--- title: "Book Reviews — paper by Wu (1994) — summary and ideas" description: "The project set out to build a broad-coverage English parser whose grammar is trained statistically." authors: ["Dekai Wu"] year: 1994 venue: "Computational Linguistics" source: "https://aclanthology.org/J94-3014/" 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/wu-1994-book-reviews-statistically-driven-computer/" --- [Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/wu-1994-book-reviews-statistically-driven-computer/) [home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [1994](https://md.research.elamur.ai/free/en/papers/1994/) / Book Reviews: Statistically-Driven Computer Grammars of English: The IBM/Lancaster Approach # Book Reviews: Statistically-Driven Computer Grammars of English: The IBM/Lancaster Approach **Authors:** [Dekai Wu](https://md.research.elamur.ai/free/en/authors/dekai-wu/) · **Year:** 1994 · **Venue:** [Computational Linguistics](https://md.research.elamur.ai/free/en/venues/cl/) ([Computational Linguistics 1994](https://md.research.elamur.ai/free/en/venues/cl/1994/)) · **Publisher:** [The MIT Press](https://md.research.elamur.ai/free/en/publishers/the-mit-press/) · **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/J94-3014/) · [PDF](https://aclanthology.org/J94-3014.pdf) · [OpenAlex](https://openalex.org/W4388885063) **Citations:** 47 citations (OpenAlex, 2026-10-14) ## Summary This is Dekai Wu's review of *Statistically-Driven Computer Grammars of English* (SDCGE), an edited volume describing a five-year joint effort by IBM T.J. Watson's Continuous Speech Recognition Group and Lancaster University's UCREL. The project set out to build a broad-coverage English parser whose grammar is trained statistically. It first aimed at AP newswire text and later narrowed to IBM computer manuals. The book is motivated by the weak accuracy of hand-built parsers. It covers the whole pipeline: CLAWS tagging with human correction, flat "skeleton" treebanking, a hand-written grammar of roughly 500 rules, and inside-outside training over feature-based equivalence classes. Systems are judged only by their single most probable parse. For sentences of up to 23 words, that parse has no constituent crossings against the treebank in 64–69% of cases, and some crossing-free parse exists for about 94%. Wu praises the book's candour about its modest criteria and its practical lessons. He sees it as the nearest thing to a textbook on building probabilistic CFGs, with value well beyond English. ## Key points - The motivating evidence is a 1992 comparison of 35 parsers against the Penn Treebank. Only seven groups reported results, and their average correctness was just 22%, even under a lenient measure. - Treebanking works in two steps. Annotators first correct CLAWS output, then add flat bracketings labelled from a small inventory of nonterminals. Both steps use interfaces built for annotator speed. - The design choices are the CLAWS2a tagset, 17 nonterminals and about 500 grammar rules. The grammar is stated as rewrite rules, but the running system used a more compact feature-based form, which the book does not give in full. - John Lafferty's chapter explains inside-outside training for beginners and sketches how to extend it to feature "mnemonics". It does not connect the algorithm to general EM. - In early tests, a history-based grammar clearly improved nonterminal labelling. It conditions derivations on extra context, above all the parent's head word, and uses an induced decision tree to choose which factors matter. ## Interesting ideas - Drive grammar work by error frequency: tally error types across the corpus and repair the most common first, so effort goes where it pays off most. - Score only the single best analysis, since emitting more candidate parses can inflate recall without limit; keeping the gap between top-1 accuracy (~66%) and oracle accuracy (~94%) visible separates ranking errors from coverage errors. - Design an annotation scheme by explicitly balancing four factors: annotator speed, reachable accuracy, agreement between annotators, and linguistic soundness. - Collapse surface rule variants into linguistically motivated feature classes ("mnemonics"), so fewer parameters need training and the grammar stays compact. - Condition structural decisions on richer history, such as the parent's head, and let a learned decision tree pick the few conditioning factors that matter, keeping the model sparse and inspectable. ## For Mova How these ideas could serve Mova, our small open language model — its files, stages and tests: - **I1** (learn): Self-induced repair rules could be prioritised by how often each failure pattern appears on silver data, not by how striking individual errors look. - **I2** (tree): Report gparse and perceptron LAS strictly on the 1-best tree, and log oracle-over-beam separately to show whether the scorer or the search is failing. - **I3** (learn): Silver-data and LLM-teacher labels could be checked against these four factors, cutting distinctions on which witness parsers or teachers cannot agree consistently. - **I4** (frames): VerbNet-based role assignment could tie parameters across verbs in the same class, so rare verbs borrow statistics from frequent siblings. - **I5** (tree): An explainable decision tree over head and context features could make relation-labelling choices traceable, in line with Mova's every-step-explainable requirement. *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 (1): [English](https://md.research.elamur.ai/free/en/topics/english/) · 2,263 works ## Cite Cite as: Dekai Wu. 1994. Book Reviews: Statistically-Driven Computer Grammars of English: The IBM/Lancaster Approach. Computational Linguistics. The paper itself, not this page. BibTeX from the library's record of the work. ```bibtex @article{wu1994book, title = {{Book Reviews: Statistically-Driven Computer Grammars of English: The IBM/Lancaster Approach}}, author = {Dekai Wu}, journal = {Computational Linguistics}, year = {1994}, doi = {10.1163/9789004653535}, url = {https://doi.org/10.1163/9789004653535} } ``` - [← previous RAFT/RAPR and Centering: a comparison and discussion of problems related to processing complex sentences](https://md.research.elamur.ai/free/en/papers/suri-1994-raft-rapr-centering-comparison-discussion/) - [next → XTAG System - A Wide Coverage Grammar for English](https://md.research.elamur.ai/free/en/papers/doran-1994-xtag-system-wide-coverage-grammar/) - [↑ back to papers of 1994](https://md.research.elamur.ai/free/en/papers/1994/) *Updated 31 October 2026*