EAGER: Identifying Affective Events and Situations in Text

大多数自然语言处理（NLP）系统仍然只能从字面上理解单词和短语。特别是，当前的NLP技术忽视了事件和情况对人们的心理影响。例如，NLP系统可以识别与被雇用或被解雇相关的短语，但它们不理解被雇用通常是可取的，而被解雇通常是不可取的。这项探索性研究旨在赋予NLP技术识别将对人们产生积极或消极影响的事件和情况的能力。这项拟议中的研究可能是变革性的，因为它将使人们能够从根本上更深入地理解语言，而不仅仅是单词和短语的字面意义。从这项研究中产生的新方法将有利于广泛的应用，需要理解书面文本和人类对话，文档摘要，问答，讽刺识别和识别威胁性陈述。社会也可以从这项技术中受益，因为它提供了新的机会，可以自动从网络和社交媒体上收集有关最经常影响不同类型的人的事件和个人情况的知识。识别情感状态可以帮助识别可能对自己或他人构成危险的高危个体，例如被欺负的青少年，患有创伤后应激障碍的退伍军人，感到愤怒或被剥夺权利的年轻人，以及可能孤独和抑郁的老年人。EArly探索性研究基金会研究学习识别文本中“情感”事件和情况的新方法。情感状态识别对于理解人们的动机、计划和目标、因果链以及文本和对话的叙事结构至关重要。这项研究的目标是创建弱监督学习方法，以自动识别与情感事件和对人们产生积极或消极影响的情况相对应的短语。本研究旨在开发一套独立的，弱监督的学习者，其中每个组件将利用特定类型的语言结构和话语背景，以确定与积极或消极的情感状态相关的事件和情况。然后，弱监督学习器的集合将嵌入集成架构中，并将探索多视图学习方法，以实现跨独立学习器集合的大规模自举知识获取。这项工作的关键方面探索不同类型的语言和话语结构的自举学习，不同的方法来学习多词表达的事件和情况，以及不同的方法来处理上下文歧义在学习和识别。词汇情感状态知识将提高NLP系统识别讽刺，识别威胁，识别因果关系，并实现对叙事文本和会话对话的更深入理解的能力。

Most natural language processing (NLP) systems still have only a literal understanding of words and phrases. In particular, current NLP technology is oblivious to the psychological impact that events and situations have on people. For example, NLP systems can recognize phrases associated with being hired or fired, but they do not understand that being hired is usually desirable while being fired is generally undesirable. This exploratory research aims to endow NLP technology with the ability to recognize events and situations that will have a positive or negative impact on people. The proposed research could be transformative because it would enable a fundamentally deeper understanding of language beyond the literal meaning of words and phrases. New methods resulting from this research would benefit a wide spectrum of applications that require understanding of written texts and human conversation, document summarization, question answering, sarcasm recognition, and identification of threatening statements. Society could also benefit from this technology through new opportunities to automatically harvest knowledge from the Web and social media about the events and personal situations that most frequently affect different types of people. Recognizing affective states could help to identify at-risk individuals who may be a danger to themselves or others, such as teenagers who are being bullied, veterans who suffer from post-traumatic stress disorder, young adults who feel angry or disenfranchised, and elderly people who may be lonely and depressed. This EArly Grant for Exploratory Research investigates novel methods for learning to recognize "affective" events and situations in text. Affect state recognition is essential to understand people's motivations, plans and goals, causal chains, and the narrative structure of text and conversation. The goal of this research is to create weakly supervised learning methods to automatically identify phrases corresponding to affective events and situations that have a positive or negative impact on people. This research aims to develop a set of independent, weakly supervised learners, where each component will exploit a specific type of linguistic structure and discourse context to identify events and situations that are associated with positive or negative affect states. The set of weakly supervised learners will then be embedded in an ensemble architecture, and multi-view learning methods will be explored to enable large-scale bootstrapped knowledge acquisition across the set of independent learners. Key aspects of this work explore different types of linguistic and discourse structures for bootstrapped learning, different approaches for learning multi-word expressions for events and situations, and different methods for handling contextual ambiguity during both learning and recognition. Lexical affect stateknowledge would improve the ability of NLP systems to recognize sarcasm, identify threats, recognize causal relationships, and achieve a deeper understanding of narrative text and conversational dialogue.