Evaluating the Impact of Morphosyntactic Ambiguity in Grammatical Error Detection

International Conference RANLP 2009 - Borovets, Bulgaria, pages 155–160 Evaluating the Impact of Morphosyntactic Ambiguity in Grammatical Error Detection Arantza D´ıaz de Ilarraza, Koldo Gojenola and Maite Oronoz Department of Computer Languages and Systems University of the Basque Country {jipdisaa, koldo.gojenola, maite.oronoz}@ehu.es Abstract We present a study of the impact of morpho- logical and syntactic ambiguity in the process of grammatical error detection. We will present three different systems that have been devised with the objective of detecting grammatical er- rors in Basque and will examine the influence of ambiguity in their results. We infer that the am- biguity rate in the input to an error detection tool can have a considerable impact on the qual- ity of the system. Keywords Grammatical error detection, morphosyntactic ambiguity 1 Introduction The relationship between ambiguity and error detec- tion has been mentioned in very few occasions [3, 13]. Similarly to most NLP areas, the development of tools for grammatical error detection and correction finds ambiguity as a main obstacle for the design of effi- cient and accurate systems. Typically the errors accu- mulated through the linguistic analysis make difficult the process of detecting grammatical errors. Birn [3] states the following relation between ambiguity, lin- guistic analysis and grammar error detection. “The relationship between disambigua- tion and grammar error detection is intricate. On the one hand, . . . disambiguation is a pre- requisite for any effort at precise error detec- tion. On the other hand, a grammar error may disturb the disambiguation, . . . and this in turn may disturb the error detection” In this paper we will study this statement over three systems designed for the detection of errors ranging from a restricted and limited context (errors in date expressions and complex postpositions) to the more general case of agreement errors between verb and sen- tence elements. We will concentrate on morphological and syntactic ambiguity: • Morphological ambiguity: each word-form can receive multiple morphological analyses, e.g. noun/verb is a typical example of categorial am- biguity. For agglutinative languages there are ad- ditional sources of ambiguity (number, case, ...). This poses a problem for grammatical error de- tection/correction. Level Linguistic features Method M1 POS CG + HMM M2 POS + SubPOS CG + HMM M3 POS + SubPOS + Case CG + HMM M4 All in morfeus CG Table 1: Disambiguation levels in eustagger. • Syntactic ambiguity: this is typically added on top of morphological ambiguity. For example, it is important to exactly know whether an NP is the subject or the object of a verb in order to detect agreement errors. The remainder of this paper is organized as follows. Section 2 comments on the general morphosyntactic analyzer for Basque, and its parameterization to in- vestigate the impact of ambiguity in error detection. Section 3 will present the experiments with different degrees of ambiguity in three settings. Section 4 com- pares our work with related systems, ending with the main conclusions. 2 Linguistic resources and pa- rameterization of ambiguity For the …

Content truncated to 500 words