Catálogo de publicaciones - libros

We consider the problem of finding communities in large linked networks such as web structures or citation networks. We review similarity measures for linked objects and discuss the k-Means and EM algorithms, based on text similarity, bibliographic coupling, and co-citation strength. We study the utilization of the principle of multi-view learning to combine these similarity measures. We explore the clustering algorithms experimentally using web pages and the Cite-Seer repository of research papers and find that multi-view clustering effectively combines link-based and intrinsic similarity.

This paper is in line with the principles of numerical taxonomy and with the program of holistic typology. It integrates the level of phonology with the morphological and syntactical level by correlating metric properties (such as n of phonemes per syllable and n of syllables per clause) with non-metric variables such as the number of morphological cases and adposition order. The study of crosslinguistic patterns of variation results in a division of languages into two main groups, depending on their rhythmical structure. Syllable-timed rhythm, as opposed to stress-timed rhythm, is closely associated with a lower complexity of syllables and a higher number of syllables per clause, with a rather high number of morphological cases and with a tendency to OV order and postpositions. These two fundamental types of language may be viewed as the “idealized” counterparts resulting from the very same and universal pattern of variation.

In this paper we present an approach for classifying web pages obtained from the Slovenian Internet directory where the web sites covering different topics are organized into a topic ontology. We tested two different methods for representing text documents, both in combination with the linear SVM classification algorithm. The first representation used is a standard bag-of-words approach with TFIDF weights and cosine distance used as similarity measure. We compared this to String kernels where text documents are compared not by words but by substrings. This removes the need for stemming or lemmatisation which can be an important issue when documents are in other languages than English and tools for stemming or lemmatisation are unavailable or are expensive to make or learn. In highly inflected natural languages, such as Slovene language, the same word can have many different forms, thus String kernels have an advantage here over the bag-of-words. In this paper we show that in classification of documents written in highly inflected natural language the situation is opposite and String Kernels significantly outperform the standard bag-of-words representation. Our experiments also show that the advantage of String kernels is more evident for domains with unbalanced class distribution.

Analysis and knowledge representation of linguistic objects tends to focus on larger units (e.g. words) than print medium characters. We analyse characters as linguistic objects in their own right, with meaning, structure and form. Characters have meaning (the symbols of the International Phonetic Alphabet denote phonetic categories, the character represented by the glyph ‘∪’ denotes set union), structure (they are composed of stems and parts such as descenders or diacritics or are ligatures), and form (they have a mapping to visual glyphs). Character encoding initatives such as Unicode tend to concentrate on the structure and form of characters and ignore their meaning in the sense discussed here. We suggest that our approach of including semantic decomposition and defining font-based namespaces for semantic character domains provides a long-term perspective of interoperability and tractability with regard to data-mining over characters by integrating information about characters into a coherent semiotically-based ontology. We demonstrate these principles in a case study of the International Phonetic Alphabet.

In this paper, we present our experience in applying collaborative filtering to real-life corporate data. The quality of collaborative filtering recommendations is highly dependent on the quality of the data used to identify users’ preferences. To understand the influence that highly sparse server-side collected data has on the accuracy of collaborative filtering, we ran a series of experiments in which we used publicly available datasets and, on the other hand, a real-life corporate dataset that does not fit the profile of ideal data for collaborative filtering.

This study focuses on the contribution of sentence length for a quantitative text typology. Therefore, 333 Slovenian texts are analyzed with regard to their sentence length. By way of multivariate discriminant analyses () it is shown that indeed, a text typology is possible, based on sentence length, only; this typology, however, does not coincide with traditional text classifications, such as, e.g., text sorts or functional style. Rather, a new categorization into specific discourse types seems reaonable.

We present a hybrid machine learning approach for information extraction from unstructured documents by integrating a learned classifier based on the Maximum Entropy Modeling (MEM), and a classifier based on our work on (DOP). The hybrid behavior is achieved through a voting mechanism applied by an iterative tag-insertion algorithm. We have tested the method on a corpus of newspaper articles about company turnover, and achieved 85.2% F-measure using the hybrid approach, compared to 79.3% for MEM and 51.9% for DOP when running them in isolation.

In many real world machine learning tasks, labeled training examples are expensive to obtain, while at the same time there is a lot of unlabeled examples available. One such class of learning problems is text classification. Active learning strives to reduce the required labeling effort while retaining the accuracy by intelligently selecting the examples to be labeled. However, very little comparison exists between different active learning methods. The effects of the ratio of positive to negative examples on the accuracy of such algorithms also received very little attention. This paper presents a comparison of two most promising methods and their performance on a range of categories from the Reuters Corpus Vol. 1 news article dataset.

Hypertext categorization is the task of automatically assigning category labels to hypertext units. Comparable to text categorization it stays in the area of function learning based on the bag-of-features approach. This scenario faces the problem of a many-to-many relation between websites and their hidden logical document structure. The paper argues that this relation is a prevalent characteristic which interferes any effort of applying the classical apparatus of categorization to web genres. This is confirmed by a threefold experiment in hypertext categorization. In order to outline a solution to this problem, the paper sketches an alternative method of unsupervised learning which aims at bridging the gap between statistical and structural pattern recognition (Bunke et al. 2001) in the area of web mining.

We investigate the performance of text mining systems for annotating press articles in two real-world press archives. Seven commercial systems are tested which recover the categories of a document as well named entities and catchphrases. Using cross-validation we evaluate the precision-recall characteristic. Depending on the depth of the category tree 39–79% breakeven is achieved. For one corpus 45% of the documents can be classified automatically, based on the system’s confidence estimates. In a usability experiment the formal evaluation results are confirmed. It turns out that with respect to some features human annotators exhibit a lower performance than the text mining systems. This establishes a convincing argument to use text mining systems to support indexing of large document collections.