--- title: "Restoration of Case Information in All-Cap English Broadcast" description: "The authors address a readability problem. 4 references. Among them: 0." authors: ["Yu-Ting Liang", "Jian-Chen Wu"] year: 2003 venue: "ROCLING 2003 Poster Papers" source: "https://aclanthology.org/O03-3003/" 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/liang-2003-restoration-case-information-all-cap/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/liang-2003-restoration-case-information-all-cap/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2003](https://md.research.elamur.ai/free/en/papers/2003/) / Restoration of Case Information in All-Cap English Broadcast Transcription # Restoration of Case Information in All-Cap English Broadcast Transcription **Authors:** Yu-Ting Liang and Jian-Chen Wu · **Year:** 2003 · **Venue:** [ROCLING 2003 Poster Papers](https://md.research.elamur.ai/free/en/venues/rocling/) ([ROCLING 2003](https://md.research.elamur.ai/free/en/venues/rocling/2003/)) · **Publisher:** [The Association for Computational Linguistics and Chinese Language Processing (ACLCLP)](https://md.research.elamur.ai/free/en/publishers/the-association-for-computational-linguistics-and-chinese-language-processing/) · **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/O03-3003/) · [PDF](https://aclanthology.org/O03-3003.pdf) · [OpenAlex](https://openalex.org/W2394815739) **Citations:** 0 citations (OpenAlex, 2026-10-14) ## Summary The authors address a readability problem. ICRT (International Community Radio Taipei) emailed its news scripts to listeners entirely in capital letters, and they built a prototype that converts this text back to normal mixed case. Training material came from VOA news, about 9138 articles and roughly 3 million words. They split it into words and estimated unigram, bigram and trigram statistics over case forms. For each word in an ICRT script, the system looks up these probabilities and runs a Viterbi search to choose the most likely case sequence, scored with a P-model value following Lita (2003). Two post-processing rules handle errors. Any word not seen in training is treated as a proper noun and given an initial capital. A heuristic sentence-boundary detector, which must cope with abbreviations such as "U.S.", decides which words start a sentence and capitalizes them. The paper shows example outputs but reports no quantitative accuracy. It illustrates a simple statistical recipe for truecasing with very few resources. ## Key points - Test input came from all-caps ICRT news scripts received by e-mail. Training data was VOA news: 9138 articles, about 3 million words. - Each word gets unigram, bigram and trigram case probabilities, stored as two, four and eight values respectively. - Decoding uses the Viterbi algorithm (Rabiner, 1989) together with a P-model value (Lita, 2003) to pick the best case for each word. - There are two exception rules: - Out-of-vocabulary words such as "Sampson" are assumed to be proper nouns and capitalized. - Sentence-initial words are capitalized using heuristic boundary detection that does not treat the period in "U.S." as a sentence end. - There is no numeric evaluation, only worked examples. Planned future work includes faster processing and an Outlook macro for readers. ## Interesting ideas - Treat case as a hidden label on each token and recover it with a small n-gram model plus Viterbi, instead of deciding word by word in isolation. This lets the neighbours settle ambiguous items like "Black Hawk" versus "black". - Use a fallback rule that assumes an unseen word is a name and capitalizes it. This is a cheap default that is right far more often than wrong in news text. - The paper's own output first broke "U.S." into separate letters and lower-cased the second sentence. This shows that a casing stage depends on a separate, abbreviation-aware sentence splitter. - Truecasing produces a gold-like signal for free: take ordinary cased text, upper-case it, and measure how well the original is restored. ## For Mova How these ideas could serve Mova, our small open language model — its files, stages and tests: - **I1** (tokens): A truecasing pass before tagging would let all-caps or lower-case inputs (headlines, chat, OCR) reach the tagger with the capitalization cues it relies on for proper nouns. - **I2** (tokens): The same default could feed an explicit, logged "unknown → PROPN candidate" decision. The explanation trace would then show why a word was cased and tagged as a name. - **I3** (tokens): Keep sentence segmentation as an independent, testable module, and add abbreviation-at-boundary cases like "U.S." to its regression suite. - **I4** (learn): Restoring casing destroyed on purpose could serve as a self-supervised check for the tagger and NER. Comparing parses of the restored text with parses of the original also gives a negative control. *Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.* ## Institutions - [National Tsing Hua University](https://md.research.elamur.ai/free/en/universities/national-tsing-hua-university/) — University, Hsinchu, Taiwan · 10 works · 1,850 citations 📚 Shelves (4): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 17,450 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 17,296 works 🧭 Topics (1): [English](https://md.research.elamur.ai/free/en/topics/english/) · 2,379 works ## Cite Cite as: Yu-Ting Liang and Jian-Chen Wu. 2003. Restoration of Case Information in All-Cap English Broadcast Transcription. ROCLING 2003 Poster Papers. The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @inproceedings{liang2003restoration, title = {{Restoration of Case Information in All-Cap English Broadcast Transcription}}, author = {Yu-Ting Liang and Jian-Chen Wu}, booktitle = {ROCLING 2003 Poster Papers}, year = {2003}, url = {https://aclanthology.org/O03-3003/} } ``` - [← previous Inductive coding of the Arabic lexicon](https://md.research.elamur.ai/free/en/papers/farghaly-2003-inductive-coding-arabic-lexicon/) - [next → The Old English Verbs of Smell Perception and Emission: Analysis of the Interface of their Semantic and Syntactic Representation](https://md.research.elamur.ai/free/en/papers/orta-2019-old-english-verbs-smell-perception/) - [↑ back to papers of 2003](https://md.research.elamur.ai/free/en/papers/2003/) ## References 4 references, 0 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet. 1. Lucian Vlad Lita, 2003. tRuEcasIng. [find](https://scholar.google.com/scholar?q=Lucian+Vlad+Lita%2C+2003%2E+tRuEcasIng%2E) 2. Hai Leong Chieu, Hwee Tou Ng, 2002 Upper Case Text. Teaching a Weaker Classifier: Named Entity Recognition on Andrei Mikheev, 1999 A Knowledge-free Method for Capitalized Word Disambiguation. [find](https://scholar.google.com/scholar?q=Hai+Leong+Chieu%2C+Hwee+Tou+Ng%2C+2002+Upper+Case+Text%2E+Teaching+a+Weaker+Classifier%3A+Named+Entity+Recognition+on+Andrei+Mikheev%2C+1999+A+Knowledge%2Dfree+Method+for+Capitalized+Word+Disambiguation%2E) 3. Alison Huettner, Pero Subasic, 2000 Fuzzy Typing for Document Management. [find](https://scholar.google.com/scholar?q=Alison+Huettner%2C+Pero+Subasic%2C+2000+Fuzzy+Typing+for+Document+Management%2E) 4. Christopher D. Manning and Hinrich Schutze. 2000, pp. 123-136 There Also, create a macro in Outlook so if the Foundations of statistical natural language processing, [find](https://scholar.google.com/scholar?q=Christopher+D%2E+Manning+and+Hinrich+Schutze%2E+2000%2C+pp%2E+123%2D136+There+Also%2C+create+a+macro+in+Outlook+so+if+the+Foundations+of+statistical+natural+language+processing%2C) *Updated 2 November 2026*