StanfordSenses

Sense Distinctions, Valence Alternations, Nominalizations, and Everything

Even if ‘standardization’ of meaning representations may be an evasive goal, for the time being, we might move forward by being (more) explicit about design goals for ‘deep’ (grammatical) analysis, levels of ambition, and criteria used in making these decisions. Assuming a ‘conceptual’ (abstract, semantics-oriented) interface to grammatical analysis, i.e. some kind of meaning representation, many of the decisions involved are about the (semantic) ‘lexicon’, e.g. the inventory of predicates or concepts and the associated inventories of argument positions or labels.

Following are some examples of design decisions embodied in the ERG; we have long felt ‘proud’ of some of these, while we keep debating whether and how to revise others at regular intervals. As an overall guiding principle, we assume that the interface to grammatical analysis should (a) include all information that is constrained by grammar (i.e. distinctions which can be argued to be grammaticized); (b) should be normalized, i.e. abstract abstract away from details of surface syntax, such that close paraphrases lead to comparable or identical structures; and (c) minimize ambiguity, i.e. not make ‘unnecessary’ distinctions, where these do not correspond to grammatical contrasts.

As a corollary of principle (b), valence alternations like the dative shift or passivization are normalized (in the ERG), for example. Principle (c) suggests that (semantic) word sense distinctions should only be made were they align with morpho-syntactic contrasts. For example, WordNet distinguishes ten senses for the noun bank, but all uniformly have the distribution of count nouns (although one might ask whether there is a morpho-syntactic sub-division into relational vs. non-relations senses). On the other hand, the count–mass distinction in pairs like much paper (substance) vs. many papers (publication) might seem indicative of a grammaticized sense distinction.

The current ERG makes comparatively few sense distinctions, often even affording itself predicate polymorphism in argument positions that can alternate between scopal and non-scopal values. For example

• (1a) I watched her arrival.
• (1b) I watched her arrive.
• (2a) I know the story.
• (2b) I know that she left.
• (3a) I believe the explanation.
• (3b) I believe you.
• (3c) I believe that you didn't cheat.
• (3d) I believe in deep analysis.