This paper presents a method for automatically determining the semantic orientation (positive or negative) of adjectives using indirect information from a large corpus. The approach leverages the constraints imposed by conjunctions on the semantic orientation of conjoined adjectives. A log-linear regression model is used to predict whether conjoined adjectives are of the same or different orientations, achieving 82% accuracy when each conjunction is considered independently. By combining these constraints across many adjectives, a clustering algorithm separates adjectives into groups of different orientations, and finally, adjectives are labeled positive or negative. Evaluations on real data and simulation experiments show high performance, with classification precision exceeding 90% for adjectives that occur in a modest number of conjunctions in the corpus. The method uses a corpus of 21 million words from the 1987 Wall Street Journal, annotated with part-of-speech tags. Adjectives with predetermined orientation labels are selected based on their frequency and semantic properties. A set of 1,336 adjectives (657 positive and 679 negative) is used for training and evaluation. The method identifies conjunctions between adjectives and uses them to infer the semantic orientation of the adjectives. A log-linear regression model is then used to predict the orientation of adjectives based on their conjunctional relationships. The results show that the method achieves high accuracy in classifying adjectives as positive or negative, with over 92% accuracy on the sparsest test set and up to 99% accuracy on denser sets. The clustering algorithm groups adjectives into subsets based on their semantic orientation, and the clusters are labeled positive or negative based on the frequency of the adjectives within the cluster. The method is evaluated on real data and simulation experiments, showing that it can achieve high performance even with a modest number of links per adjective. The results demonstrate that the method is effective in identifying the semantic orientation of adjectives and can be applied to other word classes. The method is also used to automatically identify antonyms from the corpus, without access to any semantic descriptions. The learned semantic categorization of adjectives can be used to help interpret the conjunctions they participate in, and the analysis is extended to nouns and verbs.This paper presents a method for automatically determining the semantic orientation (positive or negative) of adjectives using indirect information from a large corpus. The approach leverages the constraints imposed by conjunctions on the semantic orientation of conjoined adjectives. A log-linear regression model is used to predict whether conjoined adjectives are of the same or different orientations, achieving 82% accuracy when each conjunction is considered independently. By combining these constraints across many adjectives, a clustering algorithm separates adjectives into groups of different orientations, and finally, adjectives are labeled positive or negative. Evaluations on real data and simulation experiments show high performance, with classification precision exceeding 90% for adjectives that occur in a modest number of conjunctions in the corpus. The method uses a corpus of 21 million words from the 1987 Wall Street Journal, annotated with part-of-speech tags. Adjectives with predetermined orientation labels are selected based on their frequency and semantic properties. A set of 1,336 adjectives (657 positive and 679 negative) is used for training and evaluation. The method identifies conjunctions between adjectives and uses them to infer the semantic orientation of the adjectives. A log-linear regression model is then used to predict the orientation of adjectives based on their conjunctional relationships. The results show that the method achieves high accuracy in classifying adjectives as positive or negative, with over 92% accuracy on the sparsest test set and up to 99% accuracy on denser sets. The clustering algorithm groups adjectives into subsets based on their semantic orientation, and the clusters are labeled positive or negative based on the frequency of the adjectives within the cluster. The method is evaluated on real data and simulation experiments, showing that it can achieve high performance even with a modest number of links per adjective. The results demonstrate that the method is effective in identifying the semantic orientation of adjectives and can be applied to other word classes. The method is also used to automatically identify antonyms from the corpus, without access to any semantic descriptions. The learned semantic categorization of adjectives can be used to help interpret the conjunctions they participate in, and the analysis is extended to nouns and verbs.