Semantic Templates and Transitivity Alternations in the Lexicon

A systematic and principled account of verb subcategorization is important for largescale lexicon construction. A given verb may have several subcategorization frames in which its arguments appear. Starting from a lexical-semantic description of event structure, we describe a mechanism for generating subcategorization properties for a large variety of verb classes. In this paper, we motivate this mechanism by proposing a distinction between two kinds of resultative as well as the unaccusative and middle constructions. 1. Background We represent verb meanings in terms of event templates (Pustejovsky, 1995), in particular those inspired by (Rappaport Hovav & Levin, 1998), also (1991; 1995; 1999). An event template consists of a constant, the verb-specific core meaning that supplies spellout, and an event structure, common to all verbs of a class, containing a small number of primitive elements that combine in a finite number of ways. The basic semantic templates (STs) are as follows: Activity [ x ACT<manner> ] Achievement [ BECOME [ x <state> ]] State [ x <state> ] Accomplishment [ x CAUSE [ BECOME [ y <state> ]]] Canonical realization rules pair constants (verbs) with STs: manner → [ ACT<manner> ] jog, run, creak, whistle internally caused state → [ x <state> ] bloom, blossom, decay, flower, rot, rust, sprout externally caused state → [[ x ACT ] CAUSE [ BECOME [ y <state> ]]] break, dry, harden, melt STs may be simple or complex, i.e. augmented by the addition of subevent templates. The following examples are taken from (Rappaport Hovav & Levin, 1998): Phil swept (the floor) [ x ACT<sweep>y ] Phil swept the floor clean [[ x ACT<sweep>y ] CAUSE [ BECOME [ y <state> ]]] Furthermore, template augmentation is constrained by rules prohibiting the elimination of elements and non-compositionality. For example: ∗Princeton University and NEC Research Institute *Achievement → Activity (<state> can’t be eliminated) *Activity → Achievement (*[ BECOME [ x ACT<manner> ]] non-compositional) In the following sections, we propose a new semantic template theory for dealing with resultatives expressing change of state and transformation into new entities. In contrast to (Rappaport Hovav & Levin, 1998), who adopt a single template theory that grows monotonically as sub-events are added, we introduce the notion of a primary template expressing the core meaning of the verb in conjunction with a secondary template expressing secondary predication. A secondary template will be constrained in two ways: (1) via a general requirement that its elements must be checked-off, or licensed, against corresponding elements in the primary template, and (2) semantic admissibility conditions (to be described below). We also show how these admissibility conditions operate on primary templates in the case of both middle and achievement formation. 2. Two Types of Resultatives We draw a semantic distinction between Adjective Phrase (AP) and Prepositional Phrase (PP) resultatives. Both refer to a state brought about by the event denoted by the verb. The following examples illustrate how different verb types select for resultatives: (1) (a) I danced/ran *tired/*into an athlete (neither AP nor PP) (b) I broke/bent the bracelet into pieces/a U/*ruined (PP only) (c) I wiped/swept/rubbed the floor clean/*into a shiny surface (AP only) (d) I cleaned the floor *sparkling/*into a shiny surface (neither AP nor PP) (e) I wrote/conceptualized the book *interesting/*into a bestseller (same as above) (f) John pounded/crushed/hammered the metal flat/into a thin sheet (AP and PP) 3. Resultatives and Semantic Templates Our analysis rests crucially on the distinction between change of state and transformation into a new entity for AP and PP resultatives, respectively. A. An AP resultative, expressing the result of a change of state, comes with its own distinct template: (2) y be <state> B. Elements in a secondary template must be identified or checked off against elements in the main template. For example, the y in “y be <state>” must be identified with a corresponding y in the main template: (3) (a) John wiped the table clean (b) John wiped the table & the table clean (c) x act<manner> on y & y be <state> (d) John act <wip(ing)> on the table & the table be clean If y is not present in the main template, the derivation fails: (4) (a) ∗ John jogged dizzy (b) x act<manner> & y be <dizzy> (y unchecked) C. A PP resultative expresses a transformation; it comes with its own distinct template: (5) y become y′ y′ represents the new entity that y has transformed into as a result of an event: (6) (a) He spun wool into gold (b) He spun wool & wool into gold (c) x caus y xform<manner> (y ′) & y becomes y′ In (6c), y′ in the secondary template checks off against implicit y′ (shown in parentheses) in the primary template. 4. Conditions on Resultatives Both the AP and PP resultatives select for a primary ST with the following conditions: D. (i) y must be present (cf. (B) above) This condition merely states the requirement for a direct object, on which the resultant state or transformation is predicated. (ii) ∗ . . .<state>. . . Statehood, including (non-)existence, expressed as a constant cannot appear in the primary ST. (iii) “y will be y” (no transformation of y) This condition holds only for AP resultatives, where y does not get transformed into a new entity as a result of the event. 5. Consequences Let us re-examine the data from section 2 given these conditions. First, unergative verbs, such as dance, jog, laugh, run and sleep, are intransitives whose subject is an underlying external argument (semantically, an AGENT or ACTOR), (Perlmutter, 1978). These verbs cannot select for a resultative: (7) (a) John jogs/dances/sleeps (b) ∗ John jogs/dances/sleeps tired (resultative, not depictive reading) (8) (a) x act<manner> (b) x act<manner> & y be <state> (violates Di) (7a) and (7b) have STs (8a) and (8b), respectively. (8b) violates condition (Di) since y in the secondary template remains unchecked against a y in the primary template. Unergatives can have an AP resultative if a fake reflexive is introduced (Simpson, 1983). Contrast (9a) with (9b): (9) (a) ∗ I danced/ran tired (b) I danced/ran myself tired/*into an athlete The fake reflexive introduces a y into the primary template for these otherwise intransitive verbs, and thus secondary y can be correctly checked off. Next, many activity verbs, e.g. mop, paint, shoot, sweep or wipe, take either an optional or an obligatory direct object. As (10) shows, only AP resultatives are possible: (10) (a) John wiped/mopped/swept the floor clean (b) ∗ John wiped/mopped/swept the floor into a shiny surface These verbs have has ST (11a): (11) (a) x act<manner> on y (b) x act<manner> on y & y be <state> (c) x act<manner> on y & y become y ′ (violates B) Here (10b) is ruled out since y′ (the new entity) remains unchecked in (11c). Third, verbs like break, and tear express a resulting change of state in the direct object. These verbs are compatible with PP resultatives only: (12) (a) John broke/tore the painting (b) John broke/tore the painting into pieces (c) ∗ John broke/tore the painting ruined/worthless (13) (a) x caus y xform<manner> (y ′) (b) x caus y xform<manner> (y ′) & y become y′ (c) x caus y xform<manner> (y ′) & y be <state> (violates Diii) In the transform template, y′ is the (implicit) new entity that y transforms into. y′ need not be expressed in syntax. (12c) is ruled out because an AP resultative indicates the state of y after the event, but verbs like break and tear imply a transformation of y into a new entity y’, i.e. y is changed as a result of the event. Rule (Diii) encodes this intuition. Fourth, verbs that are derived from adjectives, either by zero-affix morphology e.g. clean, or by an affix, resist resultatives of any kind: It will be expressed when it is checked off by a secondary template as in (12b). (14) (a) I cleaned the floor (b) ∗ I cleaned the floor sparkling/into a shiny surface (15) (a) x caus y become <state> (b) x caus y become <state> & y be <state> (violates Dii: <state> in primary ST) (c) x caus y become <state> & y become y′ (violates Di and Dii) Finally, verbs expressing the creation or destruction of an entity such as write, paint and eat are incompatible with a resultative: (16) (a) I wrote the book (b) ∗ I wrote the book interesting/into a bestseller (17) (a) x caus y become<manner> exist(+) (b) x caus y become<manner> exist(+) & y be <state> (violates Dii) (c) x caus y become<manner> exist(+) & y become y ′ (violates Dii) The end state for the entity expressed by the direct object of a creation/destruction verb will be exist(±). End states preclude AP and PP resultatives (Dii). In summary, we saw that unergatives and activities take only AP resultatives, verbs of transformation or change take only a PP resultative, and verbs encoding a final state and verbs of creation/destruction take neither resultative. 6. Achievements According to (Rappaport Hovav & Levin, 1998), achievements are events referring to changes of state, as shown in 6 (adopting our terminology): (18) y become <state> We propose the following conditions for deriving achievement templates from a given primary template: F. (i) y must be present (ii) x[-sentient] (Fii) states that the causing entity must be non-sentient, i.e. inanimate. Here are some examples of non-sentient causers: wind, rain, storm, cold anger, greed acid, drugs, mold (natural force) (emotion, psychological motive) (substances) We adopt the following notation: Contrast (16b) with: I rewrote the book *interesting/into a bestseller. Rewrite is not a creation verb. reaffixation turns some creation verbs into verbs of transformation. G. (i) x underspecified, i.e. x[sentient] ∨ x[-sentient] (ii) x[sentient] (agent) (iii) x[-sentient] (causer) (Fi) requires an object, namely y. It excludes all unergatives like run, which have templates with x only. Verbs of transformation li