Logic, language and cognition Henriëtte de Swart and Yoad Winter Master program CAI, RMA Linguistics Utrecht, nov 2010 - jan 2011 Background and preliminaries I • BA: intro to linguistics/logic/cognition, and developing beginning expertise in reading academic literature, starting to work out research questions, developing relatively small applications. • MA: introduce advanced methodology and tools, develop higher level academic skills, leading to more advanced research projects and independent research. Background and preliminaries II • MA CAI: Balance between mandatory courses and optional courses. • Two tracks: Cognitive modelling Logic and intelligent systems • LLC is a mandatory course offered by linguistics and feeding into both tracks. • Core focus: meaning as core business of language. • Background reading on semantics for today: Schlenker (2008). Background and preliminaries III • Relevant recommended courses related to LLC in later periods of the year: Neurocognition of language (Avrutin, period 2) Semantic Structures (de Swart/Le Bruyn, 3) Het sociale semantische Web (Monachesi, 3) Conceptual Semantics (Zwarts, period 4) Learning in Computational Linguistics (Winter, period 4) Language and cognition • Language is means of communication in social community: transfer information from the head of the speaker to the head of the hearer. • Questions about language in the mind (language as part of human cognitive architecture). • Questions about meaning as content of linguistic expressions, and target of communication process. • Relation between mind and world: truth conditions. • Questions about processing meaning in real time. Logic, language and cognition • Formal approaches to meaning: role of logic, syntax-semantics interface, complexity of grammars, expressivity. • Embedding in cognitive system: what model of semantics suits the mind? Experimental approaches to semantic processing. • Applications in artificial systems: use of grammatical formalisms for computational applications, modelling meaning in intelligent systems. Academic skills acquired in LLC • Acquire knowledge, tools, methodology and approaches for doing linguistic research in the area of cognitive modelling and logic and intelligent systems. • Develop skills in reading primary texts in logic, language and cognition, and being able to extract the main points, discuss their contribution to the field, criticize weakness, suggest further research questions and alternative approaches. • Develop skills in formulating research questions, carrying out this research and reporting on results. Approach • Reading up-to-date literature exposing students to ‘hot’ questions in the area of logic, language and cognition. • Presentation and discussion of foundational questions defining the field, investigation of cutting edge tools and methodologies. • Homeworks developing expertise in tools and methodologies. • Balance between instructors’ activity and students’ activity in the classroom. • Support in developing academic skills and individual research. Overview of the course • http://www.let.uu.nl/~Henriette.deSwar t/personal/Classes/LLC10/index.html • General, Calendar, Exercices, Results • Requirements: present a paper (25%), do homeworks (25%), write a final paper (50%) on a topic related to the course (meet with one of the teachers for approval). • N.B. first paper presentation on Mo Nov 22 – distribute papers over students on We Nov 17. Compositionality of meaning as starting point of semantics • Core insight: the meaning of any complex expression is determined by the meanings of its parts, and the way they are put together. • Old insight: versions of compositionality have been around since BC. • Heuristic principle: semantic theories should respect principle of compositionality of meaning. • But: compositional analysis not always obvious. How exactly should we formulate the notion of compositionality? How do we deal with compositionality problems? Implications for architecture • Principle of compositionality has implications ranging from the lexicon via syntax-semantics interface to pragmatics and semantics processing. • Implications for theory of the lexicon: how to define the meaning of words in such a way that they can constitute the input to compositional rules? What kind of ontology do we need? Relation between linguistic knowledge and knowledge of the world. Implications for syntax-semantics interface • How does construction of syntactic trees relate to construction of complex meaning? Is there a one-one mapping? • Background: type theory, basic syntactic theory about (binary) trees, categorial grammar. • Different formalisms, balance between complicating syntax or complicating semantics. Implications for communication • Implications for semantics-pragmatics interface: what aspects of meaning are not immediately derived from lexicon+syntactic composition but come in through pragmatic reasoning? • What are the rules of conversation that come in at the level of communication between intelligent agents? • What is the position of pragmatics in linguistic theory? Implications for processing • Semantic processing by humans: how do we understand sentences in real time? • How is semantic processing affected by compositionality problems? • How is processing affected by pragmatic reasoning? • Processing by computers: how do we design intelligent systems that process languages as well as humans? • Which formal methods and technologies work best in natural language processing? Compositionality and implications • Approach taken in LLC: start with the principle of compositionality of meaning (today!). • Explore implications of compositionality of meaning for theories of the lexicon (We), syntax-semantics interface and pragmatics. • At each level, investigate consequences for semantic processing in experimental psychology (starting next week). • Also: consequences for natural language grammatical formalisms and computational applications (mostly second half of the course, focussing on syntax-semantics interface) Compositionality of meaning • Reading: Peter Pagin & Dag Westerståhl (2008). • Questions: How to formulate the principle of compositionality of meaning? What are the advantages/disadvantages of different versions? Why is it useful to assume that natural language is compositional for the relation between logic, language and cognition? Is it uncontroversial to assume compositionality? What problems arise if we do/don’t? Do problems imply we have to abandon the principle of compositionality of meaning? Learnability as an argument • A natural language has infinitely many meaningful sentences, expressing an infinite range of propositions. • It is impossible for a human speaker to learn the meaning of each sentence one by one. • It must be possible to learn the language by learning the meaning of a finite number of expressions, and a finite number of construction forms. • For this to be possible, the language must have a compositional semantics. Pagin & Westerståhl (2008: 15). Novelty as an argument for compositionality • Hearers are able to understand sentences they have never heard before. • Understanding novel sentences is possible only if the hearer can assign to new sentences the meaning that they independently have on the basis of their knowledge of individual expressions and rules of putting these meanings together. • So understanding novel sentences is possible only if language is compositional. Pagin and Westerståhl (2008: 16). Productivity and systematicity as arguments for compositionality • Generative syntax (Chomsky): creativity of language, production of infinite set of sentences. • Productivity argument relevant to compositionality of meaning in so far as new sentences need to be understood. (P&W 2008: 15-16) • Fodor: if a hearer understand a sentence of the form tRu, she will also understand the corresponding sentence uRT, and this is best explained by an appeal to compositionality. (P&W 2008: 16) Induction on synonymy • Fodor: if a hearer understand a sentence of the form tRu, she will also understand the corresponding sentence uRT. • Induction on synonymy: in case after case, we find the result of substitution synonymous with the original expression, if the new part is taken as synonymous with the old. • Inductive generalization: substitution by synonyms is meaning preserving. P&W (2008: 18) Intersubjectivity and communication • In most cases, speakers of the same language interpret new sentences similarly: intersubjective agreement in interpretation. • So the meaning of sentences is computable, not guessed. • Frege: the proposition has a structure that mirrors the structure of the sentence: isomorphism. • Without this structural correspondence, communicative success with new propositions would not be possible. Critical remarks • Arguments from learnability and novelty rely on creativity. Is it justified to assume an infinite set of propositions? • Arguments from production are more syntaxoriented than semantics-oriented. • Not all arguments require full-fledged compositionality. Often, some notion of computability is enough. And yet.. • Even if they are inference to the best explanation, compositional semantics is very simple. • Compositional semantics implies that a minimal number of processing steps are needed by the speaker to arrive at a full expression, and the hearer for arriving at a full interpretation, and there be communicative success. • If we assume compositionality as a viable starting point – how are we going to define it? Substitution and function • Frege: “Let us assume that the sentence has a reference. If we now replace one word of the sentences by another having the same reference, this can have no bearing upon the reference of the sentence.” Substitutional version, see P&W (2008: 2). • Compositionality of meaning requires the meaning (value) of a compound expression to be a function of other meanings (values) and a ‘mode of composition’. Functional version, see P&W (2008: 3) Structured expressions and grammar • Set E of linguistic expressions as a domain over which syntactic rules are defined. • E is generated by syntactic operations from a subset A of atoms (primitive expressions, words). • is a set of syntactic operations, defined as partical functions from En to E, such that E is generated from A via . • A grammar E is {E, A, }. • A grammatical term is the analysis tree associated with the expression (string). GTE is the set of grammatical terms for E. Semantics as mapping from terms to meanings • Semantics is function from GTE to a set of meanings M. • This means that the expression has meaning derivatively, relative to a way of constructing it, i.e. to a corresponding grammatical term. • Distinction between grammaticality (being derivable by the grammar rules) and meaningfulness (being in the domain of ). • No structure imposed upon M besides a relation of synonymy: u t iff (u) and (t) are both defined, and (u) = (t). Basic compositionality: functional version • Funct(): For every rule , there is a meaning operation r such that if (u1,…,un) has meaning, ((u1,…,un)) = r ((u1), …, (un)). • Domain Principle: subterms of meaningful terms are themselves meaningful. Basic compositionality: substitution version • Subst(): If s[u1,…,un] and s[t1,…,tn] are both meaningful terms, and if ui ti for 1 i n, then s[u1,…,un] s[t1,…,tn]. • The notation s[u1,…,un] indicates that the term s contains – not necessarily immediate – disjoint occurrences of subterms among u1,…,un and s[t1,…,tn] results from replacing each ui by ti. • DP is not presupposed for the substitution version. Functional and substitution version: equivalence • Under DP, Funct() and Subst() are equivalent. • Basic compositionality takes the meaning of a complex term to be determined by the meanings of the immediate sub-terms and the top-level syntactic operation. • Weaker versions of compositionality arise if we define compositionality at lower levels of syntactic structure. Stronger compositionality • Enlarging the domain of the semantic function. • Zabo (2000): same meaning operations define semantic functions in all possible human languages, not just for all sentences in each language taken by itself. • Cross-linguistic stability in semantic operations (‘universal’ semantics). Additional restrictions on meaningfulness. • Husserl property: Synonymous terms belong to the same semantic category. Synonymous terms can be intersubstituted without loss of meaningfulness. • Huss: u ~ t if, for every term s in E, s[u] dom() iff s[t] dom(). • Huss imposes more structure on the syntactic side. Additional restrictions on meaning composition operations • The function version is given by recursion over syntax, but only if the meanings are defined by recursion over meanings do we have a a recursive semantics. • Rec(): There is a function b and for every and operation r such that for every meaningful expression s, (s) = b(x) if s is atomic (s) = r((u1), .. (un), u1… un) if s = (u1… un). • Standard semantic theories are both compositional and recursive. Frege’s Context Principle • Frege’s Context Principle: The meaning of a term is the contribution it makes to the meanings of complex terms of which it is a part. • InvSubst(): If u is not t, then there is some term s such that either exactly one of s[u] and s[t] are meaningful, or both are, and s[u] is not s[t]. • That is, if two terms of the same semantic category make the same contribution to all complex terms, their meanings cannot be distinguished. Strengthening the Context Principle • Strengthening InvSubst to InvSubst, we can require that substitution of terms by terms with different meanings always changes meaning: • InvSubst(): If for some i, 0 i n, ni is not ti, then for every term s[u1,…un] it holds that either exactly one of s[u1,…un] and s[t1,…tn] is meaningful, or both are, and s[u1,…un] is not s[t1,…tn]. Implications for synonymy • InvSubst disallows synonymy between complex terms that can be transformed into each other by substitution of constituents at least some of which are non-synonymous, but it allows two terms with different structure to be synonymous. • How closely related should two terms be in syntactic structure in order to qualify as synonyms? Congruence • Two terms t and u are -congruent iff: • (i) r or u is atomic, t u, and neither is a constituent of the other, or • (ii) t = (t1,…tn), u = (u1,…un), ti and ui are -congruent for all i, 1 i n, and for all s1,…sn, (s1,…sn) (s1,…sn) if either is defined. • Cong: If t u, then t and u are congruent. Inverse Functional compositionality • Cong leads to inverse functional compositionality: • InvFunct(): The syntactic expression of a complex meaning m is determined, up to -congruence, by the composition of m and the syntactic expressions of its parts. Direct compositionality • Jacobson (2002). • In strong direct compositionality, expressions are built up from sub-expressions by means of concatenation (left or right). • In weak direct compositionality, one expression may wrap around another (as call up wraps around him in call him up). • In the weak version of direct compositionality, substrings are allowed to have discontinuous occurrences. Indirect compositionality • For indirect compositionality, syntactic operations are loosened up to include string deletion, string reordering, term substitution and insertion of elements. • Distinction between direct and indirect compositionality relevant for the definition of the syntax-semantics interface: what choice of syntactic theory is compatible with compositional semantics? • Important focus of second half course! Language signs • Abstract categorial grammar: signs. • Sign is a triple consisting of a string, a syntactic category and a meaning. • Grammar is defined in terms of partial functions from signs to signs. • Requires different way of organizing relation between grammatical terms and their meaning. • Abstract categorial grammar will be discussed in the second half of the course. Arguments against compositionality • How controversial is it to assume that compositionality holds? • Empirical arguments that certain constructions are non-compositional. Counterexamples falsify principle of compositionality of meaning for natural language. • Arguments that compositionality is trivial, not needed, or not suited to actual linguistic communication. Superfluity/unsuitability • Schiffer (1987): compositionality superfluous for mental processing. • But how to translate sentence representations to mental representations? • Recanati (2004): pragmatic compositionality, enrichment of meaning. Requires more thought about the semantics-pragmatics interface. • Radical contextualism, e.g. Sperber and Wilson (1992): evaluation in context underdetermined by literal meaning. Potential counterexamples to compositionality • Belief sentences: synonymy not preserved under verbs of mental attitude. • Quotation: brother and male sibling are synonyms, but ‘brother’ and ‘male sibling’ are not. • Idioms: kick the bucket means ‘to die’. • Ambiguities that are not clearly visible in lexical or structural differences, e.g. scope. Compositionality for logic and language • Principle of compositionality of meaning has shaped semantic theory. • Formal approaches favor use of type theory as a logical language that allows compositional analysis, cf. Schlenker (2008). • Primitive objects: Semantics usually defines interpretation in terms of a domain of individuals and a set of truth values. • Rules of interpretation: typically function application. Potential problems • What happens if we encounter natural language expressions that cannot be defined in terms of individuals (mass nouns, plurals). How to extend our ontology? What are the implications for construction rules? • What happens if type theory does not allow composition by function application? E.g. type conflicts – see coercion stuff! • Is a syntax-semantics interface using categories, type theory, and variables desirable, or should we explore different approaches – variable free semantics, ‘abstract’ categorial grammar. Theory-dependent compositionality problems • Compositionality problems that arise under particular assumptions of syntactic or semantic operations, e.g. negative concord. • Assume that our semantic composition rule is function application. • How to account for the contrast between: • Nobody talked to nobody [English: ] • Nessuno ha parlato con nessuno. [Ital: xy] • Solutions: different lexical entries (2nd nessuno is not negative), ambiguities (nessuno can mean either x or x), different semantic operations from function application (polyadic quantification). Processing compositional meaning • Do ‘repairs’ of compositionality problems (e.g. coercion in the case of type conflicts) have a reflection in real time semantic processing (e.g. slower understanding)? • How do we deal with ‘enriched’ meanings (presupposition, implicature): semanticspragmatics interface? What grammatical and cognitive mechanisms are involved in pragmatic reasoning? • How can processing evidence bear on semantic theory? Back to class overview • Organization of topics: relevance of compositionality for study of logic, language and cognition. • First half of course: start with the role of the lexicon in compositionality problems. Build up grammar, semantics-pragmatics interface. Provide processing evidence to test linguistic theory. • Second half of the course: Formal modelling of compositionality problems, scope, dependency relations. Balance between complicating syntax/complicating semantics.