Treffer: Hybrid model for sentiment analysis based on both text and audio data trained on MELD.

Among the tasks of natural language processing, at present, undoubtedly, the task of analyzing the tonality of data, namely the task of text processing, is an urgent task. In our daily activities, we encounter the most commonly used sentiment analysis applications that allow developers to improve technological solutions and products. For example, by identifying the emotional coloring of a dialogue, quickly determining the nature of a film review, or recognizing the rating of a review about a product, the developed recommendation system helps the user find the most appropriate solution. In present paper, we consider a mood analysis model for both text messages and audio recordings, which classifies an array of input data into two categories of emotions (positive and negative). More than 6000 data characteristics were extracted to execute the model, the source of which was various audio data from the Python openSMILE library. Using the technique of correlation between all the detected properties of audio tracks, it was possible to identify the most sensitive properties most accurately determined by the emotional coloring of audio messages. In addition, since the best sound characteristics are selected, the correlation threshold value was assumed to be 0.9, which ensures that the model will not be retrained. Another Python library, Sklearn, allows you to process text data. It is important to note that both audio and text data are present in enclosures that contain text messages for transcribing audio messages. The developed algorithm consists of several stages. First, the initial data is preprocessed, and stop words, additional spaces, etc. are deleted. Next, training takes place on pre-prepared data. The final step is to run the model on test data. For training, verification and testing, a multimodal open data set Multimodal EmotionLines Dataset (MELD) is used, consisting of more than 1400 dialogues and 13,000 models of utterances from the famous sitcom "Friends". Note that the data arrays are preliminarily divided into three categories of emotional coloring, such as positive, negative and neutral. The results of the computational experiment showed that the model created on two types of data, both text and audio versions, allows for better quality of binary sentiment analysis compared to separate processing of text data or audio data only. [ABSTRACT FROM AUTHOR]

Copyright of AIP Conference Proceedings is the property of American Institute of Physics and its content may not be copied or emailed to multiple sites without the copyright holder's express written permission. Additionally, content may not be used with any artificial intelligence tools or machine learning technologies. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)