Advanced Information Retrieval within Blogosphere and Micro-Blogosphere.

摘要

Social media, such as blogs and microblogs (Twitter), gain popularity in people's daily life. The volumes of blogosphere and micro-blogosphere are rapidly growing, which triggers the application of information retrieval on social media to address the information needs of users. However, simple ad-hoc information retrieval cannot meet users' information needs over social media. An ad-hoc information retrieval system retrieves the documents that are relevant to queries. However, current users usually want to retrieve the social media documents, such as blog posts and tweets, that are not only relevant to queries but also satisfy some extra conditions. In this thesis, we study two such advanced information retrieval problems: faceted blog distillation over blogosphere and real-time tweet ad-hoc retrieval over micro-blogosphere.;Faceted blog distillation aims at retrieving blogs (i.e. RSS feeds) that are not only relevant to a given query but also satisfy a facet of interest. In this thesis, the facets under consideration are opinionated vs. factual, personal vs. official and in-depth vs. shallow. Opinionated blogs provide the blog posts that contain relevant opinions to queries while factual blogs consist of the blog posts that describe the topics of queries without opinionated contents. The blog posts in personal blogs depict the topics related to the personal experiences of bloggers while those in official blogs deliver the commercial purposes of bloggers. In-depth blogs have the blog posts that provide the deep analysis about the topics of interest while the posts in shallow blogs simply mention the topics, without analyzing the implications of the provided information. For the opinionated and factual facets, we propose a classifier using syntactic and semantical features to determine whether opinions are relevant to queries in the context of blog posts. For the personal and official facets, we propose two categories of methods to identify the personal and official posts. The first category of methods is based on classification. We propose three classifiers that are established based on the different assumptions about two research issues. The first issue is whether a blog post exhibiting the personal or official facet depends on the query topic it mentions. The second issue is whether the personal or official facet of a blog post is dependent on whether it exhibits the opinionated or factual facet respectively. The second category of methods is based on generative models. In specific, we first propose a generative model that calculates the probabilistic distributions of blog posts exhibiting topics and those of exhibiting the personal and official facets. Observing that the posts from a feed are likely to exhibit the same facet, we improve the first model by proposing a second generative model that constrains the post from a feed to have the same facet distribution in its generative process. For the in-depth and shallow facets, we propose to calculate the depth of the coverage of a blog post on a given query by the occurrences of the concepts that are related to the query within the post. We also discussed the relationships among the facets. For example, we validate that the personal or official facet of a post depends on its opinionated or factual facet respectively. Experimental results on the TREC Blogs06 collection and the TREC Blogs08 collection show that the proposed techniques are not only effective in finding facet-oriented blogs (or posts) but also significantly outperform the best known results reported over both collections.;Real-time tweet ad-hoc retrieval ranks relevant tweets to queries in reverse-chronological order of their publishing times. In the context of this problem, to respond to a query with a timestamp t, the retrieved tweets should satisfy the following three conditions: (a) relevant to the query, (b) published on or before time t, and (c) ranked in reverse-chronological order of their publishing times. In this thesis, we propose a two-phase approach where we retrieve tweets in an ad-hoc way during the first phase and then utilize the temporal information of queries and tweets to enhance the retrieval effectiveness of tweets during the second phase. Tweets can be categorized into two types. One type consists of short messages not containing any URL of a web page. The other type has at least one URL of a web page in addition to a short message. These two types of tweets have different structures. In the first phase, we propose a method to rank tweets using the divide-and-conquer strategy to address the structural difference of tweets. Specifically, we first rank the two types of tweets separately. This produces two rankings, one for each type. Then we merge these two rankings of tweets into one ranking. In the second phase, we first categorize queries into several types by exploring the temporal distributions of their top-retrieved tweets from the first phase; then we calculate the time-related relevance scores of tweets according to the classified types of queries; finally we combine the time scores with the IR scores from the first phase to produce a ranking of tweets. Experimental results achieved by using the TREC 2011 and TREC 2012 queries over the TREC Tweets2011 collection show that the proposed divide-and-conquer method of ranking tweets yields better retrieval effectiveness than ranking them simultaneously and our proposed incorporation of temporal information into retrieval process yields further improvements. Our method also compares favorably with state-of-the-art methods in retrieval effectiveness.