A frequent problem when dealing with data gathered from multiple sources on the web (ranging from booksellers to Wikipedia pages to stock analyst predictions) is that these sources disagree, and we must decide which of their (often mutually exclusive) claims we should accept. Current state-of-the-art information credibility algorithms known as "fact-finders" are transitive voting systems with rules specifying how votes iteratively flow from sources to claims and then back to sources. While this is quite tractable and often effective, fact-finders also suffer from substantial limitations; in particular, a lack of transparency obfuscates their credibility decisions and makes them difficult to adapt and analyze: knowing the mechanics of how votes are calculated does not readily tell us what those votes mean, and finding, for example, that a source has a score of 6 is not informative. We introduce a new approach to information credibility, Latent Credibility Analysis (LCA), constructing strongly principled, probabilistic models where the truth of each claim is a latent variable and the credibility of a source is captured by a set of model parameters. This gives LCA models clear semantics and modularity that make extending them to capture additional observed and latent credibility factors straightforward. Experiments over four real-world datasets demonstrate that LCA models can outperform the best fact-finders in both unsupervised and semi-supervised settings.

We introduce a new probabilistic model for transliteration that performs significantly better than previous approaches, is language-agnostic, requiring no knowledge of the source or target languages, and is capable of both generation (creating the most likely transliteration of a source word) and discovery (selecting the most likely transliteration from a list of candidate words). Our experimental results demonstrate improved accuracy over the existing state-of-the-art by more than 10% in Chinese, Hebrew and Russian. While past work has commonly made use of fixed-size n-gram features along with more traditional models such as HMM or Perceptron, we utilize an intuitive notion of "productions", where each source word can be segmented into a series of contiguous, non-overlapping substrings of any size, each of which independently transliterates to a substring in the target language with a given probability. To learn these parameters, we employ Expectation-Maximization (EM), with the alignment between substrings in the source and target word training pairs as our latent data. Despite the size of the parameter space and the 2(|w|-1) possible segmentations to consider for each word, by using dynamic programming each iteration of EM takes O(m^6 * n) time, where m is the length of the longest word in the data and n is the number of word pairs, and is very fast in practice. Furthermore, discovering transliterations takes only O(m^4 * w) time, where w is the number of candidate words to choose from, and generating a transliteration takes O(m2 * k2) time, where k is a pruning constant (we used a value of 100). Additionally, we are able to obtain training examples in an unsupervised fashion from Wikipedia by using a relatively simple algorithm to filter potential word pairs.

Much of the information on the Web is found in articles from online news outlets, magazines, encyclopedias, review collections, and other sources. However, extracting this content from the original HTML document is complicated by the large amount of less informative and typically unrelated material such as navigation menus, forms, user comments, and ads. Existing approaches tend to be either brittle and demand significant expert knowledge and time (manual or tool-assisted generation of rules or code), necessitate labeled examples for every different page structure to be processed (wrapper induction), require relatively uniform layout (template detection), or, as with Visual Page Segmentation (VIPS), are computationally expensive. We introduce maximum subsequence segmentation, a method of global optimization over token-level local classifiers, and apply it to the domain of news websites. Training examples are easy to obtain, both learning and prediction are linear time, and results are excellent (our semi-supervised algorithm yields an overall F1-score of 97.947%), surpassing even those produced by VIPS with a hypothetical perfect block-selection heuristic. We also evaluate against the recent CleanEval shared task with surprisingly good cross-task performance cleaning general web pages, exceeding the top "text-only" score (based on Levenshtein distance), 87.8% versus 84.1%.