In this section appear some of the lines in which we are working at the moment and that, by their degree of occupation are generating some products that will be published soon:
- Comparison of different Text sizes with the techniques “centroid” and “Fold-in”
Resistance of both techniques comparing pseudodocuments of a variable size under the assumption that when increasing the size of the pseudocuments, the comparisons based on the cosine will increase more noticiable in centroid than in “Fold-in”. In addition, it is analyzed
the intervention of the first dimension in this effect. The validity of this type of techniques is important due the difference in consumption of resources between both.
- Different parameters to implement autotutors with LSA
a) university students: using corpus specific of domain of small and medium sizes formed by extremely structured texts or little structured or the sum of both. On these corpus, different semantic spaces formed under different parameters are put on trial. These parameters include the way of preprocess (purges and measures of importance of the terms), the reduction of dimensions in percentage, the way to construct to pseudodocuments (centroid and Fold-in) and the measures of similarity of texts (cosines or Euclidean distances). The evaluated students compose two groups, one of experts and another one of non-experts. Both groups also answer an open question that will be evaluated by LSA system under all and each of the semantics spaces and by a group of human graders. By an Analysis of Variance, those parameters of the spaces that contribute more to the correlation to the human criterion are evaluated.
b) Students of Primary, Secundary and University: summaries made by students of different academic courses. They also look for what parameters improve the evaluation of the summaries comparing the evaluations of the LSA with others done by expert judges. To see first results in [pdf].
- Discrimination between different emotional tones.
Text segments with an optimistic emotional style, and text segments with a depressive emotional style are put on approval under different techniques among them LSA and clusters.