Analyzing Argumentative Discourse Units in Online Interactions
Debanjan Ghosh, Smaranda Muresan, Nina Wacholder, Mark Aakhus and Matthew Mitsui
First Workshop on Argumentation Mining, ACLJune 26, 2014
But when we first tried the iPhone it felt natural immediately, we didn't have to 'unlearn' old habits from our antiquated Nokias & Blackberrys. That happened because the iPhone is a truly great design.
That's very true. With the iPhone, the sweet goodness part of the UI is immediately apparent. After a minute or two, you’re feeling empowered and comfortable.
It's the weaknesses that take several days or weeks for you to really understanding and get frustrated by.
I disagree that the iPhone just "felt natural immediately"... In my opinion it feels restrictive and over simplified, sometimes to the point of frustration.
User1
User2
User3
when we first tried the iPhone it felt natural immediately,
That’s very true. With the iPhone, the sweet goodness part ofThe UI is immediately apparent. After a minute or two, you’reFeeling empowered and comfortable.
I disagree that the iPhone just “felt natural immediately”… in myOpinion it feels restrictive and over simplified, sometimes to thePoint of frustration.
1. Segmentation 2. Segment Classification 3. Relation Identification
Argumentative Discourse Units (ADU; Peldszus and Stede, 2013)
That’s very true. With the iPhone, the sweet goodness part ofThe UI is immediately apparent. After a minute or two, you’reFeeling empowered and comfortable.
I disagree that the iPhone just “felt natural immediately”… in myOpinion it feels restrictive and over simplified, sometimes to thePoint of frustration.
3
Annotation Challenges• A complex annotation scheme seems infeasible– The problem of high *cognitive load* (annotators
have to read all the threads)
– High complexity demands two or more annotators
– Use of expert annotators for all tasks is costly
4
Our Approach: Two-tiered Annotation Scheme
• Coarse-grained annotation– Expert annotators (EAs) – Annotate entire thread
• Fine-grained annotation– Novice annotators (Turkers)– Annotate only text labeled by EAs
5
Our Approach: Two-tiered Annotation Scheme
• Coarse-grained annotation– Expert annotators (EAs) – Annotate entire thread
• Fine-grained annotation– Novice annotators (Turkers)– Annotate only text labeled by EAs
6
Coarse-grained Expert Annotation
Pragmatic Argumentation Theory (PAT; Van Eemeren et al., 1993) based annotation
Post1
Post2
Post3
Post4
Target
Callout
Post2
Post3
7
ADUs: Callout and Target
• A Callout is a subsequent action that selects all or some part of a prior action (i.e., Target) and comments on it in some way.
• A Target is a part of a prior action that has been called out by a subsequent action.
But when we first tried the iPhone it felt natural immediately, we didn't have to 'unlearn' old habits from our antiquated Nokias & Blackberrys. That happened because the iPhone is a truly great design.
That's very true. With the iPhone, the sweet goodness part of the UI is immediately apparent. After a minute or two, you’re feeling empowered and comfortable.
It's the weaknesses that take several days or weeks for you to really understanding and get frustrated by.
I disagree that the iPhone just "felt natural immediately"... In my opinion it feels restrictive and over simplified, sometimes to the point of frustration.
User1
User2
User3
when we first tried the iPhone it felt natural immediately,
That’s very true. With the iPhone, the sweet goodness part ofThe UI is immediately apparent. After a minute or two, you’reFeeling empowered and comfortable.
I disagree that the iPhone just “felt natural immediately”… in myOpinion it feels restrictive and over simplified, sometimes to thePoint of frustration.
That’s very true. With the iPhone, the sweet goodness part ofThe UI is immediately apparent. After a minute or two, you’reFeeling empowered and comfortable.
I disagree that the iPhone just “felt natural immediately”… in myOpinion it feels restrictive and over simplified, sometimes to thePoint of frustration.
Target
Callout
Callout
9
More on Expert Annotations and Corpus
• Five Annotators were free to choose any text segment to represent an ADU
• Four blogs and their first one-hundred comment sections are used as our argumentative corpus
– Android (iPhone vs. Android phones)– iPad (usability of iPad as a tablet)– Twitter (use of Twitter as a micro-blog platform)– Job Layoffs (layoffs and outsourcing)
10
Inter Annotator Agreement (IAA) for Expert Annotations
Thread F1_EM F1_OM Krippendorff’s
Android 54.4 87.8 0.64
iPad 51.2 86.0 0.73
Layoffs 51.9 87.5 0.87
Twitter 53.8 88.5 0.82
• P/R/F1 based IAA (Wiebe et al., 2005)
• exact match (EM) • overlap match (OM)
• Krippendorff’s (Krippendorff, 2004)
11
Issues• Different IAA metrics have different outcome
• It is difficult to infer from IAA that what segments of the text are easier or harder to annotate
12
Our solution: Hierarchical ClusteringWe utilize a hierarchical clustering technique to cluster ADUs that are variant of a same Callout
Thread # of Clusters# of Expert Annotator/ADUs per
cluster5 4 3 2 1
Android 91 52 16 11 7 5
Ipad 88 41 17 7 13 10
Layoffs 86 41 18 11 6 10
Twitter 84 44 17 14 4 5
• Clusters with 5 and 4 annotators shows Callouts that are plausibly easier to identify
• Clusters selected by only one or two annotators are harder to identify
13
Example of a Callout Cluster
14
Motivation for a finer-grained annotation
• What is the nature of the relation between a Callout and a Target?
• Can we identify finer-grained ADUs in a Callout?
15
Our Approach: Two-tiered Annotation Scheme
• Coarse-grained annotation– Expert annotators (EAs) – Annotate entire thread
• Fine-grained annotation– Novice annotators (Turkers)– Annotate only text labeled by EAs
16
Novice Annotation: task 1T
CO
T
CO
T
T
CO
CO
This is related to annotation of agreement/disagreement (Misra and Walker, 2013; Andreas et al., 2012) identification research.
Agree/Disagree/Other
But when we first tried the iPhone it felt natural immediately, we didn't have to 'unlearn' old habits from our antiquated Nokias & Blackberrys. That happened because the iPhone is a truly great design.
That's very true. With the iPhone, the sweet goodness part of the UI is immediately apparent. After a minute or two, you’re feeling empowered and comfortable.
It's the weaknesses that take several days or weeks for you to really understanding and get frustrated by.
I disagree that the iPhone just "felt natural immediately"... In my opinion it feels restrictive and over simplified, sometimes to the point of frustration.
User1
User2
User3
when we first tried the iPhone it felt natural immediately,
That’s very true. With the iPhone, the sweet goodness part ofThe UI is immediately apparent. After a minute or two, you’reFeeling empowered and comfortable.
I disagree that the iPhone just “felt natural immediately”… in myOpinion it feels restrictive and over simplified, sometimes to thePoint of frustration.
That’s very true. With the iPhone, the sweet goodness part ofThe UI is immediately apparent. After a minute or two, you’reFeeling empowered and comfortable.
I disagree that the iPhone just “felt natural immediately”… in myOpinion it feels restrictive and over simplified, sometimes to thePoint of frustration.
Target
Callout
Callout
18
More from Agree/Disagree Relation Label
• For each Target/Callout pair we employed five Turkers
• Fleiss’ Kappa shows moderate agreement between the Turkers
• 143 Agree/153 Disagree/50 Other data instance• We run preliminary experiments for predicting
the relation label (rule based, BoW, Lexical Features…)
• Best results (F1): 66.9% (Agree) 62.9% (Disagree)
19
Novice Annotation: task 2
2: Identifying Stance vs. Rationale
This is related to identification of justification task (Biran and Rambow, 2011)
CO S R
Difficulty
T
That's very true. With the iPhone, the sweet goodness part of the UI is immediately apparent. After a minute or two, you’re feeling empowered and comfortable.
It's the weaknesses that take several days or weeks for you to really understanding and get frustrated by.
I disagree that the iPhone just "felt natural immediately"... In my opinion it feels restrictive and over simplified, sometimes to the point of frustration.
User2
User3
That’s very true. With the iPhone, the sweet goodness part ofThe UI is immediately apparent. After a minute or two, you’reFeeling empowered and comfortable.
I disagree that the iPhone just “felt natural immediately”… in myOpinion it feels restrictive and over simplified, sometimes to thePoint of frustration.
That’s very true
I disagree that the iPhone just “felt natural immediately”
StanceRationale
21
Examples of Callout/Target pairs with difficulty level (majority voting)
Target Callout Stance Rationale Difficulty
the iPhone is a truly great design.
I disagree too. some things they get right, some things they do not.
I…too Some things…do not
Easy
the dedicated `Back' button
that back button is key. navigation is actually much easier on the android.
That back button is key
Navigation is…android
Moderate
It's more about the features and apps and Android seriously lacks on latter.
Just because the iPhone has a huge amount of apps, doesn't mean they're all worth having.
- Just because the iPhone has a huge amount of apps, doesn't mean they're all worth having.
Difficult
I feel like your comments about Nexus One is too positive …
I feel like your poor grammar are to obvious to be self thought...
- - Too difficult/ unsure
22
Difficulty judgment (majority voting)
Diff Number of Expert Annotators per cluster
5 4 3 2 1
Easy 81.0 70.8 60.9 63.6 25.0
Moderate 7.7 7.0 17.1 6.1 25.0
Difficult 5.9 5.9 7.3 9.1 12.5Too Difficult to code 5.4 16.4 14.6 21.2 37.5
23
Conclusion• We propose a two-tiered annotation scheme
for argument annotation for online discussion forums
• Expert annotators detect Callout/Target pairs where crowdsourcing is employed to discover finer units like Stance/Rationale
• Our study also assists in detecting the text that is easy/hard to annotate
• Preliminary experiments to predict agreement/disagreement among ADUs
24
Future Work• Qualitative analysis of the Callout phenomenon
to process finer-grained analysis• Study the different use of the ADUs on different
situations • Annotation on different domain (e.g. healthcare
forums) and adjust our annotation scheme• Predictive modeling of Stance/Rationale
phenomenon
25
Thank you!
26
Example from the discussion thread
StanceRationale
User2
User3
27
Predicting the Agree/Disagree Relation Label
• Training data (143 Agree/153 Disagree)• Salient Features for the experiments– Baseline: rule based (`agree’, `disagree’)– Mutual Information (MI): MI is used to select words
to represent each category– LexFeat: Lexical features based on sentiment
lexicons (Hu and Liu, 2004), lexical overlaps, initial words of the Callouts…
• 10-fold CV using SVM
28
Predicting the Agree/Disagree Relation Label (preliminary result)
• Lexical features result in F1 score between 60-70% for Agree/Disagree relations
• Ablation tests show initial words of the Callout is the strongest feature
• Rule-based system show very low recall (7%), which indicates a lot of Target-Callout relations are *implicit*
• Limitation – lack of data (in process of annotating more data currently…)
29
# of Clusters for each CorpusThread # of Clusters
# of EA ADUs per cluster5 4 3 2 1
91 52 16 11 7 5
Ipad 88 41 17 7 13 10
Layoffs 86 41 18 11 6 10
Twitter 84 44 17 14 4 5
• Clusters with 5 and 4 annotators shows Callouts that are plausibly easier to identify
• Clusters selected by only one or two annotators are harder to identify
30
Target
Callout2
Callout1User1
User2
User3
31
Target
Callout2
Callout1User1
User2
User3
Fine-GrainedNovice Annotation
32
T
CO
T
T
CO
T
CO
E.g., Agree/Disagree/Other
E.g., Relation Identification
Finer-Grained Annotation
E.g., Stance &Rationale
CO
33
Motivation and Challenges
Post1
Post2
Post3
Post4
1. Segmentation2. Segment Classification3. Relation Identification
Argumentative Discourse Units (ADU; Peldszus and Stede, 2013)
34
Why we propose a two-layer annotation?
• A two-layer annotation schema– Expert Annotation• Five annotators who received extensive training for the
task• Primary task includes selecting discourse units from user’
posts (argumentative discourse units: ADU)• Peldszus and Stede (2013
– Novice Annotation• Use of Amazon Mechanical Turk (AMT) platform to detect
the nature and role of the ADUs selected by the experts
35
Annotation Schema for Expert Annotators
• Call OutA Callout is a subsequent action that selects
all or some part of a prior action (i.e., Target) and comments on it in some way.
• TargetA Target is a part of a prior action that has
been called out by a subsequent action
36
Motivation and Challenges• User generated conversational data provides a
wealth of naturally generated arguments
• Argument mining of such online interactions, however, is still in its infancy…
37
Detail on Corpora• Four blog posts and the responses (e.g. first 100
comments) from Technorati between 2008-2010.
• We selected blog postings in the general topic of technology, which contain many disputes and arguments.
• Together they are denoted as – argumentative corpus
38
Motivation and Challenges (cont.)• A detailed single annotation scheme seems
infeasible– The problem of high *cognitive load* (e.g.
annotators have to read all the threads)
– Use of expert annotators for all tasks is costly
• We propose a scalable and principled two-tier scheme to annotate corpora for arguments
39
Annotation Schema(s)• A two-layer annotation schema– Expert Annotation• Five annotators who received extensive training for the
task• Primary task includes a) segmentation, b) segment
classification, and c) relation identification lecting discourse units from user’ posts (argumentative discourse units: ADU)
– Novice Annotation• Use of Amazon Mechanical Turk (AMT) platform to detect
the nature and role of the ADUs selected by the experts
40
Example from the discussion thread
41
A picture is worth…
42
Motivation and Challenges
1. Segmentation2. Segment
Classification3. Relation Identification
Argument annotation includes three tasks (Peldszus and Stede, 2013)
43
Summary of the Annotation Schema(s)
• First stage of annotation– Annotators: expert (trained) annotators– A coarse-grained annotation scheme inspired by
Pragmatic Argumentation Theory (PAT; Van Eemeren et al., 1993)
– Segment, label, and link Callout and Target
• Second stage of annotation– Annotators: novice (crowd) annotators– A finer-grained annotation to detect Stance and
Rationale of an argument
44
Expert AnnotationExpert Annotators
• Segmentation• Labeling• LinkingPeldszus and Stede (2013)
Coarse-grained annotation• Five Expert (trained) annotators
detect two types of ADUs• ADU: Callout and Target
45
The Argumentative Corpus
Blogs and comments extracted from Technorati (2008-2010)
3
1
2
4
46
Novice Annotations: Identifying Stance and Rationale
Callout
Crowdsourcing
• Identify the task-difficulty (very difficult….very easy)• Identify the text segments (Stance and Rationale)
47
Novice Annotations: Identifying the relation between ADUs
Crowdsourcing
Callout Target
… …
… …
Relation label
Number of EA ADUs per cluster
5 4 3 2 1
Agree 39.4 43.3 42.5 35.5 48.4
Disagree 56.9 31.7 32.5 25.8 19.4
Other 3.70 25.0 25.0 38.7 32.3
48
More on Expert Annotations• Annotators were free to chose any text segment
to represent an ADUSplitters
Lumpers
49
Novice Annotation: task 1
1: Identifying the relation(agree/disagree/other)
This is related to annotation of agreement/disagreement (Misra and Walker, 2013; Andreas et al., 2012) and classification of stances (Somasundaran and Wiebe, 2010) in online forums.
50
ADUs: Callout and Target
51
Examples of Clusters# of EAs Callout Target
5 I disagree too. some things they get right, some things they do not.
the iPhone is a truly great design.
I disagree too…they do not. That happened because the iPhone is a truly great design.
2 These iPhone Clones are playing catchup. Good luck with that.
griping about issues that will only affect them once in a blue moon
1 Do you know why the Pre ...various hand- set/builds/resolution issues?
Except for games?? iPhone is clearly dominant there.
52
More on Expert Annotations• Annotators were free to chose any text segment
to represent an ADU
53
Example from the discussion thread
54
Coarse-grained Expert Annotation
Target
Callout
Pragmatic Argumentation Theory (PAT; Van Eemeren et al., 1993) based annotation
55
ADUs: Callout and Target
56
More on Expert Annotations and Corpus
• Five Annotators were free to chose any text segment to represent an ADU
• Four blogs and their first one-hundred comment sections are used as our argumentative corpus
Layoffs
Android
iPad
57
Examples of Cluster# of EAs Callout Target
5
I disagree too. some things they get right, some things they do not.
the iPhone is a truly great design.
I disagree too…they do not. That happened because the iPhone is a truly great design.
I disagree too. But when we first tried the iPhone it felt natural immediately . . . iPhone is a truly great design.
Hi there, I disagree too . . . they do not. Same as OSX.
-Same as above-
I disagree too. . . Same as OSX . . . no problem.
-Same as above-
58
Predicting the Agree/Disagree Relation Label
Features Categ. P R F1
Baseline Agree 83.3 6.90 12.9Disagree 50.0 5.20 9.50
UnigramsAgree 57.9 61.5 59.7Disagree 61.8 58.2 59.9
MI-based unigram
Agree 60.1 66.4 63.1Disagree 65.2 58.8 61.9
LexF Agree 61.4 73.4 66.9Disagree 69.6 56.9 62.6
59
Novice Annotation: task 2
2: Identifying Stance vs. Rationale
This is related to identification of claim/justification task(Biran and Rambow, 2011)