.0
+			penalty = gamma * frag ** beta
+			meteors.append((1 - penalty) * fmean)
+
+			# print(word_match)
+
+		return max(meteors)
+
+
+def demo():
+	y_pred = [5, 2, 4, 1, 3, 2, 5, 6, 7]
+	y_true = [1, 3, 6, 7, 1, 5]
+
+	user = [[5, 3, 2], [9, 1, 2], [3, 5, 6], [7, 2, 1]]
+	h_pred = [[15, 6, 21, 3], [15, 77, 23, 14], [51, 23, 21, 2], [53, 2, 1, 5]]
+
+	metric = DefaultMetric()
+	print(metric.precision_at_k(y_true,  y_pred, 3))
+	print(metric.recall_at_k(y_true,y_pred, 3))
+	print(metric.hit_rate_at_k(user, h_pred, 2))
+
+	hyp='봉준호 감독이 아카데미에서 국제영화상을 수상했다.'
+	ref=['봉준호가 아카데미에서 각본상을 탔다.']
+	re = metric.meteor(ref, hyp)
+	print(re)
+
+if __name__=="__main__":
+	demo()