📄 A Corpus-based Syntactic Analysis of Two-termed Unlike Coordination

Summary

The paper asks which syntactic categories English allows in two-termed coordinations whose conjuncts differ in category, as in a noun phrase joined to an adjective phrase. Such cases contradict the Law of the Coordination of Likes, and earlier theories rested mostly on a few linguists' own judgments. The authors take an empirical route. They parsed raw COCA text with the Berkeley Neural Parser and found coordinations using three tree patterns typical of that parser. They also extracted coordinations from the Penn Treebank, using Ficler and Goldberg's coordination annotation extension. Chi-square tests then compared category frequencies across conjunct position, conjunction type and genre. The two conjunct slots behave differently: shorter constituents tend to come first and heavier ones second. Noun phrases also coordinate often with subordinate clauses. The authors read these asymmetries as support for antisymmetric accounts of coordination rather than flat, symmetric structures. The NP–clause pattern also suggests that both may be determiner phrases. The work matters because it puts corpus statistics behind a long-disputed theoretical question.

Key points

• Data: COCA has more than 450 million words in five genres (academic, fiction, magazine, newspaper, spoken), about 20 million words per genre per year from 1990 to 2012. The PTB adds 2,499 Wall Street Journal stories.
• COCA was parsed with the Berkeley Neural Parser (Kitaev and Klein, 2018). Coordinations were found with three patterns: a ternary X–CC–Z split, a neither…nor frame, and a verb-complement frame in which the first conjunct sits inside the VP.
• PTB coordination labels are noisy, so the study relied on Ficler and Goldberg's (2016) extension. It explicitly marks conjuncts, markers, connectives and shared material.
• NP+SBAR ranks among the most common unlike pairs, including in the PTB top ten. ADVP+PP is also prominent there.
• Admitted weakness: the human checkers of the parser's output saw the parser's labels. The authors propose a small gold set scored with precision, recall and F1 as future work.

Interesting ideas

• I1. Testing whether the two conjunct positions differ statistically is a cheap way to turn a debate about symmetric versus asymmetric structure into a measurable, falsifiable question.
• I2. Mining a parser's typical output shapes for a phenomenon is a practical way to extract constructions from automatically parsed raw text without gold annotation.
• I3. Pairing a large automatically parsed corpus with a smaller hand-annotated treebank lets each source offset the other's weakness, scale versus reliability.
• I4. The authors explain position preferences by weight, meaning heavier conjuncts drift rightward, linking coordination to heavy-NP shift.
• I5. The finding that human raters were not blind to parser labels is a reminder that evaluation must hide system output from judges to stay unbiased.

For Mova

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

• I1 (tree): Position-wise category statistics over conjuncts from our treebanks could show whether UD's first-conjunct-as-head convention matches real asymmetries, and could guide attachment choices in gparse.
• I2 (learn): We could write a few explicit tree patterns for coordination and run them over silver parses to collect unlike-coordination examples for targeted training and evaluation.
• I3 (learn): Statistics derived from silver data should be cross-checked against gold UD treebanks before they become repair rules, as a built-in control against parser artifacts.
• I4 (back): When generating coordinated phrases from the meaning graph, ordering the shorter conjunct first should make output sound more natural.
• I5 (learn): Any human or LLM-teacher check of parses should present unlabeled spans or shuffled candidates, so that agreement is not inflated by anchoring.