--- title: "Toward a Broad-coverage Bilingual Corpus for — Takezawa 2002" description: "Speech translation systems of that period worked only in narrow scenarios such as booking a hotel room." authors: ["Toshiyuki Takezawa", "Eiichiro Sumita", "Fumiaki Sugaya", "Hirofumi Yamamoto", "Seiichi Yamamoto"] year: 2002 venue: "Proceedings of the Third International Conference on Language Resources and Evaluation (LREC'02)" source: "https://aclanthology.org/L02-1305/" licence: "CC BY-NC 4.0 (ELRA, LREC proceedings)" licence_url: "https://creativecommons.org/licenses/by-nc/4.0/" summary_licence: "Elamur Cosmic License" canonical: "https://research.elamur.ai/free/papers/takezawa-2002-broad-coverage-bilingual-corpus-speech/" --- [Elamur Research](https://md.research.elamur.ai/) · [this page in HTML](https://research.elamur.ai/free/papers/takezawa-2002-broad-coverage-bilingual-corpus-speech/) [home](https://md.research.elamur.ai/) / [papers](https://md.research.elamur.ai/free/papers/) / [2002](https://md.research.elamur.ai/free/papers/2002/) / Toward a Broad-coverage Bilingual Corpus for Speech Translation of Travel Conversations in the Real World # Toward a Broad-coverage Bilingual Corpus for Speech Translation of Travel Conversations in the Real World **Authors:** Toshiyuki Takezawa, [Eiichiro Sumita](https://md.research.elamur.ai/free/authors/eiichiro-sumita/), Fumiaki Sugaya, Hirofumi Yamamoto, Seiichi Yamamoto · **Year:** 2002 · **Venue:** Proceedings of the Third International Conference on Language Resources and Evaluation (LREC'02) · **Licence:** [CC BY-NC 4.0 (ELRA, LREC proceedings)](https://creativecommons.org/licenses/by-nc/4.0/) **Links:** [Publisher's page](https://aclanthology.org/L02-1305/) · [PDF](https://aclanthology.org/L02-1305.pdf) · [OpenAlex](https://openalex.org/W229329885) **Citations:** 233 citations (openalex-snapshot · 2026-10-14) · 95 in our library, of the 140673 papers whose bibliographies we read ## Summary Speech translation systems of that period worked only in narrow scenarios such as booking a hotel room. ATR researchers wanted a bilingual Japanese–English resource that could support wider, real-world travel use. They contrasted two complementary corpora. The first is the spoken travel conversation corpus (TC): interpreter-mediated dialogues between a hotel clerk and a guest, totalling about 20k sentences. The second is a much larger basic expression corpus (BE): about 200k phrase-book sentence pairs tagged by situation, such as airport, restaurant, shopping or trouble. The comparison covered sentence length, mutual vocabulary coverage, cross perplexity under a multi-class composite n-gram model, and word frequency profiles. The two corpora share much of their vocabulary but differ statistically. Each model fits its own corpus well and fits the other poorly, while a mixed model handles both. When people spoke BE-type content themselves, some departed strongly from the phrase-book wording. The authors conclude that a usable corpus needs variety in three areas: speakers and speech, situations, and expressions. A pilot paraphrase collection suggests that gathering several people's phrasings of the same meaning is a practical way to supply the third. ## Key points - **BE corpus size:** 200,241 utterances, about 220k Japanese and 223k English sentences, and 1.69M Japanese words (21,329 entries). Sentences are short, averaging 7.67 Japanese and 5.51 English words. - **TC corpus size:** 16,084 utterances, about 21.8k Japanese and 22.9k English sentences. Sentences are longer, averaging 10.84 Japanese and 7.91 English words. - **Phrases per sentence:** BE averages 2.88 and TC averages 4.37. A National Language Research Institute study puts everyday Japanese conversation at 3.81, and TC's length distribution resembles it much more closely than BE's does. - **Vocabulary coverage:** - BE covers 94% of TC words on the Japanese side and 97% on the English side. - TC covers 73% of BE words on the Japanese side and 85% on the English side. - In both directions, the missing words are mostly nouns. - **Cross perplexity:** - A model tested on its own corpus scores about 16. - Across corpora, the score jumps to 58.3 and 72.3. - A model trained on both corpora stays around 17 on either test set. - **Spoken re-utterances:** Three subjects spoke 50 sentences in their own way. Their perplexities were 23.0, 19.9 and 35.5, against 18.0 for the original sentences. - **Paraphrase pilot:** Five Americans with sufficient Japanese each wrote three English renderings for 330 TC utterances. This yielded an average of 14.4 distinct paraphrases per Japanese sentence. ## Interesting ideas - **Pair a small realistic corpus with a large cheap one.** Costly transcribed dialogue in one domain can be combined with plentiful edited sentences across many situations, each compensating for the other's gaps. This strategy fits any low-resource setting where natural spoken data is expensive. - **Use cross perplexity as a distance between corpora.** Training a model on one corpus and testing it on the other quantifies how different two datasets really are. The result can show whether a cheaper data source can stand in for the target domain, or whether mixing them is needed. - **Measure vocabulary coverage in both directions.** Checking which words of A appear in B and vice versa exposes asymmetries that a single overlap figure hides. Looking at the part of speech of the missing words, mostly nouns here, points to what to collect next. - **Ask real speakers to voice scripted material in their own words.** Comparing the perplexity of their speech with that of the script reveals how far actual usage drifts from curated text. It also shows that this drift varies a lot from person to person. - **Collect several paraphrases per meaning and count repeats.** Having multiple people produce several phrasings each, while keeping duplicate counts, yields frequency-weighted alternatives. These can serve as multi-reference sets for automatic translation evaluation and as richer data for language modeling. *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: [Linguistics & languages](https://md.research.elamur.ai/free/shelves/linguistics/) ## Cite The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{takezawa2002toward, title = {{Toward a Broad-coverage Bilingual Corpus for Speech Translation of Travel Conversations in the Real World}}, author = {Takezawa, Toshiyuki and Sumita, Eiichiro and Sugaya, Fumiaki and Yamamoto, Hirofumi and Yamamoto, Seiichi}, booktitle = {Proceedings of the Third International Conference on Language Resources and Evaluation (LREC'02)}, year = {2002}, doi = {10.63317/3s3iqvy9jh6g}, url = {https://doi.org/10.63317/3s3iqvy9jh6g} } ``` - [← previous The LinGO Redwoods Treebank: Motivation and Preliminary Applications](https://md.research.elamur.ai/free/papers/bank-2002-oepen-lingo-redwoods-treebank-motivation-preliminary/) - [next → Unsupervised discovery of morphologically related words based on orthographic and semantic similarity](https://md.research.elamur.ai/free/papers/aclbulk-2002-baroni-unsupervised-discovery-morphologically-related-words/) - [↑ back to papers of 2002](https://md.research.elamur.ai/free/papers/2002/) ## References 10 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Adam Kilgarriff and Tony Berber Sardinha, editors. 2000. Proceedings of the Workshop on Comparing Corpora, Hong Kong. Held in conjunction with the 38th Annual Meeting of the Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Adam+Kilgarriff+and+Tony+Berber+Sardinha%2C+editors%2E+2000%2E+Proceedings+of+the+Workshop+on+Comparing+Corpora%2C+Hong+Kong%2E+Held+in+conjunction+with+the+38th+Annual+Meeting+of+the+Association+for+Computatio) 2. Kikuo Maekawa. 2001. Compiling the corpus of spontaneous Japanese. In Proceedings of the 1st Spontaneous Speech Science and Technology Workshop, pages 7–12. (in Japanese). [find](https://scholar.google.com/scholar?q=Kikuo+Maekawa%2E+2001%2E+Compiling+the+corpus+of+spontaneous+Japanese%2E+In+Proceedings+of+the+1st+Spontaneous+Speech+Science+and+Technology+Workshop%2C+pages+7%E2%80%9312%2E+%28in+Japanese%29%2E) 3. Tsuyoshi Morimoto, Noriyoshi Uratani, Toshiyuki Takezawa, Osamu Furuse, Yasuhiro Sobashima, Hitoshi Iida, Atsushi Nakamura, Yoshinori Sagisaka, Norio Higuchi, and Yasuhiro Yamazaki. 1994. A speech and language database for speech translation research. In Proceedings of the 3rd International Conferen… [find](https://scholar.google.com/scholar?q=Tsuyoshi+Morimoto%2C+Noriyoshi+Uratani%2C+Toshiyuki+Takezawa%2C+Osamu+Furuse%2C+Yasuhiro+Sobashima%2C+Hitoshi+Iida%2C+Atsushi+Nakamura%2C+Yoshinori+Sagisaka%2C+Norio+Higuchi%2C+and+Yasuhiro+Yamazaki%2E+1994%2E+A+speech+and) 4. Atsushi Nakamura, Shoichi Matsunaga, Tohru Shimizu, Masahiro Tonomura, and Yoshinori Sagisaka. 1996. Japanese speech databases for robust speech recognition. In Proceedings of the 4th International Conference on Spoken Language Processing, pages 2199–2202. [find](https://scholar.google.com/scholar?q=Atsushi+Nakamura%2C+Shoichi+Matsunaga%2C+Tohru+Shimizu%2C+Masahiro+Tonomura%2C+and+Yoshinori+Sagisaka%2E+1996%2E+Japanese+speech+databases+for+robust+speech+recognition%2E+In+Proceedings+of+the+4th+International+Co) 5. Fumiaki Sugaya, Keiji Yasuda, Toshiyuki Takezawa, and Seiichi Yamamoto. 2001. Precise measurement method of a speech translation system’s capability with a paired comparison method between the system and humans. In Proceedings of the Machine Translation Summit VIII, pages 345–350. [find](https://scholar.google.com/scholar?q=Fumiaki+Sugaya%2C+Keiji+Yasuda%2C+Toshiyuki+Takezawa%2C+and+Seiichi+Yamamoto%2E+2001%2E+Precise+measurement+method+of+a+speech+translation+system%E2%80%99s+capability+with+a+paired+comparison+method+between+the+system+) 6. Eiichiro Sumita. 2001. Example-based machine translation using DP-matching between word sequences. In Proceedings of the ACL-2001 Workshop on Data-Driven Methods in Machine Translation, pages 1–8. [find](https://scholar.google.com/scholar?q=Eiichiro+Sumita%2E+2001%2E+Example%2Dbased+machine+translation+using+DP%2Dmatching+between+word+sequences%2E+In+Proceedings+of+the+ACL%2D2001+Workshop+on+Data%2DDriven+Methods+in+Machine+Translation%2C+pages+1%E2%80%938%2E) 7. Toshiyuki Takezawa, Tsuyoshi Morimoto, and Yoshinori Sagisaka. 1998a. Speech and language databases for speech translation research in ATR. In Proceedings of the 1st International Workshop on East-Asian Language Resources and Evaluation — Oriental COCOSDA Workshop ’98 —, pages 148–155. [find](https://scholar.google.com/scholar?q=Toshiyuki+Takezawa%2C+Tsuyoshi+Morimoto%2C+and+Yoshinori+Sagisaka%2E+1998a%2E+Speech+and+language+databases+for+speech+translation+research+in+ATR%2E+In+Proceedings+of+the+1st+International+Workshop+on+East%2DAsi) 8. Toshiyuki Takezawa, Tsuyoshi Morimoto, Yoshinori Sagisaka, Nick Campbell, Hitoshi Iida, Fumiaki Sugaya, Akio Yokoo, and Seiichi Yamamoto. 1998b. A Japanese-to-English speech translation system: ATR-MATRIX. In Proceedings of the 5th International Conference on Spoken Language Processing, pages 2779–… [find](https://scholar.google.com/scholar?q=Toshiyuki+Takezawa%2C+Tsuyoshi+Morimoto%2C+Yoshinori+Sagisaka%2C+Nick+Campbell%2C+Hitoshi+Iida%2C+Fumiaki+Sugaya%2C+Akio+Yokoo%2C+and+Seiichi+Yamamoto%2E+1998b%2E+A+Japanese%2Dto%2DEnglish+speech+translation+system%3A+ATR%2DMA) 9. Toshiyuki Takezawa. 1999. Building a bilingual travel conversation database for speech translation research. In Proceedings of the 2nd International Workshop on East-Asian Language Resources and Evaluation — Oriental COCOSDA Workshop ’99 —, pages 17–20. [find](https://scholar.google.com/scholar?q=Toshiyuki+Takezawa%2E+1999%2E+Building+a+bilingual+travel+conversation+database+for+speech+translation+research%2E+In+Proceedings+of+the+2nd+International+Workshop+on+East%2DAsian+Language+Resources+and+Evalu) 10. Hirofumi Yamamoto, Shuntaro Isogai, and Yoshinori Sagisaka. 2001. Multi-class composite  -gram language model for spoken language processing using multiple word clusters. In Proceedings of the 39th Annual Meeting of the Association for Computational Linguistics, pages 531–538. [find](https://scholar.google.com/scholar?q=Hirofumi+Yamamoto%2C+Shuntaro+Isogai%2C+and+Yoshinori+Sagisaka%2E+2001%2E+Multi%2Dclass+composite+%01+%2Dgram+language+model+for+spoken+language+processing+using+multiple+word+clusters%2E+In+Proceedings+of+the+39th+A)