Language-Independent Approach for Morphological Disambiguation

Proceedings of the 29th International Conference on Computational Linguistics, pages 5288–5297 October 12–17, 2022. 5288 Language-Independent Approach for Morphological Disambiguation Alymzhan Toleu and Gulmira Tolegen and Rustam Mussabayev Institute of Information and Computational Technologies, 050010, 28 Shevchenko str., Almaty, Kazakhstan alymzhan.toleu@gmail.com and gulmira.tolegen.cs@gmail.com and rmusab@gmail.com Abstract This paper presents a language-independent approach for morphological disambiguation which has been regarded as extensions of POS tagging, jointly predicting complex morpho- logical tags. In the proposed approach, all words, roots, POS and morpheme tags are em- bedded into vectors, and contexts representa- tions from surface word and morphological con- texts are calculated. Then the inner products between analyses and the context’s representa- tions are computed to perform the disambigua- tion. The underlying hypothesis is that the cor- rect morphological analysis should be closer to the context in a vector space. Experimental results show that the proposed approach out- performs the existing models on seven differ- ent language datasets. Concretely, compared with the baselines of MarMot and a sophisti- cated neural model (Seq2Seq), the proposed approach achieves around 6% improvement in average accuracy for all languages while run- ning about 6 and 33 times faster than MarMot and Seq2Seq, respectively. 1 Introduction Morphological disambiguation (MD) is the task of jointly predicting lemma/root, part of speech (POS)(Toleu et al., 2020), and morpheme tags. For a Turkish word “yeni" (new), it can be analyzed as: 1)yen+Noun+[A3sg, Pnon, Acc]; 2)yen+Noun+[A3sg, P3sg, Nom]. If one counts analyses as tags, MD can be cast as a tagging prob- lem with an extremely large tagset. This fact dis- courages direct application of the state of the art approaches designed for small fixed tagsets. For instance, many approaches treat each anal- ysis as a tag, and apply sequence labeling mod- els to perform tagging (Mueller et al., 2013; Müller and Schütze, 2015; Malaviya et al., 2018). Treating each analysis as a tag leads to an over- sized tagset and corresponding data-sparsity is- sues, which can be a concern for morphologically complex languages such as Turkish and Kazakh, where the number of morphological analyses is theoretically unlimited (Yuret and Türe, 2006). To address this problem, a sequence to sequence (Seq2Seq)(Tkachenko and Sirts, 2018) based ap- proach was proposed, which treated each mor- phological analysis as a sequence of a composite tags and explicitly modeled their internal structure. This approach was inspired by the neural sequence- to-sequence models for machine translation (Cho et al., 2014). The LSTM networks were applied to model morphological analyses and context as a pair of sequences, which involves more sophisti- cated architectures, namely using double layers of biLSTM, one for characters and another for words. This approach was almost challenging to simplify the architecture if one wanted to keep the perfor- mance as the original Seq2Seq has. Because all types of architecture of recurrent neural networks are more fit to the nature of sequence to sequence problems (Sutskever et al., 2014). This paper presents a language-independent MD approach that applies an uncomplicated neural ar- chitecture and obtains comparable …

Content truncated to 500 words