Extracting prototypes from exemplars What can corpus data tell us about concept representation?

Abstract

Over the past four decades, two distinct alternatives have emerged to rule-based models of how linguistic categories are stored and represented as cognitive structures, namely the prototype and exemplar theories. Although these models were initially thought to be mutually exclusive, shifts from one mechanism to the other have been observed in category learning experiments, bringing the models closer together. In this paper we implement a technique akin to varying abstraction modelling, that assumes intermediate abstraction processes to underlie category representations and categorization decisions; we do so using familiar statistical techniques such as regression and clustering that track frequency distributions in input. With this model we simulate, on the basis of actual usage of Russian try verbs and Finnish think verbs as observed in corpora, how prototypes for near-synonymous verbs could be formed from concrete exemplars at different levels of abstraction.In so doing, we take a closer look at the cognitive linguistic flirtation with multiple categorization theories, suggesting three improvements anchored in the fact that cognitive linguistics is a usage-based theory of language. Firstly, we show that language provides support for considering single prototype and full exemplar models as opposite ends along a continuum of abstraction. Secondly, we present a methodology that simulates how prototypes can be obtained from exemplars at more than one level of abstraction in a systematic and verifiable way. And thirdly, we illustrate our claims on the basis of work on verbs, denoting intangible events that are neither stable in nor independent of time and express relational concepts; this implies that verbs are more susceptible to their meanings being influenced by the concepts they relate.

Machine Meets Man: Evaluating the Psychological Reality of Corpus-based Probabilistic Models

“…There were a total of 14 multiple-category variables amounting to 87 distinct variable categories or contextual properties. Divjak and Arppe (2013) used this dataset to train a polytomous logistic regression model (Arppe 2013a(Arppe , 2013b predicting the choice of verb. As a rule of thumb, the number of distinct variable combinations that allow for a reliable fitting of a (polytomous) logistic regression model should not exceed 1/10 of the least frequent outcome (Arppe 2008: 116).…”

The cognitive plausibility of statistical classification models: Comparing textual and behavioral evidence

“…The studies that meet the four requirements described above are Bresnan (2007;also Bresnan et al 2007, Bresnan & Ford 2010; for an overview paper of their work see Ford & Bresnan 2013a), Klavan (2014), Divjak (2010;Divjak et al 2016; also Divjak & Gries 2006;Divjak & Gries 2008;Divjak & Arppe 2013) and Arppe & Abdulrahim (2013;also Arppe 2008). While the studies by Bresnan and collaborators and Klavan are concerned with syntactic alternations, the studies by Divjak and collaborators and Arppe & Abdulrahim analyze lexical variation.…”

“…The four case studies we concentrate on are the English dative alternation (Bresnan 2007, Bresnan & Ford 2010, Ford & Bresnan 2013b, the alternation between the adessive case suffix and the postposition peal in Estonian (Klavan 2012(Klavan , 2014, 6 Russian verbs denoting the concept of try (Divjak 2003(Divjak , 2004(Divjak , 2010Divjak & Arppe 2013;Divjak et al 2016) and 4 near-synonymous verbs meaning come in Modern Standard Arabic (Abdulrahim 2013, Arppe & Abdulrahim 2013. Table 1 is an overview table of the corpus data used in the four studies; it specifies which corpora were used and whether the data comes from written texts (the Estonian and Russian studies), spoken language (the English study), or both (the Arabic study).…”

“…Building on earlier work by Divjak (2003Divjak ( , 2004Divjak ( , 2010, Divjak & Arppe (2013) used the dataset of 1,351 tokens to train a polytomous logistic regression model using the one-versus-all heuristic (Arppe 2013a/b) to predict the choice between 6 Russian verbs that, when combined with an infinitive, can all be translated with the English verb try. At the model building stage, a number of selection criteria were followed: the number of variable categories in the model should not exceed 1/10 of the least frequent outcome 6 , only variables with a broad dispersion among the six verbs were retained (the overall frequency of the variable in the data was to be at least 45 and to occur at least twice with each verb), one variable was excluded for each fully mutually complementary case as were variables with a mutual pair-wise uncertainty coefficient (UC) value larger than 0.5 (Divjak & Arppe 2013: 238).…”

“…It is a method widely applied in semantics (e.g. Arppe & Järvikivi 2007, Arppe 2008, Divjak 2010, Divjak & Arppe 2013, syntax (e.g. Gries 2003, Bresnan 2007, Bresnan & Ford 2010, Kendall et al 2011, Klavan 2012, morphology (e.g.…”