Typology-aware tokenizers are built on a specific claim: that segmenting text along morpheme boundaries keeps information that frequency-driven tokenizers lose. That assumption underpins a line of research showing such tokenizers help downstream tasks. But the new paper, "Does Morphological Correctness Matter for Tokenization? A Scrambled Control Study in Tagalog," asks whether morphological correctness itself is the cause.
To test this, the authors use a scrambled control in Tagalog. By breaking the link between segmentation and true morpheme boundaries, they aim to separate the effect of morphological correctness from other properties of the tokenizer. The design is meant to challenge the usual finding that typology-aware segmentation helps.
At the time of writing, the arXiv abstract available in the source is cut off mid-sentence, so the paper's findings are not reported here. The study's contribution, as described, is the control methodology rather than a new claim about Tagalog specifically.