Coercion: container, contents and measure readings

Coercion: container, contents and measure readings Peter Sutton & Hana Filip Heinrich Heine University, Düsseldorf Workshop on Approaches to Coercion and Polysemy University of Oslo November, 2017 Download these slides from: sfb991.uni-duesseldorf.de/en/c09/slides/

Outline Introduction and plan Small case study: wine and beer Analysis Structure of pseudo-partitive NPs Semantic analysis of pseudo-partitive NPs in TTR CD + Mass Noun expressions: why are measure readings (usually) blocked

Data 1: Readings of pseudo-partitive NPs Pseudo-partitive NPs such as two glasses of wine have at least 3 interpretations: (1) a. He turned to reach the two glasses of wine that stood on a bedside table. (BNC) b. i (sic.) should set the record straight with Clayart that two glasses of red wine a day have beneficial health results. [UKWaC] c. Two glasses of wine is equal to 3 standard drinks of any alcoholic beverage. [UKWaC] (1a) has a container reading: the verbs, reach, stand, select for solid objects, e.g., containers (glasses), not the contents (wine) (1b) has a contents reading: plural agreement, the contents (wine) of exactly two glasses has the effect on health, not the containers (1c) has a measure reading: singular agreement, the equivalence is between volume or alcoholic content of the totality of alcoholic beverage contained in the relevant containers 1/25 Container, contents and measures

Data 2: CD+Mass N constructions Type mismatch between Cardinal (CD) numerical and Mass Noun (Mass N) prompts a mass-to-count shift ( CD as in Penn Treebank tags): We get two different coerced mass-to-count shifts associated with constructions like two wines: (2) a. Winnie returned, with two wines, which was just as well as Rab had all but drained her glass. [BNC] b. At the rehearsal dinner I fake-drank two wines, slowly pouring them into other people s glasses when they weren t looking. (2a) container reading: returned, with two wines evokes implicit containers two glasses containing wine (2b) contents reading: pour selects liquid (therefore contents). Plural agreement. two portions of wine, each the contents of a glass. Measure readings are hard to get (none in BNC or UKWaC). (3) # Two wines is equal to 3 standard drinks of any alcoholic beverage. Nb: Actually, for wine and beer, sentences expressing subkinds are far more common (two wines were served with dinner: a Malbec and a Sauvignon.) 2/25 Container, contents and measures

Summary: Pseudo-partitive NPs such as two glasses of wine have at least 3 interpretations: Container Contents Measure Novel observation: For pseudo-partitive NPs (two wines), coercively re-interpreted with an implicit classifier-like concept (two glasses of wine), it is much harder (if at all possible) to get measure interpretations. Our main question: Why can t we get measure interpretations for two wines? What is it about measure interpretations that makes them so hard to access when coercing a mass N into a count interpretation? 3/25 Container, contents and measures

Plan Some empirical background. A (very) preliminary corpus study. An analysis of pseudo-partitive NPs using TTR (Cooper, 2012, a.m.o) Basic idea: Measure interpretations are derived from contents interpretations. An analysis of CD+Mass-N phrases: A type clash between a CD and a mass N triggers either: (1) A shifted count interpretation (container with mass N contents) (2) A shifted count interpretation (some apportioned contents, held in a container) Either (1) or (2) resolve the type clash. So usually no driving factor to coerce (2) into a measure interpretation But there can be such a driving factor (e.g., SG agreement: two wines is equal to...). So, a contents-to-measure shifted interpretation is in principle available. But it is not usually available in practice. Reason: high processing burden on agents. Two shifts required to derive the requisite measure interpretation (mass-to-count/contents shift and a count-to-mass/measure shift) 4/25 Container, contents and measures

Pilot BNC corpus study Two target Ns: wine and beer Two contexts: 1. Part of a measure phrase with a classifier-like head N e.g. glass, bottle. Four glasses of red wine. Two bottles of beer. 2. Directly modified by a CD ( cardinality ) expression Four red wines. Two beers. Resulting data hand-annotated with different readings. 5/25 Container, contents and measures

Search criteria and frequencies: Full NPs: [tag="cd"] [tag="jj"]* [lemma="glass bottle jug pitcher can"] [word="of"] [tag="jj"]* [tag="nn NNS" & lemma="beer"] [tag="cd"] [tag="jj"]* [lemma="glass"] [word="of"] [tag="jj"]* [tag="nn NNS" & lemma="wine"] CD+Mass-N: [tag="cd"] [tag="jj"]* [tag="nns" & lemma="beer"] [tag="cd"] [tag="jj"]* [tag="nns" & lemma="wine"] Frequencies: Full NPs CD+Mass-N Total beer 13 20 33 wine 37 35 72 Total 50 55 105 6/25 Container, contents and measures

Results Readings for beer (%) Full NP CD+Mass-N Subkind 0 60 Contents 46 10 Container 46 30 Measure 8 0 Readings for wine (%) Full NP CD+Mass-N Subkind 0 91 Contents 29.7 3 Container 56.8 6 Measure 13.5 0 Readings collated (%) Full NP CD+Mass-N Subkind 0 80 Contents 34 5 Container 54 15 Measure 12 0 Full pseudo-partitive NP Mostly contents and container readings No kind readings (not surprising) Relatively few measure readings CD+Mass-N Vast majority kind readings No measure readings beer vs. wine Easier to coerce beer into container/contents readings 7/25 Container, contents and measures

Discussion Measure readings are witnessed in pseudo-partitive NPs formed with classifier-like expressions. No evidence of measure readings of CD + Mass-N expressions. But more data needed. In general, constructed examples are pretty bad. Compare: (4) a. Two wines is equal to 3 standard drinks of any alcoholic beverage or any snack food that contains about 200 calories, such as 40 grams of chocolate. b. Two glasses of wine is equal to 3 standard drinks of any alcoholic beverage or any snack food that contains about 200 calories, such as 40 grams of chocolate. [UKWaC] 8/25 Container, contents and measures

Rothstein s syntactic analysis Landman (2003), Rothstein (2011) (i.a.): Numerical expressions in number marking languages can be NP modifiers. Freely licensed shift, approx: n e, t, e, t Rothstein (2011) argues for a syntactic distinction between container readings and measure readings (does not analyse contents readings): Container Measure Container reading much like a CD + Count-N structure, but with a complex NP (glasses of wine). Measure reading formed from a measure (three glasses) and an argument (wine) 9/25 Container, contents and measures

Our analysis: container/contents readings Adopt (roughly) Rothstein s analysis for direct numerical modification and container readings. Same syntax for contents reading (different semantics). np np numdet np numdet np num n num n pp three glasses three glasses of wine 10/25 Container, contents and measures

Our analysis continued: Measure readings 1 Reject Rothstein s analysis of measure NPs Tagline: Extensive measure functions are not expressed only by classifier-like N predicates like glass alone: (4b) Two glasses of wine is equal to 3 standard drinks of any alcoholic beverage or any snack food that contains about 200 calories, such as 40 grams of chocolate. [UKWaC] Measures are formed from e.g., glass of wine, not just containers, e.g., glass alone, to make the relevant different dimensions for measurement accessible. The contents of the glass matters (not just the container), because we can measure relative to different dimensions of the contents; for wine, the relevant dimension is not only volume (based on the container, e.g., glass alone), but also calories (4b), units of alcohol, volume,... You can t ask: how many calories is the contents of a glass (punkt)? You can ask: how many calories is the contents of a glass of wine? 11/25 Container, contents and measures

Our analysis continued: Measure readings 2 (4b) Two glasses of wine is equal to 3 standard drinks of any alcoholic beverage or any snack food that contains about 200 calories, such as 40 grams of chocolate. [UKWaC] In (4b), reading is: 2 (measure in calories(content(glass of wine))) I.e. Measure readings are derived from (intensionalised) contents readings Rothstein s analysis: Our analysis: measurenp num two measure np n glasses of pp wine 12/25 Container, contents and measures

Summary Direct counting numdet np np Numerical shifted to NP modifier. num n Container and contents readings Measure readings numdet num three num two three np n glasses measurenp n glasses glasses np of measure np of pp pp wine wine Numerical shifted to NP modifier. CL-like N shifted to either container or contents interpretation. Numerical denotes a numeral. Contents reading of NP shifted to a contextually salient measure over the relevant dimension. 13/25 Container, contents and measures

Count nouns and plural count nouns Functions from records to record types [x : Ind] RecType (abbreviation: Ppty) Record argument must be a witness to an individual (or plural individual ( Ind)) and a context (Cxt) Contexts, here, are individuating contexts: provide individuation schemas Cxt abbreviates (x : Ind PType) (x : Ind PType) Record type contains the counting base property (labelled p cbase ) An the application conditions (labelled s app ) [ ] x: Ind p cbase =λr : [y : Ind] [ s 1 : r.c(glass)(r.y) ] : Ppty glass =λr :. c: Cxt s app : [ s 2 : r.c(glass)(r.x) ] glasses =λr : [ x: Ind c: Cxt ] p cbase =λr : [y : Ind] [ s 1 :r.c(glass)(r.y) ] : Ppty. s app : [ s 2 : r.c(glass)(r.x) ] 14/25 Container, contents and measures

Numerical expressions and the NMOD shift Numerical expressions interpreted as record types. E.g. the type of witnesses for the singleton type of real numbers equal to 3 NMOD shifts such types to a Ppty modifier The effect: to add a condition that the argument individual has some cardinality relative to the counting base property. The counting base property is needed (if x = 52cards, then Card(x,card cbase,52), but Card(x,deck cbase,1)) (Link, 1983) three = [ n=3 : R ] [ ] x: NMOD = λr :[n : R] λp :Ppty λr : Ind. [ c: Cxt ] N type=p(r): RecType s : Card(r.x, P(r).p cbase, R.n) [ ] x: NMOD( three )= λp :Ppty λr : Ind. [ c: Cxt ] N type=p(r): RecType s : Card(r.x, P(r).p cbase, 3) 15/25 Container, contents and measures

Direct counting with lexically simple count Ns np NMOD( three ) = λp :Ppty λr : [ x: Ind c: Cxt ] [ N type=p(r): RecType. s card : Card(r.x, P, 3) ] numdet three glasses [ ] x: = λr : Ind p cbase =λr : [y : Ind] [ s 1 :r.c(glass)(r.y) ] : Ppty. c: Cxt [ s app : s2 : r.c(glass)(r.x) ] NMOD( three )( glasses ) [ ] p x: = λr : cbase =λr : [y : Ind][s 1 :r.c(glass)(r.y)] : Ppty Ind N type= s c: Cxt app : [ s 2 : r.c(glass)(r.x) ] s card : Card(r.x, N type.p cbase, 3) num np n glasses : RecType After flattening and relabelling: [ ] p x: = λr : cbase =λr : [y : Ind][s 1 :r.c(glass)(r.y)] : Ppty Ind. s c: Cxt app : [ s 2 : r.c(glass)(r.x) ] s card : Card(r.x, p cbase, 3) 16/25 Container, contents and measures

Count nouns versus mass nouns As we have argued elsewhere (i.a. Sutton and Filip, 2016b,a) Mass noun lexical entries specify a null counting context (an individuation schema that allows overlap). Overlap in counting bases makes counting go wrong (Landman, 2011) also assume a type Subst (not all mass nouns denote substances, however) Suppress details here. Main idea: If c : Cxt o, c(p) applies to an individual a iff, for some c : Cxt, c (P)(a). a must belong to some individuating partition of the domain of P If different, overlapping partitions are allowable, then evaluation at the null counting context allows them all in. glass =λr : [ x: Ind c: Cxt ]. wine =λr : [ x: ] Subst. p cbase =λr : [y : Ind] [ s 1 : r.c(glass)(r.y) ] : Ppty s app : [ s 2 : r.c(glass)(r.x) ] c : Cxt 0 p cbase =λr : [y : Ind] [ s 1 : c(wine)(r.y) ] : Ppty s app = [ s 2 : c(wine)(r.x) ] 17/25 Container, contents and measures

Container shifts One of two shifts licensed (in English), by of in pseudo-partitives (or by intonation as in Three boxes [pause] tomatoes). { ofcontainer of = of contents [ x: Ind c: Cxt of container = λp :Ppty λq :Ppty λr : N type=p(r) : RecType [ ] z : par : Ind Subst c : Cxt 0 s in.vessel : P(r).s app Q(par).s app [s contain : contain.each(r.x, par.z)] After some steps of flattening and relabelling: glasses of container wine = p cbase =λr :[y : Ind].[s 1 :r.c(glass)(r.y)] : Ppty [ ] [ ] z : x: λr : par : Ind Subst Ind c : Cxt 0. c: Cxt s 2 : r.c(glass)(r.x) s in.vessel : s 3 : par.c(wine)(par.z) s contain : contain.each(r.x, par.z) I.e. Glasses that contain wine 18/25 Container, contents and measures ].

Contents shifts The other of two shifts licensed (in English), by of in pseudo-partitives (or by intonation). [ x: of contents = λp :Ppty λq :Ppty λr : ] Ind Subst. c: Cxt N type =Q(r) : RecType [ z : par : ] Ind c : Cxt s in.vessel : Q(r). s app P(par).s app [s within : each.within(r.x, par.z)] After some steps of flattening and relabelling: glasses of contents wine = p cbase =λr :[y : Ind].[s 1 :r.c(wine)(r.y)] : Ppty [ z : [ x: λr : ] par : ] Ind Ind Subst c : Cxt. c: Cxt s 2 : r.c(wine)(r.x) s in.vessel : s 3 : par.c(glass)(par.z) s within : each.within(r.x, par.z) I.e. Individuated partition of wine such that each part of the partition is contained by a glass. 19/25 Container, contents and measures

Measure readings: a shift on a contents shift Evoked by either lexical or discourse context (singular verbal agreement with NP or comparison with a scale inconsistent with [ a contents (or container) ] reading) x: MEASURE = λp :Ppty λr :[n : R] λr : Ind Subst. d: Scale [ ] v: par 2 : Ind Subst c: Cxt R.n=Measure(r.x, r.d, P(par 2 )) : R [ ] x: MEASURE( glasses of container wine ) = λr :[n : R] λr : Ind Subst. d: Scale [ ] x: par 2 : Ind Subst c: Cxt p cbase =λr :[y : Ind].[s 1 :r.c(wine)(r.y)] : Ppty [ ] z : par : Ind c : Cxt R.n=Measure(r.x, r.d, ) : R s 2 : par 2.c(wine)(par 2.v) s in.vessel : s 3 : par.c(glass)(par.z) s within : each.within(par 2.v, par.z) Function from numbers n to property of individuals or substances that measure n with respect being wine, the contents of a glass on some scale (volume, calorific content etc.) 20/25 Container, contents and measures

Summary Predictions of the account for full pseudo-partitive NPs Measure readings of full pseudo-partitive NPs usually require a substantial contextual trigger and an interpretive effort on the part of the interpreter Why? because measure readings are derived from contents readings (a contents-to-measure shift is needed) Need discourse context or singular verbal agreement to force this shift. Accessibility hierarchy: cl-like concept (glass) container of / [pause] + N sg agreement / in + measure PP etc. measure Type clash resolved by shift to measure interpretation contents Type Clash 21/25 Container, contents and measures

Sketching an explanation: What is different about CD + Mass Noun expressions? 1. Type mismatches are resolved by container or contents shifts. Type mismatches: CD + Mass Noun, and (sometimes) Mass Noun + PL. Agents must coerce the Mass N into a count N reading. Requires supplying additional concepts that are salient or conventional (e.g. concept for glass), and then type shifting this into a container or a contents reading modifier. Type mismatches (CD + Mass N and/or Mass N + PL) are resolved by shifting to a container or a contents reading 2. Cognitive burden No need to shift again Measure readings are highly context sensitive (require identifying the relevant dimension (and a scale) for measurement of contents) It is usually too much to: retrieve a container concept, shift it to a contents reading, identify a dimension for measurement (possibly with an associated scale), and then shift the contents reading into a measure reading relative to the dimension (scale). 22/25 Container, contents and measures

Accessibility hierarchy (coercion case) mass n concept (wine) count syntax / PL morphology Type Clash enumerate kinds Type clash resolved See: Sutton and Filip (2017) cl-like concept + mass n concept combine mass concept with CL-like concept container Type clash resolved contents Type clash resolved SG agreement; in + measure PP etc. Type Clash measure Resolving 2 type clashes = too much cognitive burden Only resolvable if really forced 23/25 Container, contents and measures

Summary Full pseudo-partitive NP CD+Mass N mass n concept (wine) count syntax / PL morphology Type Clash cl-like concept (glass) of / [pause] + N enumerate kinds Type clash resolved cl-like concept + mass n concept container sg agreement / in + measure PP etc. measure Type clash resolved by shift to measure interpretation contents Type Clash container Type clash resolved combine mass concept with CL-like concept SG agreement; in + measure PP etc. contents Type clash resolved Type Clash measure Resolving 2 type clashes = too much cognitive burden Only resolvable if really forced 24/25 Container, contents and measures

Conclusions New observations: Measures are formed from a complex N (glass of wine), not from just a CL on its own (glass) Measure interpretations of e.g., glass of wine are often not volume interpretations CD+Mass N constructions routinely seem to lack coerced measure interpretations If measure interpretations are derived from contents interpretations, we can derive the following correct predictions: Full pseudo-partitive NP CD+Mass N One type clash resolution to get to the measure interpretation Fewer measure than container/ contents interpretations Two type clash resolutions to get to the measure interpretation Hardly any measure interpretations Not predicted: Why it is harder to coerce wine than beer into container/contents interpretation 25/25 Container, contents and measures

Selected References I Robin Cooper. Type theory and semantics in flux. In R. Kempson, T. Fernando, and N. Asher, editors, Philosophy of Linguistics, Handbook of the Philosophy of Science, pages 271 323. Elsevier, 2012. Fred Landman. Predicate-argument mismatches and the adjectival theory of indefinites. In M. Coene and Y. d Hulst, editors, From NP to DP Volume 1: The Syntax and Semantics of Noun Phrases, pages 211 237. John Benjamins, Amsterdam, 2003. Fred Landman. Count Nouns Mass Nouns Neat Nouns Mess nouns. The Baltic International Yearbook of Cognition, Logic and Communication, 6:1 67, 2011. Godehard Link. The logical analysis of plurals and mass terms: A lattice-theoretic approach. In Rainer Bäuerle, Urs Egli, and Arnim von Stechow, editors, Meaning, Use and the Interpretation of Language, pages 303 323. de Gruyter, Berlin, 1983.

Selected References II Susan Rothstein. Counting, measuring and the semantics of classifiers. The Baltic International Yearbook of Cognition, Logic and Communication, 6:1 41, 2011. Peter R. Sutton and Hana Filip. Mass/count variation, a mereological, two-dimensional semantics. The Baltic International Yearbook of Cognition Logic and Communication, 11:1 45, 2016a. Peter R. Sutton and Hana Filip. Probabilistic mereological type theory and the mass/count distinction. Under review for JLM (proceedings of Type Theory and Lexical Semantics (TYTLES) workshop at ESSLLI 2015), 2016b. Peter R. Sutton and Hana Filip. Restrictions on subkind coercion in object mass nouns. Proceedings of Sinn und Bedeutung 21, page (To appear), 2017. Preprint available at: https://sites.google.com/site/sinnundbedeutung21/ proceedings-preprints.

Thank you very much for listening! Download these slides from: sfb991.uni-duesseldorf.de/en/c09/slides/