Composing meaning

接下来几年的工作是建立在我之前的精神研究的基础上的，特别是情感分析、因果关系、对冲和（梦）情感分析。所有这些主题的统一特征是，它们都关注(a)区分事实信息与嵌入元信息在不同文本体裁中的区别；(b)解码这些元信息，并使其作为文本意义的平等组成部分，可用于自动文本分析。迄今为止，自然语言处理（NLP）系统在很大程度上忽略了文本的事实内容只贡献了其意义的一部分。但最近，从博客等丰富的在线资源中，利用情感和观点丰富的文本进行商业开发以改进产品的可能性，已将注意力转移到用肤浅的手段考虑情感甚至情感内容上。现在，价移词（可以颠倒一个小句的意思的语言结构，如否定）被更广泛地考虑。但是，例如，情绪分析的主要技术是简单地计算句子中积极和消极情绪词汇的数量，并获得所有整体情绪得分。这种过于简单化的方法的局限性是显而易见的：在电影评论中，一长串关于表演和导演的积极评价可以用简单的结束语来推翻：“然而，这部电影很糟糕。”随机技术的不足之处是：(a)语言相关的特征尚未被识别，因此还没有通用的标注方案和训练语料库；(b)它们的组成行为对于简单的标注方案来说太复杂了。随着对所有不同元命题设备的性质和行为的更好理解，将产生更好的注释模式，从而导致更精细的随机系统。在考虑句子意义的非事实方面的语言学以及它们如何对文本的意义做出贡献时，我建议为文本开发一种注释方案和组合语义，以连贯地对待不同的意义成分。

The work proposed here for the next few years builds on all of my previous research in spirit, and in particular sentiment analysis, causality, hedging, and (dream) emotion analysis. The unifying feature of all these topics is that they are concerned with (a) differentiating factual information from embedding meta-information in different text genres and (b) decoding this meta-information and making it accessible for automatic text analysis as an equal component in text meaning. That the factual content of a text contributes only one part of its meaning has been largely ignored in Natural Language Processing (NLP) systems to date. But recently, the possibility of commercial exploitation of sentiment and opinion-laden texts from abundant on-line sources such as blogs for product improvement has shifted attention to consider sentiment and even emotional content using shallow means. Now, valence shifters (language constructs that can invert the meaning of a clause, such as negation) are considered more widely. But still, the predominant technique for, for instance, sentiment analysis, is to simply count the number of positive vs. negative sentiment words in a sentence with a winner take all overall sentiment score. The limitations of the simplistic approach are obvious: a long string of positive statements about acting and directing in a movie review can be undone with the simple concluding statement: "And yet the film bombs.". Stochastic techniques fall short, because (a) the linguistically relevant features have not been identified and thus there are no common annotation schemes nor training corpora yet (b) their compositional behavior is too complex for simple annotation schemes. With better understanding of the nature and behavior of all of the different meta-propositional devices will come better annotation schemas that will lead to more refined stochastic systems. In considering the linguistics of non-factual aspects of sentence meaning and how they contribute to the meaning of the text, I propose to develop an annotation scheme and a compositional semantics for texts that treats different meaning components coherently.