--- title: "SPMT — paper by Marcu et al. (2006) — summary and ideas" description: "The authors propose SPMT, a family of translation models whose rules pair continuous… 40.84 BLEU. 3 references, cited by 96 papers here." authors: ["Daniel Marcu", "Wei Wang", "Abdessamad Echihabi", "Kevin Knight"] year: 2006 venue: "Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing" source: "https://aclanthology.org/W06-1606/" licence: "CC BY-NC-SA 3.0 (ACL Anthology, before 2016)" licence_url: "https://creativecommons.org/licenses/by-nc-sa/3.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/en/papers/marcu-2006-spmt-statistical-machine-translation-syntactified/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/marcu-2006-spmt-statistical-machine-translation-syntactified/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2006](https://md.research.elamur.ai/free/en/papers/2006/) / SPMT: Statistical Machine Translation with Syntactified Target Language Phrases # SPMT: Statistical Machine Translation with Syntactified Target Language Phrases **Authors:** [Daniel Marcu](https://md.research.elamur.ai/free/en/authors/daniel-marcu/), [Wei Wang](https://md.research.elamur.ai/free/en/authors/wei-wang/), [Abdessamad Echihabi](https://md.research.elamur.ai/free/en/authors/abdessamad-echihabi/) and [Kevin Knight](https://md.research.elamur.ai/free/en/authors/kevin-knight/) · **Year:** 2006 · **Venue:** [Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing](https://md.research.elamur.ai/free/en/venues/emnlp/) ([EMNLP 2006](https://md.research.elamur.ai/free/en/venues/emnlp/2006/)) · **Publisher:** [Association for Computational Linguistics](https://md.research.elamur.ai/free/en/publishers/association-for-computational-linguistics/) · **Licence:** [CC BY-NC-SA 3.0 (ACL Anthology, before 2016)](https://creativecommons.org/licenses/by-nc-sa/3.0/) **Links:** [Publisher's page](https://aclanthology.org/W06-1606/) · [PDF](https://aclanthology.org/W06-1606.pdf) · [OpenAlex](https://openalex.org/W7096877219) **Citations:** 0 citations (OpenAlex, 2026-10-14) · [96](https://md.research.elamur.ai/free/en/papers/marcu-2006-spmt-statistical-machine-translation-syntactified/cited-by/) in our library, of the 141,840 papers whose bibliographies we read ## Summary Phrase-based statistical MT handles lexical choice well but often produces ungrammatical word order for syntactically divergent pairs such as Chinese–English. Its distortion and n-gram language models are too weak to rearrange phrases correctly. The authors propose SPMT, a family of translation models whose rules pair continuous source phrases with target-side English syntax fragments. These are expressed as xRS tree-to-string transducer rules. A decoder then builds English parse trees directly from Chinese input. Rules are learned from a large Chinese–English parallel corpus (100M+ words). The English side is parsed with a Collins-style parser and the corpus is word-aligned with Giza++. Variants include a minimal-rule model (Model 1), a version with composed rules, a Model 2, and a combined system. On the NIST 2003 Chinese–English test set, SPMT beats a strong phrase-based baseline by 2.64 Bleu. Human judges also prefer its output, by 0.28 points on a 1–5 scale. The authors present this as the first evidence that a syntax-based system can outperform a mature phrase-based one on a large, open-domain benchmark. ## Key points - Rule types: - Minimally syntactified, lexicalized, "phrase-based-compatible" xRS rules, with continuous phrases on both sides. - Minimal non-lexicalized rules, obtained from the Galley et al. (2006) extraction algorithm with its lexicalized rules removed. - Optionally, one composed rule per source phrase. - About 28% of source phrases in the English–Chinese corpus could not be syntactified under the standard xRS constraint. This is because valid rules cannot be multi-headed, which motivated extending the rule set rather than discarding those phrases. - Human evaluation setup: 138 dev sentences under 20 words, three judges, and seven "systems" (six MT systems plus one hidden NIST reference). - Average scores were 3.10 for PBMT and 3.38 for SPMT-Combined. - The difference was significant (P = 4.04e−6, t = 4.67, df = 413). - On that dev subset, Bleu was 31.00 for PBMT, 33.42–34.05 for the single SPMT models, and 35.33 for SPMT-Combined. The human reference scored 40.84 Bleu and was judged only 4.67/5, mostly because of fluency errors. - In expressive power, SPMT sits between Chiang (2005), which uses single-level target trees, and Galley et al. (2006), which allows discontinuous source phrases. SPMT also yields far fewer rules than Galley et al. and has parameters that are easier to estimate. ## Interesting ideas - **Decorating phrase pairs with target syntax.** Existing phrase-table entries gain constituent labels, so the decoder knows what kind of unit each phrase is and where it can attach. This lets mature phrase-based features be reused while syntax handles assembly, a cheap way to add structure to any pipeline that already has a good lexicon. - **Defining "minimal" per source phrase.** A rule counts as minimal for a given source span, rather than being merely necessary to explain the whole sentence pair. This keeps phrase-level coverage comparable to phrase-based systems and is a useful design choice when bridging string-based and tree-based models. - **Limiting composition depth.** Only one composed rule is added per phrase, built from the first multi-child ancestor, on the intuition that longer compositions are unreliable. This kind of bounded composition controls rule-table size and noise in other grammar-induction settings. - **Measuring structural compatibility before choosing a fix.** The authors counted how many phrases fail the syntactic constraint (28%) before deciding how to handle them. This quantifies how lossy a structural restriction would be, and is a good habit whenever a formalism might silently drop training signal. - **Hiding a human reference among system outputs.** One human translation was shown to judges as if it were another system. This calibrates the human-score scale and exposes reference imperfections, which is a simple trick worth reusing in any human evaluation of generation systems. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* 📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 9,085 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 8,888 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 9,296 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 9,206 works 🧭 Topics (1): [🔁 Machine translation](https://md.research.elamur.ai/free/en/topics/machine-translation/) · 958 works ## Cite Cite as: Daniel Marcu, Wei Wang, Abdessamad Echihabi and Kevin Knight. 2006. SPMT: Statistical Machine Translation with Syntactified Target Language Phrases. Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing. The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{marcu2006spmt, title = {{SPMT: Statistical Machine Translation with Syntactified Target Language Phrases}}, author = {Daniel Marcu, Wei Wang, Abdessamad Echihabi and Kevin Knight}, booktitle = {Proceedings of the 2006 Conference on Empirical Methods in Natural Language Processing}, year = {2006}, url = {https://aclanthology.org/W06-1606/} } ``` - [← previous An Unsupervised Morpheme-Based HMM for Hebrew Morphological Disambiguation](https://md.research.elamur.ai/free/en/papers/adler-2006-unsupervised-morpheme-based-hmm-hebrew/) - [next → Partially Supervised Coreference Resolution for Opinion Summarization through Structured Rule Learning](https://md.research.elamur.ai/free/en/papers/stoyanov-2006-partially-supervised-coreference-resolution-opinion/) - [↑ back to papers of 2006](https://md.research.elamur.ai/free/en/papers/2006/) ## References 3 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. R. Bonnema. 2002. Probability models for DOP. In Data-Oriented Parsing. CSLI publications. [find](https://scholar.google.com/scholar?q=R%2E+Bonnema%2E+2002%2E+Probability+models+for+DOP%2E+In+Data%2DOriented+Parsing%2E+CSLI+publications%2E) 2. Peter F. Brown, Stephen A. Della Pietra, Vincent J. Della Pietra, and Robert L. Mercer. 1993. The mathematics of statistical machine translation: Parameter estimation. Computational Linguistics, 19(2):263–311. [find](https://scholar.google.com/scholar?q=Peter+F%2E+Brown%2C+Stephen+A%2E+Della+Pietra%2C+Vincent+J%2E+Della+Pietra%2C+and+Robert+L%2E+Mercer%2E+1993%2E+The+mathematics+of+statistical+machine+translation%3A+Parameter+estimation%2E+Computational+Linguistics%2C+19%282%29) 3. David Chiang. 2005. [A hierarchical phrase-based model for statistical machine translation](https://md.research.elamur.ai/free/en/papers/chiang-2005-hierarchical-phrase-based-model-statistical/). In Proceedings of the 43rd Annual Meeting of the Association for Computational Linguistics (ACL’05), pages 263–270, Ann Arbor, Michigan, June. *Updated 24 October 2026*