Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
Instances For
A disambiguation scenario with selectional and scenario constraints
- word : String
Name of the ambiguous word
- context : String
Context sentence
- selectional : C → ℚ
Selectional constraint (from predicate)
- scenario : C → ℚ
Scenario constraint (from frame/context words)
- concepts : List C
Support (list of concepts)
Instances For
Equations
- One or more equations did not get rendered due to their size.
Example 1: "A bat was sleeping" #
Context: The verb SLEEP provides a strong selectional preference.
Constraints:
- Selectional (SLEEP wants animate subject):
- P(ANIMAL | subject-of-SLEEP) = 0.95
- P(EQUIPMENT | subject-of-SLEEP) = 0.05
- Scenario (neutral, no frame activated):
- P(ANIMAL | neutral) = 0.5
- P(EQUIPMENT | neutral) = 0.5
Prediction: ANIMAL wins because selectional preference is strong.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Worked Computation #
Step 1: Unnormalized posteriors
- unnorm(ANIMAL) = 0.95 × 0.50 = 0.475
- unnorm(EQUIPMENT) = 0.05 × 0.50 = 0.025
Step 2: Partition function
- Z = 0.475 + 0.025 = 0.50
Step 3: Normalized posteriors
- P(ANIMAL) = 0.475 / 0.50 = 0.95
- P(EQUIPMENT) = 0.025 / 0.50 = 0.05
Result: Strong preference for ANIMAL (95%)
Example 2: "A player was holding a bat" #
Context: The word "player" activates a SPORTS scenario. The verb HOLD has weak selectional preference (both concepts are holdable).
Constraints:
- Selectional (HOLD wants concrete object):
- P(ANIMAL | object-of-HOLD) = 0.4 (can hold an animal)
- P(EQUIPMENT | object-of-HOLD) = 0.6 (slightly prefer inanimate)
- Scenario (SPORTS frame from "player"):
- P(ANIMAL | SPORTS) = 0.1
- P(EQUIPMENT | SPORTS) = 0.9
Prediction: EQUIPMENT wins because scenario constraint is strong and selectional is weak.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Worked Computation #
Step 1: Unnormalized posteriors
- unnorm(ANIMAL) = 0.40 × 0.10 = 0.04
- unnorm(EQUIPMENT) = 0.60 × 0.90 = 0.54
Step 2: Partition function
- Z = 0.04 + 0.54 = 0.58
Step 3: Normalized posteriors
- P(ANIMAL) = 0.04 / 0.58 ≈ 0.069
- P(EQUIPMENT) = 0.54 / 0.58 ≈ 0.931
Result: Strong preference for EQUIPMENT (93%)
Key observation: Even though HOLD doesn't strongly select for equipment, the SPORTS scenario from "player" tips the balance decisively.
Example 3: "The astronomer married the star" #
Context: Competing constraints create a pun/zeugma reading.
Constraints:
- Selectional (MARRY wants human object):
- P(CELEBRITY | object-of-MARRY) = 0.9
- P(CELESTIAL | object-of-MARRY) = 0.1
- Scenario (ASTRONOMY frame from "astronomer"):
- P(CELEBRITY | ASTRONOMY) = 0.1
- P(CELESTIAL | ASTRONOMY) = 0.9
Prediction: TIE → pun/zeugma reading emerges.
This is the signature example from the paper showing how conflicting constraints predict pragmatic ambiguity.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Worked Computation #
Step 1: Unnormalized posteriors
- unnorm(CELEBRITY) = 0.90 × 0.10 = 0.09
- unnorm(CELESTIAL) = 0.10 × 0.90 = 0.09
Step 2: Partition function
- Z = 0.09 + 0.09 = 0.18
Step 3: Normalized posteriors
- P(CELEBRITY) = 0.09 / 0.18 = 0.50
- P(CELESTIAL) = 0.09 / 0.18 = 0.50
Result: Perfect tie (50-50)
Key observation: When selectional and scenario constraints have equal strength but opposite preferences, we get a tie. This predicts:
- Pun reading (both meanings simultaneously)
- Zeugma effect (semantic clash)
- Garden path potential
Equations
- One or more equations did not get rendered due to their size.
Instances For
Example 4: "The sailor liked the port" #
Context: Both "sailor" (activates NAUTICAL scenario) and "port" (ambiguous between harbor/wine/computer) need disambiguation.
This shows how scenario constraints propagate: "sailor" activates NAUTICAL which then disambiguates "port".
Constraints for "port":
- Selectional (LIKE is neutral):
- P(HARBOR | object-of-LIKE) = 0.33
- P(WINE | object-of-LIKE) = 0.33
- P(COMPUTER | object-of-LIKE) = 0.33
- Scenario (NAUTICAL frame from "sailor"):
- P(HARBOR | NAUTICAL) = 0.7
- P(WINE | NAUTICAL) = 0.25 (sailors drink!)
- P(COMPUTER | NAUTICAL) = 0.05
Prediction: HARBOR wins, WINE is plausible, COMPUTER unlikely.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Worked Computation #
Step 1: Unnormalized posteriors
- unnorm(HARBOR) = 0.33 × 0.70 = 0.231
- unnorm(WINE) = 0.33 × 0.25 = 0.0825
- unnorm(COMPUTER) = 0.34 × 0.05 = 0.017
Step 2: Partition function
- Z = 0.231 + 0.0825 + 0.017 = 0.3305
Step 3: Normalized posteriors
- P(HARBOR) = 0.231 / 0.3305 ≈ 0.699
- P(WINE) = 0.0825 / 0.3305 ≈ 0.250
- P(COMPUTER) = 0.017 / 0.3305 ≈ 0.051
Result: HARBOR (70%), WINE (25%), COMPUTER (5%)
Key observation: With a neutral predicate (LIKE), the scenario constraint from "sailor" does all the disambiguation work. WINE remains plausible due to cultural association.
Example 5: "The coach told the star to play" #
Context: Multiple words contribute to the scenario:
- "coach" → SPORTS frame
- "play" → reinforces SPORTS (or ENTERTAINMENT)
This shows how constraints from multiple words combine.
Constraints for "star":
- Selectional (TELL wants animate recipient):
- P(CELEBRITY | recipient-of-TELL) = 0.95
- P(CELESTIAL | recipient-of-TELL) = 0.05
- Scenario (SPORTS frame from "coach" + "play"):
- P(CELEBRITY | SPORTS) = 0.8 (sports stars are celebrities)
- P(CELESTIAL | SPORTS) = 0.2
Prediction: CELEBRITY wins strongly (both constraints agree).
Equations
- One or more equations did not get rendered due to their size.
Instances For
Worked Computation #
Step 1: Unnormalized posteriors
- unnorm(CELEBRITY) = 0.95 × 0.80 = 0.76
- unnorm(CELESTIAL) = 0.05 × 0.20 = 0.01
Step 2: Partition function
- Z = 0.76 + 0.01 = 0.77
Step 3: Normalized posteriors
- P(CELEBRITY) = 0.76 / 0.77 ≈ 0.987
- P(CELESTIAL) = 0.01 / 0.77 ≈ 0.013
Result: CELEBRITY wins decisively (98.7%)
Key observation: When selectional and scenario constraints agree, they reinforce each other multiplicatively, leading to very confident disambiguation.
Example 6: Varying Constraint Strengths #
This example shows how the RATIO of constraint strengths matters.
Consider "The child saw the bat":
- CHILD activates weak DOMESTIC scenario (pets, toys)
- SAW is perceptually neutral
- How does varying the scenario strength affect disambiguation?
Parameterized bat disambiguation varying scenario strength
Equations
- One or more equations did not get rendered due to their size.
Instances For
Varying Scenario Strength #
| Scenario Strength | P(ANIMAL) | P(EQUIPMENT) | Interpretation |
|---|---|---|---|
| 0.5 (neutral) | 0.50 | 0.50 | Ambiguous |
| 0.6 | 0.60 | 0.40 | Slight animal |
| 0.7 | 0.70 | 0.30 | Prefer animal |
| 0.8 | 0.80 | 0.20 | Strong animal |
With neutral selectional constraint (0.5/0.5), the scenario constraint directly determines the posterior.
Example 7: Complete Analysis of "marry a star" #
The paper analyzes different contexts for "marry a star":
Neutral context: "Someone married a star"
- Selectional dominates → CELEBRITY
Astronomy context: "The astronomer married the star"
- Conflict → TIE
Hollywood context: "The producer married the star"
- Both agree → CELEBRITY (reinforced)
Sci-fi context: "The alien married the star"
- Weak conflict → depends on genre conventions
Neutral context: "Someone married a star"
Equations
- One or more equations did not get rendered due to their size.
Instances For
Hollywood context: "The producer married the star"
Equations
- One or more equations did not get rendered due to their size.
Instances For
Sci-fi context: "The alien married the star"
Equations
- One or more equations did not get rendered due to their size.
Instances For
Comparison of Contexts #
| Context | Sel(C) | Sel(S) | Scen(C) | Scen(S) | P(CELEBRITY) |
|---|---|---|---|---|---|
| Neutral | 0.90 | 0.10 | 0.50 | 0.50 | 0.90 |
| Astronomer | 0.90 | 0.10 | 0.10 | 0.90 | 0.50 (TIE) |
| Producer | 0.90 | 0.10 | 0.95 | 0.05 | 0.99 |
| Alien | 0.60 | 0.40 | 0.30 | 0.70 | 0.39 |
The "alien" case is interesting: even though CELESTIAL wins, it's not a clear pun because the selectional constraint is also weakened in sci-fi contexts.
Summary: Compositional Constraint Interaction #
Key Principles from @cite{erk-herbelot-2024} #
Product of Experts: Constraints multiply, they do not add.
- Both must be satisfied for high probability
- One zero kills the interpretation
Relative strength matters: The ratio determines the winner.
- Strong selectional + weak scenario → selectional wins
- Weak selectional + strong scenario → scenario wins
- Equal strengths → conflict/tie
Scenario propagation: Context words activate frames.
- "sailor" → NAUTICAL
- "coach" + "play" → SPORTS
- "astronomer" → ASTRONOMY
Conflict predicts pragmatic effects:
- Tie → pun/zeugma reading
- Near-tie → garden path potential
- Agreement → confident interpretation
Computational Pattern #
For word w in context C with concepts {c₁, c₂,...}:
P(cᵢ | C) ∝ P_sel(cᵢ | predicate) × P_scen(cᵢ | frame(C))
Where:
- P_sel comes from selectional preferences of governing predicates
- P_scen comes from scenario/frame activated by context words
- The product is normalized over all concepts