Abstract Motivated by an increasing use of social media for the expression of personal stance towards a certain target, we analyse the language used to produce such opinionated content with expressions of sentiment, which represents the main data source for sentiment analysis. We use the first manually annotated corpus for sentiment analysis of the Serbian language developed for the service sector of higher education. Our study focuses on how various linguistic constructions, used in different context, influence the sentiment polarity of a text. Our findings indicate that sentiment expressions and negation have a most significant role in determining whether the text conveys positive, neutral, or negative sentiment, while intensifiers (words which either increase or decrease sentiment) have a considerable influence on sentiment intensity. We also present an analysis of the impact of conjunctions, conditional sentences, comparative and modal verbs, and pronouns on sentiment polarity and intensity. Based on the derived observations, we propose a set of rules that could be integrated with machine learning algorithms into an automated sentiment analysis system for the Serbian language. Our findings also make a much-needed contribution to the few currently available resources for natural language processing of Serbian.

中文翻译：

---

学生自以为是的情感表达的语言建构：基于语料库的研究

摘要由于越来越多地使用社交媒体来表达对某个目标的个人立场，我们用情感表达来分析用于产生这种自以为是的内容的语言，这是情感分析的主要数据来源。我们使用第一个手动注释的语料库来分析为高等教育服务部门开发的塞尔维亚语的情感。我们的研究集中于在不同上下文中使用的各种语言结构如何影响文本的情感极性。我们的发现表明，情感表达和否定在确定文本传达正，中性还是消极情感方面具有最重要的作用，而强化词（增加或降低情感的单词）对情感强度有很大影响。我们还提出了连词，条件句，比较和情态动词以及代词对情感极性和强度的影响的分析。基于派生的观察，我们提出了一组规则，可以将这些规则与机器学习算法集成到塞尔维亚语的自动情绪分析系统中。我们的发现还为目前很少有的塞尔维亚自然语言处理资源提供了非常必要的贡献。