Empirical Methods for Compound Splitting

Summary

The paper addresses German compound splitting for machine translation. Freely formed compounds inflate the vocabulary and leave many words unseen in training. The authors first list every way a word can be divided into words observed in a 20-million-word German Europarl corpus, allowing the filler letters "s" and "es" at joints. They then compare several ways of picking one candidate:

eager maximal splitting;
a score based on the geometric mean of part frequencies;
choosing the split whose parts find translations in aligned English sentences, using Giza-derived lexicons;
that same parallel-corpus method with a part-of-speech filter that keeps only content-word parts.

Evaluation uses German–English noun and prepositional phrases. The parallel+POS method agrees best with a hand-made gold standard (99.1% accuracy). However, the best MT results come from simpler methods: frequency-based splitting wins with IBM Model 4, and eager splitting wins with a phrase-based model, gaining up to 0.039 BLEU. The central lesson is that the right splitting depends on the downstream application, not only on linguistic correctness.

Key points

Candidate generation:
- Parts must be known words of at least three letters.
- "s" and "es" are allowed as fillers between parts.
- Candidates are found by exhaustive recursive search, with known words hashed by their first three letters.
Parallel-corpus method:
- A translation is accepted as evidence only if its probability is at least 0.01.
- Each English word may support only one part.
- A second lexicon, trained on frequency-split German, captures compound-only senses (e.g., Grund→basic).
- The resulting splitting knowledge covers 75,055 word types.
POS filter: The TnT tagger is used to restrict parts to words that are mostly nouns, adjectives, adverbs, or verbs, plus PTKNEG.
Gold-standard results: The test set has 1000 NP/PPs (3498 words); training uses 650,000 NP/PPs.

Method Accuracy Precision

raw (no splitting) 94.2% —

eager 87.1% 24.8%

frequency 95.7% 57.4%

parallel 98.6% 83.3%

parallel + POS 99.1% 93.8%
BLEU results:

Method IBM Model 4 Marcu–Wong phrase model

raw 0.291 0.305

frequency 0.317 0.342

eager 0.222 0.344

Method	Accuracy	Precision
raw (no splitting)	94.2%	—
eager	87.1%	24.8%
frequency	95.7%	57.4%
parallel	98.6%	83.3%
parallel + POS	99.1%	93.8%

Method	IBM Model 4	Marcu–Wong phrase model
raw	0.291	0.305
frequency	0.317	0.342
eager	0.222	0.344

Interesting ideas

I1. Enumerate, then rank. Separating candidate generation (all covers by attested words) from selection lets different scorers be swapped and compared cheaply. This structure suits Ukrainian or other morphologically rich segmentation tasks.
I2. Use a translation as indirect supervision. Whether a segmentation's pieces show up in the aligned English sentence tells you if the split is meaningful. This yields training signal without manual annotation and could guide segmentation for other languages that have parallel data.
I3. Bootstrap a second lexicon. Re-aligning after a crude first split reveals meanings that a morpheme has only inside compounds. This general bootstrapping trick can improve any alignment-driven lexicon.
I4. Evaluate against the real use, not just a gold standard. The paper shows intrinsic accuracy can rank methods differently from end-task performance. The authors attribute the word-based MT result to the system recovering more easily from over-splitting than under-splitting, and phrase models can regroup aggressive splits. Measuring with the actual downstream system is a good safeguard for any preprocessing choice.
I5. Filter function words. Restricting segment parts to content-word POS classes cheaply removes spurious splits off prefixes, suffixes, and determiners that frequent function words would otherwise "justify."

Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.

Fields: Natural language processing

Topics: Morphology

Cited by

184 papers of our library cite it (of the 140884 whose bibliographies we read); 6 of them have a page here, 178 not yet.

Cite

The paper itself, not this page. BibTeX from the library's record of the work.

@inproceedings{koehn2003empirical,
  title = {{Empirical Methods for Compound Splitting}},
  author = {Koehn, Philipp and Knight, Kevin},
  booktitle = {10th Conference of the European Chapter of the Association for Computational Linguistics},
  year = {2003},
  doi = {10.3115/1067807.1067833},
  url = {https://doi.org/10.3115/1067807.1067833}
}

References

14 references, 4 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

Al-Onaizan, Y., Curin, J., Jahr, M., Knight, K., Lafferty, J., Melamed, D., Och, F.-J., Purdy, D., Smith, N. A., and Yarowsky, D. (1999). Statistical machine translation. Technical report, John Hopkins University Summer Workshop http://www.clsp.jhu.edu/ws99/projects/mt/. ↗
Brants, T. (2000). TnT - a statistical part-of-speech tagger. In Proceedings of the Sixth Applied Natural Language Processing Conference ANLP. find our card
Brown, P., Cocke, J., Pietra, S. A. D., Pietra, V. J. D., Jelinek, F., Lafferty, J. D., Mercer, R. L., and Rossin, P. (1990). A statistical approach to machine translation. Computational Linguistics,16(2):76-85. find
Brown, R. D. (2002). Corpus-driven splitting of compound words. In Proceedings of the Ninth International Conference on Theoretical and Methodological Issues in Machine Translation (TMI-2002). find
Finkler, W. and Neumann, G. (1998). Morphix. A fast realization of a classification-based approach to morphology. In 4. asterreichische Artificial-Intelligence-Tagung. Wiener Workshop - Wissensbasierte Sprachverarbeitung. find
Germann, U., Jahr, M., Knight, K., Marcu, D., and Yamada, K. (2001). Fast decoding and optimal decoding for machine translation. In Proceedings of ACL 39. find our card
Hedlund, T., Keskustalo, H., Pirkola, A., Airio, E., and Jarvelin, K. (2001). Utaclir @ CLEF 2001 - effects of compound splitting and n-gram techniques. In Second Workshop of the Cross-Language Evaluation Forum (CLEF), Revised Papers. find
Koehn, P. (2002). Europarl: A multilingual corpus for evaluation of machine translation. Unpublished, http://www.isi.edu/ -- koehn/ publications/europarl/. ↗
Koehn, P. and Knight, K. (2001). Knowledge sources for word-level translation models. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP). find
Langer, S. (1998). Zur Morphologie und Semantik von Nominalkomposita. In Tagungsband der 4. Konferenz zur Verarbeitung natiirlicher Sprache (KONVENS). find
Larson, M., Willett, D., Kohler, J., and Rigoll, G. (2000). Compound splitting and lexical unit recombination for improved performance of a speech recognition system for German parliamentary speeches. In 6th International Conference on Spoken Language Processing (ICSLP). find
Marcu, D. and Wong, W. (2002). A phrase-based, joint probability model for statistical machine translation. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP). find our card
Monz, C. and de Rijke, M. (2001). Shallow morphological analysis in monolingual information retrieval for Dutch, German, and Italian. In Second Workshop of the Cross-Language Evaluation Forum (CLEF), Revised Papers. find
Papineni, K., Roukos, S., Ward, T., and Zhu, W.-J. (2002). BLEU: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics (ACL). , , find our card

↑ back to top

This page in Markdown → https://md.research.elamur.ai/free/en/papers/koehn-2003-empirical-methods-compound-splitting/

📄 Empirical Methods for Compound Splitting