Browse Books

Reviewer: Robert Goldberg

Natural language processing typically concentrates on the literal aspects of language, both syntactic and semantic. However, the "meta-processing" of language, which involves sentiment, cultural style, personal opinions, and individual circumstances, has not been incorporated adequately. It is to address this need that the current volume is published. This compendium collects 24 full versions of papers that were originally presented at the American Association for Artificial Intelligence (AAAI) Symposium on Exploring Attitude and Affect in Text, at Stanford University in March 2004. While the individual papers research advanced topics, they are self contained and provide sufficient material for those interested in general exposure to the topic. Although the book is not formally divided into sections, a number of the papers (chapters) can be gathered under a common theme. Indeed, the preface suggests that the chapters were specifically grouped into three parts: linguistic and cognitive models; lexical resources and attitude/affect recognition and generation; and applications. The first part of the book concentrates on models of attitude that can change the general flow of a speech or text. The first paper pays close attention to sensational words used in speech or text. Words such as "terrible" and "wonderful" generally imply a certain negative and positive attitude toward the situation at hand. However, the authors argue that the context in which the words are found can change the very meaning of what is being conveyed. Continuing on the theme that linguistic expressions can change the meaning of a text, the second chapter divides a text into "profiles." Profiles are segments of discourse text that share certain properties, such as the source of the information and the degree of veracity of the content. The author applies this to a long story. The third paper argues that words alone are not sufficient for understanding shifts in attitude in a text. The authors suggest that these words depend on the nature of the piece itself (whether it is an editorial or opinion piece, news article, or fictional writing). The next set of chapters involves cognitive input for linguistic choices. The first of these models the choice of terminology and style that artificially intelligent systems, specifically next generation natural language processing (NLP) systems, can use for generating stylistic letters (in this case) for genetic counselors to their clients. Such a communiqué is a composition of styles, involving references, evidence, tenses, and different perspectives. The second of these reports on a study to analyze the degree to which different people view a given text in a common manner. The purpose of that study is to enable NLP systems to generate a text that most people will similarly understand. The third paper contains a strong psychological component. The authors create a "Weighted Referential Activity Dictionary" to model "referential activity," which rates the degree to which choice of language relates to body language and emotional experience. The last two chapters of this first part of the book present annotation schemes for manually labeling portions of text to create training sets for NLP systems on which to base further evaluations. The next nine chapters make up Part 2 of the book. This section provides interesting approaches to constructing lexical resources for identifying and generating attitude in speech and text. The first constructs semantic classes and extracts affecting words along the semantic class axes. The second classifies emotional word intensities for deciding the effect words have on text (here, in French). The following three chapters suggest computational models (supervised statistical classification, tagging methods, and discourse features) for deciphering attitude and affect in text. Then, these technologies are correspondingly applied for determining propositional opinions, tagging attitudes in transcribed dialogues, and automatically classifying scientific citations in a paper based on the author's intended effect. The penultimate chapter studies how personal perspective affects generations of text and summarizing, while the final chapter of this section describes a method for recognizing and generating texts with nuances within attitudinal expression. The last nine chapters of this volume explain the applications of the theories presented in the above chapters. The first two of these explore categorization of a text based on writing style instead of content. Both of these chapters are based on a computational model of linguistic theory. The next two concentrate on scientific writings. One labels sentences as to whether they participate in an argumentative portion of the text. To the opposite extreme, the second paper correlates vague expressions when a scientific reference is cited. Following these, two chapters investigate culling opinions from multiple sources¿the first from Internet forums and the second considering reports over time. The last three chapters of this final part present empirical studies: one examines the use of sentiment for predicting future stock prices; the second uses a probabilistic approach to cull opinions from movie reviews; and, the third automates three types of summaries of a given text¿opinions, events, and factual knowledge. While many compendiums on a plethora of topics have been published over the years, this volume shines as truly presenting cutting-edge research in a specific subfield within NLP. The editors have done a fine job in aggregating full-length papers that are both interesting and informative from established researchers in the field. Online Computing Reviews Service