Linglib.Studies.YoonEtAl2020

RSA model of [YTGF20] (Open Mind 4): polite speech arises from a speaker trading off three communicative goals — to be informative, to be kind, and to appear informative and kind. The experimental domain: Ann rates Bob's poem (states s₀–s₃, hearts) and chooses among eight utterances ({terrible, bad, good, amazing} × {plain, negated}).

The model stack (the paper's Figure 4): a literal listener P_L0(s|w) ∝ L(w,s)·P(s) over empirically elicited soft semantics; a first-order speaker with utility U_S1 = φ·ln P_L0(s|w) + (1−φ)·E_{P_L0}[V(s)] − c·l(w); a pragmatic listener jointly inferring state and goal weight, P_L1(s,φ|w) ∝ P_S1(w|s,φ)·P(s)·P(φ); and a second-order polite speaker with U_S2 = ω_inf·ln P_L1(s|w) + ω_soc·E_{P_L1}[V(s)] + ω_pres·ln P_L1(φ̂|w) − c·l(w).

Instantiated on the canonical pipeline: the S1 speaker is RSA.Canonical.S1 over (state × φ) speaker situations, and the pragmatic listener is RSA.Canonical.L1, the joint posterior over HeartState × Phi — the paper's eq. (4) by construction, with the U_pres marginal available as .snd. The paper's L0-gate (utterances with zero literal fit are unavailable to informativity-sensitive speakers) is the ⊥-utility branch.

Main statements #

social_prefers_indirect / social_prefers_positive — the pure-social speaker (φ = 0) prefers indirect and positive utterances (Figure 2, S1 social facet), for every α > 0: pure expected-value comparisons.
informative_prefers_direct / informative_prefers_direct_positive — the pure-informative speaker (φ = 1) prefers direct utterances (Figure 2, S1 informational facet), for every α > 0: the first is the L0-gate, the second a log-monotonicity comparison.
s2Utility — the full three-goal S2 utility over the joint listener, with the Table-2 MAP weight profiles.
socialGoalSubjectivityLevel — bridge to [TD02]'s intersubjectivity cline.

Implementation notes #

Lexicon provenance. softSemantics stipulates acceptance proportions k/49 attributed to literal_semantics.csv in the paper's repository (the paper itself prints no θ table; N = 51 recruited per the supplement). -- UNVERIFIED: the k/49 values and the N = 49-after-exclusions claim await verification against the CSV. Negated utterances use graded negation ⟦not φ⟧ = 1 − ⟦φ⟧ — a compositional construction of this file, not the paper's lexicon: the paper elicits θ per utterance, including negated forms. The φ grid discretizes the paper's continuous φ ~ Uniform(0,1) to five points.

Findings policy. S1-level preferences are stated as theorems: they are parameter-free (any α > 0) and robust to lexicon perturbation. The S2-level negation preferences and the L1 state-inference orderings are numeric facts — α-dependent and, for S2, sensitive to the third decimal of single norming proportions — and are recorded as verified prose (§S2 below), per the library's policy on findings whose truth depends on exact parameter values. Notably, independent recomputation shows the paper's headline U_S2(not terrible) > U_S2(terrible) under both-goal weights at state s₀ is TRUE at the fitted parameters (margin ≈ 0.14) — correcting the previous version of this file, which left it sorryed with a docstring wrongly claiming it fails under point-estimate semantics (the old reflection tactic's interval arithmetic merely could not separate the sides).

[YTGF20] — Polite speech emerges from competing social goals #

Main statements #

Implementation notes #

States, utterances, goals #

The soft lexicon #

The literal listener in closed form #

The S1 speaker on the canonical pipeline #

S1 findings (Figure 2, S1 facets) — structural, every α > 0 #

The pragmatic listener: the joint (state, goal) posterior #

L1 state inferences (prose) #

The S2 polite speaker #

S2 findings (prose, independently recomputed) #

Bridge to the subjectivity cline #