Persian Stemming data-set in order to evaluate new stemmers

There is no standard dataset for correctness evaluation of Persian stemming algorithms. In order to create a dataset for correctness evaluation of stemmers, we require a set of words with their stems. These datasets are automatically extracted from two manually stemmed corpora. The first dataset contains a collection of words and their stems, which has been extracted from the PerTreeBank corpus [1]. This corpus contains 4,689 distinct words. Moreover, in order to perform a better evaluation, we selected a large text corpus for the second dataset. The words and their stems of this dataset have been extracted from the Persian Dependency TreeBank corpus [2]. It contains 26,913 distinct words. These two datasets have good qualities in terms of the diversity of their Part-of-Speech tags.

You can use the evaluate.exe in order to evalute your stemming results. It generates report based on your result. It supports all the metrics of stemming evalution such as Accuracy, Percision, Recall, F-Measure, Understemming and Overstemming Errors, Commission and Ommission Errors.

Each stemming dataset is consist of three columns. The first column is the inflected word, the second is its stem and the third is its part-of-speech. You must append your stems to the fourth column. Then you can use below command.

Evaluate.exe "{your stemmed file path}" 1 3 {evaluation output file name}

[1] Ghayoomi, M. (2012) Bootstrapping the Development of an HPSG-based Treebank for Persian. Linguistic Issues in Language Technology, 7.

[2] Rasooli, M. S., Moloodi, A., Kouhestani, M., and Minaei-Bidgoli, B. (2011) A syntactic valency lexicon for Persian verbs : The first steps towards Persian dependency treebank. 5th Language & Technology Conference (LTC) : Human Language Technologies as a Challenge for Computer Science and Linguistics, pp. 227–231.