📄 Setting Syntactic Parameters with Implicit Negative Evidence: The Case of Zero-derived Causatives in English

Summary

The paper tackles a learnability puzzle in generative syntax: how a child can reach the correct parameter value when every input sentence fits both candidate grammars. The test case is English zero-derived causatives (ZDCs) such as John awoke Bill, analysed with Pylkkänen's (2008) Cause-selection Parameter. Under the Root-selecting setting, Cause takes a bare root phrase. Under the Verb-selecting setting, an extra little v layer appears, which licenses a "low" adverbial reading that modifies the caused event. English lacks that reading, but children never hear explicit evidence that it is impossible.

Gould builds a simple Bayesian learner. Its probabilistic phrase-structure choice points carry priors, and adverbial attachment to CauseP and to vP is treated as equally likely. Every ambiguous token gets a higher likelihood under the Root-selecting setting, because the richer grammar spreads its probability over attachments that never occur. Over time, the never-observed low reading therefore acts as implicit negative evidence and selects the subset grammar. This works without the Subset Principle or a default value. Gould also argues that the learning models of Gibson & Wexler, Sakas & Fodor, and Yang cannot handle this case.

Key points

• The case study is English ZDCs (e.g., awoke, √BREAK, √MELT) under Pylkkänen's (2008) Cause-selection Parameter, limited here to a binary choice: Root-selecting vs Verb-selecting Cause.
• When awoke takes a modifier describing a grumpy state, the grumpiness can only be attributed to the causer (high reading), never to the causee (low reading). The low reading requires a vP adjunction site, which exists only in the Verb-selecting structure.
• Root-selecting is a proper subset of Verb-selecting in three ways: its heads are {Cause, √} vs {Cause, v, √}, its adjunction sites are {CauseP} vs {CauseP, vP}, and its readings are {high} vs {high, low}.
• The comparison with other learners:
  • The Gibson & Wexler (1994) Triggering Learning Algorithm can get stuck in a local maximum because of its Greediness and Single Value constraints.
  • Sakas & Fodor (2001) wait for unambiguous triggers, which never occur for this parameter.
  • Yang's (2002) variational learner also fails on fully ambiguous input.
• Future directions:
  • Use a larger, more child-like corpus.
  • Explain how Verb-selecting languages are learned, likely through an overt v morpheme between CauseP and √P.
  • Note that a very high share of modified tokens could in principle push the learner toward Verb-selecting.

Interesting ideas

• I1. The learner can exploit absence as evidence: a hypothesis that predicts more structures than are ever observed loses likelihood with each ambiguous datum, so the subset grammar wins without any hard-coded preference.
• I2. Set uniform priors over attachment sites, so that "any projection is equally modifiable" is the neutral baseline and corpus statistics alone pull the decision.
• I3. Link a structural choice to a checkable interpretive consequence: the extra vP layer exists only if a low reading of the modifier exists.
• I4. Diagnose a learner by building a case where all input is ambiguous, then check whether it converges, stalls, or depends on a default.
• I5. Treat parameters per morpheme rather than per language: the same language can set the selection property differently for different causative markers.

For Mova

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

• I1 (learn): Self-induced repair rules and silver-data voting could score competing analyses by how much unobserved output each one predicts, preferring the tighter rule automatically.
• I2 (tree): The perceptron parser and gparse could get an explainable attachment-preference prior that is updated from treebank counts, rather than fixed heuristics for adverbial and PP attachment.
• I3 (frames): When assigning which event an adverbial modifies in a causative, prag could attach manner and state modifiers of zero-derived causatives to the causing event (the agent) and record the structural reason as its explanation.
• I4 (learn): Evaluation with negative controls could include deliberately underdetermined training slices, to test whether training converges to the restrictive analysis or overfits to the permissive one.
• I5 (world): Compiled verb-class knowledge could store causative or inchoative structure per lexical item, which would help track who caused which state change in the world of the text.