📄 A Conceptual Framework for Inferring Implicatures

Summary

The paper addresses a gap in sentiment analysis. Most systems detect only opinions that are stated outright, but readers also infer attitudes that are never voiced directly. The authors call these opinion implicatures: defeasible, default inferences about sentiment, modelled on Grice's generalized conversational implicature. They present a rule-based conceptual framework that starts from explicit sentiments and goodFor/badFor (gfbf) events, encoded as agent–event–object triples. From these it derives further sentiments and beliefs toward events, agents and objects. To keep parser noise out of the picture, the input is manually annotated sentiment and gfbf information. The system stores everything in a directed graph and applies rules repeatedly until nothing new follows. Inferences are kept inside nested private state spaces, which record whose view each conclusion belongs to. Worked examples come from the MPQA corpus. In one, annotators could find no clear sentiment phrase, yet the system still recovers the writer's negative stance toward an agent. The paper reports no numbers. Its value is as a blueprint for deeper, explainable interpretation of opinionated text.

Key points

• Input format. The input is annotated explicit sentiments plus gfbf triples ⟨agent, event, object⟩. The gfbf annotations come from the corpus of Deng et al. (2013). Every root node must be a writer-sourced sentiment or believesTrue node.
• Rule form. Rules follow Hobbs-style abduction: preconditions, then assumptions, then conclusions. A conclusion fires only if the preconditions hold, there is some basis for each assumption, and no evidence contradicts the assumptions or the conclusions.
• Rule inventory. About ten rules are shown, including:
  • sentiment toward an event leading to sentiment toward its object, and the reverse direction;
  • an animate agent's event being treated as intended, so the agent likes the idea of it;
  • nested sentiment leading to agreement or disagreement, and agreement in turn implying sentiment toward the other holder;
  • writer sentiment derived from a connotation lexicon (Feng et al.).
• Private state spaces. A private state space is a path of believesTrue/sentiment nodes rooted at the writer. Two paths are equivalent when attitude type, polarity and source match at every position. This lets the system hold claims such as "the writer believes Israel is negative toward Palestinians" without asserting them globally.
• Implementation status. A graph-based computational model implementing some of these rules exists (Deng and Wiebe, 2014). Joint optimization over ambiguous input is ongoing work.

Interesting ideas

• I1. Defeasible rules with explicit assumptions. Each rule names its assumptions separately, and any rule is blocked by contrary evidence, so default inferences stay cancellable and auditable.
• I2. Nested attitude spaces keyed by source chain. Propositions live inside nested attitude spaces identified by the chain of (source, attitude, polarity) leading to them. This keeps "X believes Y feels Z" separate from facts the text itself asserts.
• I3. Coercing events into ideas. Wrapping an event in an "ideaOf" node lets the system reason about attitudes toward it without first deciding whether the event really happened, will happen, or is only planned.
• I4. Propagating sentiment through goodFor/badFor events. Sentiment travels across gfbf events in both directions: object to event, event to agent, and through agreement between holders. This recovers attitudes where no sentiment word is present.
• I5. Gold input to isolate the inference layer. Running the inference layer on gold input separates its errors from upstream noise, which gives a clean way to judge the layer itself.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): crate prag could tag each implicature with the assumptions it relied on. A later contradiction would then retract exactly the conclusions that depended on it, and the trace would stay explainable.
• I2 (world): the world of a text could hold entities and states inside belief/sentiment contexts per holder. Fable QA could then tell what a character thinks apart from what actually happened.
• I3 (frames): attitude frames could attach to an idea-node of the event. This would sidestep realis classification when a planned action ("planning to trigger wars") draws sentiment.
• I4 (frames): VerbNet roles already give agent and patient. A small gfbf verb list plus these propagation rules would let prag infer the writer's stance on characters in tales.
• I5 (learn): evaluate prag's implicature rules once on gold UD trees and frames and once on gparse output. The difference shows which failures belong to the rules and which to parsing.