Documentation

Linglib.Theories.Pragmatics.RSA.Implementations.TesslerGoodman2022

@[reducible, inline]

Discretized height values (in inches, scaled). Heights range from "short" (0) to "tall" (10) in discrete steps. Now backed by the canonical Degree 10 type with LinearOrder and BoundedOrder.

Equations
Instances For

    Comparison classes for scalar adjectives.

    The key distinction in the paper:

    • SUBORDINATE: Compare to the kind itself (e.g., basketball players)
    • SUPERORDINATE: Compare to the broader category (e.g., people in general)

    Example: "tall basketball player"

    • Subordinate: tall compared to basketball players
    • Superordinate: tall compared to people in general

    The paper shows that listeners infer which comparison class is intended.

    Instances For
      Equations
      • One or more equations did not get rendered due to their size.
      Instances For
        Equations
        • One or more equations did not get rendered due to their size.

        Kinds (nominals) that can be modified by adjectives.

        Each kind has:

        • An associated HEIGHT DISTRIBUTION (expectations)
        • A default comparison class preference

        The paper focuses on kinds with ATYPICAL expectations:

        • Basketball players: expected to be TALL
        • Jockeys: expected to be SHORT

        These create asymmetries in comparison class inference.

        Instances For
          Equations
          • One or more equations did not get rendered due to their size.
          Instances For
            Equations
            • One or more equations did not get rendered due to their size.

            Utterances about height

            Instances For
              Equations
              • One or more equations did not get rendered due to their size.
              Instances For
                Equations
                • One or more equations did not get rendered due to their size.

                Height prior for generic people (baseline). Normal distribution centered at h5.

                Equations
                • One or more equations did not get rendered due to their size.
                Instances For

                  Height prior for basketball players. Shifted RIGHT - basketball players are taller on average. Peak at h7 instead of h5.

                  Equations
                  • One or more equations did not get rendered due to their size.
                  Instances For

                    Height prior for jockeys. Shifted LEFT - jockeys are shorter on average. Peak at h3 instead of h5.

                    Equations
                    • One or more equations did not get rendered due to their size.
                    Instances For

                      Comparison class prior given kind: P(c | k).

                      The paper assumes a BASELINE preference for SUBORDINATE comparison class (compare to the specific kind), with some probability of SUPERORDINATE (compare to people in general).

                      For "basketball player":

                      • Default: compare to basketball players (subordinate)
                      • Alternative: compare to people (superordinate)
                      Equations
                      Instances For

                        The EFFECTIVE THRESHOLD for "tall" depends on the comparison class.

                        • Subordinate (basketball players): threshold is HIGHER (need to be tall even for a basketball player)
                        • Superordinate (people): threshold is at POPULATION MEAN

                        This is the key semantic insight: the comparison class shifts the threshold.

                        Equations
                        Instances For

                          The EFFECTIVE THRESHOLD for "short" depends on the comparison class.

                          Equations
                          Instances For

                            Meaning of "tall" given comparison class and kind.

                            [[tall]](c, k)(h) = 1 iff height(h) > threshold(c, k)

                            Equations
                            Instances For

                              Meaning of "short" given comparison class and kind.

                              [[short]](c, k)(h) = 1 iff height(h) < threshold(c, k)

                              Equations
                              Instances For

                                For generic "person", subordinate and superordinate thresholds are identical. This means the comparison class makes no difference for "person".

                                For basketball players, subordinate threshold is higher than superordinate. This is because basketball players are tall, so "tall for a basketball player" requires a higher degree than "tall for a person".

                                For jockeys, subordinate threshold is lower than superordinate. This is because jockeys are short, so "tall for a jockey" requires a lower degree than "tall for a person".

                                "tall" with superordinate comparison is true for MORE heights than with subordinate comparison for basketball players. This makes "tall" MORE INFORMATIVE (less restrictive) with superordinate, driving the pragmatic inference.

                                "short" with subordinate comparison covers MORE heights for basketball players than with superordinate comparison. Heights below 7 (subordinate) vs heights below 5 (superordinate).

                                Why the Polarity x Expectations Interaction Works #

                                The Informativity Argument #

                                For "tall basketball player" with SUBORDINATE comparison (to basketball players):

                                1. Threshold is HIGH (theta = 7 for basketball players)
                                2. Most basketball players are already tall (peak at h7)
                                3. "Tall" is TRUE for few basketball players (heights > 7 are fewer)
                                4. So "tall" is barely informative - true of a narrow range

                                For "tall basketball player" with SUPERORDINATE comparison (to people):

                                1. Threshold is LOWER (theta = 5 for people)
                                2. Most basketball players exceed this (they're tall for people)
                                3. "Tall" is TRUE for most basketball players
                                4. Speaker prefers this -> Listener infers superordinate

                                The Opposite for "Short" #

                                For "short basketball player" with SUBORDINATE comparison:

                                1. Threshold is HIGH (theta = 7)
                                2. "Short" means height < 7 (many basketball players qualify!)
                                3. "Short" is INFORMATIVE about the individual

                                For "short basketball player" with SUPERORDINATE comparison:

                                1. Threshold is LOWER (theta = 5)
                                2. "Short" means height < 5 (almost no basketball players!)
                                3. This is TRIVIALLY FALSE for most basketball players
                                4. Speaker wouldn't use this -> Listener infers subordinate

                                Relation to @cite{lassiter-goodman-2017} #

                                What LG2017 Covers #

                                What TG2022 Adds #

                                The Key Extension #

                                LG2017: Different comparison classes = different HEIGHT PRIORS

                                P_L1(h, theta | "tall", basketball) uses P(h | basketball)
                                P_L1(h, theta | "tall", jockey) uses P(h | jockey)
                                

                                TG2022: Listener INFERS which comparison class speaker intended

                                P_L1(h, c | "tall", basketball) reasons about P(c | basketball)
                                

                                The comparison class is no longer fixed by context - it's inferred! This allows for the polarity x expectations interaction.

                                Unified Model #

                                A full model would combine both:

                                P_L1(h, theta, c | u, k) proportional to P_S1(u | h, theta, c) * P(h | k) * P(theta) * P(c | k)
                                

                                For simplicity, we assume the threshold is implicitly determined by the comparison class (theta = tallThreshold c k), avoiding the extra variable.

                                Comparison Class as Nested Restriction #

                                @cite{tessler-goodman-2022}'s comparison class hierarchy is structurally a NestedRestriction: subordinate (restricted) ⊆ superordinate (unrestricted). This connects comparison class inference to the same nesting pattern used by @cite{ritchie-schiller-2024}'s domain restriction possibilities (DDRPs).

                                The downstream applications differ — threshold derivation vs. quantifier domain filtering — but the nesting structure is identical: going up the scale from subordinate to superordinate gives a superset of the reference population.

                                A reference population: which heights are typical for a given kind. For each kind, isTypical k h returns whether height h is in the kind's typical range (nonzero prior weight).

                                Equations
                                Instances For

                                  The comparison class hierarchy for a given kind induces a nested restriction on heights: subordinate filters to heights typical of the kind, while superordinate includes all heights.

                                  This is a NestedRestriction ComparisonClass Height: going from subordinate to superordinate widens the reference population.

                                  Equations
                                  • One or more equations did not get rendered due to their size.
                                  Instances For

                                    Basketball player comparison class restriction: subordinate filters to heights with nonzero basketball prior (heights 2–10).

                                    Equations
                                    Instances For

                                      Jockey comparison class restriction: subordinate filters to heights with nonzero jockey prior (heights 0–8).

                                      Equations
                                      Instances For

                                        Person comparison class restriction: subordinate = superordinate (all heights have nonzero prior), so restriction is vacuous.

                                        Equations
                                        Instances For

                                          For basketball players, the subordinate threshold is higher than the superordinate threshold. This is consistent with the nesting: subordinate restricts to basketball players (who are tall), so "tall" requires a higher bar to be informative within that restricted population.

                                          For jockeys, the subordinate threshold is lower than the superordinate threshold. Subordinate restricts to jockeys (who are short), so "tall" requires a lower bar within that restricted population.

                                          The nesting property holds: subordinate region ⊆ superordinate region. This is true for all kinds by construction.