---
title: "MT at the paragraph level — paper by Hovy and Gerber (1997)"
description: "The authors define a deliberately minimal \"paragraph tree\": a root for the paragraph, optional intermediate grouping nodes… 12 references."
authors: ["Eduard Hovy", "Laurie Gerber"]
year: 1997
venue: "Proceedings of the 7th Conference on Theoretical and Methodological Issues in Machine Translation of Natural Languages"
source: "https://aclanthology.org/1997.tmi-1.6/"
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/hovy-1997-mt-paragraph-level-improving-english/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/hovy-1997-mt-paragraph-level-improving-english/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [1997](https://md.research.elamur.ai/free/en/papers/1997/) / MT at the paragraph level: improving English synthesis in SYSTRAN

# MT at the paragraph level: improving English synthesis in SYSTRAN

**Authors:** [Eduard Hovy](https://md.research.elamur.ai/free/en/authors/eduard-hovy/) and [Laurie Gerber](https://md.research.elamur.ai/free/en/authors/laurie-gerber/) · **Year:** 1997 · **Venue:** [Proceedings of the 7th Conference on Theoretical and Methodological Issues in Machine Translation of Natural Languages](https://md.research.elamur.ai/free/en/venues/tmi/) ([Conference on Theoretical and Methodological Issues in Machine Translation of Natural Languages 1997](https://md.research.elamur.ai/free/en/venues/tmi/1997/)) · **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/1997.tmi-1.6/) · [PDF](https://aclanthology.org/1997.tmi-1.6.pdf)

## Summary
Most machine translation systems handle one sentence at a time. Many output errors, however, come from phenomena that cross sentence boundaries. Hovy (USC/ISI) and Gerber (SYSTRAN) describe a research-to-industry project that adds a light discourse layer to the English synthesis stage of SYSTRAN's Japanese-to-English system. In Japanese the reporting verb comes after the quoted material. When a quote runs over several sentences, a sentence-by-sentence system cannot tell where the quote began, so it cannot move the verb to its normal English position before the quote. The authors define a deliberately minimal "paragraph tree": a root for the paragraph, optional intermediate grouping nodes such as QUOTE, and one leaf per sentence. Simple build rules driven by paragraph breaks and quote marks construct the tree, and transfer rules then relocate the reporting verb. On 57 test paragraphs, about ten rules handled roughly two thirds of the quotes correctly. Human fluency ratings improved only slightly, because the outputs had many other problems. The work shows that cheap paragraph-level structure can be retrofitted into a mature sentence-based system.

## Key points
- Four cross-sentence problems in SYSTRAN J-E output are named:
  - quotation scoping,
  - pronoun choice when Japanese drops subjects,
  - comma placement, which depends partly on prosody,
  - relative pronoun selection.

  Only quote scoping was tackled in this paper.
- The paragraph tree is built from nine basic tree operations and an "active node" mechanism:
  - an open quote creates a QUOTE node and makes it active,
  - a close quote returns control to the parent node.
- Results on 57 paragraphs with about 10 rules:
  - about 66% of quotes were treated correctly,
  - about 5% were treated wrongly,
  - the rest were missed.
- Human evaluation setup:
  - 8 native English speakers rated quoted passages from 36 texts on a 1–5 fluency scale, where 1 is perfect.
  - Each passage was rated by 4 people.
  - No judge saw the same passage in both forms.
- Average score per text moved from 3.41 (raw) to 3.25 (treated). This difference of 0.16 is smaller than the reported confidence margin of about 0.256.
- Related work they draw on includes Rhetorical Structure Theory (about 25 relations) and a taxonomy of about 120 relations by Hovy & Maier, built from over 300 collected relations. The paragraph tree is a much simpler stand-in for these structures.

## Interesting ideas
- Keep the discourse structure as small as possible and add a node type only when a concrete output-affecting rule needs it; this keeps the structure cheap and easy to justify.
- Build span structure from simple boundary events using an "active node" that is pushed on an opening mark and popped on a closing mark; this generalizes to any bracketed region that spans sentences.
- Define "correct" separately for each case type: a noun-phrase quote should stay unchanged, while a sentence quote should be inverted. Scoring every case as "did it change" would mislead.
- The human study was counterbalanced: each judge saw half original and half treated passages, never the same passage twice, and the authors reported a confidence margin that exposed how small the gain was.
- The paragraph module runs after the main pipeline and only writes instructions into the system's existing intermediate data fields, so it can be added without redesigning the generator.

## For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

- **I1** (world): Let the text-world gain grouping nodes (quoted speech, embedded narration) only when a downstream reader or answer rule actually consumes them, so every node stays explainable.
- **I2** (tokens): A pre-parse pass could mark multi-sentence quote spans and their speaker slot, so later stages know which sentences belong to whose speech.
- **I3** (learn): Evaluation suites for repair rules should carry per-category expected behaviour, including "no change" as a negative control, so that over-eager rules are penalized.
- **I4** (learn): Apply the same blind, split design with explicit confidence margins when checking whether a new generation or repair rule really improves output.
- **I5** (back): Discourse-level fixes (speaker placement around quotes, pronoun choice across sentences) could be added as an annotation layer on the meaning graph that the text generator reads, rather than as changes to the core generator.

*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/) · 16,667 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 16,194 works · [🎭 Arts and Humanities](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/) · 16,039 works · [🗣️ Language and Linguistics](https://md.research.elamur.ai/free/en/shelves/arts-and-humanities/language-and-linguistics/) · 15,891 works

🧭 Topics (2): [🔁 Machine translation](https://md.research.elamur.ai/free/en/topics/machine-translation/) · 1,087 works · [English](https://md.research.elamur.ai/free/en/topics/english/) · 2,263 works

## Cite

Cite as: Eduard Hovy and Laurie Gerber. 1997. MT at the paragraph level: improving English synthesis in SYSTRAN. Proceedings of the 7th Conference on Theoretical and Methodological Issues in Machine Translation of Natural Languages.

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

```bibtex
@inproceedings{hovy1997mt,
  title = {{MT at the paragraph level: improving English synthesis in SYSTRAN}},
  author = {Eduard Hovy and Laurie Gerber},
  booktitle = {Proceedings of the 7th Conference on Theoretical and Methodological Issues in Machine Translation of Natural Languages},
  year = {1997},
  url = {https://aclanthology.org/1997.tmi-1.6/}
}
```

- [← previous ALEP-Based Distributed Grammar Engineering](https://md.research.elamur.ai/free/en/papers/theofilidis-1997-alep-based-distributed-grammar-engineering/)
- [next → O'Keeffe O'Brien Katherine ed. 1997: Reading Old English Texts](https://md.research.elamur.ai/free/en/papers/garcia-2019-o-keeffe-o-brien-katherine/)
- [↑ back to papers of 1997](https://md.research.elamur.ai/free/en/papers/1997/)

## References

12 references, 1 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Dalianis, H. 1996. Aggregation in Natural Language Generation. Ph.D. dissertation, Stockholm University. [find](https://scholar.google.com/scholar?q=Dalianis%2C+H%2E+1996%2E+Aggregation+in+Natural+Language+Generation%2E+Ph%2ED%2E+dissertation%2C+Stockholm+University%2E)
2. Hovy, E.H. 1988. Planning Coherent Multisentential Text. Proceedings of the 26th Annual Conference of the Association for Computational Linguistics, Buffalo, NY (163-169). [find](https://scholar.google.com/scholar?q=Hovy%2C+E%2EH%2E+1988%2E+Planning+Coherent+Multisentential+Text%2E+Proceedings+of+the+26th+Annual+Conference+of+the+Association+for+Computational+Linguistics%2C+Buffalo%2C+NY+%28163%2D169%29%2E)
3. Hovy, E.H. and E. Maier. 1993. Parsimonious or Profligate: How Many and Which Discourse Structure Relations? Unpublished ms. [↗](https://doi.org/10.21236/ada278715)
4. Hovy, E.H. and L. Wanner. 1996. Managing Sentence Planning Requirements. In Proceedings of Workshop on New Directions in Planning and Natural Language Generation at the 12th European Conference on Artificial Intelligence, Budapest, Hungary (12-18). [find](https://scholar.google.com/scholar?q=Hovy%2C+E%2EH%2E+and+L%2E+Wanner%2E+1996%2E+Managing+Sentence+Planning+Requirements%2E+In+Proceedings+of+Workshop+on+New+Directions+in+Planning+and+Natural+Language+Generation+at+the+12th+European+Conference+on+Art)
5. Knight, K., I. Chander, M. Haines, V. Hatzivassiloglou, E.H. Hovy, M. Iida, S. Luk, A. Okumura, R. Whitney, K. Yamada. 1994. Integrating Knowledge Bases and Statistics in MT. In Proceedings of the 1st Conference of the Association for Machine Translation in the Americas, Columbia, MD (134-141). [↗](https://doi.org/10.48550/arxiv.cmp-lg/9409001)
6. Mann, W.C. and S.A. Thompson. 1988. [Rhetorical Structure Theory: Toward a Functional Theory of Text Organization](https://md.research.elamur.ai/free/en/papers/mann-1988-rhetorical-structure-theory-functional/). Text 8(3) (243-281). Also available as USC/Information Sciences Institute Research Report RR-87-190. [↗](https://doi.org/10.1515/text.1.1988.8.3.243)
7. Moore, J.D. 1989. A Reactive Approach to Explanation in Expert and Advice-Giving Systems. Ph.D. dissertation, University of California in Los Angeles. [find](https://scholar.google.com/scholar?q=Moore%2C+J%2ED%2E+1989%2E+A+Reactive+Approach+to+Explanation+in+Expert+and+Advice%2DGiving+Systems%2E+Ph%2ED%2E+dissertation%2C+University+of+California+in+Los+Angeles%2E)
8. Nirenburg, S., V. Lesser, and E. Nyberg. 1989. Controlling a Language Generation Planner. In Proceedings of the 11th International Joint Conference on Artificial Intelligence, Detroit, MI (1524-1530). [find](https://scholar.google.com/scholar?q=Nirenburg%2C+S%2E%2C+V%2E+Lesser%2C+and+E%2E+Nyberg%2E+1989%2E+Controlling+a+Language+Generation+Planner%2E+In+Proceedings+of+the+11th+International+Joint+Conference+on+Artificial+Intelligence%2C+Detroit%2C+MI+%281524%2D1530%29%2E)
9. Panaget, F. 1994. Using a Textual Representation Level Component in the Context of Discourse or Dialogue Generation. In Proceedings of the 7th International Workshop on Natural Language Generation, Kennebunkport, ME (35-42). [find](https://scholar.google.com/scholar?q=Panaget%2C+F%2E+1994%2E+Using+a+Textual+Representation+Level+Component+in+the+Context+of+Discourse+or+Dialogue+Generation%2E+In+Proceedings+of+the+7th+International+Workshop+on+Natural+Language+Generation%2C+Ke)
10. Penman Natural Language Group. 1989. The Penman Primer, User Guide, and Reference Manual. USC/Information Sciences Institute, Marina del Rey, CA. [find](https://scholar.google.com/scholar?q=Penman+Natural+Language+Group%2E+1989%2E+The+Penman+Primer%2C+User+Guide%2C+and+Reference+Manual%2E+USC%2FInformation+Sciences+Institute%2C+Marina+del+Rey%2C+CA%2E)
11. Rambow, O. and T. Korelsky. 1992. Applied Text Generation. In Proceedings of the 3rd Applied Natural Language Processing Conference, Trento, Italy (40-47). [find](https://scholar.google.com/scholar?q=Rambow%2C+O%2E+and+T%2E+Korelsky%2E+1992%2E+Applied+Text+Generation%2E+In+Proceedings+of+the+3rd+Applied+Natural+Language+Processing+Conference%2C+Trento%2C+Italy+%2840%2D47%29%2E)
12. Yamada., K. 1996. A Controlled Skip Parser. In Proceedings of the 2nd Conference of the Association for Machine Translation in the Americas, Montreal, Quebec (14-23). [find](https://scholar.google.com/scholar?q=Yamada%2E%2C+K%2E+1996%2E+A+Controlled+Skip+Parser%2E+In+Proceedings+of+the+2nd+Conference+of+the+Association+for+Machine+Translation+in+the+Americas%2C+Montreal%2C+Quebec+%2814%2D23%29%2E)

*Updated 1 November 2026*