Literal Meaning and the Comprehension of Metaphors

Steven L. Lytinen, Robert R. Burridge, Jeffrey D. KirtnerAAAI 1992

Introduction

The mechanism by which humans comprehend metaphors has long been a subject of debate in psycholinguistics. The "Standard Pragmatic Model" traditionally posited a serial process: a listener first attempts a literal interpretation of a statement; if that interpretation fails to meet pragmatic constraints (a "pragmatic slip"), the listener then triggers a search for a metaphorical meaning. However, empirical evidence from the late 20th century challenged this view, showing that metaphors are often processed as quickly as literal sentences. This led to a theoretical impasse: does literal meaning play any role in metaphor comprehension, or is it bypassed entirely? This paper proposes a computational resolution using the LINK system, arguing that literal and metaphorical meanings are computed simultaneously and automatically.

The authors implement their model within LINK, an integrated, unification-based natural language processing (NLP) system. In LINK, syntactic, semantic, and domain knowledge are represented as unification constraint rules. The primary innovation introduced is the Mapping Rule.

Unlike previous failure-driven models, LINK applies mapping rules incrementally and automatically. As the parser constructs a literal interpretation, it concurrently checks for applicable mapping rules. If a match is found, the system immediately generates a metaphorical interpretation as a competing hypothesis in the chart parser.

Formal syntax for mapping rules in the LINK system.

The formal structure of these rules, as shown in the grammar above, allows for a high degree of productivity. For instance, a general mapping rule such as CHANGE-ALTITUDE ⇒ CHANGE-VALUE allows the system to interpret a variety of related metaphors (e.g., "prices rose," "stocks plummeted," "inflation leveled off") using a single underlying logic rather than requiring separate lexical entries for every figurative sense.

Parallel Processing and Automaticity

The model suggests that the perceived "slowdown" in older metaphor processing theories was not due to the computation of literal meaning itself, but rather the delay caused by waiting for a literal failure to trigger metaphorical search. By processing both paths in parallel, the time cost is equalized.

This approach draws a parallel between metaphor comprehension and lexical ambiguity resolution. Just as multiple meanings of a word like "bank" are briefly activated before context suppresses the irrelevant one, both literal and metaphorical interpretations are activated. Contextual and domain constraints then prune the inappropriate interpretation later in the process. This explains why literal meanings do not appear "primed" at the end of a metaphorical sentence; they have been suppressed by the time the sentence is fully processed.

Results and Discussion

The LINK model successfully aligns with psycholinguistic data, demonstrating that metaphors need not be slower than literal text. Furthermore, the framework proves robust enough to handle complex linguistic phenomena beyond simple metaphors, including idioms (e.g., "go through the roof") and metonymy (e.g., using "the stock market" to refer to a numerical index).

One significant advantage of this model is its ability to account for the absence of literal priming. Because the system uses context to select the most appropriate interpretation, the non-selected literal meaning is suppressed, mirroring human cognitive behavior where the literal sense of a metaphor is not readily accessible once the figurative meaning is understood.

Limitations and Future Directions

While the model provides a compelling computational account of metaphor processing, it faces several limitations. The current implementation relies on a hand-crafted library of mapping rules, which would need to be significantly expanded to achieve universal coverage. Additionally, the model does not yet fully integrate frequency effects—the observation that more common non-literal meanings might be processed more readily. Future work should also focus on word-level activation during the middle of phrases to test the model's predictions about incremental processing more granularly.

Summary updated

BibTeX
@inproceedings{lytinen1992literal,
  author = {Steven L. Lytinen and Robert R. Burridge and Jeffrey D. Kirtner},
  title = {{Literal Meaning and the Comprehension of Metaphors}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {1992},
}
Figure