📄 Noun Classification From Predicate-Argument Structures

Summary

Hindle tackles the problem of grouping English nouns into semantically related sets without hand-built "is-a" hierarchies, which are costly to create and hard to derive from dictionaries. He tests Harris's distributional hypothesis at a much larger scale than earlier work. A 6-million-word sample of 1987 Associated Press news was parsed with the deterministic parser Fidditch, and verb–subject and verb–object pairs were extracted, using deep (underlying) arguments where the parser could recover them. Each noun–verb link is weighted by an estimate of mutual information, following Church and Hanks. Two nouns count as similar when they share verb contexts with positive weight (or share negative ones), summed over all verbs and both argument roles. The nearest neighbours that result often look semantically sensible. For example, most nouns close to boat denote vehicles, and treaty gathers words about agreements and rules. The work shows that purely syntactic co-occurrence, even from an error-prone parser, can produce a useful quasi-semantic noun classification. Possible uses include indexing, interpreting noun compounds and resolving modifier scope.

Key points

• Corpus and parser: about 6 million words of 1987 AP news stories were analysed sentence by sentence with Fidditch, a Marcus-style deterministic parser. The analysis yielded 4789 verbs in 274613 clausal structures.
• Each clause was stored as a 5-tuple: verb, surface subject, surface object, underlying subject and underlying object. Head nouns were reduced to root forms, and phrases of the form "a n1 of n2" were recoded as n1 n2.
• Association between a noun and a verb is scored as log2 of the observed co-occurrence frequency relative to chance. Chance is computed from the noun frequency, the verb frequency and the total clause count. Subject and object roles are scored separately.
• Results for individual nouns:
  • Of the ten nouns nearest to boat, nine are vehicles, and ship ranks first.
  • table gets a mixed neighbour list, with floor at the top.
  • section gets a diverse list, which the author attributes to it being a semantically "empty" noun.
• Reciprocal nearest neighbours (RNN) were tested on 319 nouns with frequency between 100 and 200. Only 36 had one, for example bomb–device, quake–earthquake and economist–analyst. Even boat–ship fails, because ship's top neighbour is plane.

Interesting ideas

• I1. Down-weighting promiscuous arguments. Raw co-occurrence counts are replaced with a mutual-information score, so frequent but uninformative arguments like it rank low. This is a simple, reusable way to make collocation or context lists meaningful.
• I2. Asymmetry-aware similarity. Two nouns are compared by taking the smaller of their shared positive weights per context, with separate handling for shared negative weights. This guarantees that a noun is most similar to itself, and the score grows with the number of shared contexts, which cosine similarity does not do. The approach suits any sparse-feature similarity where breadth of overlap matters.
• I3. A cautious parser in a noisy pipeline. The parser leaves attachments or arguments empty when unsure, so its errors mostly remove relations rather than invent wrong ones. In a statistical pipeline, this trades sample size for cleaner data, a useful design principle for building resources from imperfect parsers.
• I4. Reciprocal nearest neighbours as a quality filter. Keeping only noun pairs that are each other's top match yields a strict, high-precision set of near-substitutes. It is a cheap way to extract synonym candidates or sanity-check an embedding space.
• I5. Diagnosing failures by noun type. The author traces poor neighbour lists to polysemy (table) and to semantically empty nouns (section). He suggests ranking nouns by how empty their distribution is, which could serve as an automatic signal for flagging unreliable entries in any distributional lexicon.