TL;DR
This paper investigates the "uniform information density" (UID) hypothesis, which suggests speakers aim to make their language equally predictable over time. By analyzing Twitter data during the World Series, they show that both gradual (time-based) and rapid (event-based) changes in non-linguistic context influence linguistic complexity, supporting the idea that shared context, not just linguistic history, drives communication efficiency.
What Problem Does This Paper Address?
The core problem this paper tackles is understanding how speakers optimize their language to be efficiently transmitted, especially in noisy communication channels. The "uniform information density" (UID) hypothesis, rooted in information theory, posits that speakers aim for a constant rate of information delivery. Previous work found evidence for this by observing increasing linguistic complexity over time in texts, suggesting speakers adapt to growing contextual knowledge. However, these studies primarily focused on linguistic context and didn't directly quantify non-linguistic context or its impact on rapid changes in information density. This paper aims to bridge this gap by using microblogging data from a shared event to measure both linguistic and non-linguistic context and their influence on information density.
Methods
The authors employ a combination of statistical modeling and information theory to analyze their data.
- Information Density Measurement: They use entropy to quantify information density. Specifically, they calculate per-word entropy and per-tweet entropy.
- Contextual Modeling: They break down the joint entropy of linguistic and non-linguistic information into components related to local linguistic context, broader linguistic context, and non-linguistic context. This is formalized in Equation 2: $H(T_j, E_j | C_j) = H(T_j | C_j) - I(T_j; E_j | C_j) + H(E_j | C_j)$.
- Statistical Analysis: They use mixed-effects linear regression to model the relationship between time, tweet rate, and entropy measures. They also incorporate external metrics of game importance (Leverage Index and Win Probability Added) to control for non-linguistic event informativity.
- Language Modeling: They build domain-specific language models (trigrams with Witten-Bell smoothing) for entropy computation, using tweets from other games as training data to capture the style and vocabulary of the #WorldSeries corpus.
Datasets
- #Worldseries Corpus: A collection of 109,207 tweets containing the hashtag #WorldSeries, collected during the 2014 World Series. Tweets were binned by at-bats and synchronized with pitch-by-pitch data from the MLB Advance Media XML repository.
- Twitter API: Used to collect tweets. The paper notes that their search captured approximately 4% of all relevant tweets, a potential sampling bias addressed in control analyses.
- External Game Data: Major League Baseball Advance Media XML repository for pitch-by-pitch data and timestamps. Metrics like Leverage Index (LI) and Win Probability Added (WPA) were used as external measures of event importance.
Key Insights and Ideas
- Non-linguistic context matters: The study provides strong evidence that shared non-linguistic context (like in-game events) significantly influences linguistic complexity, not just the linguistic history of the conversation.
- Dual adaptation: Speakers exhibit both "slow" adaptation (gradual increase in complexity over time as common ground builds) and "fast" adaptation (rapid decrease in complexity in response to high-information, unpredictable events).
- Twitter as a proxy for shared events: The #WorldSeries hashtag on Twitter serves as a valuable platform to study communication around shared, real-world events where metadata can be precisely linked to linguistic output.
- UID applies beyond linguistic context: The findings suggest that the UID hypothesis is a more general principle of efficient communication, extending to situations where context is primarily external and shared.
Key Results
- Gradual Increase in Entropy: Per-word entropy significantly increased with time over the first two hours of games, then leveled off and slightly declined, consistent with previous UID findings and the context decay model.
- Rapid Decrease in Entropy with Event Intensity: Higher tweet rates (indicating more exciting or unpredictable events) were associated with significantly lower per-word and per-tweet entropy. This suggests speakers reduce linguistic complexity when events are highly informative, to maintain a constant overall information rate.
- Game Importance Metrics: Both Leverage Index (LI) and Win Probability Added (WPA) had significant negative effects on per-tweet entropy, even after controlling for tweet rate. This indicates that the inherent informativeness of an at-bat, independent of social media activity, drives linguistic adaptation.
- Speaker Normalization: Control analyses showed that these effects were not due to demographic shifts in the tweeting population but rather individual speakers adapting their behavior.
Limitations and Future Work
- Sampling Bias: The Twitter API's "gardenhose" feed captures only a subset of tweets, which could introduce bias. The authors address this in control analyses.
- Entropy Normalization: The paper notes that normalizing entropy by speaker was not straightforward, leading to analyses focusing on tweet length for speaker normalization.
- Mechanism of UID: The exact mechanisms underlying the tendency to regulate information content remain unknown. While UID aligns with "audience design," simpler production processes could also explain it.
- Future Work: The authors suggest testing the UID hypothesis across a wider variety of structures and contexts, and further exploring the prediction that information processing difficulty tracks total information content. They also highlight the potential of internet communications surrounding shared media events for future research due to quantifiable metadata.
Summary updated