What are the points? What are the stances? 
Decanting for question-driven retrieval and executive summarization 
 
Jean-François Delannoy  
University of Ottawa 
Ottawa, Ontario, K1N 6N5 
Canada 
delannoy@site.uottawa.ca 
 
 
 
Abstract 
Decanter illustrates a heuristic 
approach to extraction for information 
retrieval and question answering. 
Generic information about 
argumentative text is found and 
stored, easing user-focused, question-
driven access to the core information.  
The emphasis is placed on the 
argumentative dimension, to address 
in particular three types of questions: 
“What are the points?”, “Based on 
what?” “What are the comments?”. 
The areas of application of this 
approach include: question-answering, 
information retrieval, summarization, 
critical thinking and assistance to 
speed reading. 
1 
1.1 
1.2 
Introduction 
Decanter is a prototype to detect and display 
high-level information from argumentative text. 
The game is one of situating and 
contextualizing. 
Queries and Requests 
Information requests can be classified by types 
of questions, bearing for example on: descriptive 
knowledge (“tell me about Pakistan”), 
narratives/updates (“what happened in Camp 
David?”), know-how (“how can I replace the ink 
cartridge on my XYZ printer?”), evaluation or 
advice (“Is Netscape 6 stable?”; “Should I install 
Netscape 6?”). 
 
One can take them on face value or not. In 
explicitly argumentative, and in loaded topics 
(like politics) it is in the interest of the user to 
have elements of context in the cognitive 
modeling he/she is doing of the text contents.  
 
Paying due attention to argumentation 
contributes in two ways: 
- by giving contexts to answers, helping 
qualify them for credibility 
- By answering to questions about 
opinions and stances: what  
Levels of Answering: on topic, on 
question, with justifications (and 
references), with a stance 
Level zero is answering on topic. This has been 
the only concern of “classical IR” (and still, 
word-sense disambiguation is not quite there 
yet…). 
 
Level one of question-answering is then to 
answer to the point semantically or 
pragmatically (depending of what kind of 
information need there is, relevance is of a 
different nature: in a nutshell, answering a 
practical question can require action-oriented 
information, but answers a la AskJeeves talking 
of travel agents when one just wants the distance 
from Paris to London are waylaid). As I stress 
heavily in my IR course (Delannoy 2001c)  
answers, and summaries alike, have to address 
relations, not just concepts. Answers should not 
just be “about” the keywords, but give the right 
kind of information: the height, the name, the 
colour, the description rather than the price, etc. 
(in many cases, a wholesale description may be 
judged satisfactory, but the user incurs a post-
filtering overhead). 
 
There is another dimension, though: context, in 
a broad sense. Context includes: who gives the 
answer; on what medium; what it the answer 
based on in terms of auctorial or demonstration; 
is it convincing for other reasons. A valuable 
answer is one given with good … reason: the 
answer should be rational, i.e. plausible, 
checkable, supported by authority of source 
and/or good demonstration. There is here an 
idea of critical thinking. 
 
Critical thinking (Toulmin 84, Little et al. 89, 
Mendenhall 90; Aristotle’s Organon) is the 
study of formal argumentation, and of what can 
be accepted reasonably in not-so-formal 
argumentation (what dose of induction; what 
auctoritas). This is often met with derision by 
various brands of relativism in and outside 
academia, although people of this suasion too 
play the argumentation game: they offer 
elements of proof; they rarely fling totally non-
logical-looking rhetoric. But bottom lines there 
are, and, in the apt words of the title of Little et 
al. 1989, good reasoning matters! 
 
How to track it then? The next section is about 
the “decanting” done by our prototype; Section 
3 is about retrieving what has been decanted, via 
questioning. 
2 
3 
3.1 
Decanting 
The input consists of one or several texts, by one 
or several authors and possibly mentioning 
several “actors” (who are also often “utterers”, 
but ).  
 
The general workflow is the following. 
 
1. segment 
2. extract entities, in particular the actors 
3. detect utterances 
4. analyze them argumentatively in a 
simple way: links of claims-evidence, 
evidence-evidence (contribute, contrast) 
5. infer underlying goals and values (e.g. 
prioritizing equity over efficiency; 
immediate goals vs stability) 
6. detect polarity: for, against 
7. link authors to utterances (who said 
what), and to points of view (what are 
the stances) 
 
Topics are registered in a knowledge base (e.g. 
economics, war, elections) and issues (most 
efficient course of action, objective 
measurement of income or turnover, objectivity 
of declarations by public figures, etc.). It is 
considered to implement a module of semi-
automatic acquisition: the user, prompted with 
lists of potential entries, would select and edit 
them for incorporation into the knowledge base. 
 
The output comes in multiple forms, as selected 
by the user: 
- list of entities 
- main structure of the claims 
- quotes 
- marked-up text (entities, reasoning) 
- table of points and stances (and their 
holders) on the issues at hand. 
- extractive summary based on claims rather 
than evidence) and, classically, position, 
importance cues and keyword density 
(Delannoy et al. 98). 
Representation 
Keys  
Three keys give viewing angles on the 
information: actors, topics, and issues, 
correspond to basic factual questions a reader 
may have (Table 1). 
key attributes question 
An 
actor 
utters quotes  What did X say? 
 has stances on 
issues 
What does X think 
on I, i.e. is s/he for 
or against? 
 and stances on 
courses of action 
What does X 
advocate/propose/
support? 
 has previsions What does X think 
Will happen? 
 prioritizes/foregr
ounds some 
values over 
others 
What are 
(allegedly) X's 
foremost values? 
a 
topic 
involves issues What are the 
issues? 
 involves courses 
of action 
What are the 
possible courses 
of action? 
an 
issue 
involves actors 
situated pro and 
contra 
What are the 
comments? 
 
Table 1. Keys, attributes, and related questions 
3.2 
3.3 
4 
4.1 
4.2 
4.3 
Background knowledge  
Some knowledge is pre-encoded or reused from 
previous processing, and some is built during 
the analysis. 
 
For repeated analysis of texts on the same topic, 
the knowledge built can of course be reused. 
 
- list of topics and issues, and the 
corresponding heuristics used to 
determine which are expected to be 
relevant to a given text 
- list text types, and associated heuristics 
- values: e.g. equity, egalitarianism, vital 
minimum/income, safety, ethnic identity, 
personal freedom, access to information, 
democracy 
Knowledge built with the 
processing: 
- actors in the input text; other entities 
- quotes in the text; their association with 
actors 
- claims 
- evidence 
- association of actors with claims, evidence 
Processing 
The general working is the following. 
Situate and segment the text 
- guess text topic, from keywords situating 
know topics; this is done easily 
- segment the text into clauses (the various 
clauses of the same utterance are then 
linked) 
Extract elements 
- extract entities, in particular the actors 
- detect utterances 
Assign relations:  
- articulate utterance components (main 
relations: evidence-of, support, contrast) 
- assign entity-to-utterance relations (who 
said what, textually) 
- polarities (who is for/contra what; 
including the author of the document) 
- infer underlying goals and values (e.g. 
identifying, if possible, whether an author 
prioritizes equity over efficiency; 
immediate goals vs stability) 
- link authors to utterances (who said what), 
and to points of view (what are the 
stances). 
 
The program uses a small knowledge base 
about the known topics (e.g. economics, war, 
elections) and issues (most efficient course of 
action, objective measurement of income or 
turnover, objectivity of declarations by public 
figures, etc.).  
 
The processing uses heuristic rules and pattern-
matching to recognize syntactic-semantic 
patterns, e.g.: 
- entities regular expressions 
- cues to topic  
- syntactic patterns of direct and report 
speech, to assign quotes 
- cues to polarity 
- argumentation operators. 
 
It is being considered to implement a module of 
semi-automatic acquisition: the user, prompted 
with lists of potential entries, would select and 
edit them for incorporation into the knowledge 
base. 
4.4 
5 
Querying/Questioning 
Various questions can be asked and answered 
using the structures produced, and especially: 
 
- What? -> What are the points made? 
- Why so? ->  What are the justifications?  
- What are the points of view or comments? 
(including of the authors themselves) 
Example 1: Results of Decanting 
(actual example) 
From a simple input: 
 
Ehud Barak, the Israeli president, said "we 
want peace". 
He added: "This is our main goal." 
"We want peace too", OLP Leader Arafat 
answered. 
Arafat added that Barak said that Israel 
may pull out of Gaza. 
Because Barak and Arafat have different 
standpoints, the peace process is fragile, 
even though they both want a peaceful 
resolution. 
 
we derive the following structures. 
 
ACTORS AND QUOTES 
 
Context1 
ref: textname="text1", utterer="John Doe", date 
="19990101" 
  { 
    Barak [assert]: "we want peace" 
    Barak [assert]: "this is our main goal" 
    Arafat [assert] "we want peace too". 
 
  Context2 { utterer="Arafat" 
  Barak [assert] "Israel may pull out of Gaza" 
  } 
The peace process will take time [cause_from] 
Barak and Arafat have different standpoints. 
The peace process will take time[detract] Barak 
and Arafat want a peaceful resolution. 
} 
 
NB The utterer of the last assertions is the 
author of the input text. If we process multiple 
texts, we have to indicate it explicitly (author 
name  
 
STANCES 
peace [pro] Barak 
peace [pro] Arafat 
 
PREDICTIONS 
John Doe [predict] the peace process will take 
time 
 
As of the submission of this article, the 
prototype detects the quotes but not the stances 
and contexts (which functionalities are under 
development). 
6 Example 2: “What Are the 
Comments?” (manual study)   
This example is to indicate the kind of 
comparative output targeted (but not 
implemented as yet), and the series of linguistic 
and modeling difficulties involved in producing 
it. It is based on an excerpt from a BBC bulletin 
board linked, at the time, from news.bbc.co.uk, 
called “BBC Talking point”, at 
http://newsvote.bbc.co.uk/hi/english/talking_poi
nt. 
The case in point was the desirable attitude 
towards the participation of Jörg Haider’s 
Freedom Party (FPÖ) in Austria in a 
governmental coalition in February 2000. 
 
Notes on the table 
- ID: numbering, for convenience 
- No d-author (author of the page or article) is 
mentioned, as all the texts in this example 
are from the same page.  
- Author: author of the comment; 
identification if free (may be a pseudonym) 
- Statement:  original statement 
- Marked up statement: statement after 
insertion of argumentation tags 
- Summary, manual: freely rephrased (there is 
also a summary from the BBC editor, which 
we do not mention here)  
- Arguments: main justifications, rephrased 
- Orientation:: here, by convention, pro means 
“for” Haider’s mandate and against 
sanctions; NOT necessarily in favour of 
Haider and his party. 
 
id Author Statement Summary, 
manual 
Arguments Orie
nt. 
Notes 
1 Nico C. K., 
Austria  
The EU is neither 
justified nor allowed 
to isolate Austria. 
Austria is, after all, a 
full member in good 
standing of the EU and 
its new government 
has not actually 
committed any acts 
contrary to EU 
principles. If the EU 
starts policing its 
members over the 
outcome of due 
democratic process, 
who will police the EU 
when it gets out of 
hand? 
Sanctions are 
not justifiable, 
as Austria is a 
legitimate 
member and 
has done 
nothing 
wrong.. 
- sanctions are not 
justified nor legal 
- Austria is a member of 
the EU in good standing  
- no devious acts 
- Austria is master at 
home 
- counterfactual: if EU 
at large becomes 
devious, who will 
control it? 
p Pb: the core point maybe 
less noteworthy or quote-
worthy than a justification 
of it. 
2 Jason H., 
USA 
I do not believe the 
EU and America are 
over reacting. I think 
they see the Haider 
phenomenon as a 
"virus" that might 
infect other more 
important parts of 
Europe if it does not 
react strongly now to 
"quarantine" it.  
Haider is like a 
virus. Yes, 
isolate Austria  
virus epidemics 
metaphor 
p The author, implicitly, 
adopts the advice of 
quarantining the 
(metaphorical) virus or 
virus-bearer.  
3 Ron, USA The E.U. is over-
reacting. There are two 
basic ideas Brussels 
does not get, freedom 
and liberty. If this 
"political censorship" 
is carried out, let us 
remember it came 
from the left and not 
the right. The EU is 
doomed if these 
sanctions are carried 
out. Brussels should 
let the people decide 
for once.  
Freedom at 
national level 
has 
precedence. 
  
This amounts to 
political censorship. 
- Decisions should not 
come from outside or 
above. 
p Positively loaded terms: 
freedom, liberty. (In fact, 
rather redundant; and 
semantically pliable). 
Negatively loaded: 
political censorship; 
curiously, appears quoted. 
Paradox, from implicit 
knowledge that the left is 
normally more principled 
about liberty than the 
right. 
Implicit: The EU 
administration is often not 
heeding much other levels 
of decision.  
7 Jaya N., 
India 
I think the EU has 
reacted responsibly 
and followed through 
on its earlier 
statements. When one 
country acts in such a 
way as to promote 
leaders with outright 
prejudice, the rest of 
the Nations must do all 
in their power to 
subdue further action. 
The EU is 
right, and has 
been acting 
consistently, 
because this is 
a clear case of 
prejudice. 
Austria (or the FPÖ) is 
prejudiced 
c Loaded: “outright 
prejudice.” 
Reasoning from general 
(“when one country”) to 
particular. 
Rem: fails to distinguish 
between prejudice in the 
FPÖ’s policy and 
supposed prejudice of the 
country as such or in 
majority. 
 
7 
8 
8.1 
8.2 
8.3 
8.4 
Evaluation / Commentary 
This is prototype work, but several original 
functionalities are already giving results: 
- characterizing the topic, based on 
discriminating keywords – i.e. the system 
makes good guesses among a dozen topics 
including economics/finance, economic 
policy, conflict, social/labour relations, 
culture, electoral politics… 
- from the topic, predicting typical issues on 
which stances articulate: for example for 
economic policy, one may expect stances 
about deregulation, globalization, interest 
rates, etc. 
- extracting quotes in direct speech gives 
60% good results; on indirect speech, this 
goes down to about 40%. 
- stance assignment works at about 50% 
success (good positives).  
 
Entity-extraction is not particularly original, like 
finding entities, classifying them, detecting 
naming equivalences for the entities. 
Related work 
Philosophy and Critical Thinking 
Books on critical thinking (Little et al. 89, 
Mendenhall 90) use representations of argument 
structures (e.g. as diagrams) but give no hint as 
how to automate it, i.e to go from text to model. 
Linguistics and NLP 
While research in linguistics has addressed 
several brands of “discourse analysis” as 
dialogue pragmatics and the search for 
underlying “ideology” or values, there is little in 
general linguistics about the study of 
argumentation proper. 
 
Simone Teufel (1999) performs “argumentative 
zoning” on research papers, finding types of 
passages like: aim, background, own research, 
continuation. The result is a colour-coded 
display of the input, based on an XML markup. 
Bayes and ngrams are used to perform this 
classication task. (Interestingly, she finds good 
agreement between manual annotators, vs 
various research in summarization failing to 
detect “golden standard” summaries.) This is 
argumentation in a rather specialized (scientific 
research in AI, i.e., largely, innovation in 
problem-solving) and shallow (no collation of 
the points themselves; one-level) sense. In 
contrast, Decanter is designed to deliver a 
representation of conclusions and justifications, 
from several uttererers in parallel or in a nested 
fashion if applicable. 
 
Some work on summarization, in particular by 
Daniel Marcu  (Marcu 97) has looked at the 
"rhetoric" dimension of text, based on RST 
(Mann&Thompson 88). It produces a detailed 
and high-quality tree representing the 
articulation of the text, but it is qualitatively a 
hybrid: it does not separate argumentation from 
mere description or narration. The detailed user 
study and modeling done in (Endres-
Nieggemeyer 97) gives little place to 
argumentation tracking in the summarization 
process. 
 
(Barker et al. 94) process rules and examples 
legal text to produce a semantic output then fed 
to a machine learning system doing 
generalization and abstraction. Yet it does not 
consider contexts of utterance.  
Information retrieval 
Information has focused even less on 
argumentation. As indicated above, answering 
on-topic is useful, but often the user is in fact 
looking for information which answers a 
question, which is situated, and which may 
involve opinions. We know of no work in 
argumentation-based IR – all the overhead of 
high-level filtering of argument being left to the 
user. 
Knowledge Representation and 
automated reasoning 
Some authors in computational linguistics have 
approached contexts. Ballim & Wilks do 
knowledge representation with nested contexts 
with Fauconnier’s mental spaces. Moulin uses 
conceptual graphs to represent spatio-temporal 
contexts from text. (Recently, a student project 
in his department has addressed argumentation, 
it seems, but information is scarce). Recently, a 
contributor to the CG list, L. Misek-Falkoff, 
asked for tools to represent nested contexts in 
tort/defamation; there were some answers 
pointing to tools, but not to tools capable of 
doing this. 
 
Various studies of reasoning, on the legal 
domain like (Bench-Capon 97) or more general 
like (Zukerman et al. 99), represent 
sophisticated reasoning, without performing 
extraction from text. 
  
(Delannoy 99) proposed an XML mark-up 
scheme for argumentation as such, the idea 
being to flag it inside the text besides producing 
a separate representation. Decanter is designed 
to do both. 
9 
9.1 
9.2 
9.3 
10 
11 
Future work 
Further work is intended to address a variety of 
robustness and scope issues, including reference 
resolution (neglected in IR) and the detection of 
lexicalized irony in the expression of stances. 
More Manual Analysis 
I am currently working on digesting several 
argumentative corpora on 
- issues of drug legalization (in Delannoy 
2001b) 
- Colombia (Plan Colombia, conflict, 
violence) 
- the Digital Divide (i.e. low access to the 
Internet by segments of Canadian or world 
population). 
Reference resolution 
This is another neglected topic in IR. Even 
medium-quality reference resolution would 
enhance performance in IR, including in our 
approach. 
Indirect argumentation and irony 
Indirect argumentation, especially irony 
Irony is an ingredient of rhetoric and can be of 
use in tracking stance on topics, stances on other 
actors, and also style of course. In another study  
(Delannoy 2001b) I observe the alternating use 
of irony and indignation. Besides the direct 
interest as a study rhetoric, it shows the variance 
of one factor of enunciation the socio-
psychological attitude, while the doxastic-
epistemic attitude stays aligned (the stance). 
From an IR point of view, one could try to 
differentiate ironic from non-ironic passages; 
also to normalize them into a “just the stance” 
form – a desalination device of sorts! 
Conclusion 
IR and NLP should pay due attention to 
question-focused information of course, but to 
other textual elements participating in the value 
of the returns, both 1) when it gives a useful 
characterization of the usability of the answer – 
as plausible, corroborated, demonstrated, novel, 
etc. 2) to begin to answer questions never 
addressed in IR and CL but definitely pervasive 
in user needs, either easily phrasable, in the 
style: “Is Netscape a good tool?”, “Is it 
advisable to buy Microsoft stock soon?”, or as a 
more underlying information goal: “So, what is 
Le Monde saying about the new developments 
of Plan Colombia and about the political 
reactions?”. This second type can be useful both 
to interested layme and to professionals of 
information and politics.  
 
Moreover, a matrix presentation as in example 2 
can be quite useful and reusable. That is, to be 
even more useful, argumentation analysis should 
integrate information retrieval + analysis + 
aggregation.  
 
In a Baconian vein: The information retriever 
and questioner has to use Invention (IR 
techniques) and Judgment (critical thinking) to 
tap into Memory (writing, library science) and 
Tradition (corpus of knowledge, opinions). 
Decanter opens the way to the necessary 
contribution of Judgment in Invention. 
 
References 
Aristotle. Organon. 
Ballim A & Wilks. 1991: Artificial Believers, 
Lawrence Erlbaum Associates 
K. Barker, JF Delannoy, S. Matwin, S. 
Szpakowic. 1994: "From text to Horn clauses", 
Proceedings of the Canadian Conference on 
Artificial Intelligence (AI/GI/CV '94), Banff, 
Alberta, Canada, March 1994, pp. 9-16  
Bench-Capon, T.J.M. 1997. Argument in 
Artificial Intelligence and Law Artificial 
Intelligence and Law, Vol 5 No 4., 1997, pp249-
61.  
Delannoy JF. 1999. "Argumentation Mark-
Up: A Proposal", Workshop "Towards 
Standards and Tools for Discourse Tagging", 
Conference of the Association for 
Computational Linguistics (ACL'99), U. 
Maryland, College Park, MD, June 22, 1999 
Delannoy JF. 2001b.  "Arguing about drugs", 
OSSA 2001 (conference of the Ontario Society 
for the Study of Argumentation), Windsor, 
Ontario, May 2001 
Delannoy JF. 1200c. course material for 
CSI4107, Information Retrieval and the 
Internet, University of Ottawa: 
http://www.site.uottawa.ca/~delannoy/csi4107 
Endres-Nieggemeyer, Brigitte. 1997. 
Summarizing Information, Springer, 1997 
Fauconnier, G. 1985.  Mental Spaces: 
Aspects of meaning construction in natural 
language. MIT Press, Cambridge, MA. 
Little J, Groarke L,Tindale C. 1989. Good 
reasoning matters!, McClelland&Stewart 
Mann, W. & Thompson, S. 1988.  
“Rhetorical structure theory: Towards a 
functional theory of text organization”, Text 
8(3), 241-281 
Marcu, Daniel. 1997. The Rhetorical 
Parsing, Summarization, and Generation of 
Natural Language Texts, Ph.D. Dissertation, U 
Toronto, 1997 
Mendenhall V. 1990. Une introduction à 
l'analyse du discours argumentatif. Presses de 
l'Université d'Ottawa, 1990  
Teufeul, S.  1999.  Argumentative Zoning. 
Information Extraction from Argumentative 
Text. Ph.D. Thesis, U. Edimburg, 1999 
Toulmin S, Riek R, Janik A. 1984. An 
Introduction to Reasoning, MacMillan 
Zukerman, I., McConachy, R., Korb, K. and 
Pickett, D. 1999. Exploratory Interaction with a 
Bayesian Argumentation System. IJCAI99 
Proceedings (16th International Joint 
Conference on Artificial Intelligence), pp. 1294-
1299, Stockholm, Sweden, Morgan Kaufmann 
Publishers 
