Построение нейросетевых моделей морфологического и морфемного анализа текста

Abstract

Listen

Translate article

Morphological analysis of text is one of the most important stages of natural language processing (NLP). Traditional and well-studied problems of morphological analysis include normalization (lemmatization) of a given word form, recognition of its morphological characteristics and their morphological disambiguation. The morphological analysis also involves the problem of morpheme segmentation of words (i.e., segmentation of words into constituent morphs and their classification), which is actual in some NLP applications. In recent years, several machine learning models have been developed, which increase the accuracy of traditional morphological analysis and morpheme segmentation, but performance of such models is insufficient for many applied problems. For morpheme segmentation, high-precision models have been built only for lemmas (normalized word forms). This paper describes two new high-accuracy neural network models that implement morphemic segmentation of Russian word forms with sufficiently high performance. The first model is based on convolutional neural networks and shows the state-of-the-art quality of morphemic segmentation for Russian word forms. The second model, besides morpheme segmentation of a word form, preliminarily refines its morphological characteristics, thereby performing their disambiguation. The performance of this joined morphological model is the best among the considered morpheme segmentation models, with comparable accuracy of segmentation.