文档下载 文档下载导航
设为首页 | 加入收藏
搜索 请输入内容:  
 导航当前位置: 文档下载 > 所有分类 > A COMPARISON OF NATURAL AND MACHINE LEARNING OF STRESS
侵权投诉

A COMPARISON OF NATURAL AND MACHINE LEARNING OF STRESS

In the language acquisition literature children are reported to make typical errors in assigning stress to words. These errors are taken to be unequivocal evidence that children learn rules for stress assignment as opposed to learning on a word-by-word bas

assignment, i.e., the kind of model discussed by Dresher and Kaye.

The second argument relates to the inadequacy of word-by-word learning to account fordevelopmental phenomena. For instance, Fikkert (1993) states in the introduction of herstudy of the acquisition of main stress assignment in Dutch: “We will show that childrendo learn stress rules, because stress acquisition shows a regular pattern.” (Fikkert, 1993,p. 21) Thus, in developmental studies it is typically argued that because children'sacquisition follows a particular path, with typical errors characterizing various stages, thisacquisition process can only be accounted for in a rule-based model, and, hence, word-by-word learning is designated as empirically inadequate in this respect. The errorsreferred to include the observation that children find words with regular stress patternseasier to pronounce, i.e., produce them with less errors than words with irregular orprohibited stress patterns. It is argued that if children would simply memorize the stresspattern of words, they would find all words equally easy or equally hard to produce,irrespective of their stress pattern (Hochberg, 1988b; Nouveau, 1993, 1994). Moreoverchildren typically regularize the stress pattern of non-regular words, and they show atendency to irregularize the stress pattern of regular words. Also these types of errorscannot be explained in a model that memorizes words and uses a simple table-look-upstrategy.

In this paper we will argue that although a word-by-word acquisition strategy coupledwith a table-look-up procedure is obviously incapable of accounting for the relevantdevelopmental phenomena and lacks the necessary ability to generalize over the wordsencountered, we should not conclude that a rule-based system is called for. Morespecifically we will show that an exemplar-based approach, viz. lexical lookup extendedwith a learning mechanism working by analogical reasoning on the basis of storedexamples, does not suffer from the shortcomings of table-lookup. We will showexperimentally that the types of errors typically produced by children are also displayedby an exemplar-based model, and, hence, such a model represents a valid alternative for arule-based model. Moreover we will show that one type of error, viz. irregularization,which is also attested in children's language cannot be accounted in terms of a rule-basedmodel, while it receives an elegant explanation in a exemplar based model.

In previous research (Daelemans, Gillis & Durieux, 1994; Gillis, Daelemans, Durieux &Van den Bosch, 1993) an exemplar-based model incorporating 'lazy learning' was set tothe task of learning main stress assignment. We showed that the system was able to learn

上一页第5页

你可能喜欢

热门文档

相关文档

站点地图 | 文档上传 | 侵权投诉 | 手机版
本站所有资源均来自互联网,本站只负责收集和整理,均不承担任何法律责任,如有侵权等其它行为请联系我们.
文档下载 Copyright 2013 doc.guandang.net All Rights Reserved.  -  email