Mapping socio-political issues in the United Stateswith a large-scale analysis of local talk radio скачать в хорошем качестве

Mapping socio-political issues in the United Stateswith a large-scale analysis of local talk radio

Broadcast radio is a pervasive medium in the United States, and the talk radio format presents a unique resource for exploring social and political issues. Although there is a substantial volume of syndicated programming that airs nationwide, a great deal of talk radio content is only broadcast locally. This presents an opportunity for exploring geographic differences in the topics and issues covered. Yet by its nature, talk radio is also ephemeral, distributed, and difficult to study. In this work, we build on an extended version of the massive corpus of transcribed talk radio content described in [1] and develop a methodology for identifying discussion on social, political, health and a variety of other topics. We then perform initial exploratory analyses to characterize social and political issues and their geographic variations within the United States, using data covering 31 states, 123 radio stations, and more than 800unique radio shows spanning over 100,000 hours of transcribed radio content. Anecdotally, we find temporal and geographic “peaks” in selected topics correspond to important news events, such as the 2019 state abortion restriction bills. More generally, we suggest that tools for collecting and analyzing talk radio may serve as an important resource for journalists and social scientists in interpreting perspectives on social and political issues.Our analysis covers the 5 month time period from January to May 2019. The essence of the approach here is to identify topics in local radio content and quantify the share of each topic across states and over time. Radio stations typically air a mix of local programming and syndicated content, which may be broadcast nationwide. Since our focus here is on local, state-level topic variations we only consider the 844 “state-local” radio shows that air within a single state, which we determine by combining radio station location and program guide information. Once a set of state-local radio programs are identified, along with the airing start and end time of each episode, we obtain the corresponding transcripts which serve as the corpus for our analysis. This corresponds to 46,950 radio program episodes. However, there is a wide variation in the duration of different radio programs, with some shows airing for up to 7-8hours, and most shows containing advertisements, music and often covering a variety of topics.Although topic discovery algorithms can identify multiple topics in a single document, treating each individual episode airing in the corpus as a “document” yields poor topics, and also limits the temporal resolution for measuring the quantity of discussion on a topic. Thus, we applied a recently proposed attention-based neural text segmentation algorithm [2] to split the documents(episodes) into semantically different segments, yielding 634,851 segments. Applying Latent Dirichlet Allocation (LDA) [3], a standard topic modeling algorithm to these segments results in a coherent set of topics. We ran the Gensim implementation [4] of LDA with 100 topics, opting for a large number to bias toward a subset of focused, interpretable topics. We inspected the top distinctive words in each, obtaining topics focused on US-Mexico border issues, Abortion, US Presidential elections, Russian involvement in US politics, and so on.The topic labeling of radio segments, with each segment associated with a radio program, station, state and time, enables us to investigate the relative magnitude of a topic’s importance by state and over time. By reviewing sequences of heatmap images for different topics overlaid on a map of the United States, we could spot patterns corresponding to important social and political events. As an illustrative example, see Fig.1 showing a week-over-week view of the“Abortion” topic, revealing increasing activity in Georgia in March 2019, corresponding to the time when the state passed a new abortion restriction bill1.As next steps for our study, we plan a deeper exploration of the topics and their temporal and geographic variations, and how those patterns link to social and political events. We also intend to work on linking back to the original data in the radio transcripts to assist in interpretation.Additional metadata, such as sentiment expressed or the political leaning of the program hosts and guests may also yield useful insights. Through large scale analysis of talk radio, we can gain another perspective on social and political trends in the United States.