4
WIS
Web
Information
Systems
Topic models
• Topic model = unsupervised model to discover
hidden structures (i.e., topics) in corpora of text
– Example: Latent Dirichlet Allocation (LDA) [1]
– Topics are probability distributions over words
– If applied to a corpus of documents related to a debate,
topics could be interpreted as perspectives
• Joint topic model = adding additional components
(e.g., sentiment analysis) to a classical topic
model (e.g., LDA)
5
WIS
Web
Information
Systems
Our paper
RQ1. Can joint topic models support users in discovering
perspectives in a corpus of opinionated documents?
RQ2. Do users interpret the output of joint topic models
in line with their personal pre-existing stance?
Contributions:
1. Perspective-annotated data set
2. User study
6
WIS
Web
Information
Systems
Data
Document Stance Perspective
You cannot be a
Christian and support
abortion…
Against Abortion is the killing of a human
being, which defies the word of
God.
No one in the world has
any right to judge over
what someone else
does with their body, …
For Reproductive choice empowers
women by giving them control over
their own bodies.
Why put a child through
the pain of an unloving
mother…
For A baby should not come into the
world unwanted.
… … …
Final data set: 600 documents; 6 perspectives
7
WIS
Web
Information
Systems
Experimental setup
1
2
3
4
5
• Ran each model on the final
data set (i.e., for 6 topics)
• Between-subjects study: each
participant sees output of one
of the models
• Participants need to identify
the correct 6 perspectives from
the model output
8
WIS
Web
Information
Systems
Procedure
Step 1 Step 2 Step 3
Participants state:
• Age
• Gender
• Personal stance
towards abortion
• Familiarity with the
abortion debate
Participants state:
• Perceived usefulness
• Perceived awareness
increase
• Confidence in task
performance
9
WIS
Web
Information
Systems
Results: descriptive
• 158 participants (recruited from Prolific)
– After excluding 12 participants due to failing both honeypot topics
– 150 required according to power analysis
• 50.6% female, 49.4% male
• 33.3 years old on average (range 18 to 64)
• Most (57.8%) at least somewhat familiar with the topic
• Sample skewed towards the supporting viewpoint
10
WIS
Web
Information
Systems
Results: hypothesis tests
H1: Users find more correct perspectives when being exposed
to the output of a joint topic model compared to the output of a
regular topic model or baseline.
– We find a difference between models (p < 0.001, η2 = 0.126)
– TAM is the only one that performs significantly better than the baseline
3
4
5
TF−IDF LDA JST VODUM TAM LAM
Model
MeannCor
11
WIS
Web
Information
Systems
Results: hypothesis tests
H2: Users are more likely to identify sets of keywords as
perspectives that are in line with their personal stance compared
to perspectives that they do not agree with.
– No evidence for for such a relationship (ρ = 0.122, p = 0.163)
13
WIS
Web
Information
Systems
Discussion and future work
• Why did TAM perform better?
– It extracted more keywords that appeared explicitly in the perspective expression
Abortion is the killing of a human
being, which defies the word of God.
Reproductive choice empowers
women by giving them control over
their own bodies.
A baby should not come into the
world unwanted.
• Future work: different domains, novel topic models
14
WIS
Web
Information
Systems
Take home
• Joint topic models such as TAM can perform
perspective discovery
• No evidence for tendency of users to interpret
output in line with their personal stance
• Implications for several areas: journalism,
policy-making, generating explanations
(All supplementary materials are openly available at
https://osf.io/uns63/.)
15
WIS
Web
Information
Systems
References
[1] D. Blei, A. Ng, and M. Jordan, “Latent dirichlet allocation,” Journal of Machine Learning Research, vol. 3, pp. 993–
1022, 05 2003.
[2] M. Paul and R. Girju, “A two-dimensional topic-aspect model for discovering multi-faceted topics.” in AAAI, vol. 1, 01
2010. [Online]. Available: http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.
226.3550&rep=rep1&type=pdf
[3] C. Lin and Y. He, “Joint sentiment/topic model for sentiment analysis,” in Proceedings of the 18th ACM Conference on
Information and Knowledge Management, ser. CIKM ’09. New York, NY, USA: Association for Computing
Machinery, 2009, p. 375–384. [Online]. Available: https://doi.org/10.1145/1645953.1646003
[4] T. Thonet, G. Cabanac, M. Boughanem, and K. Pinel-Sauvagnat, “Vodum: A topic model unifying viewpoint, topic and
opinion discovery,” in ECIR, vol. 9626. Toulouse, France: Springer, 03 2016, pp. 533– 545.
[5] D. Vilares and Y. He, “Detecting perspectives in political debates,” in EMNLP. Association for Computational
Linguistics, 01 2017, pp. 1573–1582.
Editor's Notes
Introduce myself
Second year PhD
Imagine you are a journalist writing an article about the abortion debate
Abortion is a commonly debated topic with many people on both sides; and many perspectives
Explain stance-perspective difference
Naturally these debates are carried out online in news, social media, and fora
For you as a journalist it would be great to have an automatic way to distil these perspectives
Formalize
What existing techniques could be used here?
Sentiment analysis and stance detection no good because supervised
Perspectives are unstructured and different for every topic unsupervised
In sum, two research questions
To answer them, we
Created a data set (openly available)
Conducted a user study showing that some joint topic models can perform perspective discovery
Needed data set of opinionated documents with perspective annotation
Documents: around 3000 debate forum posts on abortion
Human annotator noted stance and perspective
Perspectives taken from ProCon list of 31
Then balanced data set of 600 documents
First describe joint topic models, then baselines
Ran all these models on the final corpus and then conducted user study with their output
Between-subjects design, randomly assigned each participant to one model
Topic model output on the left (6 topics + two honeypots)
Select one of 16 different perspectives for each topic
Step 3 we measured experience with the task
Interesting that they were skewed; as we performed Prolific pre-screening
Describe again why and what we did in this hypothesis test; we used ANOVA
post hoc tests: TAM is the only one that is better than the TF-IDF baseline model
Confirmation bias (ambiguous model output)
spearman correlation – no evidence
Normalized distribution over perspectives (x-axis)
P1-p6 in the corpus, rest not
Plot shows how often each perspective was selected
Some perspectives were well represented in all models, like P5 (or people are familiar with them)
TAM was good with perspectives that other models struggled with, such as P1 and P6
More exploratory results in the paper
Other topics more sentiment-related words
Future work: different domains, novel topic models
Supplementary material is available on our repository
Generating explanations to help people overcome biases