Rational Communication Shapes Morphological Composition

Technical Summary: Rational Communication Shapes Morphological Composition

Problem Statement
Human languages expand vocabularies by combining existing morphemes rather than inventing arbitrary forms. While communicative efficiency is known to shape lexical systems at multiple levels (Gibson et al., 2019), the specific mechanism behind morphological composition—the choice of one morpheme sequence over other plausible alternatives for the same concept—remains under-modeled. Previous work has focused on global lexical efficiency, word formation patterns, or the survival of introduced forms in the lexicon (Xu et al., 2024), but has not explicitly modeled the speaker's choice among competing morpheme sequences at a specific historical moment. The central question is whether a trade-off between listener recoverability (semantic informativeness) and speaker production cost can predict why a language settles on a specific attested composition (e.g., computer) over contemporaneously available alternatives (e.g., electric-brain or knowledge-machine).

Methodology
The authors extend the Rational Speech Act (RSA) framework (Frank & Goodman, 2012; Goodman & Frank, 2016) to morphological composition, treating word formation as a cooperative referential game within a time-indexed lexicon ($L_t$).

• Data: The study utilizes a dataset of 4,323 naturally occurring English compounds and derivations from WordNet, spanning emergence years from 1820 to 2019. The lexicon is constructed using the Corpus of Historical American English (COHA) and the Corpus of Contemporary American English (COCA) to track morpheme availability and frequency over time.
• Candidate Generation: For each target concept $c$ at time $t$, a candidate set $C(c, t)$ is generated. This set includes morpheme sequences derived from WordNet synsets, relational neighbors, and nearest neighbors from time-indexed word2vec spaces. Candidates are restricted to a maximum of three morphemes and filtered by a semantic gate to ensure plausibility.
• Model Architecture:
  • Literal Listener ($L_0$): Estimates listener recoverability by computing the compatibility between a candidate morpheme sequence $u$ and the target concept $c$. This is done by embedding morphemes and the target gloss (using Qwen3-Embedding-8B) and summarizing the distribution of cosine similarities via five statistics (mean, max, std, entropy, and whole-candidate similarity).
  • Production Cost: Modeled as an additive function of morpheme features (length, phoneme count, syllable count, and historical token/type frequencies) extracted from the time-indexed lexicon.
  • Pragmatic Speaker ($S_1$): Selects utterances by trading off listener recoverability against production cost. Two variants are tested:
    • Linear $S_1$: Implements a standard RSA utility function $U = \alpha \cdot \text{Semantic} - \beta \cdot \text{Cost}$.
    • Nonlinear $S_1$: Uses a Multi-Layer Perceptron (MLP) to capture interactions between semantic and cost scores.
  • Baselines: The study compares $S_1$ models against Cost-only, Semantic-only, and a Discriminative model (which concatenates all features into a single MLP without enforcing the RSA decomposition).
• Evaluation: Models are trained as rankers to place the gold (attested) morpheme sequence above sampled negative candidates. Performance is measured using Mean Reciprocal Rank (MRR) and Top-k Accuracy (Acc@k) on a year-stratified held-out test set.

Key Results

1. Superiority of Integrated Models: Models integrating semantic informativeness with production cost ($S_1$) consistently outperform single-factor baselines. The Discriminative model achieves the highest overall performance (MRR = 0.096), indicating that additional feature interactions exist beyond the simple two-factor RSA decomposition.
2. Pragmatic Advantage: The $S_1$ models (MRR = 0.050–0.053) significantly outperform the Semantic-only (MRR = 0.047) and Cost-only (MRR = 0.031) baselines. Crucially, the advantage of the $S_1$ model over the semantic-only baseline grows as the candidate set expands (higher $k$). This suggests that while semantic cues dominate when candidates are few, production cost becomes the deciding factor when many semantically plausible alternatives exist.
3. Qualitative Behavior:
   • Cost-only models favor short, high-frequency fragments (e.g., ed+ing) that are cheap but semantically uninformative.
   • Semantic-only models tend to select redundant or morphologically odd forms (e.g., cynic+cynic) that maximize coverage but ignore economy.
   • $S_1$ models successfully balance these pressures, ranking gold sequences near the top by selecting candidates that are meaning-aligned without being unnecessarily long or implausible.
4. Temporal Robustness: The relative ordering of model performance remains stable across different historical training windows (1830–1910), suggesting the communicative trade-off is a consistent factor in lexicalization.

Significance and Claims
The paper claims to provide a computational account of morphological composition as rational communication. By formalizing word formation as an RSA-style choice, the authors demonstrate that attested morphological forms are systematically ranked higher than unattested alternatives when evaluated for both listener recoverability and speaker efficiency.

The study extends the rational communication program from utterance-level choice to the internal structure of words. It argues that lexicalization reflects a communicative trade-off between expressiveness and efficiency. While the authors acknowledge that the RSA framework captures only a "systematic part" of the choice (as evidenced by the gap between $S_1$ and the Discriminative model), the results confirm that the informativeness–cost trade-off is a fundamental driver in how languages select specific morpheme combinations over time. The work offers a complementary, interpretable decomposition for candidate-level choices that global efficiency models or analogy-based approaches do not explicitly address.