In this paper, we propose a novel method to extend sequence-to-sequence models to accurately process sequences much longer than the ones used during training while being sample- and resource-efficient, supported by thorough experimentation. To investigate the effectiveness of our method, we apply it to the task of correcting documents already processed with Optical Character Recognition (OCR) systems using sequence-to-sequence models based on characters. We test our method on nine languages of the ICDAR 2019 competition on post-OCR text correction and achieve a new state-of-the-art performance in five of them. The strategy with the best performance involves splitting the input document in character n-grams and combining their individual corrections into the final output using a voting scheme that is equivalent to an ensemble of a large number of sequence models. We further investigate how to weigh the contributions from each one of the members of this ensemble. Our code for post-OCR correction is shared at https://github.com/jarobyte91/post_ocr_correction.

Topic ontologies or web directories consist of large collections of links to websites, arranged by topic in different categories. The structure of these ontologies is typically not flat because there are hierarchical and nonhierarchical relationships among topics. As a consequence, websites classified under a certain topic may be relevant to other topics. Although some of these relevance relations are explicit, most of them must be discovered by an analysis of the structure of the ontologies. This article proposes a family of models of relevance propagation in topic ontologies. An efficient computational framework is described and used to compute nine different models for a portion of the Open Directory Project graph consisting of more than half a million nodes and approximately 1.5 million edges of different types. After performing a quantitative analysis, a user study was carried out to compare the most promising models. It was found that some general difficulties rule out the possibility of defining flawless models of relevance propagation that only take into account structural aspects of an ontology. However, there is a clear indication that including transitive relations induced by the nonhierarchical components of the ontology results in relevance propagation models that are superior to more basic approaches.