| dc.description.abstract | The goal of high-recall information retrieval (HRIR) is to find all,
or nearly all, relevant documents while maintaining reasonable assessment effort.
Achieving high recall is a key problem in the use of applications such as
electronic discovery, systematic review, and construction of test collections for
information retrieval tasks. State-of-the-art HRIR systems commonly rely on iterative relevance feedback in which
human assessors continually assess machine learning-selected documents.
The relevance of the assessed documents is then fed back to
the machine learning model to improve its ability to select the next set of
potentially relevant documents for assessment. In many instances, thousands of human assessments might be required to achieve high recall. These assessments represent the main cost of such HRIR
applications. Therefore, their effectiveness in achieving high recall
is limited by their reliance on human input when assessing the relevance of
documents. In this thesis, we test different methods in order to improve the effectiveness and
efficiency of finding relevant documents using state-of-the-art HRIR
system. With regard to the effectiveness, we try to build a machine-learned
model that retrieves relevant documents more accurately.
For efficiency, we try to help human assessors make
relevance assessments more easily and quickly via our HRIR system.
Furthermore, we try to establish a stopping criteria for the
assessment process so as to avoid excessive assessment.
In particular, we hypothesize that total assessment effort to achieve high
recall can be reduced by using shorter document excerpts
(e.g., extractive summaries) in place of full documents for the assessment of
relevance and using a high-recall retrieval system based on continuous active
learning (CAL). In order to test this hypothesis, we implemented a
high-recall retrieval system based on state-of-the-art implementation of CAL. This high-recall retrieval system could display
either full documents or short document excerpts for relevance assessment.
A search engine was also integrated into our system to provide
assessors the option of conducting interactive search and judging.
We conducted a simulation study, and separately, a 50-person controlled user study to test our hypothesis.
The results of the simulation study show that judging even a single
extracted sentence for relevance feedback may be adequate for CAL
to achieve high recall. The results of the controlled user study
confirmed that human assessors were able to find
a significantly larger number of relevant documents within limited time when they used the
system with paragraph-length document excerpts as opposed to full documents.
In addition, we found that allowing participants to compose and execute their
own search queries did not improve their ability to find relevant
documents and, by some measures, impaired performance.
Moreover, integrating sampling methods with active
learning can yield accurate estimates of the number of relevant documents, and thus avoid excessive assessments. | en |
