A Phonological Encoding of Turkish for Neural Networks

The paper proposes a multi-dimensional, phonologically-aware numeric encoding of Turkish for use with neural networks. The system is evaluated and compared to PatPho (Li/MacWhinney 2002) in a test in which the network computes the shape of the past tense suffix. 0. Introduction A number of attempts have been made to convert phonemes to numbers in a way that adequately renders the distances between them. See Li and MacWhinney (2002: 408f) for an overview. In the same paper, a new encoding is proposed, but it is not tested. At least for Turkish, it would seem, its effectiveness is limited. The present paper proposes a different system (section 1.) and evaluates it using the Turkish past tense suffix, achieving a considerably higher accuracy (section 2.). Moreover, the proposed encoding is easier to port to different languages beacuse the method for chosing the specific numeric values is less arbitrary and, even more importantly, it is overt. Finally, the preliminary results of an actual application of the encoding are presented (section 3.). Turkish is particularly well suited for testing an encoding of this kind for two reasons. Firstly, its phonology is relatively simple and symmetric, and so it serves well as a model which can be viewed as a minimal example, appropriate for an initial presentation. Secondly, Turkish has vowel harmony and consonant assimilations on morpheme boundaries, both of which are regular and entirely dependent on phonetics alone. 364 KAMIL STACHOWSKI All the examples are presented phonologically in the Finno-Ugric transcription; see Stachowski K. (2011) for details. Unless specified otherwise, Turkish verbs are traditionally given in the infinitive, i.e. with the -ma ek suffix attached and separated with a dot.