Unsupervised Morphological Tree Tokenizer

TL;DR

This paper presents an unsupervised deep model that induces morphological structures within words to improve tokenization, outperforming traditional methods like BPE and WordPiece in segmentation and language modeling tasks.

Contribution

It introduces a novel deep model with MorphOverriding for unsupervised morphological structure induction, enhancing tokenization quality without annotated data.

Findings

Outperforms BPE and WordPiece in segmentation tasks

Effectively retains complete morphemes during tokenization

Improves language modeling performance

Abstract

As a cornerstone in language modeling, tokenization involves segmenting text inputs into pre-defined atomic units. Conventional statistical tokenizers often disrupt constituent boundaries within words, thereby corrupting semantic information. To address this drawback, we introduce morphological structure guidance to tokenization and propose a deep model to induce character-level structures of words. Specifically, the deep model jointly encodes internal structures and representations of words with a mechanism named $\textit{MorphOverriding}$ to ensure the indecomposability of morphemes. By training the model with self-supervised objectives, our method is capable of inducing character-level structures that align with morphological rules without annotated training data. Based on the induced structures, our algorithm tokenizes words through vocabulary matching in a top-down manner. Empirical results indicate that the proposed method effectively retains complete morphemes and outperforms widely adopted methods such as BPE and WordPiece on both morphological segmentation tasks and language modeling tasks. Code is available at https://github.com/martianmartina/TreeTokenizer.