[Paper Review] Chinese Character Decomposition for Neural MT with Multi-Word Expressions

Lifeng Han, Gareth J. F. Jones+2 more|arXiv (Cornell University)|Apr 9, 2021

Natural Language Processing Techniques4 citations

TL;DR

This paper proposes a novel neural machine translation (NMT) framework that leverages multi-level decomposition of Chinese characters—radicals, strokes, and intermediate forms—alongside Multi-Word Expression (MWE) integration to improve translation performance. It finds that decomposition at level 1 (radicals) and level 3 (strokes) preserves semantic meaning and achieves BLEU scores comparable to full characters, while level 2 (intermediate characters) harms performance due to misleading semantic units.

ABSTRACT

Chinese character decomposition has been used as a feature to enhance Machine Translation (MT) models, combining radicals into character and word level models. Recent work has investigated ideograph or stroke level embedding. However, questions remain about different decomposition levels of Chinese character representations, radical and strokes, best suited for MT. To investigate the impact of Chinese decomposition embedding in detail, i.e., radical, stroke, and intermediate levels, and how well these decompositions represent the meaning of the original character sequences, we carry out analysis with both automated and human evaluation of MT. Furthermore, we investigate if the combination of decomposed Multiword Expressions (MWEs) can enhance the model learning. MWE integration into MT has seen more than a decade of exploration. However, decomposed MWEs has not previously been explored.

Motivation & Objective

• To investigate the impact of multi-level Chinese character decomposition—radicals, strokes, and intermediate forms—on neural machine translation.
• To evaluate whether decomposed representations preserve the semantic meaning of original character sequences.
• To examine the effectiveness of integrating Multi-Word Expressions (MWEs) into NMT models using decomposed character representations.
• To develop and release an open-source toolkit for Chinese character decomposition and a bilingual MWE-annotated corpus for research use.
• To compare automatic (BLEU) and human evaluation (Direct Assessment) results across decomposition levels and MWE integration strategies.

Proposed method

• The authors design a multi-level decomposition pipeline that breaks down Chinese characters into radicals (level 1), strokes (level 3), and intermediate characters (level 2), with level 2 representing morphologically smaller but semantically independent characters.
• They train a sequence-to-sequence NMT model with attention mechanism using character-level, radical-level, and stroke-level representations as input embeddings.
• The model incorporates MWEs by treating them as single units during training and decoding, with bilingual MWEs extracted and annotated in a new open-source corpus.
• Human evaluation is conducted using Direct Assessment (DA) with standardized z-scores to assess translation quality across systems.
• Automatic evaluation is performed using BLEU, with significance testing planned for future work using METEOR and LEPOR.
• An IDS-based method is used for stroke-level extraction, though limitations in full stroke-level capability are acknowledged.

Experimental results

Research questions

• RQ1 How do different levels of Chinese character decomposition—radicals, intermediate characters, and strokes—affect the semantic representation and translation quality in NMT?
• RQ2 Does the integration of decomposed Multi-Word Expressions (MWEs) improve NMT performance compared to standard character-level modeling?
• RQ3 Why does decomposition level 2 yield worse performance than levels 1 and 3, despite being an intermediate representation?
• RQ4 To what extent do radical and stroke-level embeddings preserve the meaning of original character sequences in translation?
• RQ5 How do automatic metrics like BLEU compare with human evaluation in assessing translation quality across decomposition levels?

Key findings

• Decomposition at level 1 (radicals) and level 3 (strokes) achieves BLEU scores comparable to the full character baseline, indicating effective semantic preservation.
• Decomposition at level 2 produces significantly lower human evaluation scores than levels 1 and 3, due to the introduction of semantically misleading intermediate characters such as 從 (cóng, from) and 王 (wáng, king).
• The human evaluation shows no statistically significant difference between MWE-integrated and non-integrated models, despite BLEU improvements in the former.
• Error analysis reveals that level-2 decomposition generates characters with independent meanings not aligned with the original character’s semantics, disrupting model learning.
• The study confirms that stroke-level decomposition (level 3) is viable and meaningful for NMT, though current IDS-based extraction limits full stroke-level precision.
• The authors release an open-source toolkit and a bilingual MWE-annotated corpus to support future research in character decomposition and MWE-aware NMT.