The interplay between conceptual and referential aspects of meaning - - PowerPoint PPT Presentation

The interplay between conceptual and referential aspects of meaning

Gemma Boleda Universitat Pompeu Fabra (work in collaboration with Louise McNally)

BRIDGE Workshop ESSLLI 2018, 6–10 August 2018, Sofia, Bulgaria

1

Acknowledgements

◮ Co-authors in cited papers ◮ Laura Aina, Kristina Gulordava, Carina Silberer,

Ionut-Teodor Sorodoc, Matthijs Westera

◮ This project has received funding from the European

Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement No 715154). AMORE: A distributional Model Of Reference to Entities).

2

The problem

Modifier-noun relations bifurcate:

◮ Strong default interpretations ◮ In context, anything goes

How to explain this?

3

Example: adjectives with little/no context

Canadian visit / attack / decision. . . must denote agent (Kayne 1984, a.o.)

4

Example: adjectives with little/no context

Canadian visit / attack / decision. . . must denote agent (Kayne 1984, a.o.) (1) Yeltsin met the prospective Democratic presidential candidate Bill Clinton on June 18. His itinerary also included an official visit to Canada

4

Example: adjectives with little/no context

Canadian visit / attack / decision. . . must denote agent (Kayne 1984, a.o.) (1) Yeltsin met the prospective Democratic presidential candidate Bill Clinton on June 18. His itinerary also included an official visit to Canada/??an official Canadian visit.

4

Example: adjectives with little/no context

Canadian visit / attack / decision. . . must denote agent (Kayne 1984, a.o.) (1) Yeltsin met the prospective Democratic presidential candidate Bill Clinton on June 18. His itinerary also included an official visit to Canada/??an official Canadian visit. (2) Put the scarf in the red box.

4

Adjectives in context

(3) Prince Edward and wife begin Canadian visit

5

Adjectives in context

(3) Prince Edward and wife begin Canadian visit (4) (Context: For a fundraising sale, Adam and Barbara are sorting donated scarves according to color in different, identical, brown cardboard boxes. Barbara distractedly puts a red scarf in the box containing blue scarves.) Adam: Hey, this one belongs in the red box!

5

More specific questions

◮ Strong default interpretations

◮ Why does the default seem so strong?

◮ In context, anything goes

◮ Why/How can context ameliorate anything?

◮ What kind of theory can account for this compositional

phenomenon?

6

Previous work: Two general approaches to modification

◮ Semantic primitives ◮ Underspecification of modification relation + resolution in

context The closest thing we have seen to a mixed approach appears in Asher (2011).

7

Resolution via semantic primitives

◮ Long tradition ◮ Makes explicit how the concepts introduced by the modifier

and the head are composed

◮ Examples:

◮ Levi (1978): CAUSE, HAVE, MAKE, USE, BE, IN, FOR,

FROM, ABOUT, ACT, PRODUCT, AGENT, PATIENT

◮ Pustejovsky (1995): FORMAL, CONSTITUTIVE,

AGENTIVE, TELIC

◮ Ó Séaghdha and Copestake (2009): BE, HAVE, IN,

AGENT, INSTRUMENT, ABOUT 8

Resolution via semantic primitives

(5) Canadian visit: λx.visit(x) ∧ AGENT(x, Canada) (6) red apple: λx∃y.apple(x) ∧ CONSTITUTIVE(apple)=PART-OF(y,x) ∧ red(y)

9

Pros and cons

◮ Pros

◮ intuitions about default interpretations ◮ predicts productivity

◮ Cons

◮ too strong ◮ too weak ◮ huge methodological issues

10

Underspecification + context

◮ Also a long tradition ◮ Relations are established indexically or by valuing a

variable that stands for the relation

◮ Examples:

◮ Bosch (1983), Rothschild and Segal (2009): Adjectives

denote functions from contexts to contents

◮ McNally and Boleda (2004), Kennedy and McNally (2010):

Adjectives introduce variables over relations that are valued by context 11

Underspecification + context

(7) Canadian visit: λx.visit(x) ∧ Ri(x, Canada) (compare to λx.visit(x) ∧ AGENT(x, Canada)

12

Pros and cons

◮ Pro: appropriately flexible ◮ Con: too weak ◮ Pending: a theory of how context plays its role

13

Summing up

default context-dependent semantic primitives (✔) ✘ underspecification + context ✘ (✔)

14

Synthesis: Conceptual and referential affordance in language

McNally and Boleda 2017

Distinct aspects of language afford concept composition in different ways:

◮ The concepts described ◮ The entities referred to

Affordance (Chemero (2003), based on Gibson (1979)):

◮ relation between features of situations

and abilities of organisms

15

Synthesis: Conceptual and referential affordance in language

Assumption (semiotic models, a.o.):

Proposal

◮ the connection to concepts and to the world are distinct

features of language

◮ each of them affords distinct composition process ◮ speakers avail themselves of both

16

Conceptual affordance

The concepts contributed by the components of a phrase suggest the ways in which they should be composed

17

Conceptual affordance

The concepts contributed by the components of a phrase suggest the ways in which they should be composed → default interpretations, little or no need for context

17

Conceptual affordance

The concepts contributed by the components of a phrase suggest the ways in which they should be composed → default interpretations, little or no need for context

◮ productive: speakers use regularities in our lexical

knowledge

17

Conceptual affordance

The concepts contributed by the components of a phrase suggest the ways in which they should be composed → default interpretations, little or no need for context

◮ productive: speakers use regularities in our lexical

knowledge (8) Put the scarf in the red box.

17

Referential affordance

Independently available information about the referent indicates how the concepts should be composed

18

Referential affordance

Independently available information about the referent indicates how the concepts should be composed → Ad hoc interpretations, heavy context dependence

18

Referential affordance

Independently available information about the referent indicates how the concepts should be composed → Ad hoc interpretations, heavy context dependence

◮ plastic: speakers use information about the world

18

Referential affordance

Independently available information about the referent indicates how the concepts should be composed → Ad hoc interpretations, heavy context dependence

◮ plastic: speakers use information about the world

(9) (Context: For a fundraising sale, Adam and Barbara are sorting donated scarves according to color in different, identical, brown cardboard boxes. Barbara distractedly puts a red scarf in the box containing blue scarves.) Adam: Hey, this one belongs in the red box!

18

Conceptual vs. referential effects in composition

Asher 2011, McNally and Boleda 2017

“cafetera italiana” (Italian coffee maker)

19

Conceptual vs. referential effects in composition

Asher 2011, McNally and Boleda 2017

“cafetera italiana” (Italian coffee maker) ⇓ conceptually afforded composition

19

Conceptual vs. referential effects in composition

Asher 2011, McNally and Boleda 2017

“cafetera italiana” (Italian coffee maker) ⇓ conceptually afforded composition “cafetera italiana”, too! (Italian coffee maker)

19

Conceptual vs. referential effects in composition

Asher 2011, McNally and Boleda 2017

“cafetera italiana” (Italian coffee maker) ⇓ conceptually afforded composition “cafetera italiana”, too! (Italian coffee maker) ⇓ referentially afforded composition

19

Contribution 1

◮ Strong default interpretations

◮ Why does the default seem so strong?

→ conceptually afforded modification → model with distributional semantics

◮ In context, anything goes

◮ Why/How can context ameliorate anything?

◮ What kind of theory can account for this compositional

phenomenon?

20

Distributional semantics for conceptually afforded modification

◮ default interpretations are very sensitive to the lexical

semantics of the phrase components

◮ both coarse- and fine-grained

default semantic primitives (✔) underspecification + context ✘

◮ distributional semantics provides the necessary

information, like primitive-based accounts, without their drawbacks

21

Distributional semantics

Aka vector-space semantics, related to Neural Networks / deep learning

(See Stefan Evert’s course this week at ESSLLI for more!)

22

Distributional semantics

Aka vector-space semantics, related to Neural Networks / deep learning

(See Stefan Evert’s course this week at ESSLLI for more!)

likely) mug of bourbon in hand. Some stewed milk into a heavy mug, granules of holding his coffee mug cupped in his hands. drained his mug, dropping it over his tablespoons of coffee and a single mug of milk into the mug plus four spoons of sugar placing the empty mug on the floor picking up my mug with one hand and followed by a very hot mug of tea into which from time to time to drink a mug of tea. The briefed, relax over a mug of tea and a cake and cheese and a mug of strong, black then we had a mug of cocoa and a gingerbread and a white mug with a blurred inscription. was carrying a mug of tea and

22

Distributional semantics

Aka vector-space semantics, related to Neural Networks / deep learning

(See Stefan Evert’s course this week at ESSLLI for more!)

likely) mug of bourbon in hand. Some stewed milk into a heavy mug, granules of holding his coffee mug cupped in his hands. drained his mug, dropping it over his tablespoons of coffee and a single mug of milk into the mug plus four spoons of sugar placing the empty mug on the floor picking up my mug with one hand and followed by a very hot mug of tea into which from time to time to drink a mug of tea. The briefed, relax over a mug of tea and a cake and cheese and a mug of strong, black then we had a mug of cocoa and a gingerbread and a white mug with a blurred inscription. was carrying a mug of tea and

reasonable proxy for conceptual information (shown in a lot of work in Cognitive Science, Computational Linguistics)

22

Meaning in distributional semantics

Boleda and Erk 2015

man woman gentleman gray-haired boy person lad men girl

Words most similar to man in Baroni et al. (2014).

23

Meaning in distributional semantics

Boleda and Erk 2015

man woman gentleman gray-haired boy person lad men girl +HUMAN

Words most similar to man in Baroni et al. (2014).

23

Meaning in distributional semantics

Boleda and Erk 2015

man woman gentleman gray-haired boy person lad men girl +HUMAN +MALE

Words most similar to man in Baroni et al. (2014).

23

Meaning in distributional semantics

Boleda and Erk 2015

man woman gentleman gray-haired boy person lad men girl +HUMAN +MALE +ADULT

Words most similar to man in Baroni et al. (2014).

23

Meaning in distributional semantics

Boleda and Herbelot 2016

man chap lad dude guy woman bloke boy freakin’ bloke gentleman guy bloke woah chap gray-haired lad scouser dorky doofus boy fella lass dumbass dude person man youngster stoopid fella

Words most similar to man, chap, lad, dude, guy in Baroni et al. (2014).

24

Meaning in distributional semantics

Boleda and Herbelot 2016

man chap lad dude guy woman bloke boy freakin’ bloke gentleman guy bloke woah chap gray-haired lad scouser dorky doofus boy fella lass dumbass dude person man youngster stoopid fella

Words most similar to man, chap, lad, dude, guy in Baroni et al. (2014).

24

Composition in distributional semantics

Baroni and Zamparelli 2010, Boleda et al. 2013

◮ works pretty well for composition of content words (quite a

bit of work in Computational Linguistics)

◮ table shows/expresses results ◮ map shows/??expresses location ◮ (Grefenstette et al., 2013)

25

Composition in distributional semantics

Baroni and Zamparelli 2010, Boleda et al. 2013

◮ works pretty well for composition of content words (quite a

bit of work in Computational Linguistics)

◮ table shows/expresses results ◮ map shows/??expresses location ◮ (Grefenstette et al., 2013)

◮ our proposal: what it models is conceptually afforded

modification

25

Distributional semantics for conceptually afforded composition

Predicting productivity in adjectival modification

Vecchi et al. 2011, 2017

◮ distributional semantics distinguishes between acceptable

• vs. deviant phrases

◮ unattested phrases in huge corpus

26

Distributional semantics for conceptually afforded composition

Predicting productivity in adjectival modification

Vecchi et al. 2011, 2017

◮ distributional semantics distinguishes between acceptable

• vs. deviant phrases

◮ unattested phrases in huge corpus

◮ acceptable: vulnerable gunman, huge joystick, blind cook ◮ deviant: blind pronunciation, parliamentary potato, sharp

glue 26

Distributional semantics for conceptually afforded composition

Predicting productivity in adjectival modification

Vecchi et al. 2011, 2017

◮ distributional semantics distinguishes between acceptable

• vs. deviant phrases

◮ unattested phrases in huge corpus

◮ acceptable: vulnerable gunman, huge joystick, blind cook ◮ deviant: blind pronunciation, parliamentary potato, sharp

glue

→ good model for conceptually afforded modification

26

Distributional semantics for conceptually afforded composition

Predicting productivity in adjectival modification

Vecchi et al. 2011, 2017

◮ distributional semantics distinguishes between acceptable

• vs. deviant phrases

◮ unattested phrases in huge corpus

◮ acceptable: vulnerable gunman, huge joystick, blind cook ◮ deviant: blind pronunciation, parliamentary tomato, sharp

glue

→ good model for conceptually afforded modification

26

(So far,) distributional semantics can’t model referentially afforded composition

Boleda, Baroni, Pham, McNally IWCS 2013

easy difficult former commentator former colour likely threat likely base wide perspective wide detail

27

(So far,) distributional semantics can’t model referentially afforded composition

Boleda, Baroni, Pham, McNally IWCS 2013

easy difficult former commentator former colour likely threat likely base wide perspective wide detail conceptually referentially afforded? afforded?

27

After contribution 1

default context-dependent (conc. aff.) (ref. aff.) semantic primitives (✔) ✘ underspecification + context ✘ (✔) distributional semantics ✔ ✘

28

Contribution 2

◮ Strong default interpretations

→ conceptually afforded modification → model with distributional semantics

◮ In context, anything goes

◮ Why/How can context ameliorate anything?

→ largely due to referentially afforded composition

◮ What kind of theory can account for this compositional

phenomenon?

29

Empirical support: Relational adjectives

◮ respiratory, tropical, planetary, Canadian, . . . ◮ typically denominal; adjective morphology claimed to be

transparent (e.g. Spencer 1999) – they express a relation

◮ McNally and Boleda (2004): this relationship is

underspecified

◮ expectation: relational adjectives are used more when the

relationship is specified in the previous context

◮ explains data from two statistical corpus studies

◮ Catalan (Boleda, 2007) ◮ English (Boleda et al., 2012) - focused on ethnic adjectives

(Canadian, French) 30

Catalan

(Boleda, 2007)

◮ qualitative: tou ‘soft’, imperfecte ‘imperfect’ ◮ relational: respiratori ‘respiratory’, americà ‘American’

31

Contribution 3

◮ Strong default interpretations

→ conceptually afforded modification → model with distributional semantics

◮ In context, anything goes

→ largely due to referentially afforded composition

◮ What kind of theory can account for this compositional

phenomenon?

◮ referentially afforded modification has resisted distributional

treatments → mixed model for the two types of semantic composition 32

Towards an analysis

◮ Adaptation of Discourse Representation Theory (DRT,

Kamp, 1981); also builds on (Zamparelli, 1995) and (McNally, 2016) (10) a box a. standard: u box(u) b. (McNally, 2016) u Realize(u, − − → box)

33

Towards an analysis

◮ Conceptually-afforded composition

(11) a red box u Realize(u, comp(− − → red, − − → box))

34

Towards an analysis

◮ Conceptually-afforded composition

(11) a red box u Realize(u, comp(− − → red, − − → box))

◮ Referentially-afforded composition

(12) Adam: Hey, this one belongs in the red box! a.

• ption 1:

u Realize(u, compu(− − → red, − − → box))

34

Towards an analysis

◮ Conceptually-afforded composition

(11) a red box u Realize(u, comp(− − → red, − − → box))

◮ Referentially-afforded composition

(12) Adam: Hey, this one belongs in the red box! a.

• ption 1:

u Realize(u, compu(− − → red, − − → box)) b.

• ption 2:

u Realize(u, comp(f(u, − − → red), − − → box))

34

Summing up: McNally and Boleda 2017

default context-dependent (conc. aff.) (ref. aff.) semantic primitives (✔) ✘ underspecification + context ✘ (✔) distributional semantics ✔ ✘ mixed model ✔ (✔)

35

Conclusions

◮ Language interpretation cannot be understood without

simultaneously considering

◮ what we are referring to ◮ the concepts associated with the words we are using

◮ Linguistic expressions encode significant regularities

◮ conventions of language use (long tradition; also Westera

and Boleda, submitted, and Aina, submitted)

◮ speakers use these regularities profitably

◮ Once a linguistic expression is applied to a referent, it is

grounded in a very specific way

◮ the referential act has consequences:

◮ people will continue using the same expression for the

same referent (Clark, 1992)

◮ it influences the way we understand the expression in the

first place

◮ semantic change: e.g. narrowing deer - Tier

36

Outlook

◮ Understand the interplay between conceptual and

referential aspects of meaning . . .

◮ . . . and between cognition and language more generally

◮ conceptual structure and the lexicon ◮ conceptual structure and grammar ◮ e.g.: grammar is sensitive to how adjectives compose with

nouns McNally and Boleda (2004)

◮ Integrate into linguistic theory

37

The future?

default context-dependent (conc. aff.) (ref. aff.) semantic primitives (✔) ✘ underspecification + context ✘ (✔) distributional semantics ✔ ✘ mixed model ✔ (✔) AMORE ✔ ✔

38

The interplay between conceptual and referential aspects of meaning

Gemma Boleda Universitat Pompeu Fabra (work in collaboration with Louise McNally)

BRIDGE Workshop ESSLLI 2018, 6–10 August 2018, Sofia, Bulgaria

39

Asher, N. (2011). Lexical Meaning in Context. Cambridge University Press, Cambridge. Boleda, G. (2007). Automatic acquisition of semantic classes for adjectives. PhD thesis, Universitat Pompeu Fabra. Boleda, G., Evert, S., Gehrke, B., and McNally, L. (2012). Adjectives as saturators vs. modifiers: Statistical evidence. In Aloni, M., Kimmelman, V., Roelofsen, F., Sassoon, G. W., Schulz, K., and Westera, M., editors, Logic, Language and Meaning - 18th Amsterdam Colloquium, Amsterdam, The Netherlands, December 19-21, 2011, Revised Selected Papers, Lecture Notes in Computer Science 7218, pages 112–121. Springer. Bosch, P . (1983). “Vagueness” is context-dependence. A solution to the Sorites Paradox. In Ballmer, T. and Pinkal, M., editors, Approaching Vagueness, pages 189–210. North Holland, Amsterdam. Chemero, A. (2003). An outline of a theory of affordances. Ecological Psychology, 15:181–195. Clark, H. (1992). Arenas of Language Use. University of Chicago Press, Chicago. Gibson, J. J. (1979). The ecological approach to visual perception. Houghton Mifflin, Boston, MA. Grefenstette, E., Dinu, G., Zhang, Y.-Z., Sadrzadeh, M., and Baroni, M. (2013). Multi-Step Regression Learning for Compositional Distributional

• Semantics. In Proceedings of IWCS 2013 (10th International Conference
• n Computational Semantics), pages 131–142, East Stroudsburg PA. ACL.

39

Kamp, H. (1981). A theory of truth and semantic representation. In Groenendijk, J., Janssen, T., and Stokhof, M., editors, Formal Methods in the Study of Language, volume 1, pages 277–322. Mathematisch Centrum, Amsterdam. Kennedy, C. and McNally, L. (2010). Color, context, and compositionality. Synthese, 174:79–98. Levi, J. (1978). The Syntax and Semantics of Complex Nominals. Academic Press, New York. McNally, L. (2016). Kinds, descriptions of kinds, concepts, and distributions. In Balogh, K. and Petersen, W., editors, Bridging Formal and Conceptual Semantics: Selected Papers of BRIDGE-14, pages 37–59. dup, Düsseldorf. McNally, L. and Boleda, G. (2004). Relational adjectives as properties of

• kinds. In Bonami, O. and Cabredo Hofherr, P

., editors, Empirical Issues in Syntax and Semantics, volume 5, pages 179–196. http://www.cssp.cnrs.fr/eiss5. Ó Séaghdha, D. and Copestake, A. (2009). Using lexical and relational similarity to classify semantic relations. In Proceedings of the 12th Conference of the European Chapter of the ACL, pages 621–629, Athens, GA. Pustejovsky, J. (1995). The Generative Lexicon. MIT Press, Cambridge, MA.

39

Rothschild, D. and Segal, G. (2009). Indexical predicates. Mind and Language, 24(467–493). Spencer, A. (1999). Transpositions and argument structure. In Booij, G. and van Marle, J., editors, Yearbook of Morphology 1998, pages 73–101. Kluwer, Dordrecht. Vecchi, E. M., Baroni, M., and Zamparelli, R. (2011). (linear) maps of the impossible: Capturing semantic anomalies in distributional space. In Proceedings of the Workshop on Distributional Semantics and Compositionality, pages 1–9, Portland, Oregon, USA. Association for Computational Linguistics. Zamparelli, R. (1995). Layers in the Determiner Phrase. PhD thesis, U. Rochester.

39