Metaphor is often viewed as a linguistic ornament or a deviation from literal language. However, a long-standing research direction suggests that metaphor is a fundamental cognitive tool for structuring abstract thought, handled not by specialized modules, but by the same inferential mechanisms used for literal discourse. By treating metaphorical expressions as prompts for reasoning rather than simple substitutions, computational systems can resolve complex, novel, and "map-transcending" language.
Metaphor as Selective Inferencing
The core challenge in metaphor interpretation is the potential for an explosion of irrelevant meanings. If a person says an economic policy is a "stranglehold," a system could theoretically infer properties of physical hands, oxygen deprivation, or biological anatomy. To prevent this, Jerry R. Hobbs argues in Metaphor Interpretation as Selective Inferencing that metaphor resolution is a byproduct of maintaining discourse coherence.
Instead of generating all possible literal meanings, the system performs selective inferencing. It draws only the specific logical conclusions necessary to satisfy discourse relations like contrast, parallelism, or pronoun resolution. As noted in 77, the relevance of an inference is dynamically determined by the communicative goals of the surrounding text. A metaphor is "understood" when the system identifies the minimal set of attributes that make the statement logically coherent within its immediate linguistic environment.
Reasoning in the Pretence Space
While selective inferencing explains how we filter meanings, it does not fully explain how we handle novel metaphors that lack pre-existing mapping rules. To address this, Barnden et al. proposed the ATT-Meta approach in Reasoning in Metaphor Understanding.
This framework introduces a "reasoning-first" strategy. When a system encounters a metaphor, it creates a pretence space—a protected logical context where the metaphorical statement is temporarily treated as literally true. Within this space, the system uses its existing knowledge of the "source domain" (e.g., physical motion or biological processes) to perform literal reasoning. For example, if a narrative mentions "stumbling" in an economic plan, the system reasons within the source domain of motion to conclude that stumbling causes a temporary suspension of progress.
As Agerri et al. explain in On the formalization of Invariant Mappings for Metaphor Interpretation, this literal reasoning allows the system to extract implicit details—such as timing, causation, and intensity—before mapping the final conclusions back to the "target domain" (the abstract concept). This allows the model to understand "map-transcending" details that aren't covered by standard conceptual metaphors.
Evaluating Understanding via Entailment
To move beyond simple metaphor detection, researchers have increasingly framed interpretation as a Natural Language Inference (NLI) or Recognizing Textual Entailment (RTE) task. The intuition is simple: if a system truly understands a metaphor, it should be able to determine if that metaphor logically entails a literal paraphrase.
Rodrigo Agerri argues in Metaphor in Textual Entailment that this framework transforms figurative language into a measurable logical task. By testing whether a metaphorical premise implies a literal hypothesis, researchers can evaluate semantic comprehension rather than surface-level pattern matching.
Recent benchmarks have highlighted the difficulty of this task. Chakrabarty et al. found in Figurative Language in Recognizing Textual Entailment that even advanced models like RoBERTa struggle when forced to "read between the lines," often failing to resolve the underlying properties being compared. Similarly, Comsa et al. demonstrated in MiQA that while large language models are improving, there remains a significant gap in their ability to integrate metaphorical comprehension with deep commonsense reasoning.
Ultimately, this research direction suggests that understanding a metaphor is not about decoding a secret language, but about applying standard reasoning tools to a "pretence" of reality to maintain the flow of discourse.
Key Foundational Work
Metaphor Interpretation as Selective Inferencing (1981) — contributed 5 concept elements to this direction.
Metaphor Interpretation as Selective Inferencing (1981) — contributed 3 concept elements to this direction.
Metaphor Interpretation as Selective Inferencing (1981) — contributed 3 concept elements to this direction.