IndoBERT-based Complex Word Identification for Indonesian Text Simplification with Dyslexia-Oriented Thresholding

Abstract

This study investigates complex word identification (CWI) in Indonesian as a first step towards text simplification for dyslexic readers. This study uses CWI as a token-level classifier with IndoBERT, employing WordPiece tokenization and subword-to-word label alignment (only the first subword of each word receives a gold label; the rest are protected from label loss). To support a wide range of application needs, this study calibrates the decision threshold (τ) on development data and reports the results at multiple operating points to reveal the precision-recall trade-off. The model is fine-tuned on a token-level Indonesian corpus and evaluated on a held-out test set of 11.574 tokens (NonCWI = 10.769; CWI = 805). On this test set, IndoBERT-CWI delivers stable performance at three representative thresholds: τ = 0.40 (Precision 0.736, Recall 0.909, F1 0.814, Accuracy 0.971), τ = 0.46 (Precision 0.743, Recall 0.896, F1 0.812, Accuracy 0.971), and τ = 0.60 (Precision 0.771, Recall 0.856, F1 0.811, Accuracy 0.972). The precision-recall (PR) curve on the test set has an Average Precision of 0.783 and shows τ = 0.46 as a balanced operating point (highest F1), while τ = 0.40 is preferred when maximizing coverage (recall) is crucial to avoid missing complex tokens. A brief analysis shows that model errors occur most frequently when frequently occurring function words and proper nouns are marked as complex. This pattern is consistent with the characteristics of a news-based corpus and provides a practical suggestion that the simplification module should filter out entity names and function words from processing. Our results demonstrate that a well-tuned IndoBERT is effective for Indonesian CWI, with easily calibrated thresholds to balance precision–recall in dyslexia applications, and provide directions for improvement in further lexical simplification.