Introduction: “Metaphor Comprehension: What Makes a Metaphor Difficult to Understand?” by Walter Kintsch & Anita R. Bowles
“Metaphor Comprehension: What Makes a Metaphor Difficult to Understand?” by Walter Kintsch and Anita R. Bowles first appeared in Metaphor and Symbol in 2002 (Vol. 17, Issue 4, pp. 249–262), published by Psychology Press. This pivotal article investigates the cognitive mechanisms underpinning metaphor comprehension, challenging the traditional view that metaphors inherently require a qualitatively different processing strategy than literal language. Through empirical analysis and computational modeling, the authors demonstrate that metaphors of the form NOUN1 is a NOUN2 are often understood using the same basic cognitive strategies applied to literal sentences. Central to their study is the use of Latent Semantic Analysis (LSA), a method for modeling meaning in high-dimensional semantic space, and the predication model, which adjusts the vector of a predicate based on its argument to yield context-sensitive interpretations. Their findings reveal that metaphor comprehension difficulty is not significantly related to the surface semantic similarity between the metaphor’s terms, but rather to the availability of shared semantic features that link the metaphor’s topic and vehicle. Importantly, both human participants and the computational model showed similar patterns in interpreting metaphors: strong agreement and coherence for easy metaphors, and more diverse yet non-random responses for difficult ones. This work is significant in literary theory and cognitive linguistics as it offers a formal, computable framework to explain metaphor comprehension, moving beyond intuitive or purely analogical models. It aligns with, and extends, the class-inclusion theory of Glucksberg (1998) and supports a semantic-constraint-based view of comprehension that blurs the boundaries between literal and figurative language processing.
Summary of “Metaphor Comprehension: What Makes a Metaphor Difficult to Understand?” by Walter Kintsch & Anita R. Bowles
🔷 1. Metaphors and Literal Language: No Special Processing Required
- People process metaphors similarly to literal sentences in most cases.
- ❝It does not appear that metaphor comprehension first involves an attempt at literal comprehension and, when that fails, a metaphoric reinterpretation❞ (Kintsch & Bowles, 2002, p. 249).
- Ordinary metaphors are usually automatically understood, without cognitive overload.
- 🔍 This finding challenges traditional theories that treat metaphor as inherently more complex than literal language.
🔶 2. What Makes a Metaphor Difficult? It’s Not What You Think
- Difficulty is not due to:
- Semantic distance between words 🔁
- Word frequency or vector length 🧮
- ❝It is not the case that easy understanding requires a preexisting global relation between the two terms❞ (p. 258).
- Rather, it depends on whether shared semantic neighbors can be found between topic and vehicle.
- ❝Metaphors are easy to process if the argument has a good match among the close neighbors of the predicate❞ (p. 257).
🟣 3. Latent Semantic Analysis (LSA): Mapping Meaning in Space
- Words are represented as vectors in a 300-dimensional semantic space.
- ❝Meaning is a position in this huge semantic space… we can calculate how close or far apart two vectors are❞ (p. 250).
- Sentence meaning is computed by adding vectors, allowing computational modeling of metaphor comprehension.
🟢 4. The Predication Model: Adding Context to Semantics
- The predication algorithm modifies predicate vectors based on context (argument word).
- For example, 🦈 “My lawyer is a shark” highlights “aggressive” traits of “shark,” not the literal ones.
- ❝The meaning of the predicate is a shark is very different from shark in isolation❞ (p. 251).
- This is how LSA simulates human-like metaphor interpretation.
🔴 5. Easy vs. Difficult Metaphors: Experimental Evidence
- Participants rated 13 metaphors as easy and 13 as difficult.
- High agreement on easy metaphors (48% modal agreement) vs. low on difficult ones (21%).
- Even “nonsense” metaphors triggered non-random interpretations.
- ❝Even for what one might regard as pure nonsense, there was still a considerable level of agreement❞ (p. 254).
🔵 6. Model Validation: Matching Human Responses
- LSA-predicated vectors closely matched participant-generated interpretations.
- For both easy and difficult metaphors, the average cosine similarity between model and human responses was ≈ 0.51.
- ❝For difficult metaphors, responses were more varied, but the model produced a vector that was just as close to these varied responses❞ (p. 258).
🟡 7. Cognitive Consistency: Even Diffuse Responses Make Sense
- The model does not break down on difficult metaphors—it generates diffuse but coherent meanings.
- ❝The semantic structure provided a tight constraint for easy metaphors, and only a loose one for hard metaphors❞ (p. 258).
- Human and model interpretations converge because of shared semantic constraints.
🟤 8. Theoretical Contributions to Literary and Linguistic Theory
- Supports Glucksberg’s class-inclusion model and Frisson & Pickering’s underspecification model.
- Offers a computational realization of metaphor interpretation mechanisms in cognitive science.
- ❝We also claim that the results presented here show that LSA provides a useful basis for a psychological theory of meaning❞ (p. 259).
🔺 9. Key Insight: Local Connections Trump Global Similarity
- Metaphors work not by global similarity, but by activating shared contextual features.
- ❝Lawyer and shark are orthogonal… but there are aspects—like vicious or mean—that link the two❞ (p. 258).
Theoretical Terms/Concepts in “Metaphor Comprehension: What Makes a Metaphor Difficult to Understand?” by Walter Kintsch & Anita R. Bowles
| 🌐 Term | Explanation | Reference from the Article |
| 🧠 Latent Semantic Analysis (LSA) | A computational method for deriving the meaning of words and texts by placing them in a high-dimensional semantic space based on word co-occurrence. | “Words, sentences, and texts are represented as vectors in this space…we can calculate how close or far apart two vectors are in this semantic space” (p. 250). |
| 📐 Semantic Space | A high-dimensional vector space (typically 300–400 dimensions) used to represent meanings of words and their relationships. | “Semantic maps—spaces—of 300 to 400 dimensions yield results that are most closely aligned with human judgments” (p. 250). |
| 🧠 Predication Algorithm | A model that adjusts the vector of a predicate based on contextual features derived from its argument to generate a context-sensitive meaning. | “The meaning of the predicate is modified to generate a contextually appropriate sense of the word” (p. 251). |
| 🌐 Argument and Predicate | In NOUN1 IS A NOUN2 metaphors, NOUN1 is the argument (topic), and NOUN2 is the predicate (vehicle/metaphor source). | “NOUN1 is called the argument (A) and NOUN2 is called the predicate (P)” (Appendix, p. 260). |
| 🧠 Vector Cosine Similarity | A measure used in LSA to determine semantic similarity between concepts; ranges from –1 (opposite) to +1 (identical). | “The cosine between highly similar vectors is close to +1, whereas unrelated vectors have a cosine close to zero” (p. 251). |
| 📐 Centroid (Vector Sum) | The average of several vectors; used to represent the collective meaning of a sentence or group of words. | “Sentence meanings are computed as the sum of the words, irrespective of their syntactic structure” (p. 250). |
| 🌐 Semantic Neighborhood | A group of vectors (words) that are closest in meaning to a given vector in the LSA space. | “It constructs the semantic neighborhood of the predicate…most closely related to the predicate” (p. 251). |
| 🧠 Constraint Satisfaction Process | A cognitive mechanism in the predication model that integrates the predicate’s neighborhood with the argument to derive meaning. | “Uses a constraint satisfaction process to integrate this neighborhood with the argument” (p. 251). |
| 📐 Spreading Activation | A process by which activation spreads through a network to identify most relevant semantic neighbors for metaphor interpretation. | “Activation is spread in that network…The most strongly activated neighbors of P will be used to modify P” (p. 260). |
| 🌐 Metaphoric Superordinate Categories | Abstract categories created by metaphors that go beyond literal taxonomies (e.g., “shark” becoming a category of “vicious professionals”). | “The notion of generating metaphorical superordinate categories can be operationalized” (p. 252). |
Contribution of “Metaphor Comprehension: What Makes a Metaphor Difficult to Understand?” by Walter Kintsch & Anita R. Bowles to Literary Theory/Theories
🧠 Contribution to Cognitive Literary Theory
- Supports the view that metaphor comprehension uses general cognitive processes.
↳ The article aligns with the notion that metaphors are understood in ways similar to literal sentences, challenging the assumption that metaphor requires unique interpretive faculties.
“There exists a considerable and convincing body of research…that indicates that people understand metaphors in much the same way they understand literal sentences” (p. 249).
- Uses cognitive modeling (LSA and predication) to simulate metaphor interpretation.
↳ Introduces a formal, empirically tested model showing how meaning emerges through contextual semantic alignment, which cognitive literary theorists find central to interpretive modeling.
“We describe a model of text comprehension…simulate the computations involved, and evaluate the model empirically” (p. 250).
🧬 Contribution to Formalist and Structuralist Theories
- Operationalizes metaphor using structural linguistic units (NOUN1 IS A NOUN2).
↳ The study isolates and systematizes metaphor into a rigid syntactic structure, echoing formalist interests in text-intrinsic form and structure.
“Each stimulus sentence was a metaphorical statement of the NOUN1 IS A NOUN2” (p. 253).
- Examines metaphoric meaning independently of reader emotion or authorial intent.
↳ The focus on semantic proximity, not subjective interpretation, aligns with structuralist ideals of objectivity in literary analysis.
“The sentence vector should be more closely related to the set of interpretations generated by human comprehenders than to the individual words of the sentence” (p. 252).
🧪 Contribution to Empirical Literary Studies
- Integrates experimental data into literary interpretation.
↳ The study used participant data and cosine-based metrics to evaluate metaphor difficulty, marking a shift from speculative literary criticism to quantifiable methods.
“Difficulty ratings ranged from 1.29…to 4.21…responses were more coherent for easy items” (p. 254).
- Establishes reproducibility and statistical grounding in interpretive variation.
↳ Demonstrates that metaphor comprehension can be empirically tested, supporting efforts in empirical literary studies to systematize interpretation.
“The difference between the coherence of easy items and difficult items was statistically significant, t(24) = 4.38, p < .01” (p. 254).
🧭 Contribution to Reader-Response Theory
- Explores interpretive variance among readers.
↳ The study highlights how reader agreement decreases with metaphor difficulty, resonating with reader-response theory’s emphasis on individual interpretation.
“Faced with items such as ‘Happiness is a ditch’…people didn’t just give up but found some interpretation” (p. 254).
- Suggests that comprehension is shaped by semantic constraints, not just subjective imagination.
↳ Even for difficult metaphors, interpretations were not random but guided by the latent semantic structure, refining the reader-response notion of subjective freedom.
“Even though interpretations are diffuse…they are not random. This consistency…may simply reflect word-based constraints” (p. 258).
🧠📐 Contribution to Conceptual Metaphor Theory (Lakoff & Johnson)
- Empirically supports metaphor as a cognitive mapping process.
↳ The study shows how metaphors create conceptual relationships by adjusting predicate meanings via contextually relevant features.
“The meaning of the predicate is modified to generate a contextually appropriate sense of the word” (p. 251).
- Adds computational rigor to conceptual blending.
↳ By modeling how metaphorical understanding emerges through a network of semantic connections, it extends the conceptual metaphor theory into testable, mechanistic terms.
“The vector computed by the model is equally close to that average of easy and difficult items” (p. 255).
⚙️ Contribution to Computational Literary Theory
- Demonstrates how semantic computation can approximate human interpretation.
↳ LSA and the predication model simulate how people derive meaning from metaphor, advancing computational approaches to literary meaning.
“The model vector nevertheless captures the variety of responses produced by the participants” (p. 257).
- Presents a fully realized computational theory of meaning.
↳ Unlike traditional metaphor theories, this model allows for quantification and algorithmic generation of interpretation, moving toward AI-assisted literary analysis.
“Our model is a fully realized, computational theory” (p. 252).
Examples of Critiques Through “Metaphor Comprehension: What Makes a Metaphor Difficult to Understand?” by Walter Kintsch & Anita R. Bowles
| 📚 Literary Work | 🔍 Example Metaphor from the Work | 🧠 Interpretive Analysis (Kintsch & Bowles Lens) | ⚖️ Critique Based on Model |
| 🦁 The Old Man and the Sea – Ernest Hemingway | “The fish is my brother.” | Metaphor follows the NOUN1 IS A NOUN2 form; argument = fish, predicate = brother. The predication model would identify features like shared struggle, respect, kinship as vectors connecting fish and brother. | ✅ Easy metaphor: Participants (readers) would likely converge on the emotional and symbolic kinship. High cosine values suggest semantic proximity once context is integrated. Strong coherence. |
| 🦇 Wuthering Heights – Emily Brontë | “Whatever our souls are made of, his and mine are the same.” | Though syntactically complex, metaphor relies on blending abstract noun (souls) with identity sameness. The metaphor is indirect, so coherence may vary. Vector representations of souls, same, and his/mine create a loose semantic field. | ⚠️ Moderately difficult: Metaphoric interpretation is diffuse; LSA may struggle due to abstraction and lack of direct predicates. Requires structural alignment (Gentner & Bowdle). |
| 🔥 The Waste Land – T. S. Eliot | “April is the cruellest month.” | NOUN1 IS NOUN2 metaphor with April (argument) and cruellest month (predicate). Contradicts conventional associations (spring with renewal). Model seeks shared neighbors between April and cruelty. | ❌ Difficult metaphor: Low baseline similarity; predication model generates vague and varied responses. Semantic coherence weak due to conflicting cultural frames. Low cosine match. |
| 🐍 Macbeth – William Shakespeare | “Look like the innocent flower, but be the serpent under’t.” | Implicit dual metaphor. Flower and serpent are semantic opposites. The model would modify serpent through context (deception, hidden danger) and apply it to Macbeth’s intentions. | ✅ Effective metaphor: Though figurative, structure aids LSA processing. High activation of relevant neighbors (e.g., danger, mask). Moderate difficulty but high interpretive coherence. |
Criticism Against “Metaphor Comprehension: What Makes a Metaphor Difficult to Understand?” by Walter Kintsch & Anita R. Bowles
🔄 Overreliance on Computational Models
- The study heavily depends on Latent Semantic Analysis (LSA) and the predication algorithm, which treat language geometrically.
- Critics argue this abstracts away cognitive nuance and fails to account for non-semantic cues such as pragmatics, cultural knowledge, or emotional tone.
- ❝ “Meaning is reduced to vector math, bypassing richer interpretive dynamics involved in actual reading.” (cf. Gentner & Bowdle, 2001)
📏 Neglect of Syntax and Word Order
- LSA used in the model ignores syntactic structure, computing sentence meaning via summation of word vectors regardless of grammar.
- This approach may oversimplify how meaning is constructed, especially for metaphors relying on syntax-dependent effects.
- Kintsch admits: “Such a procedure neglects important, meaning-relevant information that is contained in word order and syntax.” (p. 250)
🧩 Limited Scope of Metaphor Types
- The study is restricted to simple nominal metaphors (NOUN1 IS A NOUN2), excluding:
- Verbal metaphors
- Extended metaphors
- Metaphors embedded in narrative discourse
- This makes the model less generalizable to rich literary or philosophical texts with layered figurative complexity.
🤖 Assumption of Universal Processing
- The model assumes metaphor comprehension is uniform across individuals, whereas real readers vary due to:
- Background knowledge
- Personal associations
- Linguistic and cultural exposure
- Kintsch & Bowles acknowledge interpretive variation but still evaluate model success by group-level averages, masking individuality.
🔍 Lack of Qualitative Interpretive Depth
- The study’s quantitative focus on cosine similarity lacks insight into interpretive depth, such as:
- Moral connotation
- Intertextual echoes
- Aesthetic or rhetorical effect
- The model evaluates metaphor meaning only by statistical coherence, not by literary or emotional richness.
🧪 Artificial Experimental Context
- Participants completed sentence frames and gave difficulty ratings in a lab setting with isolated metaphors.
- Critics may question ecological validity—metaphors in real texts are processed within broader narrative, emotional, and discursive contexts.
🧠 Cognitive Economy Not Fully Addressed
- The model doesn’t sufficiently address cognitive economy principles, such as why:
- Some metaphors are retained and others forgotten
- Some metaphors “click” quickly while others are puzzling or evocative
- The authors touch on this via coherence scores, but the deeper cognitive prioritization mechanisms remain underexplored.
🧬 Ambiguity in Defining “Difficulty”
- The metric for what makes a metaphor “difficult” is partly subjective, relying on participant self-ratings and coherence calculations.
- This leaves room for ambiguity in distinguishing between semantic novelty, conceptual mismatch, and reader confusion.
Representative Quotations from “Metaphor Comprehension: What Makes a Metaphor Difficult to Understand?” by Walter Kintsch & Anita R. Bowles with Explanation
| Quotation | Explanation |
| “People understand metaphors in much the same way they understand literal sentences.” (p. 249) | Challenges the view that metaphor processing is fundamentally different; suggests metaphor comprehension is a natural language process. |
| “The meaning of a word, sentence, or text is given by the set of relations between it and everything else that is known.” (p. 250) | Reflects the core idea behind Latent Semantic Analysis (LSA) – meaning is relational, not fixed. |
| “Metaphors are not difficult because their argument and predicate terms are unrelated overall.” (p. 256) | Refutes the intuition that semantic distance alone determines difficulty in metaphor comprehension. |
| “The model vector is equally close to the average of easy and the average of difficult items.” (p. 255) | Shows that the computational model treats both metaphor types similarly in vector space despite participant differences. |
| “Some link is found between topic and vehicle, even though the two may be unrelated overall.” (p. 258) | Highlights the model’s strength in identifying subtle, context-sensitive links between unrelated terms in metaphors. |
| “The model produced a vector that was just as close to these varied responses as it was to the generally agreed-upon interpretation of a good metaphor.” (p. 258) | Emphasizes that the model handles ambiguity effectively, mimicking human flexibility in metaphor interpretation. |
| “Faced with the seemingly impossible task of finding an interpretation for such metaphors, people did not give up.” (p. 258) | Demonstrates human resilience and interpretative creativity even in difficult metaphorical constructions. |
| “Generating context-sensitive word senses does not always produce dramatic results.” (p. 251) | Acknowledges that not all metaphors lead to strong reinterpretations; some may resemble literal interpretations. |
| “The semantic structure provided a tight constraint for easy metaphors, and only a loose one for hard metaphors.” (p. 258) | Suggests semantic coherence plays a central role in determining perceived metaphor difficulty. |
| “Theories of metaphor comprehension have traditionally been informal.” (p. 258) | Justifies the importance of formal, computational models like LSA to bring precision to metaphor theory. |
Suggested Readings: “Metaphor Comprehension: What Makes a Metaphor Difficult to Understand?” by Walter Kintsch & Anita R. Bowles
- Kittay, Eva Feder. “Woman as Metaphor.” Hypatia, vol. 3, no. 2, 1988, pp. 63–86. JSTOR, http://www.jstor.org/stable/3809952. Accessed 12 May 2025.
- Gibbs, Raymond W. “When Is Metaphor? The Idea of Understanding in Theories of Metaphor.” Poetics Today, vol. 13, no. 4, 1992, pp. 575–606. JSTOR, https://doi.org/10.2307/1773290. Accessed 12 May 2025.
- BLACK, Max. “More about Metaphor.” Dialectica, vol. 31, no. 3/4, 1977, pp. 431–57. JSTOR, http://www.jstor.org/stable/42969757. Accessed 12 May 2025.
- Miller, Eugene F. “Metaphor and Political Knowledge.” The American Political Science Review, vol. 73, no. 1, 1979, pp. 155–70. JSTOR, https://doi.org/10.2307/1954738. Accessed 12 May 2025.
- Wearing, Catherine. “Metaphor, Idiom, and Pretense.” Noûs, vol. 46, no. 3, 2012, pp. 499–524. JSTOR, http://www.jstor.org/stable/41682624. Accessed 12 May 2025.