📄 A comparison of the XTAG and CLE Grammars for English

Summary

Rayner, Hockey and James argue that teams building large grammars gradually stop noticing phenomena they have labelled "marginal". The best way to spot these gaps is to examine one grammar through the lens of another. They compare two broad-coverage English grammars: the lexicalized Tree Adjoining Grammar XTAG, and the unification-based CLE grammar, which also produces Quasi Logical Form semantics. The comparison shows a symmetry.

• XTAG handles complement structure better. Its initial trees can state constraints on leaves deep inside a complement. This makes idiom-like verb + possessive + noun constructions easy to encode.
• CLE handles adjunct movement better. Its gap-threading mechanism covers cases such as preposition stranding in adjuncts, which XTAG cannot capture.

For each weakness the authors sketch a practical fix. CLE needs richer NP head features, added systematically through macros. XTAG needs finer-grained displacement features, with a separate set for each kind of movement. They conclude that both gaps persisted for engineering and sociological reasons, not deep ones. Their main message is that grammar developers should run such detailed cross-comparisons routinely.

Key points

• Shared coverage. Both grammars cover a wide range of NP and clause phenomena: relatives, ellipsis, clefts, passives, topicalization, WH- and yes/no questions, and conjunction. CLE additionally pairs every construction with compositional, scope-free QLF semantics.
• CLE's complement gap. Constructions like shake one's head, shrug one's shoulders, take one's time and have one's way need a possessive determiner that agrees with the subject. XTAG enforces this by linking the determiner's <ref> features to the agreement of the subject and the verb. CLE cannot express it, because its <subcat> constraints see only the NP's maximal projection, and it has no head feature for "possessive with agreement".
• Why it matters for translation. Other languages render these concepts differently. French uses fermer les yeux (transitive verb + definite NP) and Swedish uses blunda (intransitive verb). This is an argument for treating the constructions as idioms or light verbs in transfer-based machine translation.
• XTAG's adjunct-movement gap. XTAG models subject–auxiliary inversion with a feature-based version of multi-component adjunction, using a <displ_const> clash that is resolved only when both trees are adjoined. The feature records only that something moved, not its category or the type of movement. Naively extending it would let Can they go to?-style strings through.
• Proposed XTAG fix.
  • Add a <displ_cat> feature that records the category of the moved constituent.
  • Thread separate feature sets for four movement types: verb movement, WH-movement/topicalization, tough movement, and right extraposition (passivization may be a fifth).
  • This is offered instead of CLE's list-valued gaps. Those lists exist mainly to enforce the no-crossing-dependencies constraint ("sonata sentences"), which the authors find weak or doubtful for double extraction from adjuncts.

Interesting ideas

• I1. Mutual comparison as a diagnostic. Auditing one system from the viewpoint of another exposes blind spots that its own developers have come to treat as normal.
• I2. Percolate head features to the phrase level. A constraint can only see what has been passed up to the phrase, so information such as "the determiner is a possessive agreeing with the subject" has to be raised to the NP. Adding all such features at once via macros keeps the cost manageable.
• I3. Typed displacement markers. A generic "something moved" marker invites spurious filler–gap pairings, while labelling the moved category blocks the worst errors cheaply.
• I4. One channel per kind of dependency. Instead of one general list of pending gaps, thread a small fixed set of channels, one per movement type. This works because English has only a few movement types that can involve adjuncts.
• I5. Idioms as units for transfer. Treating verb + possessive + body-part expressions non-compositionally avoids awkward transfer rules, because target languages express them with quite different structures.

For Mova

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

• I1 (learn): Mova could systematically compare the error categories where the perceptron parser and gparse (or other witness parsers) disagree, and treat recurring "nobody covers this" phenomena as candidates for new repair rules rather than noise.
• I2 (frames): when building frames, Mova could attach a "possessive bound to subject" flag to object NPs, so that shake one's head or take one's time is recognised as a single light-verb or idiom frame instead of a literal transitive action.
• I3 (tree): Mova could tag recovered gaps (stranded prepositions, fronted WH-phrases) with the filler's category, and reject attachments where the category is incompatible.
• I4 (tree): Mova could keep separate, explainable slots for WH, tough-movement and extraposition links during UD post-processing, which makes each long-distance link easy to justify and debug.
• I5 (back): when Mova translates through the meta-world, it could map close one's eyes to a single event concept, so that generation can choose a definite NP (as in French) or a single intransitive verb (as in Swedish).