Temporal Summarization (TS) is a new track introduced as part of the Text REtrieval Conference (TREC) in 2013. This track aims to develop systems which can return important updates related to an event over time. In TREC 2013, the TS track specifically used disaster related events such as earthquake, hurricane, bombing, etc. This thesis mainly focuses on building an effective TS system by using a combination of Information Retrieval techniques. The developed TS system returns updates related to disaster related events in a timely manner. By participating in TREC 2013 and with experiments conducted after TREC, we examine the effectiveness of techniques such as distributional similarity for term expansion, which can be employed in building TS systems. Also, this thesis describes the effectiveness of other techniques such as stemming, adaptive sentence selection over time and de-duplication in our system, by comparing it with other baseline systems. The second part of the thesis examines the current methodology used for evaluating TS systems. We propose a modified evaluation method which could reduce the manual effort of assessors, and also correlates well with the official track’s evaluation. We also propose a supervised learning based evaluation method, which correlates well with the official track’s evaluation of systems and could save the assessor’s time by as much as 80%.

Search and summarization of streaming social media, such as Twitter, requires the ongoing analysis of large volumes of data with dynamically changing characteristics. Tweets are short and repetitious -- lacking context and structure -- making it difficult to generate a coherent synopsis of events within a given time period. Although some established algorithms for frequent itemset analysis might provide an efficient foundation for synopsis generation, the unmodified application of standard methods produces a complex mass of rules, dominated by common language constructs and many trivial variations on topically related results. Moreover, these results are not necessarily specific to events within the time period of interest. To address these problems, we build upon the Linear time Closed itemset Mining (LCM) algorithm, which is particularly suited to the large and sparse vocabulary of tweets. LCM generates only closed itemsets, providing an immediate reduction in the number of trivial results. To reduce the impact of function words and common language constructs, we apply a filltering step that preserves these terms only when they may form part of a relevant collocation. To further reduce trivial results, we propose a novel strengthening of the closure condition of LCM to retain only those results that exceed a threshold of distinctiveness. Finally, we perform temporal ranking, based on information gain, to identify results that are particularly relevant to the time period of interest. We evaluate our work over a collection of tweets gathered in late 2012, exploring the efficiency and filtering characteristic of each processing step, both individually and collectively. Based on our experience, the resulting synopses from various time periods provide understandable and meaningful pictures of events within those periods, with potential application to tasks such as temporal summarization and query expansion for search.

Search and summarization of streaming social media, such as Twitter, requires the ongoing analysis of large volumes of data with dynamically changing characteristics. Tweets are short and repetitious -- lacking context and structure -- making it difficult to generate a coherent synopsis of events within a given time period. Although some established algorithms for frequent itemset analysis might provide an efficient foundation for synopsis generation, the unmodified application of standard methods produces a complex mass of rules, dominated by common language constructs and many trivial variations on topically related results. Moreover, these results are not necessarily specific to events within the time period of interest. To address these problems, we build upon the Linear time Closed itemset Mining (LCM) algorithm, which is particularly suited to the large and sparse vocabulary of tweets. LCM generates only closed itemsets, providing an immediate reduction in the number of trivial results. To reduce the impact of function words and common language constructs, we apply a filltering step that preserves these terms only when they may form part of a relevant collocation. To further reduce trivial results, we propose a novel strengthening of the closure condition of LCM to retain only those results that exceed a threshold of distinctiveness. Finally, we perform temporal ranking, based on information gain, to identify results that are particularly relevant to the time period of interest. We evaluate our work over a collection of tweets gathered in late 2012, exploring the efficiency and filtering characteristic of each processing step, both individually and collectively. Based on our experience, the resulting synopses from various time periods provide understandable and meaningful pictures of events within those periods, with potential application to tasks such as temporal summarization and query expansion for search.