Semantic Roles for SMT: A Hybrid Two-Pass Model

Dekai Wu, Pascale Fung

Introduction

Statistical Machine Translation (SMT) has made significant strides in lexical fluency, yet it frequently struggles with semantic adequacy. A common failure mode involves "who did what to whom" errors, where the system correctly translates the words but misarranges them, leading to incorrect semantic role assignments. While semantic role labeling (SRL) offers a potential solution, its integration into standard SMT decoders is computationally prohibitive due to the non-compositional nature of semantic structures.

The Hybrid Two-Pass Model

To address these challenges without the computational overhead of deep integration, this research proposes a hybrid two-pass approach that separates lexical selection from structural re-ordering.

Pass 1: Lexical Choice

The first pass utilizes a standard phrase-based SMT model (specifically Moses) to generate an initial translation. The primary goal of this stage is to establish the lexical choices. The authors observe that modern SMT is generally proficient at selecting the correct target-language words, even if their structural arrangement is flawed.

Pass 2: Semantic Re-ordering

The second pass serves as a semantic error-correction layer. It performs the following steps:

  1. Parsing: Shallow semantic parsers are applied to both the source sentence and the baseline translation.
  2. Comparison: If the semantic frames (predicates and their associated roles) match, the baseline is accepted.
  3. Re-ordering: If a mismatch is detected, the baseline translation is chunked into coarse syntactic units (NP, VP, PP). The system then iteratively re-orders these chunks and re-parses the resulting candidates.
  4. Selection: The final output is the re-ordered hypothesis that best aligns with the semantic structure of the source sentence.

Experimental Results

The model was evaluated on a Chinese-English translation task using LDC newswire data and the Chinese-English Bilingual Propbank.

Quantitative Performance

The baseline phrase-based SMT system achieved a BLEU score of 42.99. The hybrid two-pass model improved this to 43.51. While a 0.5 BLEU increase may seem modest, the authors argue that BLEU, being a lexical metric, tends to undervalue improvements in semantic structure and adequacy.

Qualitative Analysis

The model's strength is most apparent in its ability to correct severe role confusion. In one instance, the baseline SMT incorrectly associated the "National Development Bank of Japan" with the "capital market" rather than the predicate "issued." The semantic re-ordering pass successfully restructured the sentence to correctly identify the bank as the agent (ARG0) of the issuance, bringing the output much closer to the reference translation.

Figure 2: Comparison of semantic structures across Input, SMT baseline, Reordered output, and Reference translation.

Conclusion

This work demonstrates the first successful application of semantic role labeling to enhance SMT performance. By decoupling lexical choice from semantic ordering, the hybrid model bypasses computational bottlenecks while significantly improving translation adequacy. The authors note that semantic roles are highly robust anchors for cross-lingual alignment, with approximately 84% of roles preserved between Chinese and English. Future research will focus on manual evaluation metrics to better capture the utility of these semantic improvements.

Summary updated

BibTeX
@inproceedings{wu2009semantic,
  author = {Dekai Wu and Pascale Fung},
  title = {{Semantic Roles for SMT: A Hybrid Two-Pass Model}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2009},
  url = {https://openreview.net/forum?id=HyEJMQ-_bH},
}
Figure