| dc.description.abstract | With the massive amount of textual data being produced every day,
the ability to effectively summarise text documents is becoming increasingly
important. Automatic text summarization entails the selection
and generalisation of the most salient points of a text in order
to produce a summary. Approaches to automatic text summarization
can fall into one of two categories: abstractive or extractive approaches.
Extractive approaches involve the selection and concatenation
of spans of text from a given document. Research in automatic
text summarization began with extractive approaches, scoring and
selecting sentences based on the frequency and proximity of words.
In contrast, abstractive approaches are based on a process of interpretation,
semantic representation, and generalisation. This is closer
to the processes that psycholinguistics tells us that humans perform
when reading, remembering and summarizing. However in the sixty
years since its inception, the field has largely remained focused on
extractive approaches.
This thesis aims to answer the following questions. Does knowledge
about the discourse structure of a text aid the recognition of
summary-worthy content? If so, which specific aspects of discourse
structure provide the greatest benefit? Can this structural information
be used to produce abstractive summaries, and are these more
informative than extractive summaries? To thoroughly examine these
questions, they are each considered in isolation, and as a whole, on
the basis of both manual and automatic annotations of texts. Manual
annotations facilitate an investigation into the upper bounds of
what can be achieved by the approach described in this thesis. Results
based on automatic annotations show how this same approach
is impacted by the current performance of imperfect preprocessing
steps, and indicate its feasibility.
Extractive approaches to summarization are intrinsically limited
by the surface text of the input document, in terms of both content
selection and summary generation. Beginning with a motivation
for moving away from these commonly used methods of producing
summaries, I set out my methodology for a more human-like
approach to automatic summarization which examines the benefits of
using discourse-structural information. The potential benefit of this
is twofold: moving away from a reliance on the wording of a text
in order to detect important content, and generating concise summaries
that are independent of the input text. The importance of
discourse structure to signal key textual material has previously been
recognised, however it has seen little applied use in the field of autovii
matic summarization. A consideration of evaluation metrics also features
significantly in the proposed methodology. These play a role in
both preprocessing steps and in the evaluation of the final summary
product. I provide evidence which indicates a disparity between the
performance of coreference resolution systems as indicated by their
standard evaluation metrics, and their performance in extrinsic tasks.
Additionally, I point out a range of problems for the most commonly
used metric, ROUGE, and suggest that at present summary evaluation
should not be automated.
To illustrate the general solutions proposed to the questions raised
in this thesis, I use Russian Folk Tales as an example domain. This
genre of text has been studied in depth and, most importantly, it has a
rich narrative structure that has been recorded in detail. The rules of
this formalism are suitable for the narrative structure reasoning system
presented as part of this thesis. The specific discourse-structural elements
considered cover the narrative structure of a text, coreference
information, and the story-roles fulfilled by different characters.
The proposed narrative structure reasoning system produces highlevel
interpretations of a text according to the rules of a given formalism.
For the example domain of Russian Folktales, a system is implemented
which constructs such interpretations of a tale according to
an existing set of rules and restrictions. I discuss how this process of
detecting narrative structure can be transferred to other genres, and
a key factor in the success of this process: how constrained are the
rules of the formalism. The system enumerates all possible interpretations
according to a set of constraints, meaning a less restricted rule
set leads to a greater number of interpretations.
For the example domain, sentence level discourse-structural annotations
are then used to predict summary-worthy content. The results
of this study are analysed in three parts. First, I examine the relative
utility of individual discourse features and provide a qualitative
discussion of these results. Second, the predictive abilities of these
features are compared when they are manually annotated to when
they are annotated with varying degrees of automation. Third, these
results are compared to the predictive capabilities of classic extractive
algorithms. I show that discourse features can be used to more
accurately predict summary-worthy content than classic extractive algorithms.
This holds true for automatically obtained annotations, but
with a much clearer difference when using manual annotations.
The classifiers learned in the prediction of summary-worthy sentences
are subsequently used to inform the production of both extractive
and abstractive summaries to a given length. A human-based
evaluation is used to compare these summaries, as well as the outputs
of a classic extractive summarizer. I analyse the impact of knowledge
about discourse structure, obtained both manually and automatically,
on summary production. This allows for some insight into the knock
on effects on summary production that can occur from inaccurate discourse
information (narrative structure and coreference information).
My analyses show that even given inaccurate discourse information,
the resulting abstractive summaries are considered more informative
than their extractive counterparts. With human-level knowledge
about discourse structure, these results are even clearer.
In conclusion, this research provides a framework which can be
used to detect the narrative structure of a text, and shows its potential
to provide a more human-like approach to automatic summarization.
I show the limit of what is achievable with this approach both
when manual annotations are obtainable, and when only automatic
annotations are feasible. Nevertheless, this thesis supports the suggestion
that the future of summarization lies with abstractive and not
extractive techniques. | en_US |
