Maria-Hendrike Peetz

Using temporal bursts for query modeling

We present an approach to query modeling that leverages the temporal distribution of documents in an initially retrieved set of documents. In news-related document collections such distributions tend to exhibit bursts. Here, we define a burst to be a time period where unusually many documents are published. In our approach we detect bursts in result lists returned for a query. We then model the term distributions of the bursts using a reduced result list and select its most descriptive terms. Finally, we merge the sets of terms obtained in this manner so as to arrive at a reformulation of the original query. For query sets that consist of both temporal and non-temporal queries, our query modeling approach incorporates an effective selection method of terms. We consistently and significantly improve over various baselines, such as relevance models, on both news collections and a collection of blog posts.

Adaptive temporal query modeling

We present an approach to query modeling that uses the temporal distribution of documents in an initially retrieved set of documents. Such distributions tend to exhibit bursts, especially in news-related document collections. We hypothesize that documents in those bursts are more likely to be relevant and update the query model with the most distinguishing terms in high-quality documents sampled from bursts. We evaluate the effectiveness of our models on a test collection of blog posts.

Cognitive temporal document priors

Temporal information retrieval exploits temporal features of document collections and queries. Temporal document priors are used to adjust the score of a document based on its publication time. We consider a class of temporal document priors that is inspired by retention functions considered in cognitive psychology that are used to model the decay of memory. Many such functions used as a temporal document prior have a positive effect on overall retrieval performance. We examine the stability of this effect across news and microblog collections and discover interesting differences between retention functions. We also study the problem of optimizing parameters of the retention functions as temporal document priors; some retention functions display consistent good performance across large regions of the parameter space. A retention function based on a Weibull distribution is the preferred choice for a temporal document prior.

Semantic document selection: historical research on collections that span multiple centuries

The availability of digitized collections of historical data, such as newspapers, increases every day. With that, so does the wish for historians to explore these collections. Methods that are traditionally used to examine a collection do not scale up to todays collection sizes. We propose a method that combines text mining with exploratory search to provide historians with a means of interactively selecting and inspecting relevant documents from very large collections. We assess our proposal with a case study on a prototype system.

Estimating Reputation Polarity on Microblog Posts

We find that reputation polarity of a post is different from sentiment.We model reputation polarity using feature classes from communication theory.We introduce new features based on the replies to a post.We propose different ways to operationalise the RepLab 2012 and 2013 tasks. In reputation management, knowing what impact a tweet has on the reputation of a brand or company is crucial. The reputation polarity of a tweet is a measure of how the tweet influences the reputation of a brand or company. We consider the task of automatically determining the reputation polarity of a tweet. For this classification task, we propose a feature-based model based on three dimensions: the source of the tweet, the contents of the tweet and the reception of the tweet, i.e., how the tweet is being perceived. For evaluation purposes, we make use of the RepLab 2012 and 2013 datasets. We study and contrast three training scenarios. The first is independent of the entity whose reputation is being managed, the second depends on the entity at stake, but has over 90% fewer training samples per model, on average. The third is dependent on the domain of the entities. We find that reputation polarity is different from sentiment and that having less but entity-dependent training data is significantly more effective for predicting the reputation polarity of a tweet than an entity-independent training scenario. Features related to the reception of a tweet perform significantly better than most other features.