Argumentation mining--an introduction for linguists--Fribourg--2019-09-02

Argumentation Mining: An
Introduction for Linguists
Linguistic and corpus perspectives on argumentative
discourse, SwissUniversities Doctoral Programme
Language & Cognition
University of Fribourg, Fribourg, Switzerland
2019-09-02

Today’s Lecture
• My book w/computational linguist Manfred
Stede: Argumentation Mining
• What is argumentation?
• Argumentation mining: a first look
• Argumentative language
• Challenges for argumentation mining
• Argumentation structures
• Corpus annotation
• Why study argumentation mining?

In Swissbib, search for “Synthesis
Lectures on Human Language
Technologies”:
https://www.swissbib.ch/Search/Res
ults?lng=en&type=ISN&lookfor=194
74040
Main Chapters
1. Introduction
2. Argumentative Language
3. Modeling Arguments
4. Corpus Annotation
5. Finding Claims
6. Finding Supporting and
Objecting Statements
7. Deriving the Structure of
Argumentation
8. Assessing Argumentation
9. Generating Argumentative Text
10. Summary and Perspectives
https://doi.org/10.2200/S00883ED1V01Y201811HLT040

Example argument
We really need to tear down that building.
Granted, it will be expensive, but the degree of
asbestos contamination is not tolerable
anymore. Also, it is one of the most ugly
buildings in town!
Stede & Schneider, Argumentation Mining, 1.1, page 6

Example argument
Tweet by @robeastaway
https://twitter.com/robeastaway/status/135838892694839296

An argument has a central claim
Claim: Jaffa Cakes are cakes
Justification: When stale, cakes go hard.
Was official EU ruling.

An argument supports the claim
Claim: Jaffa Cakes are cakes
Justification: When stale, cakes go hard.
Was official EU ruling.

“Argumentation is a verbal, social, and rational
activity aimed at convincing a reasonable critic
of the acceptability of a standpoint by putting
forward a constellation of propositions justifying
or refuting the proposition expressed in the
standpoint.”
van Eemeren and Grootendorst [2004, p. 1], as quoted by

A “constellation of propositions”
justifying the standpoint/central claim.

Argumentation mining: a first look

Possible applications of argumentation
mining
• Sensemaking
• Practical reasoning
• Argument Retrieval
• Web-Scale Discourse and Debate
• Sentiment Analysis
• Writing Support
• Essay Scoring
• Dialogue Systems
Stede & Schneider, Argumentation Mining, 1.3, pages 8-9 &
10.3, pages 140-143

Full-fledged argumentation mining
1. Identify argumentative text
2. Segment the text into argumentative discourse
units (ADUs)
3. Identify the central claim.
4. Identify the role/function of ADUs.
5. Identify relations between ADUs.
6. Build the overall structural representation.
7. Identify the type and the quality of the
argumentation.
Stede & Schneider, Argumentation Mining, 1.2, pages 6-7

Segment into
argumentative discourse units (ADUs)
• [We really need to tear down that building.]0
[Granted, it will be expensive,]1 [but the
degree of asbestos contamination is not
tolerable anymore.]2 [Also, it is one of the
most ugly buildings in town!]3

Identify the central claim
buildings in town!

Identify the role of the other ADUs
• [We really need to tear down that building.]0
[Granted, it will be expensive,]1 [but the
degree of asbestos contamination is not
tolerable anymore.]2 [Also, it is one of the
most ugly buildings in town!]3

Role: Anticipating a counterargument
buildings in town!

Role: Rebutting a counterargument
buildings in town!

Role: Supporting the central claim
buildings in town!

Identify the role of the other ADUs
• 0: main claim [We really need to turn down
that building.]
• 1: counterargument [Granted, it will be
expensive,]
• 2: rebuttal [but the degree of asbestos
contamination is not tolerable anymore.]
• 3: support [Also, it is one of the most ugly
buildings in town!]

Identify relations between ADUs
• 0: main claim [We really need to turn down that
building.]
• 1: counterargument [Granted, it will be
expensive,] ATTACKS 0
• 2: rebuttal [but the degree of asbestos
contamination is not tolerable anymore.]
UNDERCUTS THE COUNTERARGUMENT ATTACK
• 3: support [Also, it is one of the most ugly
buildings in town!] SUPPORTS 0

Build the overall structural
representation
Stede & Schneider, Argumentation Mining, Figure 1.1, page 7

Full-fledged argumentation mining
2. Segment the text into argumentative discourse
units (ADUs)
4. Identify the role/function of ADUs.
5. Identify relations between ADUs
argumentation.

Subjectivity
• ‘Objective’ statements are a matter of
intersubjective agreement or disagreement:
– There is a cat on the mat.
– Winston Churchill came to office in 1940.
• Private states are not
Stede & Schneider, Argumentation Mining, 2.1, page 11-12,
following [Wiebe et al., 2005], [Quirk et al., 1985]

Various types of private states
• Reveal an emotion: Hooray!
• Give an opinion: That’s a really bad wine.
• Make a judgment: You don’t deserve the
prize.
• Make a prognosis: There will be snow
tomorrow.
• Give an estimate or speculation: I guess that’s
a llama over there.

Sentiment & Opinion
• Sentiment: Various subjective utterances
(opinion, judgement, emotion)
• Opinion is narrower for us: Subjective
evaluations of some entity.

Arguments & Opinions
• Not all opinions are arguments. But if a reason
for an opinion is provided, we can analyze it as
an argument:
“I never enjoyed Wagner’s operas, as they are
so enormously overloaded.”

Verifiability
Verifiable: objective assertions (not about personal
feelings or interpretations)
– Verifiable-private: statements that concern the
speaker’s personal state or experience.
– Verifiable-public: public information is sufficient
to verify the statement: no personal state or
experience is involved.
Unverifiable: statements that cannot be proven
with objective evidence.
Park and Cardie [2014], cited by Stede & Schneider, Argumentation
Mining, 2.2, page 13-14

Verifiable-private examples
(Park and Cardie [2014])
My son has hypoglycemia.
They flew me to NY in February.
The flight attendant yelled at the passengers.

Supported differently in
argumentation
• Verifiable-public: can provide evidence:
“I tell you Winston Churchill came to office in 1940. I
saw it on Wikipedia!”
• Verifiable-private: may provide evidence, but this is
optional.
“I have a headache”.
Hearers are not in a position to question this, but the
speaker may go on to explain: “Maybe I had too much
wine last night”.
• Unverifiable: can give a reason as support:
“I don’t like this wine, because it has so much tannin”.

Implicit Opinions
• [camera review] The viewfinder is somewhat
dark.
• [hotel review] The rooms turned out to be
small.
Rajendran et al. [2016] Stede & Schneider, Argumentation
Mining, 2.2, page 15

Need domain knowledge
(“polar facts” imply an evaluation)
• The viewfinder is somewhat dark (+ A dark
viewfinder is considered bad).  I am not in
favor of the camera.
• The rooms turned out to be small (+ A small
room is considered bad).  I am not in favor
of the hotel.

Hedging
• “There is no clear symptom of diabetes.”
Stede & Schneider, Argumentation Mining, 2.1, page 12 & 2.2 page
14. See also Farkas et al. 2010.

Stance classification
Given a topic under debate, determine whether
a contribution is pro or con.
For example:
• “vaccination of children (should/should not)
be mandatory”
• “behavior X (is/is not) acceptable”
• “legislation X (should/should not) be dropped”

Classifying speech acts
• Representatives: Speaker commits to the truth of
an assertion.
• Directives: Speaker tries to make addressee
perform some action.
• Expressives: Speaker expresses an emotional
state.
• Declaratives: Speaker changes the state of the
world by means of performing the utterance.
• Commissives: Speaker commits to doing some
action in the future.
Stede & Schneider, Argumentation Mining, 2.2, page 17. See also
Searle [1976]

Performative verbs
• You shouldn’t read that book.
• I advise you not to read that book.
• What is your name?
• I ask you to tell me your name.
• Yesterday I met a philosopher.
• I assert that I met a philosopher yesterday.
• Max is the smartest kid in the world.
• I claim that Max is the smartest kid in the world.
• It’s about 15 degrees out there.
• I estimate that it is about 15 degrees out there.
Stede & Schneider, Argumentation Mining, 2.2, pages 16-17. See
also Austin [1975]

Components of a speech act
• Locutionary act: producing the linguistic
utterance by speaking, writing, gesturing.
• Illocutionary act: the intention or goal the
speaker has in mind when performing the act.
• Perlocutionary act: the effect that the
performance of the act has on the addressee.
Stede & Schneider, Argumentation Mining, 2.2, page 18. See also
Austin [1975] & Searle [1976]

Illocutionary force indicating devices
(IFID) [Searle, 1969]
• Performative verbs (e.g. advise, ask, assert,
claim, estimate, …)
• Sentence mode (declarative, interrogative,
imperative)
• Modal verbs
• Word order
• Intonation
• Stress
also Austin [1975] & Searle [1969] & Searle [1976]

Politeness & indirect speech acts
• “Can you pass the salt?”
• Apologies, with varying levels of indirectness
– I apologize for being late.
– Sorry I’m late.
– I’m a bit late, unfortunately.
– I’m afraid I didn’t quite make it on time.
e.g. Ogiermann [2009]

Argument, Explanation, and
Justification
• Using airplanes is really a bad idea because they
are among the worst air polluters we have ever
created. (Argumentation/Persuasion)
• An airplane is able to take off because the shape
of the wings produces an upward force when the
air flows across them. (Explanation)
• I need to use airplanes a lot because my job
requires me to be in different parts of the country
every week. (Justification)
Stede & Schneider, Argumentation Mining, 2.2, page 20.

Discourse Mode [Smith 2003]
• Narrative
• Description
• Report
• Information
• Argument
Stede & Schneider, Argumentation Mining, 2.2, page 20-22, See also
Smith [2003], Werlich [1975].

Discourse Mode [Smith 2003]
• Narrative: The passengers landed in New York in the middle
of the night and then moved on to Hoboken immediately.
• Description: Hundreds of people occupied the square. In
front of them, the speaker was standing on a small podium.
• Report: My sister visited the new exhibition yesterday.
• Information: Krypton is one of the noble gases. It is one of
the rarest elements on earth.
• Argument: The award was given to Paul, but he did not
deserve it. His work is very shallow.
Stede & Schneider, Argumentation Mining, 2.2, pages 20-22, See
also Smith [2003], Werlich [1975].

Situation Entity types help distinguish
argumentative vs. non-argumentative text
[Becker 2016]
• State: Armin has brown eyes.
• Event: Bonnie ate three tacos.
• Report: The agency said applications had increased.
• Generic sentence: Scientific papers make arguments.
• Generalizing sentence: Fei travels to India every year.
• Fact: Georg knows that Reza won the competition.
• Proposition: Georg thinks that Reza won the competition.
• Resemblance: Reza looks like he won the competition.
• Question: Why do you torment me so?
• Imperative: Listen to this.
Becker et al. [2016], Stede & Schneider, Argumentation Mining,
2.2, page 22. See also Song et al. [2017]

Rhetoric
• Logos: Speakers employ rules of sound
reasoning.
• Ethos: Speakers signal their authority or
credibility (or that of their source)
• Pathos: Speakers seek to communicate
their standpoint in a manner that seeks
to evoke an emotional response
Stede & Schneider, Argumentation Mining, 1.1, page 5, based on
Aristotle.

Rhetoric
• Logos: (Logic)
That building needs to be demolished, because it is full of
asbestos, which is known to be hazardous, and there is no way
to stop its diffusion from the different parts of the building.
• Ethos: (Authority/Credibility)
That building needs to be demolished, because it is full of
asbestos, as the report by the university engineers has shown.
• Pathos: (Emotion)
That building needs to be demolished, because it is an
irresponsible source of danger to the health and indeed the
life of our children who spend so many hours in those
poisonous rooms every day!
Stede & Schneider, Argumentation Mining, 1.1, page 5, based on
Aristotle.

Rhetoric studies choices
• Close the window! It’s cold.
• It is rather cold in here and I’m already not
feeling so well. Would you be so kind as to
close the window for me?
Stede & Schneider, Argumentation Mining, 2.6, page 23. See
[Fahnestock, 2011] and others for more on these sorts of choices

Rhetoric studies choices
• Small changes in word choice, sentence
construction, and passage construction
produce different rhetorical effects.
Stede & Schneider, Argumentation Mining, 2.6, page 23. See
[Fahnestock, 2011] and others for more on these choices

Rhetorical figures
• Hyperbole: Phew, the distance from Newark
to Manhattan is a hundred miles!
• Tautology: That nice restaurant was really
great.
Stede & Schneider, Argumentation Mining, 2.6, page 24.

Rhetorical figure mining
Rhetorical figures and mining has been studied
extensively by Harris, DiMarco, and colleagues.
See also “Harnessing rhetorical figures for
argument mining”, Lawrence, Visser, and Reed
[2017].
Stede & Schneider, Argumentation Mining, 2.6, page 24. See Harris
& DiMarco [2017], Harris et al. [2018] among others.

Rhetorical moves – Swales’ Create a
Research Space
Move 1 Establishing a territory
Step 1 Claiming centrality and/or
Step 2 Making topic generalization(s) and/or
Step 3 Reviewing items of previous research
Move 2 Establishing a niche
Step 1A Counter-claiming or
Step 1B Indicating a gap or
Step 1C Question-raising or
Step 1D Continuing a tradition
Move 3 Occupying the niche
Step 1A Outlining purposes or
Step 1B Announcing present research
Step 2 Announcing principal findings
Step 3 Indicating research article structure
Swales’ Create a Research Space – see also Stede & Schneider,
Argumentation Mining, 2.6, pages 24-25.

Rhetorical moves – Teufel’s
Argumentative Zoning
Teufel, Simone. "Scientific Argumentation Detection as Limited-domain Intention
Recognition." ArgNLP. 2014.– see also Stede & Schneider, Argumentation Mining, 2.6,
pages 24-25..
Teufel and Moens [2002] is widely used. A recent hierarchy:

Hierarchical Text Structure
A text can be viewed as a hierarchy of text spans
that are recursively connected via coherence
relations:
• Elaboration: [The new Smart Watch was
introduced today.] [It costs $50 more than the old
model.]
• Motivation: [ [The new Smart Watch was
introduced today.] [It costs $50 more than the old
model.] ] [You should really buy it.]

Rhetorical Structure Theory [Mann
and Thompson, 1988]
Stede & Schneider, Argumentation Mining, 2.6, pages 25-26. Figure 2.1
from [Stede, 2011, p.115]. See also [Mann and Thompson, 1988]

Hierarchical Text Structure resources
• Penn Discourse Tree Bank – relations are
annotated independently of one another. Now
in its third version, see
https://www.seas.upenn.edu/~pdtb/
• RST Discourse Treebank [Carlston et al., 2003]
• Discourse parsers [see Stede Discourse
Parsing, 2011, Section 4.4]
Stede & Schneider, Argumentation Mining, 2.6, pages 25-26.

Challenges for argumentation mining

– Is there a standpoint on which people disagree?
• Context or domain knowledge may be needed to
determine whether or not a text is argumentative (e.g.
“The rooms turned out to be small.”)
– Prosody can impact the meaning of a text.
• “Really?” vs. ”Really.” vs. “Really!”

2. Segment the text into argumentative
discourse units (ADUs)
– Argumentative units occur at multiple levels,
including clauses, sentences, and multi-sentence
groupings.
– Additionally, there are choices: arguments can be
analyzed at different granularities.

– Sometimes there are multiple plausible
interpretations.
– Argument structure is recursive, so there may be
multiple different arguments.

4. Identify the role/function of ADUs
– Support/attack may be implicit.
– Domain knowledge or context may be needed.
5. Identify relations between ADUs.
– Support/attack may be implicit.
– Domain knowledge or context may be needed.

– Possibility of missing information because
arguments are rarely fully specified.

argumentation.
– Current approaches look at relevance,
acceptability, sufficiency.
See also, Stede & Schneider, Argumentation Mining, 8.3.2,
pages 116-120

Kinds of argument
• epistemic: some proposition is true or false;
• ethical or esthetical: something is good or bad
(or: beautiful or ugly);
• deontic: some action should be done or not
done.
Eggs [2000], as quoted in Stede & Schneider,
Argumentation Mining, 1.1, page 3

Esthetical argument
• Of all Greenaway’s works (...), this is probably the British
filmmaker’s least effective. As with all of his films, the
choreography of people and objects before the camera (...)
is elaborate and splendid. The film also marks some of
Greenaway’s favorite thematic obsessions, including (in no
particular order) spiritual and corporeal rotting, Sir Isaac
Newton and arcane mullings on things historic, classical
and numerical.
• But Greenaway’s narrative and his direction of actors—two
elements which only recently has he concerned himself
with—are without foundation. After the effects of the
visual presentation have worn off, the film becomes rather
tiresome to follow.
Desson Howe, The Washington Post. Review of ‘The Belly of an Architect’, June 29, 1990.
http://www.washingtonpost.com/wpsrv/style/longterm/movies/videos/thebellyofanarchitectnrhowe_a0b289.htm ,
as cited in Stede & Schneider, Argumentation Mining, 1.3, page 3

Argument structures can be analyzed
at different levels and different
granularities

Argument structures are varied
Rahwan [2008] as cited in Stede & Schneider, Argumentation Mining, Figure 3.1, page 28

Argument structures can be analyzed
at different levels
Stede & Schneider, Argumentation Mining, Table 3.1, page 33, based on
Bentahar, Moulin, and Bélanger’s taxonomy of argumentation models, modified from
Bentahar et al. [2010, p. 215]

Arguments can be studied at different
granularities.
• Arguments are recursive, zippered structures.
• The conclusion of one argument may be the
premise of the next argument.

Granularity

Argument, Case, Debate
• Argument - “a one-step reason for a claim”
cannot be subdivided into any other parts that
are still arguments
• Case - “a chain of reasoning leading toward a
claim” has only supporting evidence for a
claim
• Debate - “reasons for and against a claim” has
both supporting and conflicting evidence
[Wyner et al., 2015, p. 51], as cited in Stede & Schneider, Argumentation Mining, 3.1, pages 28-29

Case example: only supporting
evidence for a claim

Debate example: has both supporting
and conflicting evidence
buildings in town!

Defeasible reasoning is an extension of
classical inference

Argumentation is defeasible.
Argumentation considers tentative conclusions,
which can be revised when new information
comes to light.
This is different from classical inference. In
classical logic, the conclusion is guaranteed to
hold whenever the premises hold.
Stede & Schneider, Argumentation Mining, 3.2 page 29

Classical inference rule:
syllogism
Socrates is a man; (Minor premise)
Every man is mortal; (Major premise)
Therefore, Socrates is mortal. (Conclusion)

modus ponens
Given if p, then q.
Given p.
Therefore, q.

modus tollens
Given if p, then q.
Given not q.
Therefore, not p.

Argumentation schemes
An argumentation scheme expresses a defeasible
inference rule for showing the acceptability of a
standpoint.
There are multiple approaches, including:
• Walton/Reed/Macagno: 60 schemes
• Pragma-dialectic school: 3 schemes (sign,
comparison, cause) with variations and subtypes
• Argumentum model of topics (AMT): maxims
(from topoi & loci) activate rules.

Argumentation Schemes – Position to
Know
• Major Premise: Source a is in a position to
know about things in a certain subject domain
S containing proposition A.
• Minor Premise: a asserts that A (in Domain S)
is true.
• Conclusion: A is true.
Walton, Reed, Macagno [2008] as cited in Stede & Schneider, Argumentation Mining, 3.3, page 30

Argumentation Schemes – Position to
Know
Critical Questions:
1. Is a in a position to know whether A is true?
2. Is a an honest (trustworthy, reliable) source?
3. Did a assert that A is true?
Walton, Reed, Macagno [2008] as cited in Stede & Schneider, Argumentation Mining, 3.3, page 30

AMT
Rigotti and Greco Morasso [2010, p. 499], as cited in Stede & Schneider, Argumentation Mining, Fig 3.2, page 32

Toulmin diagram
Toulmin [2008] as cited in Stede & Schneider, Argumentation Mining, Fig 3.3, page 34

Annotation scheme example:
Microtext scheme
[Peldszus and Stede, 2013] as cited by Stede & Schneider, Argumentation Mining, 4.1, page 45-47

Microtext scheme explained
•Grey boxes: text
•Round: proponent
•Square: opponent
•Arrow: support
•Circle/square: attack
[Peldszus and Stede, 2013] as cited by Stede & Schneider, Argumentation Mining, 4.1, page 45-47

Argumentation schemes on Wikipedia
Schneider, Samp, Passant, Decker, Arguments about Deletion: How Experience Improves the Acceptability of Arguments in Ad-hoc Online Task Groups. CSCW 2013.

Factors on Wikipedia
Schneider, Passant, Decker. Deletion Discussions in Wikipedia: Decision Factors and Outcomes. WikiSym 2012

Kirschner’s science scheme
Kirschner et al. [2015, p. 1]
as cited by Stede & Schneider, Argumentation Mining, 4.1, page 47-48

Annotation scheme example: Modified
Toulmin scheme
Habernal and Gurevych [2017, p. 144]

Annotation scheme example: Cornell
eRulemaking
Niculae et al. [2017, p. 985]

Inference Anchoring Theory
Budzyńska and Reed [2011]
IF=illocutionary force
RA=inference between
propositions
TA=dialogue rule
relating utterances

Callout/target annotation
[Ghosh et al., 2014, p. 39]
as quoted by Stede & Schneider, Argumentation Mining, Figure 4.8, page 54

Example Corpora: Online Interactions
• Internet Argument Corpus
• Agreement by Create Debaters
• Agreement in Wikipedia Talk Pages
• ComArg
• Technorati technical blogs
• Web Discourse

Agreement by Create Debaters corpus
example
[Rosenthal and McKeown, 2015, p. 169]

ComArg comment/argument
[Boltužić and Šnajder, 2014, p. 54]
s=implicit support; S=Explicit support

Other examples of corpora
• Araucaria
• Argumentative Microtext Corpus
• Cornell eRulemaking corpus
• Webis Editorial corpus
• Persuasive Essay Corpus

Why study argumentation mining?

Example work
• Quality and consistency assessment
• Visualization & summarization
• Writing assessment & support

Quality and consistency assessment

Quality and consistency assessment
• Assess:
– Is the argument sound?
– Is the argument convincing?
• When information conflicts, what is the
maximally consistent subset?
– Argumentation solvers (e.g. ASPARARTIX,
Carneades, GrappaViz, etc.) are designed to do
this, given a manual analysis or controlled natural
language
Stede & Schneider, Argumentation Mining, 1.3, pages 8-9 &
10.3, pages 140-143

Get argument structure from text
(1) Households should pay tax for their
garbage.
115
Arrow: premise
(4) (1)
Paying tax for garbage increases
recycling, so households should pay.
(3) (1)
Recycling more is good, so people
should pay tax for their garbage.
Adam Wyner and Tom van Enger. Towards Web-based Mass Argumentation in Natural
Language. (long version of EKAW 2010 poster).
http://wyner.info/research/Papers/WynerVanEngersEKAW2010.pdf

Recycling Debate manually extracted
from BBC “Have your Say”
116
Arrow: premise
Dashed arrow: attacks

Maximal consistent sets
117

Visualization
• Help an individual structure their thoughts
• Show the flow of topics in a debate
• Visualize the points of agreement and
disagreement

Dana Khartabil, Jessie Kennedy, & Simon Wells
http://paltry-ship.surge.sh

Summarization
• Automatically summarize:
– An online debate
– Government consultation
– A radio debate
– Voters’ preferences
– Relevant factors from customer reviews

Travis Kriplean. Consider.it Strategic Planning example:
https://consider.it/examples/strategic_planning

Diane Litman & colleagues & students. ArgRewrite: PITT Revision Writing Assistant
http://argrewrite.cs.pitt.edu/demo.html (see also http://argrewrite.cs.pitt.edu )

Argumentation mining--an introduction for linguists--Fribourg--2019-09-02

Recommended

Recommended

More Related Content

What's hot

What's hot (20)

Similar to Argumentation mining--an introduction for linguists--Fribourg--2019-09-02

Similar to Argumentation mining--an introduction for linguists--Fribourg--2019-09-02 (20)

More from jodischneider

More from jodischneider (20)

Recently uploaded

Recently uploaded (20)

Argumentation mining--an introduction for linguists--Fribourg--2019-09-02

Editor's Notes