Korean Journal of Buddhist Studies
DOI : 10.21482/jbs.86..202603.235
Summary
Language is generally understood as a vessel that contains meaning. From a Buddhist perspective, however, this vessel is merely an expedient means (upāya) and cannot transmit object as they truly are. The Buddha’s “unanswered” questions (avyākata, Skt. avyākṛta) deliberately interrupt the patterns carried by language, thereby disclosing that language is not the essence of “reality.”
This study analyzes Buddhist insight into the limitations of language through the Buddha’s avyākata. The Buddha’s silence in response to metaphysical questions is interpreted not as a refusal to answer, but as an epistemological act that transcends the structure of language. The study seeks to present the emptiness (suññatā, Skt. śūnyatā) of language thus disclosed not as an abstract concept but as a formally visualized structure. To this end, it employs a simplified Word2Vec model, based on principles of AI pattern recognition, as a formal conceptual simulation. By quantifying patterns of word co-occurrence within selected Buddhist texts, the study diagrammatically demonstrates that meaning is constituted not as a self-subsisting essence but as an interdependent relation (paṭiccasamuppāda).
Ultimately, avyākata is not a negation of language but a non-linguistic expedient means that reveals the emptiness of linguistic representation through action, while Word2Vec functions as a non-linguistic expedient means that formally demonstrates the structure of that emptiness in informational space. In this way, the study proposes a methodological expansion in Buddhist studies through the use of formalized models for conceptual clarification.
Keywords
avyākata, unanswered, articulation, emptiness, dependent origination, pattern recognition, distributed representation
I. Introduction
The problem of “what cannot be spoken” (avyākata)[1][2] has occupied a crucial position in the Buddhist tradition. In the Cūḷa-Māluṅkyaputta Sutta (MN 63), the Buddha does not answer metaphysical questions concerning the finitude of the universe, the post-mortem existence of the Tathāgata, and the identity of the self. This silence does not stem from ignorance or evasion, but from the recognition that such questions presuppose the fixity of linguistic categories (paññatti),[3] which misrepresent the conditional and interdependent nature of phenomena.
The Buddha’s silence is not a negation of language but an act that reveals the limitations of language. Language generates meaning by pointing to objects, but simultaneously obscures relational continuity by fixing them. At this point, through the interruption of linguistic expression, one transcends the structure of language, and the limitations of the structure upon which cognition depends are revealed.
Interestingly, modern artificial intelligence research has arrived at similar insights through an entirely different route. “Distributed representation theory”, which models language not as a bearer of fixed meanings but as a network of context-dependent patterns, shows that meaning is not an essential property of individual words but a function of relational configuration. This study views this convergence not as mere coincidence but as two expressions that lead to fundamental insight into ‘patterns.’ The Buddha’s silence operates as a non-linguistic expedient means through action, while AI pattern recognition operates as a non-linguistic expedient means through visualization via computational form; though employing different methods, both disclose the emptiness of language.
Accordingly, this study raises the following question:
How does the Buddha’s avyākata reveal the emptiness of language, and how can this emptiness of language be formally exemplified through modern pattern recognition models—specifically Word2Vec?
The methodology of this study operates on two levels. First, it analyzes the manner in which the limitations of language are revealed through examples of avyākata recorded in the Nikāyas (Chapter III). Second, it conceptually demonstrates how this emptiness of language can be formalized through modern AI pattern recognition models (Chapter IV).
These two approaches, which may appear heterogeneous at first glance, are in fact mutually complementary. Literary analysis alone can discuss what the relational nature of language is, but finds it difficult to visually present how it structurally manifests. Conversely, pattern recognition models alone cannot capture the depth of Buddhist insight that transcends mechanical correlation. Therefore, this study combines classical textual interpretation with modern formal models to illuminate the emptiness of language from both practical and formal demonstration perspectives.[4]
This methodological choice represents an attempt to expand the scope of Buddhist studies. While several scholars (Varela 1991; Thompson 2007) have already pointed out that Buddhist philosophical insights structurally correspond with modern cognitive science, concrete demonstrations through formalized models are difficult to find.[5] This study is an experimental attempt to bridge that gap.
II. Literature Review and Research Methodology
1. Literature Review
Interpretations of avyākata may be traced through two main directions.
The first is the practical-methodological interpretation, which itself encompasses two related emphases. On the more strictly practical side, Conze (1959) argued that the Buddha’s teachings function as practical directives rather than metaphysical claims, and Jayatilleke (1963) similarly contended that avyākata reflects the Buddha’s pragmatic refusal to engage with questions that do not conduce to liberation. On the methodological side, Rahula (1974) emphasized that avyākata reveals the categorical limitations of certain metaphysical questions, and Gombrich (2006) proposed a shift from ontological inquiry to processual understanding. This interpretation as a whole views avyākata as a reorientation rather than doctrinal evasion. In Korean scholarship, Hwang (2004) and Kwon (2009) view avyākata as the cessation of theoretical speculation, emphasizing its practical significance.[6]
The second is the philosophical-linguistic interpretation. Murti (1955) explained avyākata as the point where linguistic concepts fail to grasp ‘reality’—that is, the gap between designated representation (saññā-paññatti) and the ultimate dimension (paramattha)— while Garfield (1995) interpreted Madhyamaka’s approach as an epistemological act in which language operates within its conventional limits while indicating what lies beyond conceptual elaboration. These approaches are significant in viewing both avyākata and Madhyamaka negation as methods that expose the conditional nature of linguistic categories.[7]
Meanwhile, information theory and cognitive science since the 20th century present interesting insights that structurally correspond with Buddhist views on language. Shannon (1948) formalized the amount of information as a function of probability distribution, establishing a statistical model independent of meaning.[8] Bateson (1972) defined information as “a difference that makes a difference,” emphasizing that meaning arises only through relational contrast.[9] Friston’s (2010) predictive processing theory explains cognition as a conditional inference process that minimizes prediction error, revealing that perception is not an independent entity but an interdependent relational network.[10]
Of course, these studies do not directly discuss Buddhism. However, they can be seen as resonating with Buddhist thinking in that language and cognition operate as relational patterns. In particular, Bengio (2013) and Mikolov’s (2013) distributed representation theory[11] shows that meaning is not a fixed definition but is constituted from co-occurrence patterns. The Word2Vec model formally demonstrates that meaning is expressed not as an isolated essence but as relational distance and direction by coordinating these distributional patterns in vector space.[12][13]
Through this comparison, this study sees that Buddhist insight—language is established by relational conditions—and the structural principles of AI models form a structural correspondence. If the Buddha’s avyākata revealed the conditionality of language through the act of silence, Word2Vec reveals that conditionality through formalization.
2. Research Methodology
The first stage of this study is philological and philosophical analysis based on Nikāya materials. The primary text is the Cūḷa-Māluṅkyaputta Sutta (MN 63), with reference to commentaries in the Visuddhimagga (Vism. XVIII.28). The focus of analysis is to clarify that the Buddha’s silence is not simply the absence of an answer, but an epistemological act that exposes the inadequacy of the linguistic framework presupposed by the question. The key concept used here, linguistic category (paññatti), is not a dharma of the ultimate dimension (paramattha-dhamma) but a nominal designation for the convenience of cognition (vohāra).[14] Therefore, language is not a transparent medium that transmits “things as they are,” but a representational device that operates within dependently arisen conditions.
The second stage uses AI pattern recognition principles in the form of formal conceptual simulation.[15] The Word2Vec model is not an empirical verification tool but a thought experiment to formalize the Buddhist thesis that “language is a relational pattern.”[16] This study uses key terms from the Heart Sūtra as a corpus[17] and projects word co-occurrence patterns into vector space. This demonstration is not statistical verification but heuristic visualization—converting abstract concepts into visible structures to aid understanding.[18]
Distributed representation theory, which forms the structure of Word2Vec, presents a cognitive scientific perspective that does not correspond meaning to a single symbol but understands it as a relational pattern distributed across multiple units (neurons·features). This perspective structurally corresponds with Buddhist dependent origination (pratītyasamutpāda). According to dependent origination, all existence is established not as independent entities but only within mutually dependent relationships, and possesses no inherent nature (“svabhāva”) in “itself”. Similarly, in distributed representation, meaning is not a unique property of individual words but is revealed only within a statistical relational network with surrounding words. Therefore, meaning is not “existence” but “pattern,” not a self-subsisting essence but a function distributed according to relationships.
On the basis of this structural correspondence, Word2Vec is a concrete neural network model that applies the principles of distributed representation to actual language data, formalizing word co-occurrence relationships into multidimensional vector space. In other words, if distributed representation corresponding to dependent origination is the theoretical framework that “meaning exists as relational distribution,” Word2Vec is an expedient means (upāya) of this era that mathematically implements that principle.
Finally, to ensure reproducibility and transparency, this study includes the entire demonstration code, the execution environment and results to a public repository (GitHub).[19]
III. The Buddha’s Avy¯kata[20]
1. The Structure of Avyākata and the Limitations of Language
Language mediates thought but simultaneously limits it. Language generates meaning by designating objects, but that act of designation necessarily leaves residual indeterminacy. In the Cūḷa-Māluṅkyaputta Sutta (MN 63), when Māluṅkyaputta questioned the finitude of the world, the post-mortem existence of the Tathāgata, and the identity of self and body, the Buddha responded in the form of silence.[21]
These ten avyākatapañhā (“undeclared questions”)[22] all presuppose ontological fixity. That is, questions such as “Is the world eternal?” and “Does the Tathāgata exist after death?” arise from linguistic attachment (diṭṭhi-gata) that regards what has been named (paññatti) as ‘real.’ The Buddha explains this through the simile of the poisoned arrow: while asking about the material of the arrow or the identity of the shooter, the actual suffering (dukkha) does not disappear.[23] However, this simile is not simply pragmatic evasion but means the deconstruction of the linguistic structure upon which the question depends. The Buddha’s avyākata is not negation of linguistic expression but an act that reveals in a non-linguistic manner that language has reached its own limitations.[24]
2. The Limitations of Linguistic Designation (Paññatti)
What the Buddha’s silence reveals is the limitation of language. Language explains phenomena but simultaneously simplifies, categorizes, and fixes them. The Saṃyutta- nikāya (SN II 114) states: “Dependent on name-and-form (nāma-rūpa) is consciousness (viññāṇa), and dependent on consciousness are name (nāma) and form (rūpa).”[25] This means that language [nāma] and phenomena [rūpa] arise interdependently. Language is not a transparent medium overlaid on “reality” but a relational pattern that arises conditionally together with phenomena.
The Visuddhimagga defines paññatti as “conceptual designation,” emphasizing that it is merely a name used for the convenience of cognition (vohāra), not a dharma of the ultimate dimension (paramattha-dhamma) [Vism. XVIII.28]. For example, the word ‘chariot’ is merely a name given when wheels, axles, body, and other components form a specific arrangement; it is not an independently existing entity.[26] Similarly, metaphysical categories such as ‘world’, ‘self’, and ‘Tathāgata’ are also linguistic constructs, and attempting definitive statements about them is to misunderstand the function of language.
Therefore, metaphysical questions commit a double error. First, they assume that linguistic categories designate “real” independent entities. Second, they presuppose that positive or negative binary judgments are possible regarding those categories. The Buddha’s silence simultaneously deconstructs both premises and operates as a device that reflexively reveals the structural conditions of linguistic expression by suspending the use of language.[27]
3. The Epistemological Function of Silence—The Gate to Emptiness
Avyākata is not the negation of language but a self-referential act in which language reveals its own limiting conditions. Murti (1955) called this “language’s self- transcendence,” and this study views this transcendence as the manifestation of language’s structure. By not answering the question “Is the world eternal?”, the Buddha practically shows that the concepts ‘world’ and ‘eternal’ do not designate ‘reality.’
When language is thus interrupted, only then does its relational nature—the emptiness of language—manifest. Emptiness is not non-existence (無) but lack of inherent nature (niḥsvabhāva), and what Nāgārjuna said, “Emptiness and dependent origination are not different (śūnyatā pratītyasamutpādasyāparyāyaḥ)” (MMK 24.18), means precisely this complete recognition of relational dependence.[28]
Therefore, avyākata is the gate to emptiness that opens at the end of language. By silencing language, the Buddha reveals the relational conditionality of language—that is, the structure of dependent origination. This silence is a ‘non-linguistic expedient means,’ a performative device that enables recognition of language’s structure and a self-transcending operation of linguistic expression.
IV. Formal Demonstration of Pattern Recognition
1. The Significance of Formal Demonstration
As seen in the previous chapter, the Buddha’s avyākata is a non-linguistic representation that interrupts language’s tendency to reduce “reality” into fixed categories. The Buddha’s avyākata does not mean the negation of language. Rather, it is a formal act that makes language reveal its limiting conditions. The purpose of this chapter is to formalize this epistemological structure—the insight that language is merely a relational pattern—through the pattern recognition structure of modern artificial intelligence. To this end, this study presents the core principles of the Word2Vec model as a conceptual demonstration.
This demonstration is not statistical verification but a thought experiment. Just as correct understanding of dependent origination can enable awareness of its emptiness, this demonstration provides a ‘diagram’ through which the relational structure of language can be formally observed.
2. The Basic Structure of Word2Vec: Spatializing Relational Meaning
Word2Vec is a model that analyzes the frequency with which words appear together in a corpus and coordinates word co-occurrence patterns in vector space. Each word is converted into a numerical vector, and the distance and direction between vectors indicate the degree of semantic similarity. That is, the meaning of a word is expressed not as the essence of the word itself but as a distribution of relations with other words.
This structure formally corresponds with what Buddhism calls lack of inherent nature (niḥsvabhāva)—the insight that objects do not exist independently but are established only within mutually dependent relationships.
In this respect, Word2Vec is a geometric expression of a relational worldview. The logic of dependent origination that “when this exists, that exists” (imasmiṃ sati idaṃ hoti) is replaced in this model with the statistical interdependence that “when this word appears, there is a probability that that word will appear together.” Of course, this does not simply mathematicize the Buddha’s performative insight, but is significant in that it provides a framework through which relational ontology can be thought visually and formally.
3. The Structure of Emptiness Revealed in Semantic Space
The vector space constructed by the model shows that relationships between words are not fixed but are constantly rearranged according to usage context. The meaning of a particular word is expressed as a position in vector space, but that position changes according to changes in surrounding words. This follows the same principle as what Buddhism calls the ontology of dependent origination—the structure of ‘arising and ceasing interdependently.’
Within this relational network, ‘meaning’ is not a fixed entity but a function of interdependent position. Language is no longer a transparent mirror reflecting “reality” but is revealed as a field in which meaning continuously moves. Just as the Buddha’s avyākata suspended the fixation of language, Word2Vec’s vector space also shows language as a moving relational structure. Language is not substance but pattern, and meaning is not self-subsisting essence but interdependent relationship.
4. Epistemological Comparison: Action and Formalization
The Buddha’s avyākata and Word2Vec reveal the same structure through different methods: action and formalization.
• Avyākata enables enacted experience of language’s conditionality by stopping linguistic expression.
• Word2Vec enables observed experience of that conditionality by converting linguistic expression into data.
That is, the former is a performative expedient means, the latter a demonstrative expedient means. Both make visible that language does not point to “reality” but operates within relational conditions. This comparison means a modern expansion of the Buddhist perspective. The Buddha’s avyākata is no longer a simple doctrinal maxim but is reinterpreted as a cognitive act that has insight into the structure of language.[29] Word2Vec functions as a formal experiment that reproduces this cognitive act in information space.
5. Limitations and Methodological Cautions
This demonstration is based on a formal model, not actual data analysis.
First, the corpus used (Heart Sūtra) is small-scale and cannot secure statistical significance.[30]
Second, vector projection (t-SNE, PCA) simplifies multidimensional relationships into two dimensions, and some of the original semantic distances or directions may change in the process.[31]
Third, the semantic depth of Buddhist terms cannot be reduced to numerical approximation.
Therefore, the demonstration in this chapter is not to present empirical conclusions but to formally and heuristically explore the relational structure of language. This point also corresponds with the Buddha’s avyākata. If the Buddha’s silence arose not from ‘the absence of truth’ but ‘awareness of the limitations of representation,’ this study’s formalization leads not to ‘making data into truth’ but to ‘awareness of language structure.’
|
<Table 1> Structural Correspondence Between Two Expedient Means |
|||
|
Category |
Buddhist Structure |
AI Model Structure |
Convergence Point |
|
What |
Avyākata (Silence) |
Word2Vec |
Conditionality |
|
How |
Performative |
Demonstrative |
Structural |
|
Why |
Inducing Recognition of Language’s Emptiness through Non-linguistic Expression |
Recognition of Language’s Emptiness |
|
|
What is Deconstructed |
Subjective |
Categorized |
Substantialist View |
|
Note: Table 1 summarizes the structural correspondence between the two expedient means. Avyākata discloses emptiness through performative action that induces experiential realization, while Word2Vec discloses it through demonstrative formalization that reveals structural patterns. Though operating at different epistemic layers with distinct ways, both reveal the relational nature of language and deconstruct the substantialist view of language—the assumption that words possess inherent meaning. |
|||
6. Summary
The Buddha’s avyākata is a non-linguistic expedient means that reveals language’s conditionality by stopping language, while Word2Vec’s demonstration is a non-linguistic expedient means that visualizes that conditionality by formalizing language’s relational structure. Both approaches reveal the emptiness of language—the fact that meaning is not self-subsisting essence but relational arising. Ultimately, the epistemology of avyākata implements the same insight in different dimensions: in the performative manner of silence, and Word2Vec in the demonstrative manner of numbers.
V. Conclusion—Avy¯kata and Its Formalization
This study has analyzed the Buddha’s avyākata and AI pattern recognition models as two expedient means that reveal the emptiness of language. The Buddha’s avyākata is not the refusal of linguistic expression but an act that reveals the conditionality of language itself at the linguistic boundary where language’s operation reaches. In the Cūḷa-Māluṅkyaputta Sutta (MN 63), by suspending his answer, the Buddha deconstructed the linguistic structure presupposed by the question itself—the tendency of concepts (paññatti) to fix ‘reality’. This silence is not negation but insight. The content of that insight is precisely the recognition of the emptiness of language—that language has no self-subsisting essence and has meaning only within relational conditions.
Modern AI pattern recognition, such as the Word2Vec model, reveals the same structure in an entirely different manner.[32] Words acquire meaning not as fixed essences but through relational positions, formally corresponding with dependent origination—“when this exists, that exists” (imasmiṃ sati idaṃ hoti, SN 12.1).
Therefore, the relationship between the two expedient means can be summarized as follows:
• Avyākata: Discloses emptiness as an act so that one can experience its emptiness by deconstructing the presuppositions of language through suspension of articulation.
• Word2Vec: Discloses emptiness as another representation so that one can experience its emptiness by revealing the structure of language through transformation of articulation’s form.
The former is a performative expedient means, the latter a demonstrative expedient means. Both reveal that meaning arises not as self-subsisting essence but within relational configuration—that is, the emptiness of language.
The significance of this study for Buddhist studies is as follows. First, by reinterpreting the Buddha’s avyākata not as a simple doctrinal maxim but as a cognitive act that reveals language’s operation, it restores the avyākata of early Buddhism at a philosophical and methodological level. Second, by using AI pattern recognition models as formal demonstration to visualize the “relational” and “layered” structure of language, it presents the possibility of modern reproduction of Buddhist perspectives. Third, it proposes that formal tools (statistical models·information structures) can be used as hermeneutic auxiliary means in Buddhist studies. This is significant as a heuristic methodology for conceptual clarification, not empirical verification.
However, a crucial difference also remains between the two. AI models recognize patterns but do not become aware that they are patterns. In contrast, Buddhist insight lies precisely in that awareness—“knowledge and vision of things as they really are” (yathābhūta-ñāṇa-dassana, 如實知見). Word2Vec “shows” emptiness but does not “realize” emptiness. For this reason, this formal experiment functions not as a substitute for Buddhist practice but as a diagrammatic device for formally understanding Buddhist insight within modern language structure. In this study, this upāya is utilized as one of the fingers pointing at the moon, enabling humans to experience in visual form their tendency to cling to language and to transcend it.[33]
Ultimately, the Buddha’s avyākata and modern pattern recognition arrive at the same insight through the languages of different eras—that language is not the bearer of “reality” but the expression of relational patterns. The Buddha revealed this through the act of silence, and modern science visualizes it through the representation of numbers. That meeting shows that Buddhist perspectives are not merely religious beliefs but profound insights into the universal structure of language and cognition.
[1] Note on Terminology. For consistency, Pāli
forms (e.g., avyākata, paṭiccasamuppāda) are used for Nikāya-based concepts,
while Sanskrit forms (e.g., upāya) are retained for Mahāyāna contexts where the term carries a distinct philosophical
development.
MN 63: The ten undeclared questions (avyākatapañhā) include whether the world is eternal, whether the Tathāgata exists after death, whether the self is identical with the
body, etc.
[2] avyākata(Pāli, Ch. 無記) derives from the Sanskrit past passive participle avyākṛta formed from the verbal root ā-vi-√kṛ, literally meaning “not distinguished, not differentiated, not made manifest.” In early Buddhist literature, the term designates the Buddha’s silence concerning metaphysical questions (MN 63; SN 44.10), yet its etymological structure carries broader philosophical implications. The Bṛhadāraṇyaka Upaniṣad (1.4.7) employs this term to describe a primordial undifferentiated state: “At that time this [world] was indeed undifferentiated (tad dhedaṃ tarhy avyākṛtam āsīt); it became differentiated precisely by name and form (tan nāmarūpābhyām eva vyākriyata).” In this Upaniṣadic context, avyākṛta indicates the condition prior to the operation of nāmarūpa—the categorical framework organizing experience into conceptual (nāma) and perceptual (rūpa) structures.
[3] MN I.426-431 (PTS). Vism. XVIII.28.
[4] This study is based on the “process-oriented understanding of Buddhism” proposed by Gombrich (2009). Gombrich suggested shifting the interpretive framework from “what exists” to “how it operates.” In a similar context, Rahula (1959) clarified that the five aggregates function not as ontological components but as epistemological tools for deconstructing the representation of self, and Hwang (2004) reinterpreted avyākata not as mere silence but as a methodological transition: the Buddha does not reject metaphysical questions but deconstructs the subjective presuppositions that enable such questions to arise (SN 12.12).
[5] This study does not claim to offer an entirely novel philosophical breakthrough, but demonstrates structural correspondence between Buddhist epistemology and computational linguistics. Word2Vec serves as a heuristic device for observing relational structures, not as empirical proof of Buddhist doctrine. The model provides a diagrammatic tool—what classical Indian logic might call a dṛṣṭānta (illustrative instance), a term used here in its analogical rather than strict pramāṇa-theoretic sense.
[6] In particular, Hwang (2004) does not regard avyākata merely as “not answering,” but interprets it through the Buddha’s yoniso manasikāra. He translates this concept, commonly rendered as “right thought,” as “thinking according to the cause,” reading the etymological meaning of manasikāra not as “attention” but as “making mind,” and understanding yoniso as “from the source” or “methodologically.” Accordingly, avyākata is interpreted as methodological silence based on yoniso manasikāra, redefined as a meta-cognitive act that transforms the very mode of thinking consciousness. According to him, the Buddha abandoned metaphysical questions asking “What is it?” and shifted the framework of thinking to causal questions asking “By what does this come to be?”, and this shift in thinking itself constitutes the meaning of avyākata. Ultimately, avyākata is understood not as the “impossibility of articulation” but as a “methodological transition that redirects the direction of articulation.”
[7] These two interpretive directions represent a classical tension in Buddhist studies, originating with the works of Murti (1955) and Jayatilleke (1963) respectively. Jayatilleke explicitly criticized Murti’s interpretation on the grounds that it over-metaphysicizes the pragmatic orientation of early Buddhism (Jayatilleke 1963, 474-475, §815). This study does not aim to adjudicate between these two interpretive directions, but takes as its analytical foundation the formal property they share: both require the suspension of linguistic conceptual operation. The question of which interpretation more accurately reflects the Buddha’s original intent is left as a matter for separate scholarly inquiry.
[8] Shannon is generally regarded as the founder of information theory. In his 1948 paper “A Mathematical Theory of Communication,” he redefined “information” as a formal concept independent of semantics. He shifted the study of communication from the question of “what is being said” to that of “how much uncertainty is reduced,” expressing the quantity of information mathematically as a function of probability distribution. In other words, information is understood not as the semantic content carried by language, but as the degree of reduction in uncertainty among possible alternatives. By replacing “meaning” with relational structure and discriminability, this definition made it possible to understand language not as content but as a network of formal relations (Shannon 1948). In this respect, it structurally corresponds to what Buddhism describes as niḥsvabhāva.
[9] Bateson (1904-1980), in Steps to an Ecology of Mind (1972), defined information as “a difference that makes a difference.” He understood the essence of information not as substantial content but as relational contrast. For a stimulus or a sign to function as information, it must generate a differential distinction among alternative possibilities; this very “difference” becomes the fundamental unit of cognition. Thus, information is not an independent entity but a formal pattern of relations arising within a network of contrasts (Bateson 1972). This view structurally resonates with the conditional logic of dependent origination expressed as “When this exists, that exists” (imasmiṃ sati idaṃ hoti, DN 15).
[10] Friston (1959-) is a neuroscientist and theoretical biologist who proposed the predictive processing theory based on the ‘free-energy principle’ (Friston 2010). This theory conceives the brain not as a passive receiver of sensory stimuli but as an active inference system that continuously generates and revises predictive models of the world. The brain does not directly perceive the external world; rather, it functions to minimize the difference between its predictions and actual sensory inputs—that is, the prediction error. In this view, perception is not the apprehension of fixed entities but a continuous process of conditional updating. Cognition, therefore, is understood as probabilistic pattern inference that temporarily stabilizes meaning within relational interdependence, rather than as a replication of external objects (Friston 2010; Thompson 2007).
[11] In artificial intelligence and cognitive science, this refers to the theory that the meaning of a concept or word is not represented as an isolated symbol, but as a relational distribution across multiple dimensions within a representational space. Whereas traditional symbolism assumes a one-to-one correspondence between a concept and a symbol, distributed representation holds that meaning is distributed across the entire activation pattern of a neural network. For example, the word “king” is not an independent symbol but a statistical composition of numerous contextual elements such as male, rule, power, and palace (Bengio et al. 2013; Mikolov et al. 2013). This understanding is structurally analogous to the Buddhist notion of dependent origination, which regards language as a network of relational patterns.
[12] Word2Vec is a language model proposed in 2013 by a research team at Google led by Thomas Mikolov. Unlike complex deep neural networks, Word2Vec represents relationships between words directly in a multidimensional vector space—where distance and direction correspond to semantic relations—making it an interpretable model that humans can analyze structurally. It is therefore well suited to formally illustrating the Buddhist insight that meaning arises as a relational pattern. In this framework, the semantic similarity between words is expressed through distances and directions among points, so that meaning appears not as a fixed essence but as a pattern of relational configuration. If distributed representation theory in cognitive science conceives meaning not as a one-to-one correspondence between a concept and a symbol but as a relational pattern distributed across multiple units (neurons or features), Word2Vec is the concrete neural network model that applies this theory to actual linguistic data, learning co-occurrence patterns among words within a multidimensional vector space. In other words, if distributed representation is the theory that “meaning exists as relational distribution,” then Word2Vec is the mathematical implementation of that principle (Bengio et al. 2013; Mikolov et al. 2013).
[13] The representation of words as vectors in Word2Vec does not mean that meaning has been reduced to numbers. The term “vector” comes from the Latin vectōr (‘carrier’), originally referring not to fixed content but to something that enables movement and arrangement. This sense remains in modern mathematics and programming. In Word2Vec, vectors do not contain meaning but serve as coordinate structures through which relational patterns from word usage can be compared and transformed. Individual vector components have no meaning alone; they function only through distance and directional relations with other vectors. In machine learning practice, vectors are treated not as representations of meaning but as computational formats enabling operations. Therefore, vector representation should be understood not as objectified meaning but as a formal structure through which semantic relations are arranged to manifest.
[14] Vism. XVIII.28.
[15] A ‘demonstration’ is not verification (empirical) that tests data, nor a pilot that trials procedures, but an act that formally reveals the structure of a concept through simplification. It focuses not on precisely reproducing detailed numbers, but on observing the invariant structure that persists across relational patterns and transformations. In this sense, demonstration resembles the mode of understanding found in topology. Just as topology does not distinguish between a doughnut and a mug but regards features such as the number of holes and the connectedness of surfaces as essential, demonstration likewise simplifies complexity to reveal the underlying formal structure. The Word2Vec demonstration in this study, in this very sense, exemplifies a topological understanding of language—one that discloses meaning not as statistical detail but as a pattern of relation.
[16] Word2Vec demonstrates how stable structures can reveal relational patterns. This study formalizes the structural dimension of emptiness—the principle that meaning arises through relations rather than from inherent essence (svabhāva)—not its experiential dimension (yathābhūta-ñāṇa-dassana, direct realization of emptiness). In the vector space, ‘king’ occupies a fixed position, yet its meaning emerges entirely from relational distances: proximity to ‘queen,’ ‘crown,’ ‘power,’ ‘rule,’ and contrast with ‘peasant,’ ‘subject,’ ‘servant.’ No single vector contains meaning; meaning is the configuration of all relations. This exemplifies Nāgārjuna's insight that emptiness and dependent origination are two descriptions of the same reality (MMK 24.18).
[17] A ‘corpus’ (sometimes translated as ‘malmuŋchi’ in Korean) refers to a collection of digitized texts prepared for analysis. In linguistics and digital humanities, it serves as a data unit for observing contextual usage patterns rather than individual sentences. In this study, the Chinese text of the Heart Sūtra was treated as a single corpus, and the co-occurrence patterns among words within it were formalized through the Word2Vec model. This is not intended to statistically verify the doctrinal content of the text but to heuristically visualize how meaning arises within relational structures. The Heart Sūtra was selected as the corpus because (1) it most concisely expresses the relational structure of emptiness, (2) its repetitive and parallel syntax allows the Word2Vec model to clearly capture co-occurrence patterns, and (3) its brevity combined with structural richness makes it suitable for an experimental study of this scale. Meanwhile, although the Pāli canon maintains linguistic stability through its recitational tradition, its relatively low lexical variation limits comparative analysis of relational patterns among words, and thus it was not used for this demonstration.
[18] The reason for adopting Word2Vec is twofold. (1) The distributed representation theory that forms the structure of Word2Vec can formalize the relational structure operating beneath the surface fixity of language, making it suitable for visualizing the emptiness of language. (2) In terms of context dependence, Word2Vec is lower than BERT or GPT, but considering the scope of a short paper and accessibility to understanding structure for humanities readers, it is more suitable than these. What is important is that Word2Vec can sufficiently reveal relational structure compared to the usual way of understanding language that only objectifies the surface, and its operating layer is clear.
[19] https://github.com/samtustadguackr/avyakata-supplementary
[20] This study adopts avyākata as its central analytical model because it represents a sophisticated instance where the Buddha’s philosophy of language reveals its essence at an epistemological threshold. The Buddha’s diverse linguistic teachings each function as essential “rafts” (kullā) in their own right for crossing the ocean of suffering. Avyākata, by contrast, is a rare pattern that emerges in response to questions concerning the state “beyond the river,” where linguistic representations can no longer find any substantial referent to point toward. At this juncture, by suspending the use of the “raft” (language), the Buddha reveals the limits of linguistic representation itself — indirectly pointing to what language cannot reach, namely, the ‘structural emptiness of linguistic representation’.
[21] MN I.426-31 (PTS). In Indian thought, silence (mauna) often functions not as the absence of speech but as a non-linguistic mode of communication. In the Chāndogya Upaniṣad (6.1-6.2), when Śvetaketu asks Uddālaka Āruṇi about the nature of Brahman, Uddālaka conveys the inexpressibility of ultimate “reality” not through verbal definition but through analogy and intentional silence. Likewise, in the early Buddhist Saṅgha, collective assent was ritually “expressed” through silence: according to the Cullavagga IV (Vin II 88-93), a proposal was announced three times—“Let those who approve remain silent; let those who do not approve speak”—and when the assembly remained silent (tuṇhī-bhāva), the motion was regarded as approved. Within this interpretive context of Indian traditions, the Buddha’s “response in the form of silence” is understood not as refusal or ignorance but as a culturally intelligible mode of communication.
[22] While the abstract and keywords employ the more commonly used translation “unanswered,” given this etymological structure, avyākata may also be understood as marking questions that cannot be addressed through categorical frameworks—that is, “undeclared.” If nāmarūpa constitutes the categorical apparatus organizing experience, then avyākatapañhā would indicate questions whose presuppositions fall prior to—or beyond—the operation of that apparatus.
[23] MN I.430.
[24] Here, defining silence as a ‘non-linguistic act that reveals the limits of language’ does not imply that silence intentionally points toward a certain metaphysical entity (such as śūnyatā). Rather, this silence constitutes a ‘structural interruption’ that refuses to operate within the frame of linguistic representation set by the questioner, thereby exposing the conditioned nature and limitations of that very frame. In this sense, the practical intent and the epistemological disclosure of emptiness are not separate; they come to coincide within the structure created by the act of silence itself.
[25] SN II.114; DN II.63.
[26] Vism. XVIII.28. Although paññatti does not appear frequently in the Nikāyas, there is a passage like this: “puggaloti paññattimattaṃ, khandhaparamatthaṃ” (AN 9.20)—“‘A person (puggala)’ is merely paññatti (concept), and the aggregates (khandha) are ultimate ‘reality.’” The chariot example is also presented in the Milindapañha as an instance of vohāra-matta.
[27] Walpola Rahula (1974, 62-64) characterizes avyākata as the Buddha’s response to “ill-formed questions”—questions that, as analyzed above, rest on category mistakes and erroneous binary presuppositions.
[28] MMK XXIV.18 Nāgārjuna identifies emptiness with dependent origination, yet whether this identification reflects the Buddha’s original intention remains a matter requiring independent argumentation. This study adopts the Madhyamaka interpretive framework as a hermeneutical premise, noting its explanatory power in analyzing the relational structure of language. The independent demonstration of this premise, however, falls beyond the scope of the present paper and is left as a task for future research.
[29] In the Diṭṭhi Sutta (AN IV.68), the Buddha classifies all metaphysical assertions such as “I exist” and “I do not exist” as attachments to views (diṭṭhi), and teaches that the state beyond such assertions is “liberation from views” (vimutti diṭṭhigatānaṃ). This reinterpretation aligns with contemporary philosophy of language, particularly Wittgenstein’s insight that “the limits of my language mean the limits of my world” (Tractatus 5.6) and his later recognition of language as use within forms of life (Philosophical Investigations).
[30] Despite clear limitations, the reason for adopting this methodology is to reveal more visibly the layers at which language operates. Humans do not intuitively recognize the dualistic structure of surface and depth in the process of using language, and this distinction is not presupposed in everyday language experience. However, by approaching language models based on information theory in artificial intelligence, especially distributed representation methods like Word2Vec, we have become able to confirm structural layers that transcend the operations of categories and concepts revealed on the surface of language. Such operations are implemented through programming languages and formulaic structures, which, while difficult to grasp intuitively without understanding programming languages, clearly exist as structures in specific technical contexts. Therefore, the structural layers of language models can be understood not as limited to mere surface-level symbolic systems, but as forms that reveal how cognition is implemented in language. In this respect, despite the limitation of not being a large language model (LLM), there is sufficient methodological value worth attempting in a research environment at the level of a short paper.
[31] t-SNE (t-distributed Stochastic Neighbor Embedding) and PCA (Principal Component Analysis) are statistical techniques used to reduce dimensionality in order to visualize relationships within high- dimensional data. However, during this process of simplification, the original multidimensional distances and directional relationships cannot be perfectly preserved, resulting in certain semantic “distortions” in the projected plane. In other words, when complex relational networks are diagrammatically represented, the proportional details among relationships may change (Van der Maaten and Hinton 2008). This reduction process is not purely statistical but formal: it resembles a topological transformation in which quantitative detail is altered while qualitative relations—connectivity and continuity—are preserved. Thus, dimensional projection functions as a heuristic act of simplification that, like topology, seeks the invariance of relational structure beneath numerical change.
[32] While avyākata and Word2Vec operate at different epistemic layers—the former nullifying representations at the subjective-presuppositional level, the latter demonstrating the emptiness of language at the objective-categorical level—they share a common structural direction: the collapse of representational apparatus. Emptiness manifests differently across these layers: in avyākata, through practical cessation of presupposing frameworks; in Word2Vec, through formal disclosure of relational dependencies. Yet both reveal the same insight: that meaning is not self-subsisting essence but arises through interdependent relations. This structural correspondence does not equate the two methods, but shows that they address the same target—the illusory stability of representation—from complementary perspectives.
[33] The finger pointing at the moon is a well-known expression frequently cited in Buddhist thought. It symbolizes that verbal expressions do not directly indicate truth but merely serve as expedient means to point toward it. The Śūraṅgama Sūtra (首楞嚴經, T945, No. 19) states: “If one points to the moon with a finger, another person follows the finger and sees the moon; but if one looks only at the finger, one misses the moon.” In other words, language is merely the pointing gesture, while truth lies in what is being pointed to—the moon itself. Accordingly, the Buddha’s silence can be understood as the act of cutting through the illusion that the pointing finger (language) is identical with the moon (truth).
