# analogical_generalization_of_linguistic_constructions__0a76d613.pdf

Analogical Generalization of Linguistic Constructions

Clifton Mc Fate

Qualitative Reasoning Group, Northwestern University c-mcfate@northwestern.edu Abstract Human language is extraordinarily creative in form and function, and adapting to this ever-shifting linguistic landscape is a daunting task for interactive cognitive systems. Recently, construction grammar has emerged as a linguistic theory for representing these complex and often idiomatic linguistic forms. Furthermore, analogical generalization has been proposed as a learning mechanism for extracting linguistic constructions from input. I propose an account that uses a computational model of analogy to learn and generalize argument structure constructions.

Introduction Humans easily interpret novel uses of known linguistic items. Consider caused motion constructions or denominal verbs (Goldberg, 2003; Kaschak & Glenberg, 2000). 1 a) Pat sliced the carrots into the salad. b) Lyn crutched Tom her apple so he wouldn t starve. In 1a, we easily interpret the sentence such that slicing caused the carrots to fall into the salad, though one would hardly want to create a new sense of the word slice for this usage. Similarly, in 1b, the usage of crutch conveys that Lyn used a crutch to pass Tom the apple. Certainly this isn t a pre-defined semantics for crutch. Instead, construction grammar views the fundamental motion and transfer semantics of these sentences as properties of the grammatical forms, namely the transitive and dative constructions. Constructions pair form and meaning. This approach is also useful in explaining more idiomatic constructions such as the comparativecoorelative ( The X-er, The Y-er) as in The bigger they are, the harder they fall (Cullicover & Jackendoff, 1999). In both examples, the structure of the sentence provides extra semantic meaning beyond that provided by the head verb. The semantics is enriched by the verb, not derived from it. This is one way construction grammar differs from traditional approaches. Additionally, construction grammar does not posit underlying grammatical transformations. Finally, a core principle of construction grammar is that constructions are learnable from input. Indeed, Casenhiser

& Goldberg (2005) demonstrate that children readily learn the semantics of novel argument structure constructions given few examples, and constructions have been invoked to explain a wide array of developmentary linguistic phenomena (Tomasello, 2003). One proposal for how humans learn these constructions is through analogical generalization.

Analogical Generalization Our approach to analogical generalization is based on Gentner s (1983) structure mapping theory in which human similarity judgments are viewed as a process of alignment between hierarchical structured representations. Falkenhainer et al s (1989) structure mapping engine (SME) is a computational implementation of structure mapping. SME takes a base and target case of structured representations and maps entities and relations. It prefers matches with a shared higher-order structure, so cases with high structural similarity but low feature similarity match. Relationships that are present in the base but not in the target can be hypothesized as candidate inferences. SME provides the matching algorithm for the Sequential Analogical Generalization Engine (SAGE), a computational model of human analogical generalization that has been used in tasks such as learning concepts from maps (Mc Lure et al, 2015). Given a new example and a case-library, SAGE compares it to existing examples using SME. If over threshold, the new example is added to an existing case to create a generalization with a probability distribution governing the features of mapped entities. Over time, SAGE produces schema-like constructs that still retains high-probability attributes and relationships.

Proposed Work and Progress

My thesis has two goals: (1) support the claim that analogical generalization is a driving mechanism in construction learning. (2) Demonstrate that analogically learned constructions can improve language understanding. Copyright 2016, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved.

Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence (AAAI-16)

In pursuing the first goal, I plan to model several existing human studies. Casenhiser & Goldberg (2005) presented children with nonce verbs in a novel construction (S-O-V) which described videos of appearance scenarios. They found that children quickly applied the novel construction to new appearance scenarios and that this was facilitated when examples frequently shared a verb. This parallels a finding in the analogical generalization literature called progressive alignment, where highly similar cases pave the way for more difficult alignments, which SAGE models. The same phenomena explains the results of Childers and Tomasello (2001) who found that frequent pronoun usage in child-directed speech facilitated their production of a transitive construction for a nonce verb. I aim to replicate these studies with my model. I will use SAGE to generalize over the syntactic surface forms of the utterances. For Casenhiser & Goldberg (2005), the semantics of the video will be automatically encoded through a series of drawings in the sketching software Cog Sketch (Forbus et al, 2011). Modeling these studies provides a framework for interactive construction learning. Secondly, replicating the progressive alignment effects supports the theory that analogy is used in human construction learning. Furthermore, I plan to show that construction generalization can improve the linguistic competence of cognitive systems at scale. To do so, I will use SAGE to generalize semantic annotations from Fillmore et al s (2001) Frame Net. I will use the resulting representations to augment our NLU system. To assess performance, I plan to use Kaschak & Glenberg s (2000) denominal verb stimuli. Using these representations, it should produce a transfer interpretation of a sentence like 1b even though no such semantic frame exists for the verb crutch. This provides needed flexibility for an interactive cognitive system.

Prior Work and Progress by 2/16 In Mc Fate (2015) I used SAGE to learn pragmatic constraints on preposing constructions. However, the syntax was limited and the representation collapsed syntactic information. I ve also started generalizing constructions from Frame Net. Initial generalizations were made based on Frame Net annotations using a representation based on the ordering of phrase-level nodes. My thesis makes use of several existing resources. Cog Sketch and SAGE are both existing systems, as is the EA NLU language understanding system (Tomai & Forbus, 2009). Frame Net provides a large corpus of semantic annotations. EA NLU uses the Research Cyc1 ontology for semantic representations.

1 http://www.cyc.com/platform/researchcyc/

By the date of the consortium I ll have a proof of concept version of our language system running which uses retrieval over Frame Net generalizations to guide interpretation. By this point, I also hope to have much of the framework in place for associating a linguistic form with a series of sketched events in Cog Sketch.

Future Work & Technical Challenges

Much remains to be done. One crucial element is deciding on formalisms for representing linguistic input. This will be influenced by ongoing research into developmental language studies. Additional challenges come from incorporating generalizations into an existing parsing pipeline, as well as integrating Cog Sketch into my existing framework. Finally, evaluating my model of Childers and Tomasello (2001) ideally involves a language generation mechanism which is still in development.

Casenhiser, D., & Goldberg, A. E. (2005). Fast mapping between a phrasal form and meaning. Developmental Science, 8(6), 500-508.

Childers, J., & Tomasello, M. (2001). The role of pronouns in young children s acquisition of the English transitive construction. Developmental Psychology, 37, 730 748.

Culicover, P.W. & Jackendoff, R. (1999) The view from the periphery: the English comparative correlative. Linguist. Inq. 30, 543 571

Fillmore, C. J., Wooters, C., & Baker, C. F. (2001). Building a Large Lexical Databank Which Provides Deep Semantics. Proceedings of PACLIC-2001, Hong Kong, China.

Forbus, K., Usher, J., Lovett, A., Lockwood, K. & Wetzel, J. (2011). Cog Sketch: Sketch understanding for cognitive science research and for education. Topics in Cognitive Science, 3(4), 648-666.

Gentner, D. (1983). StructureǦMapping: A Theoretical Framework for Analogy. Cognitive Science, 7(2), 155-170

Goldberg, A. E. (2003). Constructions: a new theoretical approach to language. Trends in cognitive sciences, 7(5), 219-224.

Kaschak, M. P., & Glenberg, A. M. (2000). Constructing Meaning: The Role of Affordances and Grammatical Constructions in Sentence Comprehension. Journal of Memory and Language, 43(3), 508 529.

Mc Fate, C.J. (2015). An Analogical Approach to Learning Constraints on Non-canonical Constructions. Advances in Cognitive Systems Student Research Workshop, Atlanta, GA.

Mc Lure, M.D., Friedman S.E. and Forbus, K.D. (2015). Extending Analogical Generalization with Near-Misses. Proceedings of the Twenty Ninth AAAI Conference on Artificial Intelligence, Austin, Texas

Tomai, E. and Forbus, K. (2009). EA NLU: Practical Language Understanding for Cognitive Modeling. Proceedings of the 22nd International Florida Artificial Intelligence Research Society Conference. Sanibel Island, Florida.

Tomasello, M. (2003). Constructing a language: a usage-based theory of language acquisition. Cambridge, Mass.: Harvard University Press.