Campus Event Calendar
Event Entry
What and Who
Incorporating Positional and Contextual Information into a Neural IR Model
Andrew Yates
MMCI
Joint Lecture Series
AG 1, AG 2, AG 3, AG 4, AG 5, SWS, RG1, MMCI
Public Audience
English
Note: We use this to send email in the morning.
Date, Time and Location
Wednesday, 7 March 2018
12:15
60 Minutes
E1 5
002
Saarbrücken
Abstract
Ad-hoc information retrieval models consider query-document interactions to produce a document relevance score for a given query and document. Traditionally, such interactions have been modeled using handcrafted statistics that generally count term occurrences within a document and across a collection. Recently, neural models have demonstrated that they provide the instruments necessary to consider query-document interactions directly, without the need for engineering such statistics.
In this talk, I will describe how positional term information can be represented and incorporated into a neural IR model. The resulting model, called PACRR, performs significantly better on standard TREC benchmarks than previous neural approaches. This improvement can be attributed to the fact that PACRR can learn to match both ordered and unordered sequences of query terms in addition to the single term matches considered by prior work. Using PACRR's approach to modeling query-document interactions as a foundation, I will describe how several well-known IR problems can be addressed by incorporating contextual information into the model; the resulting Co-PACRR model significantly outperforms the original PACRR model. Finally, I will provide a brief look inside the model to illustrate the interpretability of the learned weights and to investigate how match signals are combined by the model to produce a query-document relevance score.
In this talk, I will describe how positional term information can be represented and incorporated into a neural IR model. The resulting model, called PACRR, performs significantly better on standard TREC benchmarks than previous neural approaches. This improvement can be attributed to the fact that PACRR can learn to match both ordered and unordered sequences of query terms in addition to the single term matches considered by prior work. Using PACRR's approach to modeling query-document interactions as a foundation, I will describe how several well-known IR problems can be addressed by incorporating contextual information into the model; the resulting Co-PACRR model significantly outperforms the original PACRR model. Finally, I will provide a brief look inside the model to illustrate the interpretability of the learned weights and to investigate how match signals are combined by the model to produce a query-document relevance score.
Contact
Jennifer Müller
2900
--email hidden
passcode not visible
logged in users only
Anna Rossien, 02/05/2018 13:58 -- Created document.