Metaphor is pervasive in everyday language: we "grasp" ideas, "attack" arguments, and "navigate" conversations. Computational metaphor processing encompasses both metaphor detection (identifying which words or phrases are used metaphorically) and metaphor interpretation (determining the intended meaning). The field draws on Lakoff and Johnson's Conceptual Metaphor Theory (CMT), which holds that metaphors are not merely linguistic decorations but reflect systematic mappings between conceptual domains, such as ARGUMENT IS WAR or TIME IS MONEY.
Conceptual Metaphor Theory
ARGUMENT IS WAR:
attack → criticize, defend → support, win → persuade
Detection features:
Selectional preference violation: P(w | context) is low for literal sense
Domain incongruence: topic(w) ≠ topic(context)
Abstractness shift: concrete word in abstract context
According to CMT, metaphorical expressions are surface manifestations of deeper conceptual mappings between a source domain (typically concrete and embodied) and a target domain (typically abstract). The mapping is systematic: if LOVE IS A JOURNEY, then lovers are travelers, the relationship is a vehicle, difficulties are obstacles, and goals are destinations. This systematicity means that metaphor processing is not simply a matter of listing idiomatic expressions but requires understanding the underlying conceptual structure.
Computational Approaches
Computational metaphor detection has been approached through several paradigms. Selectional preference violation methods identify metaphors as words that violate the expected semantic type of their argument position: "drink" expects a liquid object, so "drink in the scenery" signals metaphorical use. Clustering and topic models detect domain incongruence by identifying words whose topic cluster differs from that of their context. Neural models fine-tuned on metaphor-annotated corpora (VUA Metaphor Corpus, TroFi) achieve state-of-the-art detection, with BERT-based classifiers reaching F1 scores above 75% on token-level metaphor identification.
Metaphor awareness improves several NLP tasks. Sentiment analysis systems that recognize "This movie is a roller coaster" as metaphorical can better assess its evaluative import. Machine translation must handle metaphors that differ across cultures: English "raining cats and dogs" has no literal counterpart in most languages. In clinical NLP, metaphor use patterns have been explored as potential markers for mental health conditions. Metaphor generation, producing apt novel metaphors, is an emerging challenge for creative text generation systems.
Challenges and Resources
Key challenges in metaphor processing include the continuum between literal and metaphorical usage (many expressions are conventionalized to varying degrees), the difficulty of distinguishing metaphor from metonymy and other figurative devices, and the dependence of metaphoricity on context and speaker intent. The VUA Metaphor Corpus, based on the BNC-Baby corpus, provides the largest manually annotated dataset for metaphor detection, using the MIPVU procedure (Metaphor Identification Procedure Vrije Universiteit) for systematic annotation.
Recent advances leverage pre-trained language models and multimodal information. Models that combine linguistic features with conceptual knowledge from resources like ConceptNet and framenet improve detection of novel metaphors. Cross-lingual metaphor detection exploits the fact that many conceptual metaphors are shared across languages while their linguistic realizations differ. The interaction between metaphor and other aspects of figurative language -- irony, hyperbole, simile -- is an active area of research requiring integrated models of non-literal language processing.