Informatica

In this paper, a new approach has been proposed for multi-label text data class verification and adjustment. The approach helps to make semi-automated revisions of class assignments to improve the quality of the data. The data quality significantly influences the accuracy of the created models, for example, in classification tasks. It can also be useful for other data analysis tasks. The proposed approach is based on the combination of the usage of the text similarity measure and two methods: latent semantic analysis and self-organizing map. First, the text data must be pre-processed by selecting various filters to clean the data from unnecessary and irrelevant information. Latent semantic analysis has been selected to reduce the vectors dimensionality of the obtained vectors that correspond to each text from the analysed data. The cosine similarity distance has been used to determine which of the multi-label text data class should be changed or adjusted. The self-organizing map has been selected as the key method to detect similarity between text data and make decisions for a new class assignment. The experimental investigation has been performed using the newly collected multi-label text data. Financial news data in the Lithuanian language have been collected from four public websites and classified by experts into ten classes manually. Various parameters of the methods have been analysed, and the influence on the final results has been estimated. The final results are validated by experts. The research proved that the proposed approach could be helpful to verify and adjust multi-label text data classes. 82% of the correct assignments are obtained when the data dimensionality is reduced to 40 using the latent semantic analysis, and the self-organizing map size is reduced from 40 to 5 by step 5.

This paper presents the corpus-driven approach in building the computational model of fundamental frequency, or ${\mathit{F}}_0$, for Lithuanian language. The model was obtained by training the HMM-based speech synthesis system HTS on six hours of speech coming from multiple speakers. Several gender specific models, using different parameters and different contextual factors, were investigated. The models were evaluated by synthesizing ${\mathit{F}}_0$ contours and by comparing them to the original ${\mathit{F}}_0$ contours using criteria of root mean square error (RMSE) and voicing classification error. The HMM-based models showed an improvement of the RMSE over the mean-based model that predicted ${\mathit{F}}_0$ of the vowel on the basis of its average normalized pitch.