📄 Topicalization and Truth Conditions: A Categorial Account

Summary

The paper argues that in Japanese, choosing the topic particle WA or the nominative GA for a subject is not only a matter of pragmatics. The choice can change whether a sentence is true. Nakamura separates four ways to mark a subject: plain thematic WA, stressed contrastive WA, neutral GA and stressed exhaustive-listing GA. He tests them on a Japanese version of Rooth's orthography example about 'U' following 'Q', where some markings give a true sentence and others a false one.

The meanings are drawn with Lewis Carroll's four-cell logic diagrams. Each cell is marked as occupied, empty or left unspecified, which makes the system effectively three-valued. Focal stress is treated as the thing that makes a set of alternatives relevant.

Each particle then receives a categorial-grammar lexical entry that works like a generalized quantifier. Truth-conditional differences therefore come straight from the lexicon, and strong compositionality is kept. The main result is a lexicalist syntax-semantics account in which:

• thematic WA produces a tripartite, universal-like structure;
• focused GA produces exhaustivity;
• contrastive WA produces an underinformative statement that can take narrow scope.

Key points

• Data:
  • The Japanese counterpart of Rooth's example "a U always follows a Q" is true with U-ga but false with thematic U-wa.
  • With contrastive WA, the same sentence can become true again.
• Against a common claim: Kuno's view that WA marks only given referents is challenged. Examples from the old text Nihon-Reiiki and its modern translation show WA introducing a new referent where GA is impossible.
• Lexical entries proposed:
  • Thematic WA is (S_Top/(S\X))\X, with X being NP, PP or S. It carries an implication-style TOP quantifier.
  • Focused GA[+F] is S/(S\NP), with an exhaustivity meaning.
  • Contrastive WA is S|(S|NP). Its non-directional slashes reflect that contrastive topics can occupy any position.
• WA as a concord particle: WA is treated as a kakari-joshi (concord/coherence particle). It gathers the rest of the sentence into a comment that serves as the nuclear scope of the quantifier.
• Stated limit: the account does not cover the scalar implicature that contrastive topics often convey.

Interesting ideas

• I1. Carroll-style cells distinguish "known to be occupied", "known to be empty" and "not asserted", so what a sentence leaves open is kept apart from what it denies.
• I2. Alternatives become relevant only when focus marks them, rather than through free existential generalization, which keeps the set of alternatives small and linguistically licensed.
• I3. All quantificational force (topic, exhaustivity, contrast) is packed into the lexical entry of the marker itself, so composition stays mechanical and every truth-conditional effect can be traced to a single word.
• I4. Minimal pairs in which one marker change flips the truth value make a sharp test of whether a system computes meaning or just matches words.
• I5. The topic phrase acts as a restrictor and the rest of the sentence as the nuclear scope, giving generic statements a tripartite logical form.

For Mova

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

• I1 (world): Store asserted facts as explicitly three-valued (true / false / unstated), so the reader never treats silence as negation when answering.
• I2 (frames): Trigger alternative sets only from overt focus cues such as "only", "even" or cleft constructions, and record them in the frame instead of reasoning over every possible contrast.
• I3 (world): Give focus particles like "only" and "always" explicit lexical semantics (for example, swapping antecedent and consequent for "only"), so each derived fact can be explained by pointing to the word responsible.
• I4 (learn): Build negative-control test items that differ only in focus or quantifier wording ("only U follows Q" vs "U always follows Q") and check that answers flip accordingly.
• I5 (frames): Map English topic-like or generic subjects ("Dogs bark", "As for X, ...") to restrictor/scope structures so that generic claims are not stored as facts about specific individuals.