Catálogo de publicaciones - libros

Internet sources, such as newsgroups, message boards, and blogs, are an under-exploited resource for developing analyses of community and market responses to everything from consumer products and services, to current events and politics. In this paper, we present an overview of our exploration of effective ways of characterizing this large volume of information. In our approach, we first create time-series that represent the subjects, opinions, and attitudes expressed in the Internet sources, and then generate “Linguistic Summaries” that provide natural and easily understood descriptions of the behaviors exhibited by these time-series.

A simple and novel method for generating labeled examples for sentiment analysis is introduced: news stories about publicly traded companies are labeled positive or negative according to price changes of the company stock. It is shown that there are many lexical markers for bad news but none for good news. Overall, learned models based on lexical features can distinguish good news from bad news with accuracy of about 70%. Unfortunately, this result does not yield profits since it works only when stories are labeled according to cotemporaneous price changes but does not work when they are labeled according to subsequent price changes.

One approach to the assessment of overall opinion polarity (OvOP) of reviews, a concept defined in this paper, is the use of supervised machine learning mechanisms. In this paper, the impact of lexical feature selection and feature generalization, applied to reviews, on the precision of two probabilistic classifiers (Naïve Bayes and Markov Model) with respect to OvOP identification is observed. Feature generalization based on hypernymy as provided by WordNet, and feature selection based on part-ofspeech (POS) tags are evaluated. A ranking criterion is introduced, based on a function of the probability of having positive or negative polarity, which makes it possible to achieve 100% precision with 10% recall. Movie reviews are used for training and testing the probabilistic classifiers, which achieve 80% precision.

An interactive information retrieval system that provides different types of summaries of retrieved documents according to each user’s information needs, situation, or purpose of search can be effective for understanding document content. The purpose of this study is to build a multi-document summarizer, “”, which produces summaries according to such viewpoints. We tested its effectiveness on a new test collection, , which contains human-made reference summaries of three different summary types for each of the 30 document sets. Once a set of documents on a topic (e.g., documents retrieved by a search engine) is provided to , it returns a list of topics discussed in the given document set, so that the user can select a topic or topics of interest as well as the summary type, such as fact-reporting, opinion-oriented or knowledge-focused, and produces a summary from the viewpoints of the topics and summary type selected by the user. We assume that sentence types and document genres are related to the types of information included in the source documents and are useful for selecting appropriate information for each of the summary types. “Sentence type” defines the type of information in a sentence. “Document genre” defines the type of information in a document. The results of the experiments showed that the proposed system using automatically identified sentence types and document genres of the source documents improved the coverage of the system-produced fact-reporting, opinion-oriented, and knowledge-focused summaries, 13.14%, 34.23%, and 15.89%, respectively, compared with our baseline system which did not differentiate sentence types or document genres.