Abstract
Contrary to the long-held belief that language simplifies as people develop "common ground," this paper reveals that in complex collaborative tasks, language complexity actually increases over time. By analyzing the CerealBar game, the authors show that as participants become more expert, instructors use longer, syntactically richer instructions to pack multiple goals into a single message, prioritizing efficiency and utility over brevity.
Introduction: The Utility of Complexity
In psycholinguistics, a classic finding is that as two people interact repeatedly and form "common ground," their language becomes more efficient and concise. This phenomenon has been studied extensively using reference games, where the goal is simply to identify an object. In such settings, a descriptive phrase like "blue circle with a bite taken out of it" eventually shortens to "the bitten one."
However, real-world collaboration is rarely just about referencing objects; it is about achieving complex goals. The authors argue that reference games lack "flexible utility." In a task where multiple objectives can be accomplished simultaneously, does language still simplify? Or do participants use their shared expertise to communicate more information in a single breath? This paper investigates how community language evolves in a situated, goal-oriented environment where the utility of an instruction can scale.
The CerealBar Environment
The study utilizes the CerealBar dataset, which captures a two-player collaborative game. In this environment, a "Leader" has a global view of the map and provides natural language instructions to a "Follower." The Follower has a limited first-person view but possesses more movement points to collect sets of matching cards. The researchers analyzed 795 human-human interactions, totaling 17,524 instructions. To ensure they were measuring community-level convention-forming rather than individual learning, the data was pruned to focus on experienced players and split chronologically into ten "deciles."
Methodology
To quantify linguistic complexity, the researchers employed several metrics:
- Surface Metrics: Tracking utterance length and vocabulary size.
- Syntactic Complexity: Using dependency trees to measure structural density, including maximum depth (nesting of clauses) and maximum width (number of modifiers), normalized for length.
- Language Effort: Calculating the number of tokens and instructions required per "card event" (the primary game goal).
- Part-of-Speech (POS) Analysis: Specifically monitoring conjunctions (e.g., "and," "then") that facilitate multi-step instructions.
Key Results: Packing More into Less
The analysis revealed a significant shift in how the community communicated over time. Instead of shortening, instructions became more robust and information-dense.
Increasing Complexity and Structural Density
Mean instruction length grew from 11.9 to 14.1 tokens, while the vocabulary expanded from 752 to 1,070 unique tokens. Structural density also increased; both the normalized branching factor and maximum width of dependency trees rose significantly, indicating more descriptive and modifier-heavy language.
Efficiency through Conjunctions
The primary engine of this complexity was the increased use of conjunctions. Rather than issuing sequential commands like "Go to the tree" followed by "Pick up the card," leaders began "packing" tasks: "Go to the tree and pick up the card then turn around..." This shift allowed the number of instructions per goal to decrease from 0.879 to 0.783, even as individual messages grew longer.
Performance Gains
This linguistic evolution correlated with a dramatic improvement in game success. Mean scores per game rose from 3.8 in the first decile to 12.3 in the tenth, suggesting that the move toward complex, information-dense communication was highly effective for achieving goals.
Discussion and Future Work
The findings suggest that common ground acts as an enabler for more sophisticated communication rather than just a shortcut. Because the follower understands the environment better, the leader can afford to be more structurally complex without causing confusion.
However, the authors note some limitations. The objects in CerealBar are relatively unambiguous; in environments with more abstract referents, the "shorthand" behavior typical of reference games might reappear. Future work could involve manipulating "instruction utility" to determine precisely when speakers prioritize complexity over brevity.
Summary updated