Questions for Jean Clipperton and Ali Sanaei

[qingruexu-create]

Questions for Jean Clipperton and Ali Sanaei's talks:

It's a new soundtrack: Candidates, Campaigns, and Rally playlists As political candidates are increasingly using music at rallies and releasing or revealing playlists, placing their selected songs into this framework can provide clear and concise opportunities to study how they construct their public image, build political brands, and signal values to voters. I find both cross-party and within-candidate patterns: all candidates increased the use of pro-social language in their playlists, while Democratic candidates increasingly had more negative emotions and more moral language in their songs. Additionally, trust was the most common emotion present in all front-runner candidate playlists. Playlists can send signals about the candidates and how they view their responsibility: as a delegate acting on behalf of the people or as a brand in which voters invest. The dataset covers three presidential elections from 2016 to 2024, and over 2,000 songs. Trump, in each of his three campaigns, evidenced more 'brandidate'-type language in his playlists, with 'I' pronounced being used very heavily. In contrast, Democratic candidates tended toward delegate-type language and more inclusive 'we' language.

Reconstructing Pahlavi Governance: Leveraging Oral Histories with Retrieval-Augmented Generation Oral histories provide valuable insights that are often impossible obtain with any other methods. If we are able to analyze oral histories at the corpus level, instead of focusing on one or a few interviews, we can benefit by getting an automatic triangulation of narratives and perspectives, and also by being able to query about the details that may not be possible to obtain from any single interview. We can leverage large language models by retrieval-augmented generation (RAG) techniques to accomplish this. We first divide the corpus into small snippets of text, then, for any given query, we retrieve the most relevant ones by semantic similarity using word-embeddings, then we show that while the generative phase fails if it is done in one pass, we can divide this phase into multiple tasks and obtain high quality results. We extract lessons and excerpts from each snippet, and finally, we synthesize the lessons and piece them together with the excerpts to create a verifiable narrative that answers the query by reference to the source. We apply this technique to Harvard Iranian Oral History Project corpus with queries about economic governance in the late Pahlavi era.

[Mengyang-Cao]

Thanks for sharing the research. How do U.S. presidential candidates use campaign rally playlists to construct their public image, build political brands, and signal values to voters, and what linguistic and emotional patterns emerge across parties and individual candidates from 2016 to 2024?

[AnnieZzz01]

Thank you for sharing. Your research does a great job decoding the messages in these playlists. Stepping back, I always wonder about the voter's perspective. Since songs come with their own pre-existing meanings, how often do you think the candidate's intended message is at odds with what the audience actually hears?

[PaulWang-Uchicago]

Thanks for sharing. Does the style or rhythm of the music in candidate playlists also shape how voters see them — for example, whether they come across as energetic, inclusive, or more brand-focused — beyond just the lyrics?

[gabereichman]

Very interesting!
Is there any cross-party song usage, or are songs avoided once they've been used by another party, even during different elections?
For these oral histories, are you also including stop words? I imagine there could be some insights in addition to the semantics here because they are not initially written accounts but oral accounts.

[nitomanto]

Thanks for sharing your research. With RAG, is there concern of bias with reconstructed narratives, either from the corpus itself or from the LLM used to generate responses to queries?

[tyeddie]

Thanks for sharing. Regarding the paper Reconstructing Pahlavi Governance, I'm interested in why use retrieval-augmented generation (RAG) for analyzing oral histories instead of relying on standard large language model generation?

[bkehoechicago]

Thanks for sharing your research. I would think that it would be possible for a very large proportion of within party music to be selected by a handful of people per party. How do you ensure that these results aren't artifacts of the individual preferences of the campaign managers?

[zixuanzzx01]

Thank you very much for sharing! I'm very interested in the first topic on soundtracks and candidate behavior. The results seem very intuitive, but I wonder what is the mechanism for the discrepencies in word preferences for candidates from different parties? Would you say songs as well as many other artistic products are a good way into collective consciousness? In addition, is it possible to define pro-social languages without introducing bias?

[gtriberti9]

Thanks for sharing! My first idea would be that while oral histories provide unique insights, they also require careful methods to detect, cross-check, and contextualize possible distortions or biases. Can AI methods like retrieval-augmented generation help us identify and cross-check biases or inconsistencies in oral histories, making them more reliable sources for reconstructing governance under the Pahlavis?

[AnqiW02]

Thanks for sharing your insightful research! While the first research relies on quantitative textual analysis of lyrics, I wonder how we can integrate the emotional nuances of the music itself - such as rhythm, genre, and chord progressions - with lyrical analysis to gain a more comprehensive understanding of the emotional signals conveyed by candidates? For example, would the message of a song with positive lyrics change if it were performed in an intense, "anthem-like" style?

[ZheruiLU]

Thank you for sharing! For the playlists study, I wonder how much of the music selection reflects the candidate’s own preferences versus the choices of campaign staff, and how you disentangle the personal from the strategic in interpreting these signals?

[VeraMao]

Thanks for sharing these! For It's a new soundtrack: Candidates, Campaigns, and Rally playlists, do you think playlists are more effective in reinforcing existing partisan identities or in persuading undecided voters? And how might the emphasis on “trust” across candidates’ playlists reflect broader dynamics of political polarization and credibility crises?

For Reconstructing Pahlavi Governance: Leveraging Oral Histories with Retrieval-Augmented Generation Oral histories often reflect memory biases, how does RAG handle contradictions or conflicting accounts across different interviews?

[yilinx-10]

Thank you for sharing! Regarding the political candidates’ playlists, do you think other musical characteristics might also play a role? For instance, is there a relationship between political position and features like beat or song structure? And for using RAG techniques to analyze oral history records, how do you assess the quality of the results? Are there established benchmarks for comparison?

[jwrampino]

Sanaei: Have you incorporated geographical information from the Project dataset or in your queries? For example, a primary source from someone born and raised in Tehran might have a very different account than someone who was a rural agricultural worker. Furthermore, how do you verify the narratives historically? Do you contextualize the outputs with secondary source literature on the Pahlavi era?

Clipperton: Beyond language, have you considered Spectral Clustering or other methods to analyze elements like the music's timbre, melody, harmony, or other elements that have a history of analysis in music theory?

[skbyrne]

Thank you both so much for sharing your work! I am especially curious about your finding that Democratic candidates' playlists increasingly incorporated more negative emotions. Was this difference between parties present at the beginning of the period you studied or did it emerge over time? Also, are you aware of any research looking at campaign music across earlier elections, or have you considered extending your analysis further back historically?