Jun'ichi Tsujii

Ideas Mova took (1)

From the ideas ledger: an idea tried, kept or queued, or measured with no gain, and what it gave.

• tried — Attach, do not align. In the standard encoder (profile.rs), write a target function word (case marker, article, auxiliary) with the chunk of the content word it binds to, not through the word alignment. Measure the function-word class of the residual in… (Effective Use of Function Words for Rule Generalization in… · translate r4-fw: function words bound to the head concept were much worse (186.4→278.2); the per-pair insert/drop table (r5-fwt) is kept instead)

Papers

1. Effective Use of Function Words for Rule Generalization in Forest-Based Translation (2011)
   English-to-Japanese forest-to-string translation on about one million abstract pairs. In a hand check of 100 GIZA++ alignments, 309 of 1,324 links were wrong, and 237 of those were target function words (particles and auxiliaries), which tend to align to a few English words such as "of" and "the". The fix: drop the…
2. Exploring Difficulties in Parsing Imperatives and Questions (2011)
   Parsers trained on the Wall Street Journal fail on imperatives and questions even when the words are known, because the word order and the missing subject change the structure.
3. Descriptive and Empirical Approaches to Capturing Underlying Dependencies among Parsing Errors (2009)
   Parser errors are usually counted one by one, but one wrong decision often causes several errors. The authors look for these underlying dependencies in two ways. The descriptive approach defines combinations of error patterns, such as argument inconsistencies around a verb, and extracts them from the errors. The…
4. Effective Analysis of Causes and Inter-dependencies of Parsing Errors (2009)
   F-scores hide where a parser fails and how its errors depend on each other. The authors analyse errors of the Enju HPSG parser on predicate-argument triplets in two ways. First, 14 hand-written pattern rules map sets of mismatched triplets to cause categories such as PP-attachment, coordination, relative-clause…
5. Feature Forest Models for Probabilistic HPSG Parsing (2008)
   Probabilistic models for lexicalized grammars are hard because their structures are typed feature structures, not trees or lattices.
6. Task-oriented Evaluation of Syntactic Parsers and Their Representations (2008)
   Parsers are usually compared on their own accuracy, and parse representations differ from framework to framework, so the scores are hard to compare.
7. Probabilistic Disambiguation Models for Wide-Coverage HPSG Parsing (2005)
   The paper builds log-linear disambiguation models for a wide-coverage HPSG grammar, trained on 20 sections of the Penn Treebank.
8. Machine Translation without a source text (1990)
   The authors propose an MT system that works without a prepared source text. A monolingual writer begins with a rough sketch of what they want to say, and the system acts as an «intelligent secretary» with knowledge of the foreign language. It asks the writer questions in a keyboard dialogue to collect what the target…

Co-authors (8)

• Yusuke Miyao ×6
• Tadayoshi Hara ×3
• Takuya Matsuzaki ×3
• Danny Jones
• Harold L. Somers
• Kenji Sagae
• Rune Sætre
• Xianchao Wu

This page in Markdown → https://md.mova.elamur.ai/free/en/docs/papers/authors/jun-ichi-tsujii/