Passage retrieval revisited

By Marcin Kaszkiel, Justin Zobel. 1997

Ranking based on passages addresses some of the shortcomings ofwhole-document ranking. It provides convenient units of text toreturn to the user, avoids the difficulties of comparing documentsof different length, and enables identification of short blocks ofrelevant material amongst otherwise irrelevant text. In this paperwe explore the potential of passage retrieval, based on anexperimental evaluation of the ability of passages to identifyrelevant documents. We compare our scheme of arbitrary passageretrieval to several other document retrieval and passage retrievalmethods; we show experimentally that, compared to these methods,ranking via fixed-length passages is robust and effective. Ourexperiments also show that, compared to whole-document ranking,ranking via fixed-length arbitrary passages significantly improvesretrieval effectiveness, by 8% for TREC disks 2 and 4 and by18%-37% for the Federal Register collection.

Subfields: Information Retrieval & Search

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Information Retrieval & Search

Citations in the corpus

Pretrained Transformers for Text Ranking: BERT and Beyond (2020, methodology): “trieval literature. There is a long thread of work, dating back to the 1990s, that leverages passage retrieval techniques for document ranking [Hearst and Plaunt, 1993, Callan, 1994, Wilkinson, 1994, Kaszkiel and Zobel, 1997, Clarke et al., 2000]—that is, aggregating passage-level evidence to estima”
Context-Aware Document Term Weighting for Ad-Hoc Search (2020, methodology): “Meanwhile, prior research shows that fixed-size passages of 200-300 wordsmore are effective than natural passages [20].”
Modeling Diverse Relevance Patterns in Ad-hoc Retrieval (2018, methodology): “Note here we focus on document retrieval using passage-level signals, and will not include the work taking passages as retrieval units [16].”
UMass at TREC 2004: Novelty and HARD (2004, background): “Previous work has been inconclusive as to whether there is beneﬁt to retrieving passages of different lengths [13, 3, 25].”
Relevance ranking for one to three term queries (1997, background): “Apart from few exceptions (Kaszkiel & Zobel, 1997; Knaus, Mittendorf, SchaÈ uble, & Sheridan, 1995), most passage retrieval methods divide a document into passages at the time it is added to the database (Allan, 1995; Callan, 1994; Hearst & Plaunt, 1993; Salton, Allan, & Buckley, 1993; Wilkinson,…”
Cross-Domain Modeling of Sentence-Level Evidence for Document Retrieval (2019, background): “This approach, in fact, is well motivated: There is a long thread of work in the information retrieval literature, dating back decades, that leverages passage retrieval techniques for document ranking (Hearst and Plaunt, 1993; Callan, 1994; Kaszkiel and Zobel, 1997; Clarke et al., 2000).”
A Critique and Improvement of an Evaluation Metric for Text Segmentation (2002, background): “Text segmentation issues are also important for passage retrieval, a subproblem of information retrieval (Hearst and Plaunt, 1993; Salton, Allan, and Buckley, 1993; Callan, 1994; Kaszkiel and Zobel, 1997).”
Overview of Touché 2022: Argument Retrieval (2022, unknown): “Using the TREC CAsT tools 19 ), we split the documents at the sentence boundary into fixed-length passages of approximately 250 terms since working with fixed-length passages is more effective than variable-length original passages [41].”
When will information retrieval be "good enough"? (2005, methodology): “Although our truth data used variable-length passages, fixed-length passage retrieval is at least as good as variablelength, and much easier to implement [8, 2, 12].”
Is searching full text more effective than searching abstracts? (2009, background): “As a start, I have experimented with two different indexing granularities (full articles and spans), but alternative approaches include sliding windows [18,20], multi-para-graph segments [19], hierarchically-overlapping segments [38,39], and segments based on topic shifts [17].”