Description
TitleCorroborating information from multiple sources
Date Created2016
Other Date2016-10 (degree)
Extent1 online resource (xii, 127 p. : ill.)
DescriptionInformation on the Internet is abundant but often inaccurate. Given a query that has a unique answer (as opposed to a Web query against a search engine), different Web sources might provide multiple conflicting answers. As a result, users are left with the burden of validating the correctness of the answer from each source. In order to tackle this problem, corroboration techniques have been proposed in order to identify the correct answer given a set of candidate answers extracted from the sources. Corroboration is the technique that evaluates the quality of the answers by considering the trustworthiness of the sources from which the answers are extracted. Intuitively, an answer extracted from a trustworthy source is more likely to be the correct answer. In return, the more correct answers it reports, the more trustworthy a source is. Unfortunately, several challenges arise before we can successfully apply a corroboration technique to find the correct answer to a query. First of all, the prime challenge is how to evaluate the trustworthiness of the sources and henceforth derive the quality of an answer based on the sources reporting it. Secondly, in a case where all the sources agree on a single candidate answer, how to validate the correctness of the answer. Third, in an application in which each source only provides a partial answer and the final answer is a combination of partial answers from multiple sources, how to evaluate the quality of answers and how to efficiently compute the correct answer. This thesis investigates several real world problems and proposes novel corroboration techniques that address each of the challenges presented above.
We first studied the problem of using corroboration for the task of question answering. With many web sources providing conflicting information on the Internet, users often have to rummage through a large number of different sites to both retrieve the information and ascertain the correctness of the retrieved information. While a naive approach that returns the most frequent answer can eliminate outlier answers such as typos, it fails to consider the fact that answers extracted from different pages are rarely equally important in answering the query. By ranking the answers based on the number, relevance and similarity of the web sources reporting them, as well as the prominence of the answers within the sources, our algorithm is able to efficiently identify accurate answers for most queries. We investigated the problem of verifying the correctness of claims that are unanimously agreed upon among all sources. Intuitively, a claim supported by all the sources must be true, simply because there is no other source rebutting it. However, it might not be the case in real world scenarios since agreeing sources might be out-dated or due to copy/paste. In such a scenario, existing corroboration approaches tend to reach consensus quickly and conclude that all claims are true since there is little conflict among the sources. We studied this problem in a real world scenario (restaurant listings) and proposed a novel corroboration algorithm that evaluates the claims on a gradual basis. More specifically, our approach divides the claims into multiple sets and evaluates each set of claims using a different trust score from each source. Different from existing algorithms that assign a single trust score to each source, our approach computes a set of finer-grained trust scores for each source that is used to evaluate different set of claims. In real world scenarios there often exist queries in which a single source is insufficient to provide a candidate answer. To answer these queries, users have to fetch and combine information from multiple sources and derive a potential final answer. Such cases are similar to the case of finding air ticket between two points without direct flight, and differs in that there is no centralized source (e.g., Expedia.com) that provides the information of all connecting flights. The process of combining information from two sources is similar to the join operation in relational databases and therefore this problem can be viewed as a join query processing over multiple web-accessible databases. The main bottleneck of join query processing is tuple accessing of web databases, which typically exhibit high and variable latency. In order to find the top-k answers for a join query, a branch-and-bound algorithm has to be developed to avoid computing scores of all candidates exhaustively. Our method efficiently computes bounds for partial query results and determines a good order in which it accesses the tables so as to minimize wasted efforts in the computation of top-k answers. In summary, this thesis studies real world problems that involve information from multiple sources. We demonstrate that using information from a single source is often of low quality and in some cases insufficient. We discuss the challenges in each individual problem and present novel corroboration algorithms that efficiently compute scores for the answers by taking into consideration of the trustworthiness of the sources.
NotePh.D.
NoteIncludes bibliographical references
Noteby Minji Wu
Genretheses, ETD doctoral
Languageeng
CollectionGraduate School - New Brunswick Electronic Theses and Dissertations
Organization NameRutgers, The State University of New Jersey
RightsThe author owns the copyright to this work.