TL;DR
This paper introduces a novel computational model to quantify linguistic influence between speakers in group conversations. By adapting De Groot's model of social influence, they propose a convex model that estimates pairwise influence directly from speech and language data, offering a more nuanced understanding of conversational dynamics than previous methods.
What Problem Does This Paper Address?
Understanding how individuals influence each other's language use in conversations is a fundamental aspect of human interaction. While linguistic accommodation—the tendency for conversational partners to converge their communication styles—is well-documented, computational models for quantifying this influence, especially in multi-speaker group settings, have been limited. Existing work often focuses on specific linguistic markers or requires complex inference methods. This paper aims to bridge this gap by developing a computationally efficient and versatile model that can infer pairwise influence levels directly from conversational data, applicable to various real-world scenarios.
Methods
The core of the proposed method is a convex model adapted from De Groot's model of social influence. The fundamental assumption is that a speaker's language model at any given time can be represented as a convex combination of the language models of the individuals they interact with. Specifically, a speaker's language model in the next epoch ($L_{i}^{(t+1)}$) is influenced by the language models of other speakers in the previous epoch ($L_{j}^{(t)}$) and their own default language model ($L_{i}$), weighted by influence coefficients ($W_{ij}$).
$L_{i}^{(t+1)} = \sum_{j \in \mathcal{E}^{(t)}} W_{ij} L_{j}^{(t)} + W_{ii} L_{i}$
These influence coefficients ($W_{ij}$) form a linguistic influence matrix (LIM), where $W_{ij}$ represents the probability that speaker $i$ will imitate speaker $j$. The model estimates these coefficients by minimizing the difference between the actual language model of a speaker at time $t+1$ and the predicted model based on the convex combination of other speakers' models. This optimization is performed subject to constraints that the influence coefficients are non-negative and sum to one for each speaker. The authors use a simple unigram model for language representation and employ an L2 norm for error measurement, making the optimization problem convex and solvable using standard tools.
Datasets
The paper evaluates the proposed model on three distinct datasets:
- Synthetic Corpus: A simulated conversation with a pre-defined influence matrix, used to test the algorithm's accuracy under controlled conditions.
- Early Development Corpus: Transcripts of a child interacting with family members over time, used to study language development and identify influential family members.
- Supreme Court Cases Corpus: Transcripts of U.S. Supreme Court oral arguments, used to analyze influence dynamics between judges, defendants, and prosecutors.
Key Insights and Ideas
- De Groot's Model as a Foundation: Adapting a well-established model for opinion fusion to linguistic influence provides a theoretically grounded and computationally tractable approach.
- Direct Inference from Text: The model's ability to infer influence directly from language samples, without relying on external metadata like social status, makes it broadly applicable.
- Convexity for Tractability: Framing the problem as a convex optimization problem ensures efficient and reliable solutions for estimating influence coefficients.
- Versatility Across Domains: Demonstrating the model's effectiveness on diverse datasets (synthetic, child development, legal discourse) highlights its generalizability.
- Influence vs. Interaction Frequency: The child development experiment suggests that while interaction frequency plays a role, a child's own language model development can eventually lead to a more distinct linguistic identity.
Key Results
- Synthetic Data Performance: The model's accuracy in inferring the LIM improves with larger vocabulary sizes and longer conversation lengths. For parameters mimicking real-world conversations, the model significantly outperforms random chance in estimating influence.
- Child Language Development: The model identified that in early stages, a child's language is heavily influenced by their mother and sister. Over time, the child's own language model becomes more dominant, suggesting a developmental trajectory away from direct imitation.
- Supreme Court Dynamics: Analysis of Supreme Court transcripts revealed that judges exert the most influence on attorneys and defendants, with statistically significant differences compared to the influence exerted by attorneys on judges. This aligns with previous findings suggesting a hierarchical influence structure in legal settings.
Limitations and Future Work
- Divergence Measures: The authors acknowledge that while the L2 norm is used for simplicity, other divergence measures (e.g., Kullback-Leibler, Dp divergence) could be explored for comparing language models, potentially offering more nuanced insights.
- Joint Estimation: Future work could involve methods for jointly estimating the influence matrix and the language models, rather than estimating them separately.
- Acoustic Data Integration: The current model focuses on text data. The authors suggest extending the approach to incorporate acoustic features (e.g., pitch, speaking rate) to model prosodic influence, which would require adapting the language model representation.
Summary updated