Skip to content

An entity-name guard built on a verb list does not survive a new corpus #193

Description

@WaylandYang

is_entity_name rejects a clause by looking for a finite verb in CLAUSE_MARKERS, a hand-written list of 19 words. A different corpus walked straight past it:

The removal could slow down the artificial intelligence industry as a whole.   12 words
Shares in Microsoft fell nearly three percent following the announcement.      10 words

The list has would but not could, went and came but not fell or hits. The first sentence is also exactly MAX_NAME_WORDS, so the length check misses it by one.

A closed verb list cannot generalise; English has thousands of finite forms, and the current list was drawn from whatever one corpus happened to produce. Structural signals travel better: a trailing period after a lowercase word (Safe Superintelligence Inc. keeps its capital abbreviation), a colon, a determiner opening a long string.

The guard has 5 negative and 6 positive cases, all from one corpus. Whatever replaces the rule needs a labelled set spanning at least two, or it fails the same way.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions