Understanding Narratives through Sentiment, Emotions and Events

Automatically understanding narratives is key for AI systems processing documents from diverse domains such as healthcare, journalism or teaching. However, the narratives' heterogeneity across and within domains poses a significant challenge. To be successful, the understanding process requires robustness in several domain-specific aspects, such as the jargon, linguistic style, and background knowledge. This is particularly evident in personal narratives, i.e. recounts of emotionally charged personal events. In personal narratives, each narrator has a unique lexicon and style for recounting personal experiences grounded in personal backgrounds. Another challenge is the ambiguity caused by possible mismatches between what the narratee, i.e. the addressee, understands and the actual narrator's perspectives and feelings. These issues are not only reflected in the development of narrative understanding models but also in the design of collection and annotation protocols for constructing corpora of narratives. In this thesis, we study a range of methodologies for automating the analysis and acquisition of narratives. Initially, we test and improve the robustness of sentiment analyser models on different genres, including personal narratives and financial news. Then, we investigate the extraction and representation of the fluctuations of the emotional state conveyed by the unfolding events of personal narratives. The findings of these studies have revealed that deep neural models struggle to distinguish between positive or negative emotion levels from neutral ones. This has driven us to design and implement a protocol for acquiring multimodal personal narratives to leverage implicit and explicit signals. Another challenge encountered during the development of this protocol concerns the support of recounting emotionally charged personal events. In this regard, we investigate the application of large language models by assessing their abilities in assisting narrators with a dialogue. To understand the progression of a narrative, it is necessary to consider the temporal relations among events. For this, we investigate the performance of large language models in recognising temporal relations among events and compare it with encoder-only models.