Sentence retrieval is an important step in many question-answering (QA) technologies. However, characteristics of sentences and of the question-answering task itself often make it difficult to apply document retrieval techniques to sentence retrieval. The use of translation dictionaries offers one potentially useful approach to sentence retrieval, but training such dictionaries using QA corpora often introduces noise that can negatively impact retrieval performance. In this study, we experiment with using data elicited from assessors during interactions as training data for a translation dictionary. We employ two different interactions that elicit two types of data: data about assessors’ topics and data about retrieved sentences. Results show that using sentence-level relevance feedback to adjust the translation dictionary improved retrieval for about half the topics, but harmed it for the other half.
[1]
W. Bruce Croft,et al.
A Translation Model for Sentence Retrieval
,
2005,
HLT.
[2]
Xin Fu,et al.
The loquacious user: a document-independent source of terms for query expansion
,
2005,
SIGIR '05.
[3]
Mounia Lalmas,et al.
A survey on the use of relevance feedback for information access systems
,
2003,
The Knowledge Engineering Review.
[4]
Jimmy J. Lin,et al.
Overview of the TREC 2006 ciQA task
,
2007,
SIGF.
[5]
Robert Krovetz,et al.
Viewing morphology as an inference process
,
1993,
Artif. Intell..
[6]
Vanessa Murdock,et al.
Aspects of sentence retrieval
,
2007,
SIGF.
[7]
David Yarowsky,et al.
Statistical Machine Translation: Final Report
,
1999
.
[8]
John Cocke,et al.
A Statistical Approach to Machine Translation
,
1990,
CL.