DeReKo/tiger_turku_evaluate.py

from CoNLL_Annotation import *
from collections import Counter
import pandas as pd


def eval_lemma(sys, gld):
    match, err, symbol = 0, 0, []
    mistakes = []
    for i, gld_tok in enumerate(gld.tokens):
        if gld_tok.lemma == sys.tokens[i].lemma:
            match += 1
        elif not sys.tokens[i].lemma.isalnum(): # This was added because Turku does not lemmatize symbols (it only copies them) => ERR ((',', '--', ','), 43642)
            symbol.append(sys.tokens[i].lemma)
            if sys.tokens[i].word == sys.tokens[i].lemma:
                match += 1
            else:
                err += 1
        else:
            err += 1
            mistakes.append((gld_tok.word, gld_tok.lemma, sys.tokens[i].lemma))
    return match, err, symbol, mistakes
    


if __name__ == "__main__":

    # Read the Original TiGeR Annotations
    gld_filename = "/home/daza/datasets/TIGER_conll/tiger_release_aug07.corrected.16012013.conll09"
    gld_generator = read_conll_generator(gld_filename, token_class=CoNLL09_Token)
    
    # Read the Annotations Generated by the Automatic Parser [Turku] 
    sys_filename = "/home/daza/datasets/TIGER_conll/tiger_turku_parsed.conllu"
    sys_generator = read_conll_generator(sys_filename, token_class=CoNLLUP_Token)
    
    lemma_all_match, lemma_all_err, lemma_all_mistakes = 0, 0, []
    lemma_all_symbols = []
    
    for i, (s,g) in enumerate(zip(sys_generator, gld_generator)):
        assert len(s.tokens) == len(g.tokens), "Token Mismatch!"
        lemma_match, lemma_err, lemma_sym, mistakes = eval_lemma(s,g)
        lemma_all_match += lemma_match
        lemma_all_err += lemma_err
        lemma_all_mistakes += mistakes
        lemma_all_symbols += lemma_sym
    
    print(f"Lemma Matches = {lemma_all_match} || Errors = {lemma_all_err} || Symbol Chars = {len(lemma_all_symbols)}")
    print(f"Lemma Accuracy = {lemma_all_match*100/(lemma_all_match + lemma_all_err)}%")
    lemma_miss_df = pd.DataFrame(lemma_all_mistakes, columns =['Gold_Word', 'Gold_Lemma', 'Sys_Lemma']).value_counts()
    lemma_miss_df.to_csv(path_or_buf="LemmaErrors.tsv", sep="\t")
        
#         
#     the_count = Counter(lemma_all_mistakes).most_common(100)
#     for x in the_count:
#         print(x)