A coordination is not a description, and one word is not a rebind - #596
Merged
Conversation
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Signed-off-by: WaylandYang <wayland0916@gmail.com>
WaylandYang
force-pushed
the
fix/span-grammar-edges
branch
from
September 10, 2026 23:47
d69a519 to
05d7461
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Closes #595.
Two grammatical edges of the span check, each from one row of the recall-bench ledgers after #583.
A coordination is one item of a list
"SB Energy and SoftBank will build at least 10 GW of new energy generation" came out as subject
SB Energywithsubject_span: "SB Energy and SoftBank". Tier 4 reads a bound name followed by lowercase words as a modifier inside a larger phrase;andis lowercase, so the span was a description and a truepartner_withfact was dropped.andand&now join the articles as grammar words: after the bound name, a conjunction is skipped before the "do the following words look like names" test.SB Energy and SoftBankbound to SB Energy is kept;OpenAI and its investorsbound to OpenAI is still a description; the span bound to the later item is the prefixed tier, as before.One word is not a rebind
"Stockholders approved the election of each of our ten (10) director nominees" came out as subject
NVIDIAwithsubject_span: "Stockholders", and the response also declared2026 Annual Meeting of Stockholders of NVIDIA Corporation. Tier 3 rebinds a span whose words all sit inside another declared name and allowed a single word when it looked like a proper name. A sentence-initial capital is not that evidence, and the fact moved from NVIDIA to the meeting.A single-word span no longer rebinds by containment; it rebinds only by exact or stem match, as
slot_matchesalready does (Claude,GPT-3). Multi-word spans still rebind by containment (OpenAI's board→OpenAI's board of directors). No single-word containment rebind has been right in any ledger so far (#583's six runs, the two bench rounds);Altmanbound to OpenAI becomes a counted coreference instead of a rebind, and the test says so.Tests
a_name_in_a_coordination_is_one_of_the_list: the SB Energy sentence withandand with&; bound to the later item; a conjunction followed by a description.a_span_that_names_a_description_is_not_the_entity:AltmanandStockholdersas single words are coreferences, not rebinds.🤖 Generated with Claude Code