Probabilistic Topic Models are widely applied in many NLP-related tasks due to their effective use of unlabeled data to capture variable dependencies. Analytical solutions for Bayesian inference of such models, however, are usually intractable, hindering the proposition of highly expressive text models. In this scenario, Variational Auto-Encoders (VAEs), where an inference network (the encoder) is used to approximate the posterior distribution, became a promising alternative for inferring latent topic distributions of text documents. These models, however, also pose new challenges such as the requirement of continuous and reparameterizable distributions which may not fit so well the true latent topic distributions. Moreover, inference networks are prone to component collapsing, impairing the collection of coherent topics. To overcome these problems, we propose two new text topic models based on the categorical distribution Gumbel-Softmax (GSDTM) and on mixtures of Logistic-Normal distributions (LMDTM). We also provide a study on the impact of different modeling choices on the generated topics, observing a trade-off between topic coherence and document reconstruction. Through experiments using two reference datasets, we show that GSDTM largely outperforms previous state-of-the-art baselines when considering three different evaluation metrics.

Word recognition is a challenging task faced by many applications, specially in very noisy scenarios. This problem is usually seen as the transmission of a word through a noisy-channel, such that it is necessary to determine which known word of a lexicon is the received string. To be feasible, just a reduced set of candidate words are selected. They are usually chosen if they can be transformed into the input string by applying up to k character edit operations. To rank the candidates, the most effective estimates use domain knowledge about noise sources and error distributions, extracted from real use data. In scenarios with much noise, however, such estimates, and the index strategies normally required, do not scale well as they grow exponentially with k and the lexicon size. In this work, we propose very efficient methods for word recognition in very noisy scenarios which support effective edit-based distance algorithms in a Mor-Fraenkel index, searchable using a minimum perfect hashing. The method allows the early processing of most promising candidates, such that fast pruned searches present negligible loss in word ranking quality. We also propose a linear heuristic for estimating edit-based distances which take advantage of information already provided by the index. Our methods achieve precision similar to a state-of-the-art approach, being about ten times faster.

Memory-based Collaborative filtering solutions are dominant in the Recommendation Systems domain, due to their low implementation effort and service maintenance, when compared to Model-based approaches. Memory-based systems often rely on similarity metrics to compute similarities between items (or users) using ratings, in what is often named neighbor-based Collaborative filtering. This paper applies Fuzzy Fingerprints to create a novel similarity metric. In it, the Fuzzy Fingerprint of each item is described with a ranking of users ratings, combined with words obtained from the items' description. This allows the presented similarity metric to use fewer neighbors than other well-known metrics such as Cosine similarity or Pearson Correlation. Our proposal is able to reduce RMSE by at least 0.030 and improve NDCG@10 by at least 0.017 when compared with the best baseline here presented.

In this paper we present two new algorithms designed to reduce the overall time required to process top-k queries. These algorithms are based on the document-at-a-time approach and modify the best baseline we found in the literature, Blockmax WAND (BMW), to take advantage of a two-tiered index, in which the first tier is a small index containing only the higher impact entries of each inverted list. This small index is used to pre-process the query before accessing a larger index in the second tier, resulting in considerable speeding up the whole process. The first algorithm we propose, named BMW-CS, achieves higher performance, but may result in small changes in the top results provided in the final ranking. The second algorithm, named BMW-t, preserves the top results and, while slower than BMW-CS, it is faster than BMW. In our experiments, BMW-CS was more than 40 times faster than BMW when computing top 10 results, and, while it does not guarantee preserving the top results, it preserved all ranking results evaluated at this level.

State‐of‐the‐art search engine ranking methods combine several distinct sources of relevance evidence to produce a high‐quality ranking of results for each query. The fusion of information is currently done at query‐processing time, which has a direct effect on the response time of search systems. Previous research also shows that an alternative to improve search efficiency in textual databases is to precompute term impacts at indexing time. In this article, we propose a novel alternative to precompute term impacts, providing a generic framework for combining any distinct set of sources of evidence by using a machine‐learning technique. This method retains the advantages of producing high‐quality results, but avoids the costs of combining evidence at query‐processing time. Our method, called Learn to Precompute Evidence Fusion (LePrEF), uses genetic programming to compute a unified precomputed impact value for each term found in each document prior to query processing, at indexing time. Compared with previous research on precomputing term impacts, our method offers the advantage of providing a generic framework to precompute impact using any set of relevance evidence at any text collection, whereas previous research articles do not. The precomputed impact values are indexed and used later for computing document ranking at query‐processing time. By doing so, our method effectively reduces the query processing to simple additions of such impacts. We show that this approach, while leading to results comparable to state‐of‐the‐art ranking methods, also can lead to a significant decrease in computational costs during query processing.

Search engines are essential tools for web users today. They rely on a large number of features to compute the rank of search results for each given query. The estimated reputation of pages is among the effective features available for search engine designers, probably being adopted by most current commercial search engines. Page reputation is estimated by analyzing the linkage relationships between pages. This information is used by link analysis algorithms as a query‐independent feature, to be taken into account when computing the rank of the results. Unfortunately, several types of links found on the web may damage the estimated page reputation and thus cause a negative effect on the quality of search results. This work studies alternatives to reduce the negative impact of such noisy links. More specifically, the authors propose and evaluate new methods that deal with noisy links, considering scenarios where the reputation of pages is computed using the PageRank algorithm. They show, through experiments with real web content, that their methods achieve significant improvements when compared to previous solutions proposed in the literature.

The currently booming search engine industry has determined many online organizations to attempt to artificially increase their ranking in order to attract more visitors to their web sites. In the same time, the growth of the web has also inherently generated several navigational hyperlink structures which have a negative impact on the importance measures employed by current search engines. In this paper we propose and evaluate algorithms for identifying all these noisy links over the web graph, may them be spam or simple relationships between real world entities represented by sites, replication of content, etc. Unlike prior work, we target a different type of noisy link structures, residing at the site level, instead of the page level. We thus investigate and annihilate site level mutual reinforcement relationships, abnormal support coming from one site towards another, as well as complex link alliances between web sites. Our experiments with the link database of the TodoBR search engine show a very strong increase in the quality of the output rankings after having applied our techniques.