📄 Analyzing Coordinate Structures Including Punctuation in English

Summary

Coordinate structures (CSs) make English parsing hard. Their scope is ambiguous, they tend to appear in long sentences, and commas can mark coordination, enclose parentheticals, or simply separate constituents. Kurohashi adapts his earlier parallelism-based conjunct finder, originally built for Japanese, into a joint analyzer for English that decides comma roles and CS scopes together.

The method works in three steps:

• A small phrase-structure grammar over key expressions (commas, and, or, but) lists every "general structure" (GS), meaning every combination of comma roles and CS nesting.
• For each GS, CSs are resolved from the innermost outward. A dynamic-programming search over a word-similarity matrix finds the most similar left and right word sequences. After each step the sentence is reduced so that outer CSs are compared on simplified text.
• GSs are ranked by summed, weighted scores.

On held-out sentences checked against the Penn Treebank, about 80% of commas were interpreted correctly, and about 80% of CS scopes were detected correctly. This matters because it can guide a full parser on long, real sentences without the explosion of complete parse enumeration.

Key points

• Commas fall into four roles: strong connector (before and/or in lists of three or more), weak connector (coordination with no conjunction), parenthesizor, and delimiter.
• Word similarity combines three signals: POS match, exact word match, and closeness in the WordNet hierarchy.
• The path score adds element points, penalizes unbalanced conjunct lengths and boundary symbols inside a conjunct, rewards markers such as both, and normalizes by a power of the span length.
• The key-expression pattern ", , and and" alone yields 135 candidate GSs. This growth is still much slower than enumerating full parses.
• n-best POS tags used alone hurt scope detection badly because they produce too many spurious similarities. Merging one-best and n-best tags matched one-best accuracy and fixed some tagger errors.

Interesting ideas

• I1. Treat punctuation and conjunctions as terminals of a tiny grammar, which lets you cheaply list all possible comma-role and nesting skeletons before doing any detailed analysis.
• I2. Resolve the innermost coordinations first, then collapse them to their first conjunct and delete parentheticals, so that outer conjuncts are compared on a simplified sentence.
• I3. Use parallelism between candidate conjuncts as the main scoring signal, mixing POS, lexical identity and thesaurus distance with a penalty for lopsided lengths.
• I4. Choose the correct coordination scope and the correct POS of nearby words at the same time, feeding the scorer a mix of the best tag and alternative tags.
• I5. Weight evidence by how explicit the signal is: a conjunction counts more than a list comma, which counts more than a bare comma joining clauses.

For Mova

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

• I1 (tree): Before gparse or the perceptron parser commits, enumerate comma/conjunction skeletons and use them as constraints, or as an explainable reason for each conj/parataxis/appos decision.
• I2 (tree): A reduce-then-compare pass could re-check conj attachments in long sentences and catch cases where nested lists confuse the parser.
• I3 (learn): A parallelism score could act as an independent witness when silver parsers disagree on coordination scope, breaking ties with a transparent reason.
• I4 (tokens): When the tagger is unsure near and/or (e.g. NN vs VB), keep the alternatives and let coordination parallelism pick one, instead of trusting the single best tag.
• I5 (frames): When splitting coordinated clauses into separate frames, rank splits licensed by weak commas lower and flag them as less certain in the explanation.