The development and evaluation of an automatic clitic generator for Pashto language

Clitics are unstressed and unaccented words or particles that are phonetically dependent on adjacent words, accented in nature. They are available in many languages around the world including the Pashto language, which is spoken in Pakistan and Afghanistan. The native Pashto speakers use clitics extensively in their everyday conversation and writings. There are two basic types of clitics in Pashto called: Second Position (2P) clitics and Endoclitics. 2P clitics fall into three different categories: Proclitics, Enclitics (Modal) and Adverbial. Proclitics and enclitics are further grouped into context-free and context-dependent clitics. Endoclitics among all the clitics are the most challenging type of clitics, as their generation faces many restrictions. Clitics play a vital role in text generation systems. In general, these systems must be understandable, coherent, and accurate. In addition to this, Pashto Language is a low resource language lacking corpus, parser, tagger, and semantic analyzer. It also lacks syntactic-morphology interaction. Furthermore, there is no automatic clitic generation tool. These challenges make the generation of Pashto cliticised sentences a more challenging task. To overcome these challenges, the linguistic behavior of Pashto clitics is studied and formalized into rules to support the automatic generation of cliticised text in this paper. In particular, it used nine different clitic generation procedures and produced 80 clitic generation rules. The proposed clitic generation system is developed in Python, which generates cliticised sentences from the semantic representation of the sentences. To evaluate the efficiency of the developed system, a corpus of 256 syntactically annotated sentences was developed and used. It used syntactic pattern matching rules for the identification and generation of clitics at the sentence level. The sentences produced were checked against the correct responses to declare them correct or incorrect. The proposed system successfully developed both types of clitics with an overall accuracy of 89.72%. In particular, it produced Proclitics and Enclitics with an accuracy of 91.75%. However, its efficiency for modal clitics (87.95%) reduced the overall efficiency of 2P clitics to 89.85%. The precision of the endoclitic system developed was 89.47%.