|Title: ||Generating the Conserving Effect without Language Acquisition. An agent-based model|
|Authors: ||Pijpops, Dirk ×|
Beuls, Katrien #
|Issue Date: ||8-Jun-2015 |
|Conference: ||Poster presentation at the BKL-CLK Thematic Conference Computational Construction Grammar and Constructional Change location:Brussels date:07-09 June 2015|
|Abstract: ||The Conserving Effect, one of three well-known effects of frequency, states that frequently used expressions are resistant to language change (Bybee and Thompson 1997: 380–381; Bybee 2006: 715). While this effect has been demonstrated in several corpus studies (Krug 2000; Corbett et al. 2001; Lieberman et al. 2007; Carroll, Svare and Salmons 2012; Cuskley et al. 2014), it is harder to disentangle the influence of its possible causes, using solely corpus techniques (cf. Landsbergen et al. 2010: 367–368). Several authors hold that its cause should be sought in language acquisition: highly frequent exceptions are easier to learn, and are thus better capable of withstanding regularization pressure (Lieberman et al. 2007: 713; Cuskley et al. 2014: 4). This has been shown to hold true in computational models of language learning in connectionism and the Principles & Parameters framework (resp. Hare and Elman 1995; Yang 2002). However, the claim is somewhat at odds with the usage-based outlook on language change, which views language usage, not acquisition, to be the primary driving force behind change (Croft 2000; Bybee 2010; Petré and Van de Velde 2014). Moreover, Bybee herself, as well as Diessel, only mention usage as a causal factor (Bybee 2006: 715; Diessel 2007: 118).
As such, we investigated the minimal requirements for generating a Conserving Effect through agent-based modelling. The employed case study is the past tense inflection of Dutch (Salverda 2006; Knooihuizen and Strik 2014). Our model only assumes that for transparent language forms, not only the individual forms are stored, but also a more abstract construction, with both representations being susceptible to the influence of frequency (cf. Bybee 2006: 713–714). The model does not incorporate a learning component: all agents start with knowledge of all possible verb forms, and no agents are replaced during simulation. Evaluation is done by the following criteria. First, the verbs should show sustained co-existence at the type-level, i.e. both the regular and irregular inflection as a whole should be able to continuously co-exist, with some verbs being irregular and others regular. Second, at the token level, fixation should take place. This means that the regular and irregular form of a particular verb should not typically co-exist (Cuskley et al. 2014: 3). Third, the deciding factor whether a verb should remain irregular or regularize, should be its frequency. High-frequency verbs should become strongly entrenched in their irregular forms, while low-frequency verbs should regularize. The proposed model complies with these criteria.
We do not mean to claim that language acquisition is not instrumental in causing the Conserving Effect. What we do hope to show however, is that – in the spirit of Landsbergen et al. (2010: 386) – both acquisition and usage are independently capable of generating the Conserving Effect.
|Publication status: ||published|
|KU Leuven publication type: ||IMa|
|Appears in Collections:||Quantitative Lexicology and Variational Linguistics (QLVL), Leuven|