blob: 11ffd51220fb88a2c0a6765ce82894c33cac185f [file] [log] [blame]
dazae3bc92e2020-11-04 11:06:26 +01001from lib.CoNLL_Annotation import *
2from collections import Counter, defaultdict
3import pandas as pd
4import numpy as np
5from sklearn.metrics import precision_recall_fscore_support as eval_f1
6from tabulate import tabulate
7import logging, argparse, sys
8from datetime import datetime
9
10
11tree_tagger_fixes = {
12 "die": "der",
13 "eine": "ein",
14 "dass": "daß",
15 "keine": "kein",
16 "dies": "dieser",
17 "erst": "erster",
18 "andere": "anderer",
19 "alle": "aller",
20 "Sie": "sie",
21 "wir": "uns",
22 "alle": "aller",
23 "wenige": "wenig"
24}
25
26
27def save_evaluated(all_sys, all_gld, out_path, print_gold=True):
28 with open(out_path, "w") as out:
29 if print_gold:
30 out.write(f"ORIGINAL_CORPUS_TAGS\n\nTAG\tGLD_COUNT\tSYS_COUNT\n")
31 for g_tag,g_count in sorted(all_gld.items()):
32 s_count = all_sys.get(g_tag, 0)
33 out.write(f"{g_tag}\t{g_count}\t{s_count}\n")
34
35 out.write("\n\nSYSTEM_ONLY_TAGS\n\nTAG\tG_COUNT\tSYS_COUNT\n")
36 for s_tag,s_count in sorted(all_sys.items()):
37 g_count = all_gld.get(s_tag, 0)
38 if g_count == 0:
39 out.write(f"{s_tag}\t{g_count}\t{s_count}\n")
40
41
42
43def eval_lemma(sys, gld):
44 match, err, symbol = 0, 0, []
45 y_gld, y_pred, mistakes = [], [], []
46 for i, gld_tok in enumerate(gld.tokens):
47 # sys_lemma = tree_tagger_fixes.get(sys.tokens[i].lemma, sys.tokens[i].lemma) # Omit TreeTagger "errors" because of article lemma disagreement
48 sys_lemma = sys.tokens[i].lemma
49 y_gld.append(gld_tok.pos_tag)
50 y_pred.append(sys_lemma)
51 if gld_tok.lemma == sys_lemma:
52 match += 1
53 elif not sys.tokens[i].lemma.isalnum(): # Turku does not lemmatize symbols (it only copies them) => ERR ((',', '--', ','), 43642)
54 symbol.append(sys.tokens[i].lemma)
55 if sys.tokens[i].word == sys.tokens[i].lemma:
56 match += 1
57 else:
58 err += 1
59 else:
60 err += 1
61 mistakes.append((gld_tok.word, gld_tok.lemma, sys.tokens[i].lemma))
62 return y_gld, y_pred, match, err, symbol, mistakes
63
64
65def eval_pos(sys, gld):
66 match, mistakes = 0, []
67 y_gld, y_pred = [], []
68 for i, gld_tok in enumerate(gld.tokens):
69 y_gld.append(gld_tok.pos_tag)
70 y_pred.append(sys.tokens[i].pos_tag)
71 # pos_all_pred[gld_tok.pos_tag] += 1
72 # pos_all_gold[sys.tokens[i].pos_tag] += 1
73 if gld_tok.pos_tag == sys.tokens[i].pos_tag:
74 match += 1
75 elif gld_tok.pos_tag == "$." and sys.tokens[i].pos_tag == "$":
76 match += 1
77 y_pred = y_pred[:-1] + ["$."]
78 else:
79 mistakes.append((gld_tok.word, gld_tok.pos_tag, sys.tokens[i].pos_tag))
80 return y_gld, y_pred, match, mistakes
81
82
83
84if __name__ == "__main__":
85 """
86 EVALUATIONS:
87
88 ********** TIGER CORPUS ALL ************
89
90 python systems/evaluate.py -t Turku --corpus_name Tiger\
91 --sys_file /home/daza/datasets/TIGER_conll/tiger_turku_parsed.conllu \
92 --gld_file /home/daza/datasets/TIGER_conll/tiger_release_aug07.corrected.16012013.conll09
93
94 python systems/evaluate.py -t SpaCy --corpus_name Tiger\
95 --sys_file /home/daza/datasets/TIGER_conll/tiger_spacy_parsed.conllu \
96 --gld_file /home/daza/datasets/TIGER_conll/tiger_release_aug07.corrected.16012013.conll09
97
98 python systems/evaluate.py -t RNNTagger --corpus_name Tiger\
99 --sys_file /home/daza/datasets/TIGER_conll/tiger_all.parsed.RNNTagger.conll \
100 --gld_file /home/daza/datasets/TIGER_conll/tiger_release_aug07.corrected.16012013.conll09
101
102 python systems/evaluate.py -t TreeTagger --corpus_name Tiger\
103 --sys_file /home/daza/datasets/TIGER_conll/tiger_all.parsed.TreeTagger.conll \
104 --gld_file /home/daza/datasets/TIGER_conll/tiger_release_aug07.corrected.16012013.conll09
105
106 ********** UNIVERSAL DEPENDENCIES TEST-SET ************
107
108 python systems/evaluate.py -t Turku --gld_token_type CoNLLUP_Token --corpus_name DE_GSD\
109 --sys_file /home/daza/datasets/ud-treebanks-v2.2/UD_German-GSD/de_gsd-ud-test.conllu.parsed.0.conllu \
110 --gld_file /home/daza/datasets/ud-treebanks-v2.2/UD_German-GSD/de_gsd-ud-test.conllu
111
112 python systems/evaluate.py -t SpaCyGL --gld_token_type CoNLLUP_Token --corpus_name DE_GSD\
113 --sys_file /home/daza/datasets/ud-treebanks-v2.2/UD_German-GSD/de_gsd-ud-test.parsed.germalemma.conllu \
114 --gld_file /home/daza/datasets/ud-treebanks-v2.2/UD_German-GSD/de_gsd-ud-test.conllu
115
116 python systems/evaluate.py -t SpaCy --gld_token_type CoNLLUP_Token --corpus_name DE_GSD\
117 --sys_file /home/daza/datasets/ud-treebanks-v2.2/UD_German-GSD/de_gsd-ud-test.parsed.conllu \
118 --gld_file /home/daza/datasets/ud-treebanks-v2.2/UD_German-GSD/de_gsd-ud-test.conllu
119
120 python systems/evaluate.py -t RNNTagger --gld_token_type CoNLLUP_Token --corpus_name DE_GSD\
121 --sys_file /home/daza/datasets/ud-treebanks-v2.2/UD_German-GSD/de_gsd-ud-test.RNNtagger.parsed.conll \
122 --gld_file /home/daza/datasets/ud-treebanks-v2.2/UD_German-GSD/de_gsd-ud-test.conllu
123
124 python systems/evaluate.py -t TreeTagger --gld_token_type CoNLLUP_Token --corpus_name DE_GSD\
125 --sys_file /home/daza/datasets/ud-treebanks-v2.2/UD_German-GSD/de_gsd-ud-test.treetagger.parsed.conll \
126 --gld_file /home/daza/datasets/ud-treebanks-v2.2/UD_German-GSD/de_gsd-ud-test.conllu
127
128 """
129
130 # =====================================================================================
131 # INPUT PARAMS
132 # =====================================================================================
133 parser = argparse.ArgumentParser()
134 parser.add_argument("-s", "--sys_file", help="System output in CoNLL-U Format", required=True)
135 parser.add_argument("-g", "--gld_file", help="Gold Labels to evaluate in CoNLL-U Format", required=True)
136 parser.add_argument("-t", "--type_sys", help="Which system produced the outputs", default="system")
137 parser.add_argument("-c", "--corpus_name", help="Corpus Name for Gold Labels", required=True)
138 parser.add_argument("-gtt", "--gld_token_type", help="CoNLL Format of the Gold Data", default="CoNLL09_Token")
139 parser.add_argument("-cs", "--comment_str", help="CoNLL Format of comentaries inside the file", default="#")
140 args = parser.parse_args()
141
142 # =====================================================================================
143 # LOGGING INFO ...
144 # =====================================================================================
145 logger = logging.getLogger(__name__)
146 console_hdlr = logging.StreamHandler(sys.stdout)
147 file_hdlr = logging.FileHandler(filename=f"logs/Eval_{args.corpus_name}.{args.type_sys}.log")
148 logging.basicConfig(level=logging.INFO, handlers=[console_hdlr, file_hdlr])
149 now_is = datetime.now().strftime('%Y-%m-%d %H:%M:%S')
150 logger.info(f"\n\nEvaluating {args.corpus_name} Corpus {now_is}")
151
152 # Read the Original GOLD Annotations [CoNLL09, CoNLLUP]
153 gld_generator = read_conll_generator(args.gld_file, token_class=get_token_type(args.gld_token_type), comment_str=args.comment_str)
154 # Read the Annotations Generated by the Automatic Parser [Turku, SpaCy, RNNTagger]
155 if args.type_sys == "RNNTagger":
156 sys_generator = read_conll_generator(args.sys_file, token_class=RNNTagger_Token, comment_str="#")
157 elif args.type_sys == "TreeTagger":
158 sys_generator = read_conll_generator(args.sys_file, token_class=RNNTagger_Token, sent_sep="</S>", comment_str="#")
159 else:
160 sys_generator = read_conll_generator(args.sys_file, token_class=CoNLLUP_Token, comment_str="#")
161
162 lemma_all_match, lemma_all_err, lemma_all_mistakes = 0, 0, []
163 lemma_all_symbols, sys_only_lemmas = [], []
164 pos_all_match, pos_all_err, pos_all_mistakes = 0, 0, []
165 pos_all_pred, pos_all_gld = [], []
166 lemma_all_pred, lemma_all_gld = [], []
167 n_sents = 0
168
169 for i, (s,g) in enumerate(zip(sys_generator, gld_generator)):
170 # print([x.word for x in s.tokens])
171 # print([x.word for x in g.tokens])
172 assert len(s.tokens) == len(g.tokens), f"Token Mismatch! S={len(s.tokens)} G={len(g.tokens)} IX={i+1}"
173 n_sents += 1
174 # Lemmas ...
175 lemma_gld, lemma_pred, lemma_match, lemma_err, lemma_sym, mistakes = eval_lemma(s,g)
176 lemma_all_match += lemma_match
177 lemma_all_err += lemma_err
178 lemma_all_mistakes += mistakes
179 lemma_all_symbols += lemma_sym
180 lemma_all_pred += lemma_pred
181 lemma_all_gld += lemma_gld
182 # POS Tags ...
183 pos_gld, pos_pred, pos_match, pos_mistakes = eval_pos(s, g)
184 pos_all_pred += pos_pred
185 pos_all_gld += pos_gld
186 pos_all_match += pos_match
187 pos_all_err += len(pos_mistakes)
188 pos_all_mistakes += pos_mistakes
189
190 logger.info(f"A total of {n_sents} sentences were analyzed")
191
192 # Lemmas ...
193 logger.info(f"Lemma Matches = {lemma_all_match} || Errors = {lemma_all_err} || Symbol Chars = {len(lemma_all_symbols)}")
194 logger.info(f"Lemma Accuracy = {(lemma_all_match*100/(lemma_all_match + lemma_all_err)):.2f}%\n")
195 lemma_miss_df = pd.DataFrame(lemma_all_mistakes, columns =['Gold_Word', 'Gold_Lemma', 'Sys_Lemma']).value_counts()
196 lemma_miss_df.to_csv(path_or_buf=f"outputs/LemmaErrors.{args.corpus_name}.{args.type_sys}.tsv", sep="\t")
197 save_evaluated(Counter(lemma_all_pred), Counter(lemma_all_gld),
198 f"outputs/Lemma-Catalogue.{args.corpus_name}.{args.type_sys}.txt", print_gold=False)
199
200 # POS Tags ...
201 logger.info(f"POS Matches = {pos_all_match} || Errors = {pos_all_err}")
202 logger.info(f"POS Tagging Accuracy = {(pos_all_match*100/(pos_all_match + pos_all_err)):.2f}%\n")
203 pos_miss_df = pd.DataFrame(pos_all_mistakes, columns =['Gold_Word', 'Gold_POS', 'Sys_POS']).value_counts()
204 pos_miss_df.to_csv(path_or_buf=f"outputs/POS-Errors.{args.corpus_name}.{args.type_sys}.tsv", sep="\t")
205 save_evaluated(Counter(pos_all_pred), Counter(pos_all_gld), f"outputs/POS-Catalogue.{args.corpus_name}.{args.type_sys}.txt")
206
207 ordered_labels = sorted(set(pos_all_gld))
208 p_labels, r_labels, f_labels, support = eval_f1(y_true=pos_all_gld, y_pred=pos_all_pred, labels=ordered_labels , average=None)
209 scores_per_label = zip(ordered_labels, [x*100 for x in p_labels], [x*100 for x in r_labels], [x*100 for x in f_labels])
210 logger.info("\n\n")
211 logger.info(tabulate(scores_per_label, headers=["POS Tag","Precision", "Recall", "F1"], floatfmt=".2f"))
212 p_labels, r_labels, f_labels, support = eval_f1(y_true=np.array(pos_all_gld), y_pred=np.array(pos_all_pred), average='macro', zero_division=0)
213 logger.info(f"Total Prec = {p_labels*100}\tRec = {r_labels*100}\tF1 = {f_labels*100}")
214
215