📄 Book Reviews: Corpus Linguistics and the Automatic Analysis of English

Summary

This is Ted Briscoe's review of Nelleke Oostdijk's 1991 book on the Nijmegen TOSCA project, where a formal generative grammar was built to analyze a balanced one million-word corpus of written English. Briscoe focuses on how the grammar was developed and how the parser performs. The grammar uses an extended affix grammar formalism. Its analyses are constituency-based, and the category labels are enriched with grammatical-function information. It covers hard constructions such as gapping, nonconstituent coordination, apposition and partitives. Each input is first given unambiguous, fine-grained lexical categories, and an analyst then interactively selects the correct parse.

Briscoe praises the coverage and the empirical, data-driven method. He also criticizes several things:

• The book does not compare TOSCA with other broad-coverage grammars or formalisms.
• It assumes that rule-based and statistical methods cannot be combined.
• Linguistically real ambiguity was removed from the grammar to keep parsing tractable.
• Some odd analyses chosen to fit corpus examples are then propagated through the whole annotated corpus.

The review matters because it sets out the trade-offs of corpus-driven grammar engineering.

Key points

• The TOSCA project at Nijmegen, led by Jan Aarts, had done corpus analysis for nearly two decades. The book documents a grammar for a one million-word written English corpus.
• The formalism is extended affix grammar (Koster 1971). Briscoe judges it roughly comparable in power to PATR-II or ANLT, with useful extensions for agreement in coordinate structures.
• A separate step assigns unambiguous lexical categories that encode subcategorization. These are finer-grained than CLAWS2 tags. Parsing then runs over that category sequence.
• Parse selection is manual and interactive, but the analyst can only pick among analyses the grammar licenses. Even so, the parser cannot always compute every analysis.
• One criticized design choice: noun phrase complements of ditransitive verbs are grouped into a single constituent so that gapped sentences can be handled.

Interesting ideas

• I1. Limit human annotation to choosing among analyses the grammar already licenses, rather than letting annotators invent new structures, so the database stays consistent and correction stays fast.
• I2. If every sentence must get some analysis, a bad local decision can become fixed and spread across the whole annotated corpus.
• I3. Removing genuine ambiguity, such as fixing coordination attachment at specific nodes, for the sake of efficiency hides real differences in semantic scope.
• I4. Run a separate pre-parsing stage that fixes rich lexical categories, including subcategorization, so the parser works over an unambiguous and informative sequence.
• I5. Rule-based grammars and statistical methods work well together, for example by using probabilities to rank the parses a grammar licenses.

For Mova

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

• I1 (learn): LLM teacher tiers and human reviewers could be asked only to pick among candidate trees proposed by the witness parsers, which keeps silver data consistent and makes the choices auditable.
• I2 (learn): self-induced repair rules and treebank-wide conversions should be checked against negative controls, so that one questionable rule is not silently propagated into all training data.
• I3 (tree): when coordination attachment is uncertain, the parser should keep alternative attachments or flag them, so downstream stages can resolve scope instead of inheriting an arbitrary choice.
• I4 (tokens): the tagger could output subcategorization or valency hints, for example from VerbNet classes, alongside POS tags, which would narrow the parser's search and explain later attachments.
• I5 (tree): explicit, explainable rule constraints could filter candidates while the perceptron or gparse scores rank what remains, combining transparency with robustness.