A unified framework for event related information seeking
| Author(s) | Lu, Kuang | |
| Date Accessioned | 2021-03-25T12:18:01Z | |
| Date Available | 2021-03-25T12:18:01Z | |
| Publication Date | 2020 | |
| SWORD Update | 2021-02-08T17:03:55Z | |
| Abstract | Events of different types and scopes are happening every day and influencing nearly every aspect of people's life. In order to obtain information about the events to understand them, search is usually the tool to use. Therefore, it is crucial to provide adequate support for event related information. News and microblogs are two popular information sources to be searched. In this work, we focus on providing effective search techniques for event related queries. The challenges of event related information seeking on these two sources, as well as the inadequacy of the existing methods in terms of overcoming the challenges, are discussed. Based on this, we propose a unified framework to assist event related search interests on news and microblogs that not only includes novel techniques to address the challenges, but also bridge the retrieval on them to better satisfy the search interests, meaning to leverage background information from news to refine the results on microblogs. ☐ On news retrieval, we argue the importance of background information and that it is essential to provide dedicated background information retrieval support. Through analyzing background information from temporal and semantic perspective, a time filter and an aspect based background retrieval model are proposed. Experiments on the TREC News Track data set illustrate that these two methods can work in concert to provide statistically significant improvements over a competitive baseline. Further analysis demonstrates the usefulness of the individual components of the model, such as aspect identification and aspect language model estimation. ☐ On microblog retrieval, due to the lack of long term search interest support and the unsatisfactory pseudo relevance feedback performance, a novel relevance signal called query collectivity is proposed, which measures relevance by using the collective presence of multiple query terms. For long term search interests, this measure is used to detect the time periods (e.g. days) when there is no relevant information and informs the system to not return any results. However, when the measure is employed for pseudo relevance feedback, expansion terms are selected in a conservative fashion, and the measure is able to only select them for queries when there is a potential of improving the retrieve effectiveness. Experiment results on TREC Microblog data sets show that the measure achieves advantageous performances on both tasks. ☐ Besides concentrating on enhancing the retrieval for different genres individually for event related searches, we also propose to bridge the two types of retrieval by leveraging the background aspects from news to re-rank the initial retrieval results on microblogs on a per aspect basis to provide an option of exploring microblogs in finer granularity. We carefully design an experiment and its results suggest that the news aspects can represent meaningful event related search interests and, with the help of language models of the aspects that can be obtained from the background news retrieval, initial microblog results can be effectively re-ranked for the aspects. | en_US |
| Advisor | Fang, Hui | |
| Degree | Ph.D. | |
| Department | University of Delaware, Department of Electrical and Computer Engineering | |
| DOI | https://doi.org/10.58088/tnvb-bb15 | |
| Unique Identifier | 1243269035 | |
| URL | https://udspace.udel.edu/handle/19716/28860 | |
| Language | en | |
| Publisher | University of Delaware | en_US |
| URI | https://login.udel.idm.oclc.org/login?url=https://www.proquest.com/dissertations-theses/unified-framework-event-related-information/docview/2492271022/se-2?accountid=10457 | |
| Keywords | Unified framework | en_US |
| Keywords | Event related | en_US |
| Keywords | Information seeking | en_US |
| Title | A unified framework for event related information seeking | en_US |
| Type | Thesis | en_US |