Myths about Construction Grammar

1. Laura A. Michaelis Department of Linguistics, Institute of Cognitive Science University of Colorado at Boulder Myths about Construction Grammar Fifth International Conference on Construction Grammar

2. The Good News • Language scholars from varied backgrounds concur that knowledge of language includes grammatical generalizations of varied grains . • CxG informs models of acquisition (Tomasello 2006, Diessel & Tomasello 2002), aphasia (Gahl et al. 2000), sentence processing (Hare & Goldberg 2000, Glenberg & Kaschak 2002, Bencini & Goldberg 2005) concept learning by autonomous agents (Steels & De Beule 2006). • CxG is compatible with statistical parsers (Jurafsky & Narayanan 1998), exemplar-based syntax/phonology (Bybee 2001), statistical genre studies (Gries et al. 2005). • CxG provides new insights into language evolution: we ask not Is grammar adaptive? (Pinker & Bloom 1990) but Are the constructions that collectively constitute the grammar adaptations? (Jackendoff 2003). Fifth International Conference on Construction Grammar

3. The Bad News • CxG has affected neither the practice nor ideology of mainstream syntax: • Chomsky is widely regarded as having retained his place at the center of the discipline. It’s his theories that you’ll find today in most linguistics textbooks. “When the intellectual history of this age is written, Chomsky is the only linguist whom anybody will remember,” says Geoffrey Nunberg, an adjunct professor at the School of Information at the University of California at Berkeley and a consulting professor of linguistics at Stanford University. (Evan Goldberg, ‘Who Framed George Lakoff?’, The Chronicle of Higher Education, 8/15/08) Fifth International Conference on Construction Grammar

4. Why is this the Case? • The focus of syntactic theory has long been on defining a possible human language. • For Chomsky(1995: 435) and adherents, this justifies a reduction in the range of linguistic facts that the theory should seek to cover. • Construction grammarians retain a commitment to thorough description of individual language grammars. • CxG therefore appears to be a demonstration of the infinite diversity of language and not an explanatory theory of language. Fifth International Conference on Construction Grammar

5. What to do? • Where possible formalize generalizations about constructions, using Sign-Based Construction Grammar (SBCG; Sag 2007, 2008). • Explain what makes this formalism flexible, elegant and principled: • It is localist. • It is the one generative theory that doesn’t overgenerate. • Family-resemblance relations are encoded in ‘constructional DNA’. Fifth International Conference on Construction Grammar

6. What to do? • Show that theories without constructions founder on the facts: • There is no sensible conception of compositionality without constructions. • Constructions license complements and complement realization. • Core and periphery are interleaved during production. Fifth International Conference on Construction Grammar

7. What to do? • Counteract the major myths about CxG: • It is anti-formal. • It fails to capture generalizations in grammar. • Its practitioners are obsessed with linguistic marginalia. • It is opposed to compositional semantics. • It is not constrained (“anything can be a construction”). • It does not provide a universal framework for syntax. Fifth International Conference on Construction Grammar

8. Today’s plan • Develop a toolkit of CxG talking points based on the three Fs: • Functionality • Three reasons for people to adopt the SBCG formalism (whether they’ve used a formalism before or not) • Facts • Three lines of linguistic evidence for a construction-based view of grammar • Fighting back • Rebuttals to six entrenched myths about CxG Fifth International Conference on Construction Grammar

9. Functionality: localism • (Combinatoric) constructions are descriptions of constructs. • These descriptions are type constraints: • x-cxt ⟹ […] • Constructs (intuitively) are local trees (mother-daughter configurations) with feature structures (specifically, signs) ‘at the nodes’. • But a construct is not actually a tree: it is a feature structure that contains a mother (MTR) feature and a daughters (DTRS) feature. Fifth International Conference on Construction Grammar

10. Brief excursus: signs • A sign is a type of feature structure (AVM). • It contains values for 5 features (Sag 2007): • PHON • FORM: a list of formatives comprising the expression • ARG-ST: a ranked list of a lexical expression’s arguments • SYN: CAT and VAL(ENCE) • SEM: INDEX and FRAMES • CNTXT: TOPIC and FOCUS Fifth International Conference on Construction Grammar

11. Functionality: localism • Early CxG (Fillmore & Kay 1995) saw the grammar as an inventory of trees (nested boxes). • Argument-structure constructions were represented by feature structures, e.g., the Transitive construction. • But it was not made clear how a lexical entry qualified as a construction. Fifth International Conference on Construction Grammar

12. Functionality: localism • SBCG captures properties common to combinatory constructions and lexical-class constructions: • Lexeme, word and phrase are all subtypes of the type sign. • Both words and phrases are licensed by the Sign Principle (Sag 2007, (36)): • Every sign must be lexically or constructionally licensed, where: • A sign is lexically licensed only if it satisfies some lexical entry. • A sign is constructionally licensed only if it is the mother of some construct. • So we neutralize the difference between words and phrases by describing a phrase type according to the properties of its mother sign. Fifth International Conference on Construction Grammar

13. Lexical-class and combinatoric constructions (Sag 2007) Fifth International Conference on Construction Grammar

14. Functionality: variable grains • Some language generalizations are not very general: • ISIS (Brenier & Michaelis 2005): The real question was is are we getting a reasonable return on our investment. • BE1 may be was, is or being, but BE2 is invariantly is. • Independent-clause (IC) exclamatives (Sag 2008): God, *( I can’t believe) who they hired/where they went! • IC and subordinate-clause exclamatives differ with regard to the syntactic category of the filler daughter. • “A grammar that provides no mechanism for imposing […] category restrictions will overgenerate” (Sag 2008). Fifth International Conference on Construction Grammar

15. Functionality: variable grains • SBCG avoids overgeneration by treating nodes as feature structures—not category labels. • A description of a feature structure is a set of properties. • As property sets, feature-structure descriptions follow the logic of set inclusion: • The more properties in the description, the smaller the class of language objects it picks out. • For example, the feature set that describes an IC exclamative (e.g., What fools!) includes that which defines the filler-head construction. • Inclusion relations among FS descriptions allow us to model constructs at each step along the idiomaticity continuum. • With an array of constructions of correspondingly graded generality. Fifth International Conference on Construction Grammar

16. Functionality: variable grains • The filler-head construction: Fifth International Conference on Construction Grammar

17. Functionality: Inheritance • A leading insight of CxG from inception (e.g., Lakoff 1987): grammar rules are not procedures but objects, and as such, subject to taxonomic organization. • Goldberg (1995) describes two major inheritance relations: • Instance link • Subpart link • But it has remained unclear just how these relations are to be represented: • A stipulation in a construction x, ‘inherit construction y’ (Fillmore & Kay 1995) • Typed links in radial-category diagrams (Lakoff, 1987, Goldberg 1995, Michaelis & Lambrecht 1996) Fifth International Conference on Construction Grammar

18. Functionality: Inheritance • In SBCG, inheritance relations are integrated into descriptions of MTR and DTR signs. • Recall that a construction is a conditional statement defining the properties that are common to all instances of a given feature-structure type: • x-cxt ⟹ […] • Instance link: A construction defines a construct type A, and another construction defines a subtype of A by mentioning A in the consequent clause. • Subpart link: The HD-DTR of a construct type is the same as the MTR of some other construct type. Fifth International Conference on Construction Grammar

19. Functionality: Inheritance Instance link: inverted exclamative (Was I wrong!) MTR’s type is found in two constructions. Subpart link: negative-adverb inversion (Never have I seen one.) The HD-DTR is [INV+], as is the MTR sign of the YNQ construction. Fifth International Conference on Construction Grammar

20. Facts • We have looked at three major assets of SBCG: • locality • variable granularity • integration of inheritance into type constraints • We will now look at three major lines of evidence for construction-based syntax: • compositionality needs constructions. • constructions license complements. • core and periphery are interleaved during production. Fifth International Conference on Construction Grammar

21. Fact: Composition requires constructions • Say a class of expressions can be viewed as licensed by a PS rule accompanied by a rule that composes the semantics of the mother from the semantics of the daughters. • In such cases, CxG proposes a construction that is functionally equivalent to such a rule-to-rule pair. • But CxG can also describe linguistic structures in which the mother of a given local tree may have more than one interpretation. • Can proponents of strict composition? Fifth International Conference on Construction Grammar

22. Fact: Composition requires constructions • Z. Szabó, Stanford Encyclopedia of Linguistics (2007): • If a language is compositional, it cannot contain a pair of non-synonymous complex expressions with identical structure and pairwise synonymous constituents. • But what about syntactically regular idioms (Fillmore et al. 1988)? • Pseudoimperative. Now watch me be somehow unable to make it out there. • Pseudoconditional. If you're Obama, you might not like the idea of the Clintons in the White House. Fifth International Conference on Construction Grammar

23. Fact: Composition requires constructions • From these interpretive affordances we must conclude either of two things (Kay & Michaelis to appear): • English is not compositional. • Pseudoimperatives and pseudoconditionals are syntactically distinct from their respective vanilla counterparts. • A conception of composition based on X’-syntax leads to an absurd result. • We need composition by construction: • Two different meanings for the same syntactic form are licensed by two different collections of constructions. Fifth International Conference on Construction Grammar

24. Fact: Complements are licensed by constructions • There are two lines of evidence supporting the contention that constructions license complements: • Valence variation. The full range of verb-valence variability, including zero complement realization, cannot be described by augmentative operations on event structure (as per Rappaport Hovav & Levin 1998 [RHL]). • Weird sisterhood. Many verb frames license sisterhood relations not predicted by the general-purpose rule that combines heads and complements. Fifth International Conference on Construction Grammar

25. Fact: Complements are licensed by constructions • The RHL model assumes that verbs unify with event-structure templates based on Aktionsart class. • It makes three predictions about null complements (Ruppenhofer 2004, Goldberg 2005, 2006): • As nonstructural arguments, the second arguments of bivalent state, achievement and activity verbs should always be omissible, e.g., Have you eaten? • Nonstructural participants are subject only to a recoverability condition based on prototypicality (p. 115); therefore all null complements should have existential (indefinite) interpretations, as in, e.g., She reads. • As structural arguments, patient arguments of accomplishment verbs, e.g., kill and break, should never be omissible. Fifth International Conference on Construction Grammar

26. Fact: Complements are licensed by constructions • But each prediction proves false: • False: ARG2 omission always possible: She resembles *(people). We discussed *(issues). *I prefer (things). • False: Exclusively indefinite null ARG2: I remember (that). I prepared (for that event) for weeks. She arrived (there). • False: Accomplishment verbs cannot omit ARG2: • Existential null complements licensed by verbs of emission/ingestion (spit, rain, swallow) . • Almost any verb in an iterated-event context: Using a low-flying plane, they kill in winter, You just take and take, She has never failed to impress. • And null complements of nouns, prepositions, adjectives remain unaccounted for. Fifth International Conference on Construction Grammar

27. Fact: Complements are licensed by constructions • Instead of defining a class of ‘structurally intransitive’ verbs, SBCG proposes that derivational constructions license null complements. • These constructions remove arguments from verbs’ valence lists. • Evidence that null complementation is a construction-based affordance: override of null complementation restrictions on verbs in generic/habitual contexts: They kill in winter, She has always impressed. • Constructions determine existential versus anaphoric construal of zero arguments. Fifth International Conference on Construction Grammar

28. A Null Instantiation Construct Fifth International Conference on Construction Grammar

29. Fact: Core and periphery are interleaved • Stretches of speech licensed by idiomatic constructions can contain within them stretches licensed by ‘regular rules’ and vice versa. • Example: A politician pull the leg of a philosopher? No way. • This is an example of the Incredulity Response (IR) construction. • The IR semantico-pragmatic properties: • a property predicate (e.g., pull the leg of a philosopher) • an entity (e.g., a politician) • an expression of incredulity concerning the entity’s membership in the class of individuals named by the property predicate (e.g., No way). Fifth International Conference on Construction Grammar

30. Fact: Core and periphery are interleaved • Formally, the entity is expressed by a NP and the predicate by a nonfinite VP or other phrase. • Lambrecht (1990) argues that the IR is a topic-comment construction: the entity and predicate are detached topics. • The NP and XP are distinct intonation units. • They can be reordered with respect to one another: Pull the leg of a philosopher? A politician? Fifth International Conference on Construction Grammar

31. Fact: Core and periphery are interleaved • This construction performs a basic communicative function—commenting on the validity of someone’s prior attribution. • But it does so in a way that owes little or nothing to the ordinary English syntax of predication and subordination. • It is also obvious, however, that the entity constituent and the predicate constituent are licensed by regular or “core” constructions of English. • Only their combination in the IR construction is idiomatic. Fifth International Conference on Construction Grammar

32. Fact: Core and periphery are interleaved • Further, coterminous with the syntactically transparent VP try to pull the leg of a philosopher, we find the VP idiom pull the leg of a philosopher, licensed by the idiomatic pull-someone’s-leg construction. • Going further inside the NP the-leg-of-a-philosopher, (licensed by the idiomatic pull-someone’s-leg construction), we find the transparent genitive PP of a philosopher. • Thus, it is unlikely that grammar consists of a set of productive rules, a lexicon and a collection of frozen phrasal idioms. • Instead, these ‘modules’ appear to be permeable. Fifth International Conference on Construction Grammar

33. Fighting back • Having looked at three major facts favoring CxG (composition, complementation, interleaving), we will discuss rebuttals to six major myths about CxG: • It is nonrigorous • It does not offer generalizations • It is obsessed with linguistic marginalia • it is opposed to compositional semantics • it is not constrained • it does not provide a universal framework for syntax Fifth International Conference on Construction Grammar

34. Fighting back: CxG is formal • Myth 1: CxG is anti-formal and therefore nonrigorous. • Rebuttal • Not all work in CxG is formal, nor should it be. • Without descriptive work, there would be nothing to formalize. • But CxG also provide tools for formalists. • In SBCG the basic units of grammatical description are signs (licensed either by lexical entry or construction). • The grammar comprises a set of lexical entries and a set of constructions, structured by a type hierarchy. Fifth International Conference on Construction Grammar

35. Fighting back: CxG captures generalizations • Myth 2: Construction-based generative syntax failed back in the 60s; there’s no reason to believe it should work now. • Rebuttal • Yes, construction-based TG lacked cross-constructional generalizations: • Each movement rule specified the same operation operating over the same unbounded context as every other such transformation (Ginzburg & Sag 2000:4). • But constructional TG lacked generalizations over patterns because it modeled patterns (e.g., relative clauses, WH questions, topicalization) as recursive processes rather than grammar objects. • TG was designed to represent one type of relationship—that between tree structures—and tree structures are not in grammar. • SBCG abandons trees in favor of feature structures, which can be organized taxonomically. Fifth International Conference on Construction Grammar

36. Fighting back: CxG upholds standards of data coverage • Myth 3: CxG is a theory of linguistic marginalia. • Rebuttal: • Admittedly, constructionists have favored the case-study format and have focused on minor patterns. • But constructions are intended to capture all combinatoric patterns of a language—from the most idiomatic to the most general. • The general patterns include the specifier-head pattern and the head-complement pattern. • We view the core-periphery distinction as a gradient phenomenon, and do not assume a cut-off point within a language at which our descriptive obligations end.  Fifth International Conference on Construction Grammar

37. Fighting back: CxG has composition • Myth 4: CxG is opposed to compositional semantics. • Rebuttal: • In putting semantics in the MTR that is not in the DTRs, CxG indeed allows conceptual content to come from outside the lexicon (Jackendoff 1997). • But strict composition may not be tenable anyway: • Szabó 2007: a compositional language does not contain any paired non-synonymous phrases with identical structure and pairwise synonymous constituents. • If this definition is valid, either: • English is noncompositional, or • any phrase with both idiomatic and composed readings (e.g., spill the beans) has two different syntactic analyses. Fifth International Conference on Construction Grammar

38. Fighting back: CxG has composition • Composition without constructions leads to an absurd result. • For constructionists, each meaning is licensed by a distinct construction.  • Construction-based grammars are therefore intuitively compositional: • If you know the meanings of the words, • and all the rules that combine words and phrases into larger formal units, • then you know the meanings of all the larger units.  Fifth International Conference on Construction Grammar

39. Fighting back: ‘construction’ means something • Myth 5: CxG is not constrained; anything can be a construction. • Rebuttal: • This may be a valid criticism of earlier incarnations of CxG (Fillmore & Kay 1995, Goldberg 1995), but construction in SBCG is precisely defined: • A construction is a type constraint. • It describes a sign that is either the mother of a construct or a lexical entry. Fifth International Conference on Construction Grammar

40. Fighting back: CxG is universalist • Myth 6: CxG has no theory of syntactic universals • Rebuttal: • Type hierarchies allow us to both capture potentially universal constraints on structure and describe entrenched exemplars of particular structures.  • In fact, SBCG makes strong universal claims, captured by the Sign Principle and the Head-Feature Principle. • Other putatively universal constraints on grammatical architecture, in particular those advanced by proponents of Principles and Parameters, appear to confuse representational conventions with linguistic facts.  Fifth International Conference on Construction Grammar

41. Fighting back: CxG is universalist • Syntactic procedures (movement rules) and hierarchical structure are not discoveries (pace Pesetsky 1997: 337) but mutually reinforcing assumptions: • The head assumption: dependency relationships are syntactic sisterhood relationships. • The derivation assumption: when there are two different syntactic expressions of a given semantic role, there is a procedure that changes the position of the expression in hierarchical structure. • These conventions subserve the autonomy thesis, but they are not language facts. • We know this because, e.g., HPSG/SBCG describe case alternations without recourse to derivations and binding constraints with ‘flat’ dependencies (ARG-ST). Fifth International Conference on Construction Grammar

42. Conclusion • Careful grammatical description is the constructionist’s stock in trade. • But let us not be afraid to ‘go meta’. • This means: • Rearticulating the vision in each paper. • Explaining why the insights being offered would not be available in another framework. • Anticipating objections and exposing the assumptions on which they’re based. Fifth International Conference on Construction Grammar