Frustration Level Annotation in Latvian Tweets with Non-Lexical Means of Expression

Proceedings of Recent Advances in Natural Language Processing, pages 814–823 Sep 1–3, 2021. https://doi.org/10.26615/978-954-452-072-4_093 814 Abstract We present a neural-network-driven model for annotating frustration intensity in customer support tweets, based on representing tweet texts using a bag-of- words encoding after processing with subword segmentation together with non- lexical features. The model was evaluated on tweets in English and Latvian languages, focusing on aspects beyond the pure bag-of-words representations used in previous research. The experimental results show that the model can be successfully applied for texts in a non-English language, and that adding non-lexical features to tweet representations significantly improves performance, while subword segmentation has a moderate but positive effect on model accuracy. Our code and training data are publicly available1. 1 Introduction Dramatically increasing data storage and processing capacities have resulted in an explosion of available data, and of potential uses for the many kinds of knowledge or insights that could theoretically be extracted from that data. The development of Web 2.0 has created unprecedented amounts of text, and, in recent decades, images and videos. But the sheer volume of the available data is problematic due to a shortage of human resources (time and attention) available for analyzing or even just browsing through it all, as described in (Verma et al, 2016). From the very start, as soon as machine learning techniques appeared, they were immediately applied to text analysis. Neural networks have proven particularly suitable for such tasks, but to learn to approximate human judgements, these 1 Source code is available at https://github.com/Lynx1981/dfrustration/ methods generally need a large amount of manually annotated training data to obtain a viable model. Creating such corpora is a very time- consuming task, especially if it has to be done from scratch. For languages with a relatively low number of speakers, such as Latvian, this is a particularly pressing problem, because of the small number of textual corpora available, so that researchers generally need to create their own datasets. Much of the information of importance to be captured from the Internet is by its nature related to emotions. Recommendation systems, conversational “chat bots”, automated customer support assistants, various tools for analyzing, monitoring or enhancing personal well-being or mental health, not to mention systems for targeting or crafting advertising or marketing messages — all can benefit from understanding the emotional state of the recipient or sender of a given message. But recognition of emotions in text generally requires considerable investment of human resources; and is difficult to automate in principle. Perception of emotions is largely subjective, and the emotions perceived from the same source by different people differ. Further, in written communication, and especially communication in social networks, the words alone do not capture all of the emotional information available in the message. For example, the emotional tone of a sentence may vary depending on whether it was followed by period as the absence of periods is a characteristic feature of internet jargon usage, as shown in (Khalifa, 2020), which may carry a different emotional …

Content truncated to 500 words