Fuzzy Language in Action
Transcription
Fuzzy Language in Action
Fuzzy Language in Action Michael Franke October 6, 2012 Main Idea language use constitutes basic meaning Main Idea compositional semantic values influences language use determines constitutes basic meaning Main Points of Interest 1 optimal use of fuzzy language 2 vagueness 3 extreme-value use 4 use of borderline contradictions Sim-Max Games Experimental Data Vagueness Fuzzy → Action Sim-Max Games Experimental Data Vagueness Fuzzy → Action Lewis-Style Signaling Games ts ∈ T m∈M tr ∈ T ts = tr m success 5 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Lewis-Style Signaling Games ts ∈ T m∈M tr ∈ T ts 6 = tr m failure 6 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Signaling Games for Similarity Maximizing ts ∈ T m∈M tr ∈ T success ∝ similarity(ts , tr ) 7 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Signaling Games for Similarity Maximizing success(ts, tr ) = 1 − |ts − tr | (c.f. Jäger, 2007; Jäger and van Rooij, 2007) 8 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Utilities Sim-Max Games linear Gaussian 9 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Utilities Lewis-Style Signaling Games 10 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Strategies Sender σ — | T | × | M | row-stochastic matrix Receiver ρ — | M | × | T | row-stochastic matrix Example 0.2 0.5 σ = 0.1 0.0 1.0 | T | = 5 and | M | = 2 0.8 0.5 0.9 1.0 0.0 ρ= 0.1 0.5 0.8 0.1 0.1 0.0 0.0 0.3 0.0 0.1 ! 11 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Example Simulation Run 12 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Example Simulation Run 12 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Example Simulation Run 12 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Example Simulation Run 12 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Example Simulation Run 12 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Example Simulation Run 12 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Example Simulation Run 12 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Example Simulation Run 12 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Example Simulation Run 12 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Analytical Results esss ⊂ “Voronoi languages” • speaker strategy (≈ declarative meaning): partitions states into cells • receiver strategy (≈ imperative meaning): “Bayesian estimators” of cells Voronoi Language Example (1 dimensional space, 2 messages) R ( m1 ) | • {z S ( m1 ) R ( m2 ) } | • {z S ( m2 ) } (Jäger et al., 2011) 13 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Solt and Gotzner (2012): Material 14 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Solt and Gotzner (2012): Results 15 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Alxatib and Pelletier (2011): Material x is tall x is not tall x is tall and not tall x is neither tall nor not tall True False Can’t tell 16 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Alxatib and Pelletier (2011): Results P not-P both neither percentage true judgements 1 0.8 0.6 0.4 0.2 0 0 1 2 3 4 5 suspect 17 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Ripley (2011): Material Disagree • is near and it isn’t near . • both is and isn’t near . • neither is near nor isn’t near . • neither is nor isn’t near . Agree 1 2 3 4 5 6 7 18 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Ripley (2011): Results 19 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Shortcomings of Sim-Max Predictions 1 vagueness 2 extreme-value use 3 use of borderline contradictions R ( m1 ) | • {z S ( m1 ) R ( m2 ) } | • {z S ( m2 ) } 20 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Stochastic choice Best Response 1 P(ai ) = | argk uk =maxj uj | 0 if ui = maxj uj otherwise Quantal Response P(ai ) ∝ exp(λui ) Motivation • real people are not perfect utility maximizers • they make mistakes sub-optimal choices • still, high utility choices are more likely than low-utility ones (Goeree et al., 2008) 21 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Quantal Response Equilibria in Sim-Max Games 6 1 ·10−2 0.8 4 ρ(t | m) σ(m | t) 0.6 0.4 2 0.2 0 0 0 0.2 0.4 0.6 t 0.8 1 0 0.2 0.4 0.6 0.8 1 t (Franke et al., 2011) 22 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Where QRE goes wrong 23 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Noisy Best Response Confusion Matrix C — | T | × | T | row-stochastic matrix Cij ∝ N (ti , sn ) is the probability of confusing j for i Noisy Best Response NBR(ρ) = C BR(ρ) NBR(σ ) = BR(σ ) CT Motivation perceptual mistakes anti-proportional to similarity perhaps also plausible: probability of suitable context 24 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Noisy Response Equilibrium 25 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Fuzzy Logical Semantics Language L • basic expressions {P} or {P, Q} • degrees x ∈ R • if A is a basic expression, then Ax is a formula • if ϕ and ψ are formulas, then so are ¬ ϕ and ϕ ∧ ψ Semantics V (Ax) ∈ [0; 1] V (¬ ϕ) = 1 − V ( ϕ) V ( ϕ ∧ ψ) = min(V ( ϕ), V (ψ)) 26 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Message Space M ≈ {{ψ ∈ L | V ( ϕ) = V (ψ)} | ϕ ∈ L} M1 = {P , ¬P , P ∧ ¬P} M2 = {P , Q , ¬P , ¬Q , P ∧ Q , ¬P ∧ ¬Q , P ∧ ¬P , Q ∧ ¬Q} 27 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action The Dynamics ρ0 = 0 1 ... ... 1 0 ! σk = Extend(NBR(ρk−1 )) ρk = NBR(σk ) Extend(sigma): for i <= |M|: if i is basic: s_i = sigma_i else: s_i = normal fuzzy values based on use of basic predicates as given in sigma 28 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Basic Results with P and Q Pr = N (1/2, 1/4), noise level = .25 29 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Comparison with Solt and Gotzner (2012): Baseline noise level = .25 30 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Comp. with Solt and Gotzner (2012): Left-Skewed noise level = .25 31 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Comp. with Solt and Gotzner (2012): Right-Skewed noise level = .25 32 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Comparison with Alxatib and Pelletier (2011) noise level = .25 33 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action Summary compositional semantic values influences language use determines constitutes basic meaning Main Result availability of complex forms shapes meaning of basic forms 34 / 37 Sim-Max Games Experimental Data Vagueness Fuzzy → Action To Do 1 charter dynamics 2 model fitting 3 more data 1 2 antonym pairs unrelated properties 35 / 37 References Alxatib, Sam and Francis Jeffry Pelletier (2011). “The Psychology of Vagueness: Borderline Cases and Contradictions”. In: Mind & Language 26.3, pp. 287–326. Franke, Michael et al. (2011). “Vagueness, Signaling & Bounded Rationality”. In: JSAI-isAI 2010. Ed. by T. Onoda et al. Springer, pp. 45–59. Goeree, Jacob K. et al. (2008). “Quantal Response Equilibrium”. In: The New Palgrave Dictionary of Economics. Ed. by Steven N. Durlauf and Lawrence E. Blume. Basingstoke: Palgrave Macmillan. Jäger, Gerhard (2007). “The Evolution of Convex Categories”. In: Linguistics and Philosophy 30.5, pp. 551–564. Jäger, Gerhard and Robert van Rooij (2007). “Language Structure: Psychological and Social Constraints”. In: Synthese 159.1, pp. 99–130. Jäger, Gerhard et al. (2011). “Voronoi Languages: Equilibria in Cheap-Talk Games with High-Dimensional Types and Few Signals”. To appear in Games and Economic Behavior. Ripley, David (2011). “Contradictions at the Borders”. In: Vagueness in Communication. Ed. by Rick Nouwen et al. Springer, pp. 169–188. Solt, Stephanie and Nicole Gotzner (2012). “Who Here is Tall? Comparison Classes, Standards and Scales”. In: Pre-Proceedings of the International Conference Linguistic Evidence, pp. 79–83.