On constituent chunking for Turkish

Abstract

Chunking is a task which divides a sentence into non-recursive structures. The primary aim is to specify chunk boundaries and classes. Although chunking generally refers to simple chunks, it is possible to customize the concept. A simple chunk is a small structure, such as a noun phrase, while constituent chunk is a structure that functions as a single unit in a sentence, such as a subject. For an agglutinative language with a rich morphology, constituent chunking is a significant problem in comparison to simple chunking. Most of Turkish studies on this issue use the IOB tagging schema to mark the boundaries. In this study, we proposed a new simpler tagging schema, namely OE, in constituent chunking for Turkish. "E" represents the rightmost token of a chunk, while "O" stands for all other items. In reference to OE, we also used a schema called OB, where "B" represents the leftmost token of a chunk. We aimed to identify both chunk boundaries and chunk classes using the conditional random fields (CRF) method. The initial motivation was to employ the fact that Turkish phrases are head-final for chunking. In this context, we assumed that marking the end of a chunk (OE) would be more advantageous than marking the beginning of a chunk (013). In support of the assumption, the test results reveal that OB has the worst performance and OE is significantly a more successful schema in many cases. Especially in long sentences, this contrast is more obvious. Indeed, using OE means simply marking the head of the phrase (chunk). Since the head and the distinctive label "E" are aligned, CRF finds the chunk class more easily by using the information contained in the head. OE also produced more successful results than the schemas available in the literature. In addition to comparing tagging schemas, we performed four analyses. Along with the examination of window size, which is a parameter of CRF, it is adequate to select and accept this value as 3. A comparison of the evaluation measures for chunking revealed that F-score was a more balanced measure in contrast to token accuracy and sentence accuracy. As a result of the feature analysis, syntactic features improves chunking performance significantly under all conditions. Yet when withdrawing these features, a pronounced difference between OB and OE is forthcoming. In addition, flexibility analysis shows that OE is more successful in different data.