The paper presents a system for determining the sentiment of opinions. The system identifies people who hold opinions about a given topic and the sentiment of each opinion. It includes modules for determining word sentiment and combining sentiments within a sentence. The system experiments with various models for classifying and combining sentiment at the word and sentence levels, achieving promising results. The paper defines an opinion as a quadruple [Topic, Holder, Claim, Sentiment], where the Holder believes a Claim about the Topic and associates a Sentiment with the belief. Sentiment is defined as an explicit or implicit expression of the Holder's positive, negative, or neutral regard toward the Claim. The system aims to automatically identify these sentiments. The system operates in four steps: selecting sentences containing the topic and holder candidates, delimiting opinion regions, classifying sentiment-bearing words, and combining them to determine the Holder's sentiment for the whole sentence. The system uses a word sentiment classifier and a sentence sentiment classifier. The word sentiment classifier uses a seed list of words, expanded using WordNet, to classify words as positive, negative, or neutral. It also assigns a strength of sentiment polarity to words. The sentence sentiment classifier identifies the Holder and defines a sentiment region, then combines the sentiments of words within that region using different models. The paper discusses the challenges of sentiment classification, including the difficulty of determining sentiment for words with both positive and negative connotations, and the challenge of identifying the Holder in a sentence. The system's performance is evaluated using human annotations and shows that it achieves reasonable accuracy, especially when using manually annotated Holders. The paper concludes that sentiment recognition is a challenging part of understanding opinions. The authors plan to extend their work to more difficult cases, such as sentences with weak-opinion-bearing words or multiple opinions about a topic. They also plan to use a parser to associate regions more reliably with Holders and explore other learning techniques. The experiments show that encouraging results can be obtained even with relatively simple models and minimal manual seeding effort.The paper presents a system for determining the sentiment of opinions. The system identifies people who hold opinions about a given topic and the sentiment of each opinion. It includes modules for determining word sentiment and combining sentiments within a sentence. The system experiments with various models for classifying and combining sentiment at the word and sentence levels, achieving promising results. The paper defines an opinion as a quadruple [Topic, Holder, Claim, Sentiment], where the Holder believes a Claim about the Topic and associates a Sentiment with the belief. Sentiment is defined as an explicit or implicit expression of the Holder's positive, negative, or neutral regard toward the Claim. The system aims to automatically identify these sentiments. The system operates in four steps: selecting sentences containing the topic and holder candidates, delimiting opinion regions, classifying sentiment-bearing words, and combining them to determine the Holder's sentiment for the whole sentence. The system uses a word sentiment classifier and a sentence sentiment classifier. The word sentiment classifier uses a seed list of words, expanded using WordNet, to classify words as positive, negative, or neutral. It also assigns a strength of sentiment polarity to words. The sentence sentiment classifier identifies the Holder and defines a sentiment region, then combines the sentiments of words within that region using different models. The paper discusses the challenges of sentiment classification, including the difficulty of determining sentiment for words with both positive and negative connotations, and the challenge of identifying the Holder in a sentence. The system's performance is evaluated using human annotations and shows that it achieves reasonable accuracy, especially when using manually annotated Holders. The paper concludes that sentiment recognition is a challenging part of understanding opinions. The authors plan to extend their work to more difficult cases, such as sentences with weak-opinion-bearing words or multiple opinions about a topic. They also plan to use a parser to associate regions more reliably with Holders and explore other learning techniques. The experiments show that encouraging results can be obtained even with relatively simple models and minimal manual seeding effort.