Iterative Topic Taxonomy Induction with LLMs: A Case Study of Electoral Advertising
Authors: Alexander Brady, Tunazzina Islam
Organizations: Department of Computer Science, ETH Zürich, Zurich, Switzerland · Department of Computer Science, Purdue University, West Lafayette, IN 47907, USA
Social media platforms play a pivotal role in shaping political discourse, but the scale and rapid evolution of online content make systematic analysis difficult. We introduce an end-to-end framework for inducing an interpretable topic taxonomy from unlabeled text corpora. The framework combines embedding-based clustering with iterative large language model (LLM) inference to construct a topic taxonomy without requiring predefined labels or seed topics. It first synthesizes candidate topics from document clusters and then uses the resulting taxonomy to assign consistent topic labels across clusters. We evaluate the approach through a case study of political advertising ahead of the 2024 U.S. presidential election. We use the induced taxonomy to support downstream analyses of issue prevalence, moral framing, advertising spend, and demographic exposure patterns. These results suggest that iterative taxonomy construction can provide a scalable and interpretable approach to organizing large unlabeled text corpora while supporting substantive downstream analysis.
Figures & tables
Figure 1: Overview of the proposed framework. Documents are first embedded and clustered. (a) Cluster representatives are processed sequentially by the LLM to induce a shared topic taxonomy. (b) The LLM assigns one topic label to each cluster from the completed taxonomy.
Issue
#Clusters
Example Ad
economy
15
Molly Buck’s agenda is failing Iowa families. We can’t afford Molly Buck in the State House.
voting rights
13
Make Your Vote Count. RE-register to Vote in the District of Your Second Home.
crime/justice
12
Orange County Firefighters Trust Dave Min To Keep Our Communities Safe.
personal freedom
4
Tim Sheehy will always fight for Montana Values in the Senate!
voting
4
Vote Rebecca for State Rep by Nov 5th!
education
4
Vote for students. Vote for teachers. Support our public schools.
Table 1: Topics identified by the LLM in the political ads dataset. ‘#Clusters’ column gives the number of unique clusters assigned to each topic by the LLM.
Model
Average Score
Best Label
BERTopic
1.2
3
TopicGPT-style
2.7
5
Our Method
2.8
12
Table 2: Annotation results. The average score is out of 5, and the best label is the number of times that model’s label was selected as the most fitting (among all annotators).
Model
Macro F1
Accuracy
Logistic Regression
0.37
0.55
XGBoost
0.31
0.53
RoBERTa
0.32
0.53
SetFit
0.36
0.60
Table 3: Document-level classification performance for topic assignment.
Figure 2: Associations between moral-foundation labels and topic labels for (a) the proposed taxonomy and (b) LDA topics. Cells show Pearson correlations between one-hot topic and moral-foundation assignments.
Figure 3: Average advertising spend and impressions per ad by annotated topic.
Figure 4: Spend by the top five funders of economy, crime/justice, and abortion ads, segmented by moral foundation.
Figure 5: Distribution of advertising spending across moral foundations within economy, crime/justice, and abortion advertisements. Percentages indicate within-topic spending shares; total spending is shown beneath each chart.
Figure 6: PPMI associations between political issue or moral-foundation labels and observed audience demographics. Higher values indicate that a label occurs disproportionately among advertisements delivered to the corresponding demographic group.
Appendix figures & tables1 asset
Supplementary material from the paper’s appendix.
Appendix
Care/Harm: It suggests that someone other than the speaker is worthy of compassion or is experiencing harm, grounded in the values of kindness, tenderness, and care.
Fairness/Cheating: Emphasizes justice, personal rights, and independence; involves comparing with other groups. Advocates for equal opportunity and resists those who benefit without contributing (“Free Riders").
Loyalty/Betrayal: Based on the values of loyalty to one’s country and willingness to sacrifice for the group. Activated by a sense of unity and collective responsibility—“one for all, and all for one".
Authority/Subversion: Centers on showing respect (or resistance) toward established authority and following long-standing traditions. It includes maintaining social order and fulfilling the duties tied to hierarchical roles, such as obedience, respect, and role-based responsibilities.
Sanctity/Degradation: Beyond religion, this value highlights respect for human dignity and aversion to moral or physical corruption, promoting purity, self-control, and the belief that the body is sacred and vulnerable to defilement.
Liberty/Oppression: Captures the feelings of reactance and resentment people experience when their freedom is restricted, often leading to collective disdain for authoritarian figures and motivating unity and resistance against oppression.
Appendix
Table 4: Six basic moral foundations Iyer et al. (2012) ; Haidt and Graham (2007) ; Haidt and Joseph (2004) .