Position-based contextualization for passage retrieval

Abstract

We present a novel contextualization approach for passage retrieval. The core principle is to let any occurrence of a query term in a document affect the passage retrieval score, whether the occurrence is in the passage or not. This effect is controlled by the distance between the term occurrence and the passage. Empirical evaluation demonstrates the merits of our approach; the resultant retrieval performance substantially transcends that of previously proposed passage retrieval methods, including those that use various contextualization approaches.

Retrieving Passages and Finding Answers

“…In this experiment we use a Gaussian kernel function with σ = 2000, as it is shown to be an effective choice [4]. Beside the tf.idf scoring model, we also use the estimated pseudo-frequencies in a query likelihood model using Dirichlet smoothing.…”

On the Effectiveness of Contextualisation Techniques in Spoken Query Spoken Content Retrieval

“…The kernel function is commonly parametrised by a propagation parameter σ which adjusts the influence that a term occurrence has over distant positions. Among these, the Gaussian kernel exp(−(j − i) 2 /2σ 2 ) has been shown effective in previous studies [10,6,8].…”

“…In recent work, PMs were proposed as a contextualisation technique for passage retrieval [6]. In this work, a standard TFIDF approach was used to compute the relevance score of a passage p within document D where the frequency of term t in p is given by its pseudo-frequency estimate:…”

“…The process of taking context into account when computing the relevance score of an element is known as contextualisation [3]. Various contextualisation techniques have been proven effective in XML retrieval [3], passage retrieval [6,8], and SCR [12,15] tasks. In these techniques, elements are scored depending not only on the query terms occurring within the element itself but also on those occurring in other positions within the document.…”

“…Two simple and widely adopted approaches are interpolating the scores of an element with those of its document to consider global context [12], or with those from a fixed number of surrounding elements to consider local context [15]. In contrast, techniques based on positional models (PM) [6,8] allow consideration of longer-spans of context ignoring element boundaries.…”