Mani, Inderjeet (2001) Automatic Summarization. John Benjamins,
xi+285pp, paperback ISBN 1-58811-060-5, Natural Language Processing, 3.
Michael McQuaid, Center for the Management of Information, University
of Arizona.
The publisher's announcement appeared in
http://linguistlist.org/issues/12/12-1735.html.
SUMMARY
Inderjeet Mani provides a concise textbook for automatic summarization
(AS). Summarizing a textbook on summarization offers a worthy challenge
and shows how useful the book can be. A detailed description of the
introductory chapter and a brief description of each subsequent chapter
follows.
CHAPTER 1, Preliminaries
The first chapter of this textbook merits considerable attention for
the framework developed there and followed throughout the book.
Mani avoids the cognitive view of all documents as summaries of the
world, viewing documents to be summarized as reflections of the world.
He further narrows his scope to automatic document summarization by
treating human summarization only for its value in informing
techniques, and establishes a model of condensation for a particular
reader or reader type in a specific role or task.
Summarizer systems produce condensed representations for human
consumption. Defining summarizers in this way distinguishes AS from
related fields including text compression, document retrieval,
indexing, information extraction, text mining, and question answering.
A discussion of summarizers must address the extent of human
participation in summarizing, ranging from supervision to no
involvement (fully automatic), where Mani focuses on fully automatic
summarization.
Kinds of condensation include extraction, where a summary contains only
material from the source, and abstraction, where new material may be
introduced. Abstraction may be performed to replace reading the source
(informative abstraction) or to guide selection of source material for
complete reading (indicative abstraction).
Condensation may be characterized by measuring its information content,
either by a semantic approach, concentrating on the relationship
between propositions in the summary and propositions in the source, or
by an information theoretic approach, measuring how fully the source
can be reconstructed from the summary.
Other ways to characterize a summarizer include coherence (or tolerance
of incoherence), span (distinguishing single document summarization
from multi-document summarization (MDS)), language (which may include
monolingual, multilingual, cross-language, and sub-language), genre
(whether different strategies are used for genres such as news, email,
editorials), and media (whether the source contains media other than
text and the summarizer output uses these media).
Mani's structure for the remainder of the book can best be understood
in terms of a generic model for a summarizer, containing three stages
between source and summary: analysis, transformation, and synthesis.
Each stage has parameters: analysis is generic or user focused,
transformation is indicative or informative, and synthesis is
fragmentary or connected. This model reflects an emphasis on
abstraction over extraction, since the transformation stage matters
only for the former operation.
Any stage of the model may employ three condensation operators,
selection, aggregation, and generalization, whereby more complicated
operations can be expressed. These operators can be applied to any
elements, including traditional linguistic elements such as sub-word,
word, phrase, and sentence, as well as paragraphs and documents.
Elements may be represented at different levels, including
morphological, syntactic, semantic, and discourse.
These levels imply classification of summarizers into shallow
(syntactic analysis, extracts, integration for coherence), and deep
(semantic or discourse, typically using structured data from specific
domains).
Mani finishes the new material in the first chapter of the textbook
with an appropriately brief discussion of current applications of
automatic summarization. I say "appropriately" because the increasing
pace of progress in this area dooms any attempt at a definitive summary
of current applications. Indeed, this section might make a
good springboard for a student project in reviewing current
applications.
Each chapter provides a conclusion and a review. These sections are of
widely differing lengths and detail, and make a good first pass at the
book and its viewpoint.
Subsequent chapters treat the following topics (number of pages in
parentheses): professional (human) summarizing (16), extraction (29),
revision (13), discourse-level information (35), abstraction (36),
multi-document summarization (MDS) (37), multimedia summarization (11),
and evaluation (35). In general, these page lengths provide good
indications of the book's emphases.
CHAPTER 2, Professional summarizing
Mani motivates the discussion of abstraction by examining professional
human summarizers, and includes the most extensive empirical model to
date, Endres-Niggemeyer (1998), for human summarizer abstraction:
document exploration, relevance assessment, and summary production.
What features of human abstraction might be relevant for AS? The
above-mentioned study notes some important behaviors, guiding later
chapters toward examining discourse-level representations, constructed
by human abstractors from structural information. Human abstractors
use representations at syntactic, semantic, and discourse levels in
specific ways, so all three levels are important.
CHAPTER 3, Extraction
Mani situates extraction within its limitations, including the variable
compression rate needed for sources of differing lengths, comparison in
multi-document summarization, and the prevalence of abstracting in
human (presumably benchmark) summarization.
He then describes extraction in enough detail for students to construct
some useful software, building on the framework of Edmundson (1969),
which scores the "extract - worthiness" of each sentence of a source
text based on the presence of four independent features. The original
research called these features: cue words (words indicating extract -
worthiness based on analysis of source texts), title words (words from
titles, subtitles, and headings of the source text), key words
(frequently occurring words in a source text that were not cue words or
in a stop word list), and sentence location.
Generalizations of these four features and their close relatives
dominate extraction models ever since. Mani describes various
outgrowths of this model, both in a technical and evaluative sense, and
includes a case study. He concludes this chapter by addressing the
important problem of coherence of extracts.
CHAPTER 4, Revision
This brief chapter treats three shallow methods for addressing the
coherence problem in extraction. These methods involve simple
procedures, easily implemented, but not providing solutions that easily
stand alone. For instance, Mani asks how much revision can be
accomplished by methods that inherently contain no machinery for
building an understanding of what's being talked about.
CHAPTER 5, Discourse-level information
This difficult chapter draws two concepts from the discourse
literature, text cohesion and text coherence, Morris and Hirst (1991);
and describes techniques for their use, as well as techniques for using
genre-specific document structure.
Text cohesion can be thought of as analogous to the word frequency
measures mentioned earlier, except that instead of measuring the
frequency of occurrence of a unit of text, the summarizer mentions the
"frequency" with which it is connected to other units of text, either
by grammatical or lexical means.
Text coherence requires much more of the chapter to treat. Mani begins
with an example of a coherence relation between sentences (although
these relations may occur between clauses as well) and some desiderata
for a theory of such relations. He then lists some of the existing
theories, such as rhetorical structure theory, coherence relations,
discourse grammar, macrostructures, and Toulmin structure. Some of the
names of relations from these theories, such as contrast, elaboration,
and evaluation, give a flavor of the kind of relations described.
Mani details coherence relations and, to a greater extent, rhetorical
structure theory, including a case study using rhetorical structure
theory for summarization and showing how to work with cue phrases,
described in earlier chapters.
CHAPTER 6, Abstraction
Mani compares abstraction methods using a three stage framework
beginning with building semantic representations of text (analysis),
using the previously established framework of selection, aggregation,
and generalization operations to develop new representations
(transformation), and finishing by rendering the new representations in
natural language (synthesis).
Much of the chapter consists of detailing template, term-rewriting,
event-relation, and concept hierarchy approaches. Of particular
interest are the sections on natural language generation of summaries.
CHAPTER 7, Multi-document summarization
The book's longest chapter treats the summarization of collections of
related documents, occasioned in large part by the advent of the World
Wide Web. This specialization of summarization adds the burden of
removing redundancy and dealing with similarities and differences
between sources. Compression is a bigger issue for collections, as are
the differences in source sizes. Summary may not be the best strategy
to depict some relations between documents, and sentence extraction may
well use up available compression too quickly to be useful.
Cross-document fusion, combining the previously mentioned elimination,
aggregation, and generation operations across the collections, forms an
umbrella for much of the discussion. Mani lists numerous possible
relationships across documents, such as contradiction, summary, and
fulfillment.
Mani introduces a generic five-stage multi-document summarization
algorithm, more specialized than the generic model presented in the
first chapter, and catalogs refinements of it by linguistic level, e.g.,
morphological or syntactic. The algorithm stages are (1) identify text
elements to be extracted, (2) match instances of elements across
sources, (3) filter matched elements using some salience criterion, (4)
reduce matched elements using aggregation and generalization
operations, and (5) present results using generation or visualization.
Mani delves into examples at various levels, including a case study.
CHAPTER 8, Multimedia summarization
The books shortest chapter provides a brief overview of some current
developments in summarization of video, dialog, and diagrams, as well
as production of multimedia briefings from multi-document
summarization.
CHAPTER 9, Evaluation
Evaluation was the topic for Mani's invited talk at the Automatic
Summarization Workshop at NAACL 2001, suggesting a certain stature in
the community, valuable for a discussion that necessarily includes
subjective judgments and issues in the philosophy of science.
This chapter includes some evaluation of different evaluation methods
(!), including intrinsic methods (judging a system), such as quality
evaluation, comparison against reference output, fidelity to source
comparison, and summarizer component testing, as well extrinsic methods
(judging the use of system) such as relevance assessment, reading
comprehension, presentation strategy evaluation, and mature system
evaluation.
CRITICAL EVALUATION
Mani's stature in evaluation has been noted, so of all the parts of a
summary of a summarization book, this is the most humbling.
There are not other obvious candidates for comparison as a textbook on
automatic document summarization. It's possible to use collections of
papers, but the goals differ and the price of the paperback edition
(29.95 USD) compares favorably with, for instance, Mani's previous
compendium, Mani and Maybury, Eds. (1999), and price means a good deal
to students.
The structure and layout of the book works well, although there are few
screenshots and little feel for user interaction with summarization
systems. The review sections accompanying each chapter provide an easy
way to skim before a full reading, and the materials seem very
complete, including useful case studies.
Evaluation here is then less a comparison and more a list of what
instructors and students who use this text will need in addition.
Happily, most of the cited research is well represented on the Web. The
briefest area, multimedia summarization, is, not coincidentally, the
easiest to augment from web-based sources. This reviewer, and possibly
many students would like to see code and screenshots, but would also
admit that adding such ephemera might consume energy disproportionate
to the added value.
One minor suggestion would be that this reviewer believes that the
brief treatment of TextTiling, Hearst (1997), in Chapter 5, section
2.3.1, suffers by ignoring the role of the paragraph in producing
results and comparing the algorithm to human judgment. Further, this
section and section 2.3.3 may have benefited by mentioning the work of
Allan et al. (2000), characterizing the difficulty in recognizing new
topics in a topic segmentation task.
BIBLIOGRAPHY
Allan, J., Lavrenko, V., and Jin, H. (2000) "First story detection in
TDT is hard". Proceedings of the Ninth Conference on Information
Knowledge Management 374--381.
Edmundson, H. P. (1969) "New methods in automatic abstracting". Journal
of the Association for Computing Machinery 16 (2):264--285.
Endres-Niggemeyer, B. (1998) Summarizing Information. Berlin: Springer.
Hearst, M. (1997) "TextTiling: Segmenting text into multi-paragraph
subtopic passages". Computational Linguistics 23(1):33--64.
Mani, I. and Maybury, M. T., Eds. (1999) Advances in Automatic Text
Summarization. Cambridge, Massachusetts: MIT Press.
REVIEWER BIOGRAPHY
Michael McQuaid is a research scientist at the Center for the
Management of Information, where he investigates knowledge management
questions, especially information visualization for managerial
sensemaking using domain-specific ontologies and symbologies. He is
ABD at the Department of Management Information Systems, the University
of Arizona.
|
