Dependency parsing with stacked conditional random fields for Turkish

Creative Commons License

Bilgin M., Amasyalı M. F.

JOURNAL OF THE FACULTY OF ENGINEERING AND ARCHITECTURE OF GAZI UNIVERSITY, vol.32, no.2, pp.385-392, 2017 (SCI-Expanded) identifier identifier


In the most general form Sequence Labelling is the production of an output sequence in response to an input sequence. Many of natural language processing problems such as (entity name recognition, machine translation, morphological analysis, separation of the elements of sentence etc.) can be defined as a sequence labelling. Dependency parsing is to determine the relationship and the type of the relationship between words within a sentence and it is essential to perform semantic analysis of a sentence. When dependency parsing is defined as a sequence labelling problem, production of two outputs (relationship type, related words) is required. Our recommendation is to use the Conditional Random Fields (CRF) which is commonly used in sequence labelling problems. However CRF is a method that produces a single output. To overcome this difficulty we propose to divide Dependency Parsing which is a problem with two outputs into two parts. The overall solution is provided by combining the results of these parts. With the performed operation we reached the best dependency parsing results for Turkish language.