Homographs are words that share identical spelling but differ in meaning and pronunciation. In Tagalog, these distinctions are disambiguated in writing through tuldik (diacritical marks), which are rarely preserved in contemporary Tagalog text. This scarcity of diacritics limits diacritic restoration, which is useful for downstream tasks such as prosody selection for text-to-speech synthesis. To address this gap, this study introduces TDR, an annotated corpus of 40 frequent Tagalog homographs constructed from news text. It also introduces a collocation-driven retrieval method to expand the dataset and mitigate severe class imbalance among minor variants. Finally, we show that TDR supplies a capability absent from pretrained models, and that this capability can be learned by small, inexpensive models as well as by fine-tuned LLMs.
Figures & tables
Workable
Moderate
Extreme
bukas*
abala*
bagong
paano
galing
baka
bilang
pala
galit*
buhay
gabi
pamahalaan
gamit*
butas*
kaibigan
panalo
huli*
dating
kasama
pangulong
ika
huling
kaya
patuloy
Table 1: The 40 target homographs in the Pre-TDR and TDR corpus, grouped by asymmetry type. Asterisks indicate additions from [ 16 ] . Cell shading denotes improvements relative to Pre-TDR (Yellow: extreme → moderate; Green: moderate → workable; Blue: extreme → workable).
AM
Description
Value
LogDice [ 39 ]
Log-scaled Dice coefficient, measures strength of association regardless of corpus size.
≥7.0
MI-2 [ 40 ]
Bias-corrected mutual information, distinguishes exclusivity of homographs.
≥6.0
ΔP2 [ 41 ]
Metric of collocation directionality, shows how strongly a collocate predicts the node.
≥0.005
t -score [ 42 ]
Co-occurrence measure with adjusted variance, gives reliability that a collocate linkage is not due to chance.
≥2.0
Table 2: Association measures used for collocate identification, with retention cutoffs. A candidate must meet all four.
Homograph
Collocate
Completed Phrase
abála
malaking
malaking abála
abála
mabawasan
mabawasan [ang] abála
báka
kalabaw
kalabaw [at] báka
buháy
manatiling
manatiling buháy
butás
bulsa
butás [ang] bulsa
datíng
lakas
lakas [ng] datíng
Table 3: Representative collocates extracted for minor-variant homographs, used as search leads for corpus expansion.
Figure 1: The decision-making process in the targeted article retrieval. Note that the brainstorming based on dictionary meaning relied on native speaker’s intuition.
Corpus
Articles
Sentences
Homograph Counts
Pre-TDR
8,800
100,087
38,096
TDR
10,065
146,310
48,459
Table 4: Corpus before and after collocation-driven expansion.
Homograph
Variant 1
Variant 2
Variant 3
Ratio
κ
galit
galít: 197
gálit: 191
0.4923
0.939
kilala
kilalá: 254
kilála: 238
0.4837
0.956
puno
punô: 170
punò: 158
0.4817
0.983
taong
táong: 666
taóng: 608
0.4772
0.954
galing
gáling: 368
galíng: 280
0.4321
0.975
huli
hulí: 333
húli: 232
0.4106
0.959
Table 5: Variant counts, asymmetry ratio, and inter-annotator agreement (free-marginal κ ) for all homographs in the TDR corpus. Rows in red have the lowest agreement value.
Model
Macro-F1
W+M F1
Weighted F1
Majority
0.558
0.396
0.568
Logistic regression
0.800
0.817
0.794
RoBERTa-base
0.654
0.601
0.646
RoBERTa-large
0.559
0.505
0.558
RoBERTa-Tagalog-base
0.786
0.817
0.754
RoBERTa-Tagalog-large
0.815
0.844
0.826
Table 6: Multiple-choice disambiguation results. W+M F1 considers workably and moderately asymmetric homographs only. Weighted: per-homograph F1 weighted by test instances.
Model
Exact Match
Recovery
Majority †
85.20
85.20
LR †
93.73
93.73
Llama, zero-shot
0.00
0.02
Llama, fine-tuned
95.50
97.04
SEA-LION, zero-shot
0.18
0.74
SEA-LION, fine-tuned
95.47
96.99
Table 7: Sentence reconstruction results, shown in percentages. For simulated baselines ( † ), exact match equals recovery.
Input Text
N
HPER
Undiacritized
129
35.66
LLM-restored
129
12.40
Gold (diacritized)
130
11.54
Table 8: HPER by input text, based on the majority vote of annotators. Clips with invalid LLM output and marked as unintelligible are excluded.