TL;DR
This paper introduces MavenRank, a novel graph-based method for identifying influential speakers in a discussion by analyzing the lexical similarity of their contributions. By applying it to the US Congressional Record, they demonstrate that MavenRank can capture speaker influence beyond traditional committee structures, particularly in ideologically charged topics.
What Problem Does This Paper Address?
In any large-scale discussion or debate, such as legislative proceedings, identifying the most influential participants is crucial for understanding the dynamics of idea propagation and decision-making. Traditional methods often rely on formal structures like committee assignments or seniority. However, these might not fully capture influence derived from rhetorical skill or the ability to shape discourse through language. The authors aim to bridge this gap by developing a data-driven approach that quantifies influence based on the linguistic content of speeches themselves.
Methods
The core of the paper is the MavenRank algorithm, which is inspired by the concept of "mavens" from Malcolm Gladwell's The Tipping Point. It's a graph-based approach that models a discussion as a network where nodes are speeches and edges represent lexical similarity.
- Graph Construction: Speeches are linked if they exhibit a high degree of lexical similarity, measured using tf-idf cosine similarity. This metric quantifies how much two speeches share important, distinguishing terms.
- Lexical Similarity: The paper uses the standard tf-idf (term frequency-inverse document frequency) weighting scheme to represent speech documents as vectors. The idf values are calculated per topic to capture context-specific word importance.
- MavenRank Algorithm: MavenRank then performs a random walk on this similarity graph. The score of a speech is proportional to the scores of its similar neighbors, recursively. This is mathematically equivalent to finding the left eigenvector of the row-normalized similarity matrix, similar to Google's PageRank algorithm.
- Speaker Score Aggregation: A speaker's overall influence is calculated by summing the MavenRank scores of all their speeches within a given topic.
Datasets
- US Congressional Speech Corpus: The primary dataset used is an XML-formatted version of the electronic United States Congressional Record from the Library of Congress. The study focuses on speeches from the 105th and 106th Senates.
- Topic Clusters: The speeches were pre-clustered into 42 topics using a model from Quinn et al. (2006).
- Committee Membership Information: Data on Senate committee assignments, including rank and seniority, was obtained from Stewart and Woon's codebook.
Key Insights and Ideas
- Lexical Centrality as Influence: The paper posits that speakers whose rhetoric is linguistically similar to many others are likely influential, acting as "mavens" who shape the discourse.
- Beyond Formal Structures: MavenRank offers a way to identify influence that might not be apparent from formal roles like committee membership or seniority, especially in topics where ideology or persuasive rhetoric plays a significant role.
- Topic-Specific Influence: Calculating idf values per topic is crucial, as word importance varies greatly across different subject matters (e.g., "abort" is rare in general but significant in abortion-related topics).
- Committee vs. Rhetorical Influence: The study highlights that while committee status often correlates with influence, there are specific topics (e.g., highly politicized or "distributive politics" topics) where rhetorical centrality (MavenRank) is a better indicator of influence.
Key Results
- The study found that MavenRank scores are largely driven by committee status in many topics, validating the hypothesis that formal roles contribute to influence.
- However, for topics where speeches are used to express ideological positions rather than directly influence legislation (e.g., abortion, workers' rights), MavenRank captures speaker centrality that is not strongly correlated with committee rank or seniority.
- Examples are provided showing how Senators like Rick Santorum and Barbara Boxer ranked highly in topics related to their known policy stances (abortion, Social Security reform), aligning with their MavenRank scores.
- The paper also notes that "distributive politics" topics (e.g., agriculture, public lands) show broad distribution of influence with no clear mavens, and some topics (e.g., post-9/11 reforms) had insufficient data for robust analysis.
Limitations and Future Work
- Topic-Committee Mapping: The mapping between topics and committees is not one-to-one and can be ambiguous, limiting the direct comparison in some cases.
- Smin Parameter: The paper used a fixed similarity threshold ($s_{min} = 0.25$) without extensive experimentation to test its effect on MavenRank scores.
- Future Work: The authors plan to apply their methods to the US House of Representatives and parliamentary speech from other countries (UK, Australia). They are also developing a dynamic version of MavenRank to track how ideas propagate over time in evolving networks, including congressional records, blogs, and newsgroups. They are also interested in exploring other similarity measures beyond tf-idf cosine.
Summary updated