A Discriminative Latent Variable-Based “DE” Classifier for Chinese-English SMT

Jinhua Du, Andy WayCOLING 2010

TL;DR

This paper tackles the challenge of accurately translating Chinese sentences with the ubiquitous "的" (DE) construction, a major source of word order errors in machine translation. They introduce a novel discriminative latent variable model (DPLVM) and a "tree-pattern" feature to better classify and reorder these constructions, leading to significant improvements in Chinese-English machine translation quality.

What Problem Does This Paper Address?

Chinese and English have fundamentally different word orders, particularly concerning noun phrases and their modifiers. The Chinese "的" (DE) particle plays a crucial role in these structures, often marking possessives, attributives, or relative clauses. However, its flexible usage and varied translation equivalents make it a persistent challenge for machine translation (MT) systems. Existing approaches often struggle to capture the nuances of DE constructions, leading to incorrect reordering and, consequently, poor translation quality. This paper aims to improve MT by specifically addressing the accurate classification and reordering of DE constructions.

Methods

The core of this paper is the introduction of a Discriminative Latent Variable Model (DPLVM) for classifying DE constructions. Unlike traditional sequence labeling models, DPLVM incorporates latent variables that can capture complex dependencies between observations and labels, potentially leading to more robust classifications.

A key innovation is the proposed "tree-pattern" feature. Instead of relying on manually crafted rules for reordering, this feature automatically learns reordering patterns by analyzing the syntactic parse tree structure surrounding the DE construction. This allows the model to generalize and discover reordering rules directly from the data.

The authors also employ other features, including DE's part-of-speech tag (DEPOS), POS-grams, lexical features, and semantic class information. They adapt the DPLVM algorithm for this specific task, treating DE classification as a sequence labeling problem where only the DE word itself needs to be labeled within a noun phrase structure.

Datasets

The experiments utilize the Chinese Treebank 6.0 (CTB6) and the English-Chinese Translation Treebank 1.0. These datasets are used for both training and evaluating the DE classifier. For the machine translation experiments, a large parallel corpus of 2,159,232 sentence pairs is used. The NIST MT2006 test set serves as the development set, and the NIST MT2008 "current" test set is used for final evaluation.

Key Insights and Ideas

  • DE is a critical bottleneck: The paper strongly emphasizes that the DE construction is a major source of word order errors in Chinese-English MT, and addressing it directly can yield substantial improvements.
  • Latent variables unlock deeper understanding: The use of DPLVM with latent variables proves effective in capturing subtle dependencies that simpler models might miss, leading to better classification accuracy.
  • Automatic rule learning is powerful: The "tree-pattern" feature demonstrates the advantage of learning reordering rules automatically from syntactic structures, bypassing the limitations of manual rule creation.
  • Improved classification directly benefits MT: The paper provides strong evidence that higher accuracy in DE classification translates directly into better overall translation quality.
  • DE reordering aids word alignment and phrase tables: Beyond direct translation improvements, the DE reordering process positively impacts downstream MT components like word alignment and phrase table generation.

Key Results

The proposed DPLVM classifier, especially with the "tree-pattern" feature, significantly outperforms a log-linear classifier in DE classification accuracy. For the 5-class classification, the DPLVM achieves 76.8% accuracy compared to 75.1% for the log-linear model. For the 2-class classification, DPLVM reaches 88.3% accuracy versus 86.7% for the log-linear model.

In terms of machine translation quality, the DPLVM-based reordering leads to notable improvements:

  • PB-SMT: 6.42% relative improvement in BLEU score.
  • Hierarchical Phrase-Based SMT (HPB): 3.08% relative improvement in BLEU score.

Qualitative analysis shows that DE reordering also leads to slightly better word alignment (e.g., 1.46% improvement in F1 score) and potentially better phrase tables.

Limitations and Future Work

The authors acknowledge that the "topicality" feature, used in prior work, was excluded due to its perceived limited utility in randomly stored sentence corpora. They also note that the DE classification is currently focused on noun phrases.

Future work includes:

  • Extending the DE classification and reordering approach to other syntactic structures like Verb Phrases (VP) and Left-Complex Phrases (LCP).
  • Applying the DE-annotated approach to syntax-based MT systems.
  • Improving the DE classifier's accuracy by incorporating richer features.

Summary updated

BibTeX
@inproceedings{du2010discriminative,
  author = {Jinhua Du and Andy Way},
  title = {{A Discriminative Latent Variable-Based “DE” Classifier for Chinese-English SMT}},
  booktitle = {International Conference on Computational Linguistics},
  year = {2010},
}
Figure