[論文レビュー] A Rate-Distortion view of human pragmatic reasoning

Noga Zaslavsky, Jennifer Hu+1 more|arXiv (Cornell University)|May 13, 2020

Natural Language Processing Techniques参考文献 21被引用数 60

ひとこと要約

本論文は Rational Speech Act (RSA) を交互最大化として再定義し、RSA が期待効用と伝達努力のトレードオフを最適化すること、そして RSA を Rate–Distortion theory に根拠づけて RD-RSA を形成することを示し、ダイナミクスと挙動予測を比較する。

ABSTRACT

What computational principles underlie human pragmatic reasoning? A prominent approach to pragmatics is the Rational Speech Act (RSA) framework, which formulates pragmatic reasoning as probabilistic speakers and listeners recursively reasoning about each other. While RSA enjoys broad empirical support, it is not yet clear whether the dynamics of such recursive reasoning may be governed by a general optimization principle. Here, we present a novel analysis of the RSA framework that addresses this question. First, we show that RSA recursion implements an alternating maximization for optimizing a tradeoff between expected utility and communicative effort. On that basis, we study the dynamics of RSA recursion and disconfirm the conjecture that expected utility is guaranteed to improve with recursion depth. Second, we show that RSA can be grounded in Rate-Distortion theory, while maintaining a similar ability to account for human behavior and avoiding a bias of RSA toward random utterance production. This work furthers the mathematical understanding of RSA models, and suggests that general information-theoretic principles may give rise to human pragmatic reasoning.

研究の動機と目的

• Clarify the optimization principle underlying RSA recursive reasoning.
• Show RSA dynamics as an alternating maximization process balancing utility and communicative effort.
• Ground RSA in Rate–Distortion theory to derive RD-RSA and compare predictions with RSA.
• Assess whether RD-RSA preserves RSA’s explanatory power while reducing bias toward non-informative random utterances.

提案手法

• Formulate RSA as an optimization of G_alpha = H_S(U|M) + alpha E_S[V_L], showing RSA recursion implements alternating maximization (S_t and L_t updates).
• Derive RD-RSA by minimizing F_alpha[S,L] = I_S(M;U) - alpha E_S[V_L] and derive self-consistent update rules S(u|m) ∝ S(u) exp(alpha V_L(m,u)), S(u) = sum_m P(m) S(u|m), L(m|u) ∝ S(u|m) P(m)/S(u).
• Analyze asymptotic behavior of RSA and RD-RSA as a function of alpha, including a critical point at alpha = 1.
• Compare RSA and RD-RSA predictions to human data from a reference game experiment, evaluating fit across recursion depths.
• Discuss implications for information-theoretic accounts of pragmatics and potential links to related frameworks like Optimal Transport.

実験結果

リサーチクエスチョン

• RQ1 Does RSA recursion maximize expected utility, or a tradeoff that includes communicative effort?
• RQ2 Can RSA be grounded in Rate–Distortion theory, yielding RD-RSA with altered speaker updates?
• RQ3 How do RSA and RD-RSA dynamics behave as recursion depth grows and as alpha varies?
• RQ4 Do RD-RSA predictions better avoid non-informative random utterances while preserving explanatory power for human data?
• RQ5 How do RSA and RD-RSA compare to human pragmatic reasoning in reference-game data?

主な発見

• RSA recursion implements an alternating maximization that optimizes a tradeoff between expected utility and communicative effort, not strictly expected utility maximization.
• RD-RSA provides a principled RD-theoretic grounding of RSA with a slight but meaningful modification to the speaker's update rule.
• Both RSA and RD-RSA show alpha-dependent dynamics with a critical point at alpha = 1, and RD-RSA can converge globally at alpha = 1 under certain conditions.
• RD-RSA yields similar predictive accuracy to RSA on human data but avoids RSA’s bias toward non-informative random utterance production.
• Empirical comparison with a reference-game dataset shows RSA and RD-RSA both improve predictions over literal listeners in early recursion, with performance leveling or declining at greater depths.