We show how to consider similarity between features for calculation of similarity of objects in the Vector Space Model (VSM) for machine learning algorithms and other classes of methods that involve similarity between objects. Unlike LSA, we assume that similarity between features is known (say, from a synonym dictionary) and does not need to be learned from the data. We call the proposed similarity measure soft similarity. Similarity between features is common, for example, in natural language processing: words, n-grams, or syntactic n-grams can be somewhat different (which makes them different features) but still have much in common: for example, words "play" and "game" are different but related. When there is no similarity between features then our soft similarity measure is equal to the standard similarity. For this, we generalize the well-known cosine similarity measure in VSM by introducing what we call "soft cosine measure". We propose various formulas for exact or approximate calculation of the soft cosine measure. For example, in one of them we consider for VSM a new feature space consisting of pairs of the original features weighted by their similarity. Again, for features that bear no similarity to each other, our formulas reduce to the standard cosine measure. Our experiments show that our soft cosine measure provides better performance in our case study: entrance exams question answering task at CLEF. In these experiments, we use syntactic n-grams as features and Levenshtein distance as the similarity between n-grams, measured either in characters or in elements of n-grams.

No abstract

Abstract. Opinion mining deals with determining of the sentiment orientation-positive, negative, or neutral-of a (short) text. Recently, it has attracted great interest both in academia and in industry due to its useful potential applications. One of the most promising applications is analysis of opinions in social networks. In this paper, we examine how classifiers work while doing opinion mining over Spanish Twitter data. We explore how different settings (n-gram size, corpus size, number of sentiment classes, balanced vs. unbalanced corpus, various domains) affect precision of the machine learning algorithms. We experimented with Naïve Bayes, Decision Tree, and Support Vector Machines. We describe also language specific preprocessing-in our case, for Spanish language-of tweets. The paper presents best settings of parameters for practical applications of opinion mining in Spanish Twitter. We also present a novel resource for analysis of emotions in texts: a dictionary marked with probabilities to express one of the six basic emotionsProbability Factor of Affective use (PFA)Spanish Emotion Lexicon that contains 2,036 words.

No abstract

The paper presents a new corpus for fake news detection in the Urdu language along with the baseline classification and its evaluation. With the escalating use of the Internet worldwide and substantially increasing impact produced by the availability of ambiguous information, the challenge to quickly identify fake news in digital media in various languages becomes more acute. We provide a manually assembled and verified dataset containing 900 news articles, 500 annotated as real and 400, as fake, allowing the investigation of automated fake news detection approaches in Urdu. The news articles in the truthful subset come from legitimate news sources, and their validity has been manually verified. In the fake subset, the known difficulty of finding fake news was solved by hiring professional journalists native in Urdu who were instructed to intentionally write deceptive news articles. The dataset contains 5 different topics: (i) Business, (ii) Health, (iii) Showbiz, (iv) Sports, and (v) Technology. To establish our Urdu dataset as a benchmark, we performed baseline classification. We crafted a variety of text representation feature sets including word n-grams, character n-grams, functional word n-grams, and their combinations. After applying a variety of feature weighting schemes, we ran a series of classifiers on the train-test split. The results show sizable performance gains by AdaBoost classifier with 0.87 F1Fake and 0.90 F1Real. We provide the results evaluated against different metrics for a convenient comparison of future research. The dataset is publicly available for research purposes.