Narrative texts and personal conversations typically revolve around situations that people find themselves in. Current natural language systems extract only the literal meaning of events, failing to recognize how people are impacted by them. For example, if a man says that he has been laid off or diagnosed with cancer, then a system should understand that he is in a negative situation. If a woman says that she just graduated from college or has been promoted at work, then a system should understand that she is in a positive situation. This project develops technology to automatically identify situations that positively or negatively impact people. Affective knowledge is essential for applications such as sentiment and social media analysis, for example to recognize at-risk individuals experiencing adverse situations that may make them a danger to themselves or others.
This research develops natural language processing technology to recognize implicit affective states associated with events. Events and states are represented as situation frames and automatically harvested from large text corpora. Implicating situations are learned from blogs using semi-supervised label propagation with context graphs to propagate positive/negative evidence based on topic clustering, event-event co-occurrence, and discourse relations. Implicating situations are learned from tweets using bootstrapping methods applied to affective and sarcastic tweets. Each affective situation is automatically assigned a polarity and connotative strength. This research advances human language technology toward fundamentally deeper language understanding about affective states, motivations, and goals for narrative text and conversational dialogue.
