Search

Text mining and topic identification models are becoming increasingly relevant to extract value from the huge amount of unstructured textual information that companies obtain from their users and clients nowadays. Soft approaches to these problems are also gaining relevance, as in some contexts it may be unrealistic to assume that any document has to be associated to a single topic without any further consideration of the involved uncertainties. However, there is an almost total lack of reference documents allowing a proper assessment of the performance of soft classifiers in such soft topic identification tasks. To address this lack, in this paper a method is proposed that generates topic identification reference documents with a soft but objective nature, and which proceeds by combining, in random but known proportions, phrases of existing documents dealing with different topics. We also provide a computational study illustrating the application of the proposed method on a well-known benchmark for topic identification, as well as showing the possibility of carrying out an informative evaluation of soft classifiers in the context of soft topic identification.

Traditionally, the last step of edge detection algorithms, which is called scaling-evaluation, produces the final output classifying each pixel as edge or nonedge. This last step is usually done based on local evaluation methods. The local evaluation makes this classification based on measures obtained for every pixel. By contrast, in this work, we propose a global evaluation approach based on the idea of edge list to produce a solution that suits more with the human perception. In particular, we propose a new evaluation method that can be combined with any classical edge detection algorithm in an easy way to produce a novel edge detection algorithm. The new global evaluation method is divided in four steps: in first place we build the edge lists, that we have called edge segments. In second place we extract the characteristics associated to each segment: length, intensity, location, and so on. In the third step we learn the characteristics that make a segment good enough to become an edge. At the fourth step, we apply the classification task. In this work we have built the ground truth of edge list necessary for the supervised classification. Finally, we test the effectiveness of this algorithm against other classical algorithms based on local evaluation approach.

Most edge detection algorithms deal only with grayscale images, and the way of adapting them to use with RGB images is an open problem. In this work, we explore different ways of aggregating the color information of a RGB image for edges extraction, and this is made by means of well-known edge detection algorithms. In this research, it is been used the set of images from Berkeley. In order to evaluate the algorithm’s performance, F measure is computed. The way that color information -the different channels- is aggregated is proved to be relevant for the edge detection task. Moreover, post-aggregation of channels performed significatively better than the classic approach (pre-aggregation of channels).