seq2seq-arabic-dialect-lemmatization/utils.py at main · CAMeL-Lab/seq2seq-arabic-dialect-lemmatization · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
from camel_tools.disambig.bert import BERTUnfactoredDisambiguator
from camel_tools.morphology.database import MorphologyDB
from camel_tools.morphology.analyzer import Analyzer
import pandas as pd

def load_disambiguators(dialect, back_off='NOAN_PROP'):
    if dialect == 'egy':
        egy_db = MorphologyDB('morphological DBs/calima-egy-c044_0.1.1.utf8.db')
        egy_calima_analyzer = Analyzer(egy_db, back_off)

        egy_bert = BERTUnfactoredDisambiguator.pretrained(
            'egy', top=5000, pretrained_cache=False
        )
        egy_bert._analyzer = egy_calima_analyzer
        disambig_list = [egy_bert]

    elif dialect == 'glf':
        egy_db = MorphologyDB('morphological DBs/calima-egy-c044_0.1.1.utf8.db')
        egy_calima_analyzer = Analyzer(egy_db, back_off)

        egy_bert = BERTUnfactoredDisambiguator.pretrained(
            'egy', top=5000, pretrained_cache=False
        )
        egy_bert._analyzer = egy_calima_analyzer
        disambig_list = [egy_bert]

        glf_bert = BERTUnfactoredDisambiguator.pretrained(
            'glf', top=5000, pretrained_cache=False
        )
        glf_bert._analyzer._backoff = back_off

        lev_bert = BERTUnfactoredDisambiguator.pretrained(
            'lev', top=5000, pretrained_cache=False
        )
        lev_bert._analyzer._backoff = back_off
        disambig_list = [glf_bert, egy_bert, lev_bert]

    elif dialect == 'lev':
        egy_db = MorphologyDB('morphological DBs/calima-egy-c044_0.1.1.utf8.db')
        egy_calima_analyzer = Analyzer(egy_db, back_off)

        egy_bert = BERTUnfactoredDisambiguator.pretrained(
            'egy', top=5000, pretrained_cache=False
        )
        egy_bert._analyzer = egy_calima_analyzer
        disambig_list = [egy_bert]

        glf_bert = BERTUnfactoredDisambiguator.pretrained(
            'glf', top=5000, pretrained_cache=False
        )
        glf_bert._analyzer._backoff = back_off

        lev_bert = BERTUnfactoredDisambiguator.pretrained(
            'lev', top=5000, pretrained_cache=False
        )
        lev_bert._analyzer._backoff = back_off
        disambig_list = [lev_bert, egy_bert, glf_bert]

    return disambig_list

def remove_sukun_and_tatweel(text):
    if isinstance(text, str):
        return text.replace('\u0652', '').replace('\u0640', '')
    return text

def evaluate_disambiguation_with_sentences(
    df,
    disambig_list):
    # Step 1: Prepare list of tokenized sentences
    sentences_list = []
    indices_list = []
    # Group and sort
    for _, group in df.groupby('sentence_index'):
        group = group.sort_values('word_index')
        sentences_list.append(list(group['word'].astype(str)))
        indices_list.append(group.index.tolist())

    # Step 2: Disambiguate with each disambiguator
    all_flattened_analyses = []

    for i, disambig in enumerate(disambig_list):
        results = disambig.disambiguate_sentences(sentences_list)

        for sentence_idx, sentence in enumerate(results):
            indices = indices_list[sentence_idx]

            for word_idx, word in enumerate(sentence):
                word_text = word.word
                original_idx = indices[word_idx]

                for analysis in word.analyses:
                    entry = {
                        'disambig_model': disambig,
                        'sentence_index': sentence_idx,
                        'word_index': word_idx,
                        'word': word_text,
                        'original_index': original_idx,
                        'diac': analysis.diac,
                        'score': analysis.score
                    }
                    entry.update(analysis.analysis)
                    all_flattened_analyses.append(entry)

    # Step 3: Create combined DataFrame
    disambig_df = pd.DataFrame(all_flattened_analyses)

    disambig_df = disambig_df.drop_duplicates(
        subset=['sentence_index', 'word_index', 'lex', 'pos', 'stemgloss']
    ).reset_index(drop=True)

    counts = disambig_df.groupby(['sentence_index', 'word_index'])['score'].transform('count')

    # Apply filtering condition
    disambig_df = disambig_df[
        (disambig_df['score'] == 1) |
        ((counts == 1) & (disambig_df['score'] == 0))
    ].reset_index(drop=True)

    # Step 2: Create s2s_lookup once
    s2s_lookup = df.set_index(['sentence_index', 'word_index'])['s2s_predicted_lemma'].to_dict()

    # Step 3: Process each word individually (like code 1)
    selected_rows = []

    grouped = disambig_df.groupby(['sentence_index', 'word_index'])

    for (s_idx, w_idx), group in grouped:
        # print(s2s_lookup.get((s_idx, w_idx), ''))
        # print(s_idx)
        # print(w_idx)
        val = s2s_lookup.get((s_idx, w_idx), '')
        predicted_lemma = str(val).strip() if not pd.isna(val) else ''

        # Look for exact match on lex
        match_idx = group[group['lex'] == predicted_lemma].index

        if not match_idx.empty:
            selected = group.loc[match_idx[0]]
        else:
            selected = group.iloc[0]

        selected_rows.append(selected)

    top1_df = pd.DataFrame(selected_rows).reset_index(drop=True)

    # Step 5: Merge predictions with gold labels
    df['lex'] = top1_df['lex']
    df['pos'] = top1_df['pos']
    df['stemgloss'] = top1_df['stemgloss']
    df['lex'] = df['lex'].apply(remove_sukun_and_tatweel)

    mask = df['pos'].isin(['punc', 'digit'])
    df.loc[mask, 's2s_predicted_lemma'] = df.loc[mask, 'word']

    return df

def get_context_window(df, sentence_index, word_index, window_size=2):
    # Get words and their indices for the sentence
    sentence_data = df[df['sentence_index'] == sentence_index][['word', 'word_index']].sort_values(by='word_index')
    words = sentence_data['word'].astype(str).tolist()
    indices = sentence_data['word_index'].tolist()
    # oracle_pos_tags = sentence_data['pos_tag_encoded'].tolist()

    # Find the position of the target word
    target_pos = indices.index(word_index)

    # Define context window
    start_idx = max(0, target_pos - window_size)
    end_idx = min(len(words), target_pos + window_size + 1)
    context_words = words[start_idx:end_idx]

    # Mark the target word with <target> tags
    target_word_idx = target_pos - start_idx
    context_words[target_word_idx] = f"<target>{context_words[target_word_idx]}<target>"

    return f"lemmatize: {' '.join(context_words)}"