Abstract. This paper describes a corpus annotation project to study issues in the manual annotation of opinions, emotions, sentiments, speculations, evaluations and other private states in language. The resulting corpus annotation scheme is described, as well as examples of its use. In addition, the manual annotation process and the results of an inter-annotator agreement study on a 10,000-sentence corpus of articles drawn from the world press are presented.

This paper presents the results of developing subjectivity classifiers using only unannotated texts for training. The performance rivals that of previous supervised learning approaches. In addition, we advance the state of the art in objective sentence classification by learning extraction patterns associated with objectivity and creating objective classifiers that achieve substantially higher recall than previous work with comparable precision.

There has been a recent swell of interest in the automatic identification and extraction of opinions and emotions in text. In this paper, we present the first experimental results classifying the intensity of opinions and other types of subjectivity and classifying the subjectivity of deeply nested clauses. We use a wide range of features, including new syntactic features developed for opinion recognition. We vary the learning algorithm and the feature organization to explore the effect this has on the classification task. In 10-fold cross-validation experiments using support vector re-gression, we achieve improvements in mean-squared error over baseline ranging from 49 % to 51%. Using boosting, we achieve improvements in accuracy ranging from 23 % to 96%.

An analysis of human planning strategies reveals that much of the knowledge that underlies intelligent planning involves commonsense psychology, the way that people think that they think. In this paper we describe our continuing effort to formalize a large-scale theory of commonsense psychology as 30 interrelated content theories in first-order logic. This paper discusses key aspects of the 16 content theories that we have completed, focusing on those that provide an account of how knowledge and intention lead to action, namely, memory, knowledge management, envisionment, goals, planning, and execution. Some of these areas present challenges to many of the simplifying assumptions that have traditionally been made in formal knowledge representation research; others are areas of commonsense knowledge where few formal treatments have previously been attempted. 1

As technology and human-computer interaction advances, there is an increased interest in affective computing. One of the current challenges in computational speech and text processing is addressing affective and expressive meaning, an area that has received fairly sparse attention in linguistics. Linguistic investigation in this area is motivated both by the need for scientific study of subjective language phenomena, and by useful applications such as expressive text-to-speech synthesis. The study makes contributions to the study of affect and language, by describing a novel data resource, outlining models and challenges for exploring affect in language, applying computational methods toward this problem with included empirical results, and suggesting paths for further research. After the introduction, followed by a survey of several areas of related work in Chapter 2, Chapter 3 presents a newly developed sentence-annotated corpus resource divided into three parts for large-scale exploration of affect in texts (specifically tales). Besides covering annotation and data set description, the chapter includes a hierarchical affect model and a qualitative-interpretive examination suggesting characteristics of a subset of the data marked by high agreement in affective label assignments. Chapter 4 is devoted to experimental work on automatic affect prediction in text. Different computational methods are explored

In the context of developing formal theories of commonsense psychology, or how peole think they think, we have developed a formal theory of goals. In it we explicate and axiomatize, among others, the goal-related notions of trying, success, failure, functionality, intactness, and importance.

This work is concerned with the automatic understanding of evaluative text. We investigate sentence level opinion polarity prediction by assigning lexical polarities and deriving sentence polarity from these with the use of contextual valence shifters. A methodology for iterative failure analysis is developed and used to refine our lexicon and identify new contextual shifters. Algorithms are presented that employ these new shifters to improve sentence polarity prediction accuracy beyond that of a state-of-the-art existing algorithm in the domain of consumer product reviews. We then apply the best configuration of our algorithm to the domain of movie reviews.

We understand discourse so well because we know so much. If we are to have natural language understanding systems that are able to deal with texts with emotional content, we must encode knowledge of human emotions for use in the systems. In particular, we must equip the system with a formal version of people’s implicit