Review Reports - Semantic Ambiguity in Japanese Architectural Terms: A Computational Contextual Analysis

Round 1

Reviewer 1 Report

Comments and Suggestions for Authors

Brief Summary:

The article examines interpretable ambiguity of specific Japanese terms in architectural discourse using fine-tuned multilingual BERT masked-language model. The paper uses rigorous analytical model to identify clear cases of ambiguity. The main contribution extends beyond Japanese architectural terms and have potential impact on interpretation of architectural discourse across different languages in general.

Reviews:

Overall well-structured paper to explain research questions, methodology and conclusion. However, the language seems to be quite technical and “heavy” with machine learning technical terms, e.g. section 2.4, that may make the paper difficult to understand by the general audience of the journal and professionals from the industry. More explanation of some of these terms and processes could be included in the text itself with simple diagrams / tables for illustration.

The phenomenon of interpretable ambiguity does not just appear in Japanese architectural terms and architectural discourse in many different languages, e.g. Chinese and Italian, exhibits similar ambiguities. The review of literature in Introduction focuses on past studies on Japanese architectural terms only and claims that “This problem is especially sharp in Japanese architectural discourse” without at least touching on investigations on interpretable ambiguity in other languages.

Justification of the eight Japanese architectural terms needs to be better explained as they are of very different nature. For example, why are these eight terms – engawa, tokonoma, ma, mu, shakkei, doma, shōji, and fusuma – chosen out of the pool of 28 terms from the authors’ previous paper? What is the sampling strategy? To show that the contextual-semantic structure condition appears across different categories of terms? Or to show the range of condition strength variations? Are these the only Japanese architectural terms that “raise[ ] a plausible ambiguity question under the FULL and CLEAN conditions”? What are the others and explain why they are not chosen.

The justification of the five registers: physical-spatial, symbolic-theoretical, represented/imagined, boundary/transition, and hybrid – is also not explained. Why these diametric pairs and what does “hybrid” mean? How about other very important concepts in architecture, such as light-dark, inside-outside, artificial-nature, etc, which are widely accepted distinctive features of Japanese architecture.

What are the actual data used in the analysis as it is not clear reading through the methodology sections. Some names are mentioned in sections 4.1-4.7. Are these the names of authors of text being included in the dataset? What are these text – papers, books, news articles, interviews, etc?

Explanations in section 4 needs to be better connected to the discussions in section 3. How are the largely qualitative statements related to, or derived from, the quantitative data showed in section 3?

Specific Comments:

Line 34-35: I wonder how Hillier and Hanson’s work on spatial analysis through space syntax that examines architecture as physical compositions of space contributes to the topic of the paper.

Figure 1: It would really help understanding if there is more description of each of the stages in the analytical workflow. Simple objectives of each of the stages would also help, e.g. “To identify ambiguities”.

Figure 4: Why do most of the right panels (coloured part) show only 4 of the 5 primary interpretive registers? For example, there is no orange colour (Symbolic-theoretical) for engawa, ma and mu.

Ambiguous conclusion: The first sentence in the first paragraph of section 4.8 reads, “Across the eight terms, ambiguity appears as a patterned but uneven phenomenon, with different terms receiving different levels of support.” This is then followed by first and second sentences of the second paragraph, “The results do not show one general type of ambiguity. Each term shows a different pattern.” Finally, the very last paragraph states, “The main conclusion is that semantic ambiguity in Japanese architectural language follows recurring contextual patterns.” So what are these “patterns”? Are they patterns within individual terms? Patterns that several terms share? Are there categories of patterns?

Author Response

We sincerely thank the reviewer for the careful reading of the manuscript and for the detailed and constructive comments. We particularly appreciate the reviewer’s recognition of the rigor of the analytical approach, the overall structure of the paper, and the potential relevance of the framework beyond the specific Japanese architectural terms examined here. The comments helped us identify several points that required clearer explanation, particularly the relationship between the computational methodology and its architectural purpose, the selection of the target terms and interpretive registers, the composition and role of the corpus, and the distinction between computational evidence and subsequent architectural interpretation. We have revised the manuscript accordingly while retaining the methodological detail necessary for reproducibility.

Comment 1: Overall well-structured paper to explain research questions, methodology and conclusion. However, the language seems to be quite technical and “heavy” with machine learning technical terms, e.g. section 2.4, that may make the paper difficult to understand by the general audience of the journal and professionals from the industry. More explanation of some of these terms and processes could be included in the text itself with simple diagrams / tables for illustration.

Response: We appreciate this concern and agree that the computational procedure should remain understandable to readers whose primary background is architecture. At the same time, we have retained the detailed technical information in Section 2.4 because reproducibility is one of the methodological contributions of the study. The active model configuration, software environment, deterministic settings, and runtime parameters are necessary for independent inspection and reproduction of the reported analysis. Substantially reducing this information would weaken that aspect of the paper.

Rather than removing the technical specification, we added a short explanatory passage at the beginning of Section 2.4 clarifying why this level of detail is retained. The revised manuscript now states that these settings are reported to document the computational environment and analytical configuration, not because they constitute architectural findings. It also directs readers to Figure 1 and the subsequent subsections for the broader analytical logic of the procedure.

We also revised Figure 1 so that the workflow can be understood before engaging with the detailed computational specifications. The figure now distinguishes the analytical resources inherited from the previous validated framework from the five analytical stages performed in the present study, and each stage includes a concise statement of its purpose. In this way, the technical detail required for reproducibility is retained while the overall logic is made more accessible to architectural readers.

Comment 2: The phenomenon of interpretable ambiguity does not just appear in Japanese architectural terms and architectural discourse in many different languages, e.g. Chinese and Italian, exhibits similar ambiguities. The review of literature in Introduction focuses on past studies on Japanese architectural terms only and claims that “This problem is especially sharp in Japanese architectural discourse” without at least touching on investigations on interpretable ambiguity in other languages.

Response: We agree with the reviewer that semantic ambiguity is not unique to Japanese architectural terminology. Our intention was not to claim that Japanese architectural discourse is uniquely ambiguous, but to use Japanese architectural terminology as a bounded empirical case through which to test whether the present computational framework can identify and support the interpretation of contextual semantic variation.

We therefore revised the Introduction to remove the earlier wording that could be read as implying that this problem is especially or uniquely characteristic of Japanese architectural discourse. The revised manuscript now states explicitly that semantic ambiguity is not unique to Japanese terminology and that the Japanese terms examined here define the empirical scope of this study rather than the limits of the analytical framework.

We recognize that the reviewer’s suggestion concerning Chinese and Italian architectural terminology refers to the broader literature context rather than necessarily requesting a cross-linguistic empirical comparison. We considered adding a brief comparative literature discussion, but ultimately chose not to expand the Introduction in this direction for several reasons. First, the present study does not make the claim that ambiguity exists only in Japanese architectural terminology. Second, its objective is not to compare ambiguity across languages, but to test how the computational framework performs when applied to a bounded set of semantically complex architectural terms. A substantive review of comparable Chinese, Italian, or other architectural terminologies would introduce a separate comparative literature problem that is not examined empirically in the present study. Conversely, mentioning such literatures only briefly would risk introducing comparative claims without developing them sufficiently.

Such an extension would also substantially broaden the scope and length of an already methodologically detailed manuscript beyond its principal objective. We therefore chose to clarify the limits of the present claim rather than add a comparative literature section that the study itself does not investigate.

The revised manuscript nevertheless makes the potential broader application explicit. It now states that, with appropriate corpus construction, model adaptation, architecture-facing interpretive criteria, and separate validation, the same analytical logic could be applied to other terminological fields and linguistic or architectural contexts. Comparative studies could therefore constitute a subsequent stage of research, including applications to other architectural languages individually or direct comparisons across linguistic traditions.

The Limitations and Scope section similarly clarifies that extension to other vocabularies, languages, or architectural traditions would require separate empirical testing. We therefore acknowledge the reviewer’s broader point while retaining the focused scope of the present paper.

Comment 3: Justification of the eight Japanese architectural terms needs to be better explained as they are of very different nature. For example, why are these eight terms – engawa, tokonoma, ma, mu, shakkei, doma, shōji, and fusuma – chosen out of the pool of 28 terms from the authors’ previous paper? What is the sampling strategy? To show that the contextual-semantic structure condition appears across different categories of terms? Or to show the range of condition strength variations? Are these the only Japanese architectural terms that “raise[ ] a plausible ambiguity question under the FULL and CLEAN conditions”? What are the others and explain why they are not chosen.

Response: We agree that the original manuscript did not explain the selection procedure sufficiently. Section 2.5 has therefore been expanded to clarify both the sampling strategy and the architectural rationale for each term.

The revised manuscript now states that the eight-term set was purposively selected from the 28-term vocabulary examined in the preceding studies and was not intended to be exhaustive or statistically representative of that wider vocabulary. Importantly, the eight terms were selected before the present computational analysis. The selection was based on prior architectural interpretation and literature review, not on the clustering stability, evidential strength, or other computational outcomes subsequently obtained.

The terms were retained because they posed different architectural ambiguity questions before computation, including materially anchored elements, boundary or transitional conditions, perceptual and representational relations, and more abstract conceptual cases. The target set was also deliberately kept small enough to permit detailed term-level contextual and architectural interpretation.

We further clarify that terms outside this eight-term set were not excluded because they were assumed to be semantically unambiguous. They simply fall outside the bounded scope of the present study. The aim was therefore neither to identify the only ambiguous terms among the original 28 nor to select terms retrospectively according to the strength of the computational results.

The revised Section 2.5 now also explains the rationale for all eight terms. Engawa was selected because of questions concerning connection and boundary-making; tokonoma because of the relation between physical location, ritual, display, and representation; shakkei because of the relation between external scenery and compositional incorporation; ma because of its relational and temporal range across interval, spacing, pause, timing, and relation; mu as a more abstract conceptual case involving nothingness, absence, and potentiality; doma because of its interior-threshold condition; shōji because of its relation to division, permeability, light, and closure/connection; and fusuma because of movable partitioning, spatial recomposition, opacity, decorated surfaces, and possible representation.

The results subsequently produced markedly different evidential strengths across these terms, including strong, cautious, abstract, and low-stability cases. This unevenness further demonstrates that the selection was not made post hoc to produce uniformly favorable results.

Comment 4: The justification of the five registers: physical-spatial, symbolic-theoretical, represented/imagined, boundary/transition, and hybrid – is also not explained. Why these diametric pairs and what does “hybrid” mean? How about other very important concepts in architecture, such as light-dark, inside-outside, artificial-nature, etc, which are widely accepted distinctive features of Japanese architecture.

Response: We agree that the role and selection of the five registers required clearer explanation. Section 2.7 has therefore been expanded.

The revised manuscript clarifies that these registers are not intended as diametric pairs and do not constitute a complete ontology of architectural meaning. They are broad architecture-facing dimensions selected so that the same interpretive frame can be applied across the heterogeneous eight-term set without requiring every term to conform to the same specific architectural opposition.

The revised text defines the registers explicitly: physical-spatial captures material, locational, and built-spatial reference; symbolic-theoretical captures conceptual and abstract architectural discourse; represented/imagined captures representational, perceptual, or image-related meaning; boundary/transition captures threshold, edge, mediation, and spatial transition; and hybrid is reserved for cases in which several primary dimensions operate together and no single register adequately describes the resulting profile.

We also addressed the alternatives mentioned by the reviewer. Relations such as inside/outside, light/dark, and artificial/natural are important architectural relations, but they are not equally applicable to every term in the target set. Adopting them as universal primary registers would risk imposing term-specific architectural themes on the full set before contextual structure had been identified. Where relevant, these relations remain visible in the representative contexts and secondary relational cues rather than being imposed as universal primary classifications.

The manuscript now also makes explicit that “hybrid” is not a residual category for uncertain cases. It is used only when several primary registers operate together in a way that cannot adequately be represented by a single register.

Comment 5: What are the actual data used in the analysis as it is not clear reading through the methodology sections. Some names are mentioned in sections 4.1-4.7. Are these the names of authors of text being included in the dataset? What are these text – papers, books, news articles, interviews, etc?

Response: We appreciate this comment because it identified a distinction that was documented in the preceding methodological study and public repository but was too compressed in the submitted manuscript.

Section 2.2 has now been expanded to describe the corpus more explicitly. The revised manuscript states that the inherited corpus is an English-language expository corpus assembled primarily from web-based encyclopedic and related explanatory sources and supplemented by peer-reviewed architectural literature. The present study does not reconstruct or resample this corpus. It reuses the archived FULL and CLEAN versions so that the ambiguity analysis remains directly comparable with the preceding studies. The raw corpus, filtered corpora, source lists, and associated archived materials remain publicly available through the repository associated with the original corpus study.

We also clarified the status of the scholarly authors cited in Sections 4.1–4.7. These works are used as interpretive literature for contextualizing the computational findings, not as the representative corpus contexts from which the reported clusters were generated. The interpretive sources cited in Section 4 were not included in the FULL or CLEAN corpora used for the study-specific model fine-tuning or contextual analysis. Consequently, the computational pipeline had no direct study-specific exposure to those interpretive texts from which to reproduce their architectural readings. The representative contexts discussed in Section 4 are instead drawn from the corpus outputs generated by the computational pipeline and archived in the study repository.

This distinction is important because the architectural literature functions as an independent interpretive layer applied after the computational evidence has been generated.

The results also do not necessarily reproduce the literature-derived interpretations. Fusuma provides a clear example: although the architectural and art-historical literature makes a representational reading plausible, the computational evidence in this study supports movable partitioning, opacity, decorated surfaces, and spatial recomposition more strongly than representational extension. This divergence further supports the distinction between the evidence generated by the computational analysis and the literature subsequently used to interpret it.

Comment 6: Explanations in section 4 needs to be better connected to the discussions in section 3. How are the largely qualitative statements related to, or derived from, the quantitative data showed in section 3?

Response: We agree and have revised the manuscript substantially to make this relationship explicit.

The opening of Section 4 now states that the computational pipeline identifies recurring contextual structures and register alignments but does not independently establish architectural concepts or final semantic categories. The architectural formulations developed in the Discussion are therefore presented as interpretations of the computational patterns in relation to representative corpus contexts and relevant architectural scholarship.

The manuscript now also states explicitly that each Discussion subsection distinguishes between the computational evidence reported in Section 3 and the architectural interpretation developed from that evidence. We also clarified the role of the representative sentence evidence. The representative contexts are themselves outputs of the computational pipeline and are archived in the representative-context CSV files. Section 3 uses only brief sentence cues because its primary purpose is to report the recovered computational structure, including screening strength, cluster recovery, register alignment, and cross-condition persistence. The fuller representative contexts are reserved for Section 4 because their principal role is interpretive: they allow the reader to inspect the textual evidence from which the architectural reading is developed. Reporting them in full in Section 3 and then interpreting the same sentences again in Section 4 would increase repetition rather than reduce it. The revised organization therefore distinguishes the reporting of computational structure from the close architectural interpretation of the representative textual evidence.

This distinction was carried through to the individual interpretations. In the case of engawa, for example, the computational analysis supports positive boundary-edge and connector-flow alignment. The formulation of engawa as a “modal littoral-circuit space” is then presented explicitly as an architectural interpretation developed from that evidence and the architectural literature, rather than as a semantic category directly generated by the model.

Conversely, where the computational evidence constrains an architectural hypothesis, the manuscript now states this clearly. Fusuma again provides an example: the representational reading remains theoretically plausible, but partitioning and spatial recomposition receive stronger computational support.

The revised argumentative sequence is therefore:

computational result → representative corpus evidence → architectural interpretation.

This preserves the architectural contribution of the Discussion while making clear where the computational evidence ends and interpretive reasoning begins.

Comment 7: Line 34-35: I wonder how Hillier and Hanson’s work on spatial analysis through space syntax that examines architecture as physical compositions of space contributes to the topic of the paper.

Response: We appreciate the opportunity to clarify this citation. Hillier and Hanson are not cited because the present study applies space-syntax analysis. Their relevance lies in the broader theoretical premise developed in The Social Logic of Space: architectural space is not treated merely as a physical composition but as an ordering of spatial relations through which relations between people and forms of social organization are structured.

We therefore retained the reference but revised the Introduction so that its relevance is stated explicitly:

“Hillier and Hanson [5] provide a related spatial argument: their account treats the ordering of architectural space not merely as physical composition, but as a means through which relations between people and forms of social organization are structured.”

The source therefore supports the spatial-relational premise of the manuscript rather than its computational methodology.

Comment 8: Figure 1: It would really help understanding if there is more description of each of the stages in the analytical workflow. Simple objectives of each of the stages would also help, e.g. “To identify ambiguities”.

Response: We agree and have redesigned Figure 1 specifically in response to this comment.

The revised figure now distinguishes the analytical resources inherited from the preceding validated framework from the five stages of the present study. Stage 00 identifies the inherited analytical resources, namely the fine-tuned multilingual BERT model and archived FULL/CLEAN corpora. It represents the starting condition supplied by the prior framework rather than an additional analytical operation performed in this study.

Stages 01–05 then constitute the present study, and each now includes a concise statement of purpose:

Screening Audit: determine whether each term has sufficient usable evidence to proceed with analysis.

Contextual Cluster Recovery: identify recurring contextual contrasts within each target term.
Architecture-Facing Register Alignment: relate the recovered contextual structure to predefined architectural dimensions.
Cluster-Aware Interpretation: preserve differences between recovered clusters rather than reducing each term to one average profile.
Bounded Engawa Concept-Alignment Module: stress-test the specific boundary-edge and connector-flow interpretation proposed for engawa.

Section 2.1 was also revised so that the same distinction is stated explicitly in the text: the fine-tuned model and paired FULL/CLEAN corpora are inherited unchanged from the preceding validated framework, whereas the five-stage workflow constitutes the ambiguity analysis conducted in the present study.

The revised figure therefore provides the simple stage-by-stage objectives requested by the reviewer without replacing the detailed methodological description necessary for reproducibility.

Comment 9: Figure 4: Why do most of the right panels (coloured part) show only 4 of the 5 primary interpretive registers? For example, there is no orange colour (Symbolic-theoretical) for engawa, ma and mu.

Response: We agree that the original caption did not explain this sufficiently.

All five primary interpretive registers are evaluated for every cluster. A missing color does not mean that a register was omitted from the analysis. Rather, a register may have no visible segment when its contribution is not retained in the resulting cluster-level profile.

We therefore revised the Figure 4 caption to state explicitly:

“All five registers are evaluated for every cluster, but a register may have no visible segment when its contribution is not retained in the resulting cluster-level profile; the absence of a color therefore indicates no displayed contribution for that register, not that the register was omitted from the analysis.”

This directly explains why, for example, symbolic-theoretical is not visibly represented in some of the right-hand panels even though it was included among the five evaluated registers.

Comment 10: Ambiguous conclusion: The first sentence in the first paragraph of section 4.8 reads, “Across the eight terms, ambiguity appears as a patterned but uneven phenomenon, with different terms receiving different levels of support.” This is then followed by first and second sentences of the second paragraph, “The results do not show one general type of ambiguity. Each term shows a different pattern.” Finally, the very last paragraph states, “The main conclusion is that semantic ambiguity in Japanese architectural language follows recurring contextual patterns.” So what are these “patterns”? Are they patterns within individual terms? Patterns that several terms share? Are there categories of patterns?

Response: We agree that the earlier wording allowed two different meanings of “pattern” to be conflated.

The intended claim concerns recurring contextual structure within individual terms, particularly where that structure remains recoverable across both FULL and CLEAN conditions. It does not mean that the eight terms share one common ambiguity pattern or one common architectural category.

Section 4.8 has therefore been revised to state explicitly:

“The recurrence identified here does not imply that all eight terms share one common type of ambiguity. Rather, ‘recurring contextual patterns’ refers to the repeated recovery of differentiated contextual structure within individual terms, particularly where that structure remains identifiable across both FULL and CLEAN conditions. The architectural content of that structure remains term-specific.”

The revised text then illustrates this distinction: for engawa, the structure concerns connection and boundary; for tokonoma, display and symbolic interiority; for shakkei, scenic incorporation; and for ma, broader relational and conceptual breadth.

We also revised the final conclusion so that it now states that several of the selected Japanese architectural terms exhibit recurring within-term contextual-semantic structure across the paired corpus conditions, while the architectural form and evidential strength of that structure differ by term.

The revised manuscript therefore uses “recurring” to describe the repeated recovery of differentiated contextual structure within individual terms and across corpus conditions, while recognizing that the architectural content and evidential strength of that structure remain term-specific.

Reviewer 2 Report

Comments and Suggestions for Authors

The paper is technically rigorous, reproducible, and well documented. The inclusion of audited workflows, public code, and reproducibility controls considerably strengthens the methodological transparency. Despite these strengths, the manuscript currently overextends several interpretive claims beyond what the computational evidence alone can support. The paper would benefit from a clearer separation between computational findings and theoretical interpretation.

Interpretation of the results

The computational analysis is carefully conducted and successfully identifies contextual differences among the selected Japanese architectural terms. However, in several places the discussion goes beyond what the results directly support. Some of the proposed architectural interpretations, such as the new conceptual descriptions of engawa and fusuma, appear to be the authors' theoretical interpretations rather than conclusions that emerge directly from the computational analysis. The manuscript would be stronger if it clearly separated the computational findings from the authors' conceptual interpretations and presented the latter more cautiously.

Validation of the semantic categories

The study relies entirely on computational clustering to identify semantic patterns. While the clustering appears stable, the manuscript does not provide any external validation of the identified categories. Including an expert review, manual annotation, or another independent validation approach would increase confidence that the recovered semantic groups genuinely reflect architectural meanings rather than computational artefacts.

Justification of the architectural registers

The five architectural registers used throughout the analysis play a central role in interpreting the results. However, the manuscript provides only limited explanation of why these particular registers were selected and whether other possible classification frameworks were considered. A clearer justification of this framework would help readers better understand the basis of the interpretations.

Scope and generalizability

The study focuses on eight Japanese architectural terms, which is appropriate for a focused investigation. However, the discussion could better explain how the proposed framework could be applied to a larger vocabulary, other languages, or different architectural traditions. A brief discussion on the broader applicability of the method would strengthen the overall contribution of the paper.

Length and organization of the discussion

The Discussion section is comprehensive but considerably longer than necessary. Several sections repeat results that have already been presented before offering interpretation. Reducing repetition and focusing more on the broader implications of the findings would improve the readability and overall flow of the manuscript without reducing its scholarly contribution.

Strengthen the conclusion by emphasising the broader significance of the methodology while avoiding claims that exceed the empirical evidence. Carefully proofread the manuscript for minor stylistic redundancies and improve transitions between methodological description, results, and interpretation.

Comments on the Quality of English Language

Some paragraphs exceed half a page and contain multiple nested clauses. Breaking longer paragraphs into shorter units would improve readability without changing content.

Author Response

We sincerely thank the reviewer for the careful and constructive assessment of the manuscript. We particularly appreciate the reviewer’s recognition of the technical rigor, reproducibility, documentation, audited workflow, and public availability of the analytical materials. The comments helped us identify where the manuscript needed a clearer separation between computational findings and architectural interpretation, a fuller explanation of the architectural-register framework, more explicit limits on generalizability, targeted condensation of the Discussion, and a more carefully qualified statement of the broader methodological contribution. We have revised the manuscript accordingly. Where a requested addition would require a separate empirical study, particularly external expert validation, we have clarified the scope and limits of the present analysis rather than claiming evidence that the current research design does not provide.

Comment 1: Interpretation of the results

The computational analysis is carefully conducted and successfully identifies contextual differences among the selected Japanese architectural terms. However, in several places the discussion goes beyond what the results directly support. Some of the proposed architectural interpretations, such as the new conceptual descriptions of engawa and fusuma, appear to be the authors' theoretical interpretations rather than conclusions that emerge directly from the computational analysis. The manuscript would be stronger if it clearly separated the computational findings from the authors' conceptual interpretations and presented the latter more cautiously.

Response: We agree that the submitted manuscript needed a clearer separation between computational findings and architectural interpretation. We have revised the manuscript throughout to make this evidential distinction explicit.

The Methods now state that the outputs of the analysis are treated as evidence for architectural interpretation rather than as proof of final semantic categories. The opening of the Discussion similarly states that the computational pipeline identifies recurring contextual structures and register alignments but does not independently establish architectural concepts or final semantic categories. The architectural formulations developed in the Discussion are therefore presented as interpretations constructed from three distinguishable elements: the computational structure reported in the Results, the representative corpus contexts generated by the pipeline, and the relevant architectural or art-historical scholarship.

The Abstract has been revised accordingly. It now states that the focused engawa test supports a combined boundary-edge and connector-flow pattern and separately identifies “modal littoral-circuit space” as the architectural interpretation developed in this paper. The formulation is therefore no longer presented as if it were a semantic category generated directly by the computational model.

Section 4.1 now makes the same distinction in greater detail. The computational analysis supports positive alignment with boundary-edge, connector-flow, derived hybridity, and the riverbank analogy, while the bootstrap differences are too weak to establish one axis as dominant. The riverbank relation is explicitly identified as an architectural analogy rather than an etymological or historical claim. The longitudinal and transverse reading of engawa, and the resulting “modal littoral-circuit” formulation, are consequently identified as architectural interpretations developed from the computational evidence and relevant architectural scholarship.

The revised tokonoma subsection follows the same evidential sequence. It first reports the two contextual modes recovered by the analysis: a materially anchored display and room-orientation mode and a more title-like or concept-framing mode. It then introduces the relevant scholarship as a separate interpretive layer. Finally, it explains that, when the computational results, representative contexts, and scholarship are read together, tokonoma can be interpreted as a physically stable interior site whose changing contents make natural, spatial, literary, philosophical, or spiritual domains present through representation. The subsection explicitly states that the clustering does not establish a discrete opposition between a physical alcove and an exterior or abstract conceptual space and does not independently establish the full representational interpretation.

Comparable qualifications have been introduced for shakkei and shōji. The shōji subsection, for example, distinguishes the computationally supported constructional and functional evidence from the secondary architectural interpretation that continuity of filtered light may weaken the perceived finality of enclosure. The manuscript now states explicitly that perceptual spatial enlargement is not tested directly by the present analysis.

Fusuma has also been revised substantially. The computational evidence supports movable partitioning, opaque separation, room division and reconfiguration, decorated surfaces, paintings, and image-bearing surface use more directly than it supports perceptual extension. The possibility that fusuma-e can alter the perceived limit of a confined interior is therefore retained as an architectural interpretation informed by the corpus evidence and pictorial and architectural scholarship, not as a semantic structure established directly by clustering.

The Conclusion has been revised in the same manner. The response to RQ4 now separates the positive boundary-edge and connector-flow alignment from the subsequent architectural formulation of engawa. The final conclusion also limits the central claim to recurring within-term contextual-semantic structure among several of the selected terms and recognizes that the architectural form and evidential strength of that structure differ by term.

Comment 2: Validation of the semantic categories

Response: We agree that a preplanned independent evaluation could strengthen a future validation study. However, we did not add a post hoc expert review or manual annotation exercise to the present manuscript because the current study was not designed to establish externally validated semantic categories.

The clustering stage identifies corpus-conditioned contextual structure. A two-cluster solution is treated as a conservative contextual contrast, not as an exhaustive inventory of dictionary senses or a canonical classification of architectural meanings. The five architectural registers are subsequently used as constrained interpretive dimensions, not as externally verified semantic labels. The revised manuscript makes this limitation clearer by consistently referring to contextual structures, contextual modes, register alignments, and architectural interpretations rather than treating the clusters as definitive semantic categories.

A credible external validation would require a separate protocol established before expert inspection of the present results. Such a study would need independently selected contexts, a defined annotation scheme, multiple architectural and linguistic experts, and an explicit measure of inter-rater agreement. A retrospective classification conducted by the authors after examining the clusters would not constitute genuinely independent validation and could introduce circularity by asking annotators to confirm categories already suggested by the computational outputs.

The present study instead provides internal computational and evidentiary controls appropriate to its stated purpose. These include exact target-span recovery, matched-context thresholds, bootstrap stability, permutation-separation diagnostics, comparison between FULL and CLEAN corpus conditions, archived representative contexts, deterministic settings, and a complete public implementation record. These controls assess whether the reported contextual structure is reproducible and sufficiently stable under the audited pipeline. They do not replace external semantic validation, and the manuscript does not claim that they do.

The architectural and art-historical scholarship introduced in the Discussion is also kept separate from the corpus evidence used to produce the computational results. It functions as an interpretive frame rather than as external validation of the clusters. This separation is now stated explicitly in the Methods and Discussion.

The Limitations and Scope section further clarifies that extension to other vocabularies, languages, or architectural traditions would require separate empirical testing and further validation. We therefore addressed this concern by narrowing the status of the identified structures and making the absence of externally validated semantic categories clear, rather than adding a limited post hoc annotation exercise that would not provide the independent validation requested by the reviewer.

Comment 3: Justification of the architectural registers

Response: We agree that the role and selection of the five architectural registers required clearer explanation. Section 2.7 has therefore been expanded substantially.

The revised section explains that physical-spatial, symbolic-theoretical, represented/imagined, boundary/transition, and hybrid are not intended as mutually exclusive oppositions or as a complete ontology of architectural meaning. They were selected as broad architecture-facing dimensions that could be applied consistently across a heterogeneous set containing materially anchored elements, representational relations, boundary conditions, and more abstract concepts.

The revised manuscript now defines each register. Physical-spatial captures material, locational, and built-spatial reference. Symbolic-theoretical captures conceptual and abstract architectural discourse. Represented/imagined captures representational, perceptual, or image-related meaning. Boundary/transition captures threshold, edge, mediation, and spatial transition. Hybrid is reserved for cases in which several primary dimensions operate together and no single register adequately describes the identified profile. The manuscript also clarifies that hybrid is not a residual category for uncertain or unclassified cases.

We additionally explain why more specific relations, such as inside/outside, light/dark, or artificial/natural, were not adopted as universal primary registers. These relations are architecturally important, but they are not equally applicable to all eight terms. Using them as universal categories would risk imposing term-specific architectural themes on the full target set before the contextual structure had been identified. Where relevant, such relations remain visible in the representative contexts and in the six secondary relational cues used by the analysis.

The registers are also applied only after contextual cluster identification. The analysis therefore begins with the corpus-derived contextual structure rather than assigning every occurrence to a predetermined architectural classification. The revised explanation makes clearer that the registers provide a shared comparative surface across the eight terms while remaining subordinate to the identified contextual evidence.

Comment 4: Scope and generalizability

Response: We agree and have revised the Introduction, Limitations and Scope section, and Conclusion to explain the broader applicability of the method while avoiding unsupported claims of generalizability.

The Introduction now states that the Japanese terminology examined in the paper defines the empirical scope of the study rather than the limits of the analytical framework. It also explains that, with appropriate corpus construction, model adaptation, architecture-facing interpretive criteria, and separate validation, the same analytical logic could be adapted to other terminological fields and linguistic or architectural contexts.

The revised Limitations and Scope section distinguishes methodological extensibility from demonstrated generalizability. Related work within the same research program shows that the broader computational-semantic framework can be applied to contextual comparisons among selected architectural terms. However, this does not by itself establish that the present findings generalize to larger vocabularies, other languages, or other architectural traditions. Those applications would require new corpora, renewed model and extraction audits, context-appropriate interpretive criteria, and separate empirical testing.

The revised limitations also identify specific directions for extension, including larger specialist corpora, alternative clustering criteria, and finer-grained examination of terms whose current structures remain broad or cautious.

Finally, the concluding paragraph now identifies one practical application of the method. By distinguishing concise material definitions from recurring symbolic, relational, perceptual, and experiential uses, the framework may inform layered architectural glossary and encyclopedia entries. The manuscript qualifies this contribution by stating that such outputs must remain subject to architectural, historical, and linguistic review. This addition explains a broader scholarly and reference-oriented benefit of the method without presenting the present eight-term analysis as universally generalizable.

Comment 5: Length and organization of the discussion

Response: We agree that genuine repetition between the Results and Discussion should be reduced. We therefore made targeted reductions throughout Section 4 while preserving the evidence needed to connect the computational results to the architectural interpretations.

The revised Discussion reduces repeated reporting of cluster counts, target-span identification, stability values, register rankings, and other diagnostics already established in Section 3. Repeated conclusions within individual term subsections have been removed or consolidated, and the extended term-by-term recap in Section 4.8 has been replaced by a shorter synthesis distinguishing the term-specific forms of contextual structure.

The final tokonoma subsection provides a specific example of this targeted condensation. A paragraph that repeated the same representative corpus contexts already given immediately above has been removed. The contexts now appear once, followed by the relevant scholarship and then the architectural synthesis. This eliminates genuine duplication while retaining the complete interpretation of tokonoma’s representational capacity.

Comparable reductions were made in the ma, doma, shōji, and fusuma discussions. In particular, repeated diagnostic summaries and repeated concluding formulations were shortened where their evidentiary function was already fulfilled elsewhere.

We did not remove the representative corpus contexts themselves. These contexts are outputs of the computational pipeline stored in the representative-context CSV files, and they provide the traceable textual evidence needed to understand why a recovered contextual structure receives a particular architectural interpretation. Section 3 reports the computational structure and uses only brief contextual cues. Section 4 uses the fuller representative evidence to evaluate what that structure can reasonably mean architecturally.

Moving all representative sentences into Section 3 would lengthen the Results and require the same evidence to be discussed again in Section 4. Deleting them from the manuscript because they are preserved in the CSV files would also weaken the evidentiary connection visible to the reader. The revised organization therefore establishes a clearer division of labor:

Section 3 reports the computational findings and brief contextual cues.

Section 4 retains only the representative evidence needed to support the architectural interpretation and avoids repeating quantitative diagnostics already reported in Section 3.

Section 4.8 provides a concise cross-term synthesis without reproducing every term-level result.

We have therefore followed the reviewer’s recommendation through targeted condensation rather than wholesale deletion. This improves the flow of the Discussion without removing the representative evidence required to make the interpretive reasoning auditable.

Comment 6: Strengthen the conclusion by emphasising the broader significance of the methodology while avoiding claims that exceed the empirical evidence. Carefully proofread the manuscript for minor stylistic redundancies and improve transitions between methodological description, results, and interpretation.

Response: We agree and have revised the Conclusion, transitions, and wording throughout the manuscript.

The response to RQ2 has been condensed substantially. It now summarizes the persistence of contextual-semantic structure across FULL and CLEAN while retaining the differences in evidential strength among the eight terms. It concludes with a qualified affirmative answer and explicitly states that persistence across corpus conditions does not imply uniform evidential strength or one common ambiguity structure.

The response to RQ4 now separates the computational finding from the architectural interpretation. It first reports positive alignment with boundary-edge and connector-flow logic and then states that, building on this evidence and the architectural literature, the paper proposes the modal littoral-circuit formulation as an architectural interpretation of engawa.

The final conclusion has also been narrowed. It no longer states broadly that semantic ambiguity in Japanese architectural language follows recurring contextual patterns. It now concludes that several of the selected Japanese architectural terms exhibit recurring within-term contextual-semantic structure across the paired corpus conditions, while the architectural form and evidential strength of that structure differ by term.

At the same time, the broader methodological significance is now stated more clearly. Contextual computational analysis can make variation within architectural terminology more explicit and testable while remaining anchored in architectural interpretation. The conclusion also identifies the potential contribution to layered architectural glossary and encyclopedia entries, subject to architectural, historical, and linguistic review.

Transitions among the analytical stages have been clarified throughout the manuscript. The Methods distinguish the inherited model and corpora from the five-stage analysis performed in the present study. The Results present the computational evidence as a cumulative sequence. The opening of the Discussion explains how representative corpus contexts connect those results to architectural interpretation. The individual subsections then distinguish the recovered computational structure from the architectural reading developed from it.

Finally, we proofread the revised manuscript for stylistic repetition, duplicated wording, punctuation, and transitions. Repeated material was removed where it did not perform a distinct evidentiary function, while evidence-bearing representative contexts were retained where necessary. These revisions strengthen the broader methodological conclusion while keeping the claims within the empirical limits of the selected terms, the audited corpus conditions, and the analytical design.

Reviewer 3 Report

Comments and Suggestions for Authors

1. The manuscript is well structured and methodologically transparent. The research questions are clearly formulated, and the analytical workflow from screening, contextual clustering, register alignment, cluster-aware interpretation, to the focused engawa analysis is systematically presented.
2. The Introduction could be more concise and synthesised. Although the manuscript provides extensive relevant literature, the relationship among architectural semantics, Japanese aesthetics, semantic ambiguity, and contextual computational modelling could be explained more directly to clarify the theoretical contribution of the study.
3. The distinction between computational findings and architectural interpretation should be made clearer. In particular, the characterization of engawa as a “modal littoral-circuit space” is an interesting conceptual interpretation, but it should be presented more explicitly as an interpretive formulation rather than as a category directly established by the computational model.
4. The methodological asymmetry of the engawa analysis should be discussed more explicitly. Only engawa receives a dedicated concept-alignment module, while the other terms are evaluated through the common analytical pipeline. This may affect the comparability of the strength of conclusions across the eight terms.
5. The authors appropriately acknowledge variation in evidential strength across the terms. This is a strength of the manuscript, particularly in the cautious interpretation of doma as a low-stability case and fusuma as being more strongly supported in terms of partitioning and spatial recomposition than representational extension.
6. Some repetition between the Results and Discussion sections could be reduced. Several descriptions of cluster characteristics, register profiles, and interpretations are repeated across subsections. Modest condensation would improve readability and strengthen the overall argument.
7. Overall, the manuscript makes a useful contribution to computational approaches to architectural semantics. The main revisions required concern sharper theoretical synthesis, clearer separation between computational evidence and higher-level interpretation, and improved concision rather than major changes to the research design.

Author Response

We sincerely thank the reviewer for the careful and constructive assessment of the manuscript. We appreciate the reviewer’s recognition of the methodological transparency, systematic analytical workflow, and potential contribution of the study to computational approaches to architectural semantics. The comments were particularly useful in identifying where the manuscript needed a clearer distinction between computational evidence and architectural interpretation, a more explicit discussion of the methodological asymmetry of the engawa analysis, stronger theoretical synthesis, and more selective condensation of repeated material. We have revised the manuscript accordingly. Where the reviewer recommended greater concision, we reduced repetition where this could be done without removing evidence necessary to connect the computational findings to their architectural interpretation.

Comment 1: The manuscript is well structured and methodologically transparent. The research questions are clearly formulated, and the analytical workflow from screening, contextual clustering, register alignment, cluster-aware interpretation, to the focused engawa analysis is systematically presented.

Response: We thank the reviewer for this positive assessment. No substantive methodological change was required in response to this comment. We have nevertheless preserved and further clarified the analytical sequence throughout the revised manuscript. The revised Figure 1 and related methodological text distinguish the analytical resources inherited from the preceding validated framework from the five analytical stages performed in the present study. The manuscript also continues to distinguish screening, contextual cluster recovery, architectural register alignment, cluster-aware interpretation, and the bounded engawa concept-alignment module as separate analytical stages.

We have retained this staged structure because it is important to the evidentiary logic of the paper: each stage constrains how strongly the subsequent stage can be interpreted. The revised Methods and Results sections make this progression more explicit while maintaining the reproducibility of the analysis.

Comment 2: The Introduction could be more concise and synthesised. Although the manuscript provides extensive relevant literature, the relationship among architectural semantics, Japanese aesthetics, semantic ambiguity, and contextual computational modeling could be explained more directly to clarify the theoretical contribution of the study.

Response: We appreciate this comment and agree that the relationship among the different strands of the Introduction needed to be made more explicit. We therefore revised the Introduction selectively to clarify the progression from architectural language and relational meaning, through Japanese architectural and aesthetic interpretation, to the specific problem of semantic ambiguity and the role of contextual computational modeling.

The revised Introduction now states the research problem more directly as two linked gaps. The first is architectural: although substantial qualitative scholarship exists on several of the focal terms, their ambiguity has rarely been examined as a distinct and reproducibly testable problem. The second is methodological: contextual corpus methods have rarely been applied to semantically dense Japanese architectural terminology in an explicit ambiguity audit. The Introduction then explains why contextual modelling is appropriate to this problem, namely because the study concerns changes in meaning across local contexts rather than frequency alone.

We also clarified the scope of the theoretical claim. Japanese architectural terminology is now identified as the bounded empirical case examined in this study rather than as a uniquely ambiguous architectural language, and the Introduction makes explicit that the analytical logic could potentially be adapted to other terminological or architectural contexts only through separate corpus construction, model adaptation, and validation.

We did not, however, respond by substantially compressing the literature review or combining citations simply to reduce its length. We considered this carefully because several adjacent references support different parts of the argument. For example, [28,29] establish the broader use of NLP and text-mining methods in construction-related research; [30,31] provide more architecture-specific precedents involving computational architecture and architectural discourse; and [32] supplies the broader digital-humanities context while also emphasizing methodological issues concerning training data and domain adaptation. Collapsing these into a single generalized statement with a citation range such as [28–32] would make it less clear which sources support which claims and would reduce evidentiary precision.

Accordingly, we interpreted “more concise and synthesised” as requiring a tighter theoretical argument rather than simply fewer sentences or fewer citation points. The revised version therefore retains distinctions between sources where those distinctions are substantively important while making the connection between the architectural problem, the semantic problem, and the computational approach more explicit.

Comment 3: The distinction between computational findings and architectural interpretation should be made clearer. In particular, the characterization of engawa as a “modal littoral-circuit space” is an interesting conceptual interpretation, but it should be presented more explicitly as an interpretive formulation rather than as a category directly established by the computational model.

Response: We agree and have revised the manuscript substantially in response to this comment.

The Abstract now separates the computational finding from the architectural interpretation. It first states that the focused engawa analysis supports a combined boundary-edge and connector-flow pattern and then identifies “modal littoral-circuit space” as the architectural formulation developed from that evidence rather than as a category generated by the model.

The distinction is also made explicit in the Methods and Discussion. The manuscript now states that the computational pipeline identifies contextual structures and register alignments but does not independently establish architectural concepts or final semantic categories. Architectural formulations are instead developed by interpreting those computational patterns in relation to representative corpus contexts and relevant scholarship.

We also clarified the status of the literature used in Section 4. The architectural and art-historical sources cited in the Discussion are separate from the corpus evidence used to produce the computational results. They are introduced after the computational evidence has been generated and function as an independent interpretive layer.

For engawa specifically, the revised Discussion distinguishes the positive boundary-edge, connector-flow, derived-hybridity, and riverbank alignments from the subsequent architectural formulation. The riverbank analogy is explicitly described as an architectural analogy rather than an etymological or historical claim, and the “modal littoral-circuit” formulation is presented as an interpretation of the recovered boundary-and-connection structure together with the architectural literature.

The same distinction has been extended beyond engawa. In the revised tokonoma subsection, the recovered contextual modes are reported first: a materially anchored display and room-orientation mode and a more title-like or concept-framing mode. The supporting scholarship is then introduced as a separate interpretive layer, after which the broader representational reading is stated explicitly as an architectural interpretation. The subsection now clarifies that the clustering does not establish a discrete opposition between a physical alcove and a represented exterior or abstract conceptual space. Rather, read together with the representative contexts and scholarship, the results support interpreting tokonoma as a stable material site whose changing displays can make natural, spatial, literary, philosophical, or spiritual domains present within the interior. The manuscript also states explicitly that this fuller representational interpretation is not independently established by the clustering. The revised shōji and fusuma subsections apply the same evidential discipline. In fusuma, the computational analysis supports partitioning, spatial recomposition, and decorated or image-bearing surfaces more directly than the further claim that fusuma-e perceptually extends a confined interior. That latter claim is therefore identified as an architectural interpretation rather than as a semantic structure directly established by clustering.

Comment 4: The methodological asymmetry of the engawa analysis should be discussed more explicitly. Only engawa receives a dedicated concept-alignment module, while the other terms are evaluated through the common analytical pipeline. This may affect the comparability of the strength of conclusions across the eight terms.

Response: We agree and have made this asymmetry explicit in both the Methods and the Limitations and Scope section.

Section 2.9 now states directly that the dedicated secondary concept-alignment module was applied only to engawa. The module tests a narrowly specified architectural hypothesis involving boundary-edge, connector-flow, derived hybridity, and a riverbank analogy. No equivalent bespoke concept-alignment module was applied to tokonoma, shakkei, doma, shōji, fusuma, or the other terms.

We also explain why this additional test was restricted to engawa. The engawa interpretation was sufficiently specific to be expressed through predefined conceptual axes and therefore could be subjected to a bounded hypothesis-stress test. Applying comparable post hoc probes to every term without similarly precise prior hypotheses would risk turning the procedure into a loose confirmation exercise.

The revised Limitations and Scope section consequently states that the additional concept-alignment evidence available for engawa is not directly equivalent to the evidential basis available for the other seven terms and should not be interpreted as showing that engawa is intrinsically more ambiguous or semantically stronger. Comparable secondary tests could be developed for other terms in future work only where the architectural hypothesis can be specified narrowly enough in advance.

Thus, the shared main pipeline remains comparable across all eight terms, while the additional engawa module is explicitly treated as supplementary and methodologically asymmetric rather than as evidence that places engawa on an automatically stronger comparative footing.

Comment 5: The authors appropriately acknowledge variation in evidential strength across the terms. This is a strength of the manuscript, particularly in the cautious interpretation of doma as a low-stability case and fusuma as being more strongly supported in terms of partitioning and spatial recomposition than representational extension.

Response: We thank the reviewer for recognizing this aspect of the analysis. We have preserved this evidence-strength scaling and have made the distinction still more explicit in the revised manuscript.

Doma remains identified as the lowest-stability case. Although its contexts provide an architecturally plausible interior-threshold interpretation, the low stability under both corpus conditions prevents it from being treated as a strong ambiguity case.

Fusuma likewise remains explicitly cautious. The revised discussion distinguishes between the strength of cluster recovery and the strength of the architectural interpretation. Fusuma retains a two-cluster solution under both FULL and CLEAN with complete target-span identification and maximum observed bootstrap stability, so it is not a weak or failed clustering case. The caution instead arises from the weaker separation diagnostics, the small secondary cluster, and the actual content of the representative contexts. Those contexts directly support movable partitioning, opaque separation, room reconfiguration, decorated surfaces, paintings, and image-bearing surface use more clearly than they establish perceptual spatial extension.

The revised Discussion therefore continues to describe fusuma as a cautious case in which partitioning and spatial recomposition receive stronger direct computational support, while representational extension remains a secondary architectural interpretation informed by the corpus evidence and supporting literature.

No change was made that would weaken the evidential caution noted positively by the reviewer.

Comment 6: Some repetition between the Results and Discussion sections could be reduced. Several descriptions of cluster characteristics, register profiles, and interpretations are repeated across subsections. Modest condensation would improve readability and strengthen the overall argument.

Response: We agree that some repetition between the Results and Discussion could be reduced, and we have made targeted condensations throughout Section 4. At the same time, we distinguished between material that was genuinely repetitive and material that is necessary to make the movement from computational evidence to architectural interpretation auditable.

The revised manuscript reduces repeated reporting of cluster counts, span-identification information, stability values, register rankings, and other diagnostic information already established in Section 3. We also removed or compressed repeated conclusions within individual term subsections and reduced the extended term-by-term recap in Section 4.8. The aim was to avoid having the Discussion simply reproduce the Results before interpretation.

We did not, however, remove the representative corpus contexts from the Discussion. This was a deliberate methodological choice. The representative contexts are computational outputs, but they are not reported as a separate quantitative result. Their fuller presentation is reserved for the Discussion because they function as traceable textual evidence for interpreting the recovered cluster structure. The representative sentences are indeed outputs generated by the computational pipeline and stored in the representative-context CSV files, but their function in the argument differs from the function of the quantitative outputs. Section 3 reports the computational structure itself: screening strength, recovered clusters, register alignment, cross-condition persistence, and brief representative cues showing that the clusters correspond to recognizable contextual uses. The manuscript explicitly states that these sentence cues are used sparingly in the Results and that their fuller architectural interpretation is deferred to the Discussion.

The more detailed representative contexts are therefore used in Section 4 as an evidentiary bridge between the computational result and the architectural interpretation. They are not treated as a separate statistical result or as independently discovered architectural categories. Instead, they allow the reader to inspect why a recovered cluster is subsequently interpreted in architectural terms such as boundary mediation, symbolic interiority, scenic incorporation, luminous permeability, or spatial recomposition. The opening of the Discussion now makes this division explicit by stating that the representative contexts come from the pipeline-generated CSV outputs and provide the textual basis for interpreting the identified cluster structure. It also states that each subsection distinguishes between the computational evidence reported in Section 3 and the architectural interpretation developed from that evidence.

We considered placing the detailed representative sentences directly in the Results section, but doing so would have created two problems. First, it would substantially lengthen Section 3 by reproducing textual evidence whose main purpose is architectural interpretation. Second, the same sentences would then have to be discussed again in Section 4 in order to explain their architectural significance, increasing rather than reducing the repetition identified by the reviewer. The revised organization therefore uses Section 3 to establish what contextual structure was computationally recovered and Section 4 to examine what that structure can reasonably mean architecturally, using the archived representative contexts as traceable qualitative evidence.

A related revision concerns the rationale for selecting the eight target terms. Section 2.5 was expanded to make the pre-analytical architectural questions behind the selection explicit. This material was not removed for concision because it performs a different methodological function: it demonstrates that the terms were selected because they posed identifiable architectural ambiguity questions before the present computational results were known and that they were not selected retrospectively because they produced favorable clusters. The manuscript now states explicitly that selection was based on prior architectural interpretation and literature review, not on the stability, clustering, or evidential strength subsequently produced by the analysis. The pre-analytical interpretations are then revisited in Section 4 and evaluated against the computational results, representative contexts, and scholarship rather than being presented as findings produced by the model.

For this reason, the revision does not pursue concision by deleting all material that appears in more than one analytical stage. Instead, it establishes a clearer division of labor:

Section 2.5 identifies the pre-analytical architectural hypotheses that motivated term selection; Section 3 reports the computational findings and only brief contextual cues; and Section 4 uses the fuller representative corpus evidence and architectural literature to evaluate and interpret those findings.

We therefore addressed the reviewer’s recommendation through modest, targeted condensation, while retaining material whose repetition is only apparent because it serves a distinct evidentiary role at a different stage of the argument. This approach reduces unnecessary duplication without weakening the traceability of the study’s central sequence from computational output to corpus evidence to architectural interpretation.

Comment 7: Overall, the manuscript makes a useful contribution to computational approaches to architectural semantics. The main revisions required concern sharper theoretical synthesis, clearer separation between computational evidence and higher-level interpretation, and improved concision rather than major changes to the research design.

Response: We thank the reviewer for this assessment and agree that the required revisions concern presentation, interpretive discipline, and synthesis rather than a redesign of the empirical study.

The revised manuscript addresses these three areas in combination.

First, the theoretical synthesis has been sharpened by clarifying the progression from architectural semantics and Japanese architectural/aesthetic interpretation to semantic ambiguity and contextual computational analysis. The Introduction now states the architectural and methodological gaps more directly and defines the role of the Japanese terminology as the empirical scope of the study rather than as a claim of unique ambiguity.

Second, the distinction between computational evidence and higher-level interpretation is now explicit throughout the manuscript. The Abstract, Methods, opening of the Discussion, individual term interpretations, Limitations, and Conclusion distinguish between contextual structures recovered by the computational pipeline and architectural formulations developed from those structures, representative corpus contexts, and scholarship. The engawa “modal littoral-circuit” formulation is the clearest example, but the same evidential distinction is now applied to tokonoma, shakkei, shōji, fusuma, and the other terms where appropriate.

Third, the Discussion has been selectively condensed. We removed or shortened repeated computational diagnostics and redundant summaries while retaining representative corpus contexts where they are necessary to justify the architectural interpretations. We did not reduce the Discussion by eliminating those evidence-bearing passages because doing so would make the interpretive claims less transparent rather than more concise in a substantively useful sense.

The tokonoma subsection provides a specific example of this targeted condensation. We removed a paragraph that repeated the same representative contexts already reported immediately above. Those contexts are now presented once, followed by the relevant scholarship and then the architectural synthesis. This removes genuine duplication without deleting the representative evidence or the interpretation of tokonoma’s representational capacity.

Section 4.8 has also been revised to clarify the overall synthesis. “Recurring contextual patterns” now refers to the repeated recovery of differentiated contextual structure within individual terms, particularly where that structure persists across FULL and CLEAN, rather than to one shared ambiguity category across all eight terms. The architectural content of those structures remains term-specific. The revised synthesis therefore avoids treating the eight terms as instances of one general ambiguity type.

Finally, the Limitations and Scope section now makes the methodological boundaries of the study clearer, including the narrow target set, the broad expository corpus, the conservative two-cluster design, and the additional engawa-specific concept-alignment test. These revisions preserve the original research design while making clearer what the analysis establishes, what remains interpretive, and where further validation would be required.

Overall, we have therefore followed the reviewer’s recommendation through sharper theoretical organization, clearer evidential boundaries, targeted condensation, and more explicit methodological limitations, without introducing unnecessary changes to the underlying research design.

Round 2

Reviewer 1 Report

Comments and Suggestions for Authors

Revisions have addressed all my concerns. Discussions and conclusions are presented in much improved and coherent form.

Author Response

Comment 1: Revisions have addressed all my concerns. Discussions and conclusions are presented in much improved and coherent form.

Response: We thank the reviewer for this positive assessment. We are pleased that the revised Discussion and Conclusions address your concerns. In response to comments from another reviewer, we made the following additional changes:

1- We clarified throughout the manuscript that recurring contextual variation is a computational finding, while interpreting that variation as semantic ambiguity requires examination of representative contexts.

2- We stated more explicitly that the corpus contains English-language explanatory material and architectural literature, and clarified the limits this places on claims about Japanese-language usage and specialist practice.

3- We explained that the architectural registers and prompts were selected from prior architectural interpretation and that their alignment results depend on those choices.

4- We described the additional engawa concept-alignment analysis as an exploratory, theory-guided probe applied only to that term, and clarified the limits of its positive similarity scores.

5- We removed duplicated wording in §2.4, streamlined overlapping statements in §4.8, and made further editorial changes to sentence order and clarity.

These additional revisions clarify the scope and presentation of the findings. The corpus, computational procedure, and reported numerical results remain unchanged.

Reviewer 2 Report

Comments and Suggestions for Authors

Thank you for the careful revision. The manuscript is substantially improved, particularly in clarifying the methodological limitations and distinguishing computational findings from architectural interpretation.

A few issues still need attention, especially the distinction between contextual variation and semantic ambiguity, the English-language nature of the corpus, the theory-guided nature of the architectural registers, and the relatively exploratory status of the engawa concept-alignment analysis.

I also recommend a final editorial check, as there are some duplicated and unfinished sentences in the revised text. Addressing these points would further strengthen the manuscript.

Author Response

We sincerely thank the reviewer for carefully reading the revised manuscript and for recognizing improvements in our discussion of methodological limitations and the distinction between computational findings and architectural interpretation. The remaining comments helped us clarify the evidential scope of the analysis and revise passages where our wording suggested a stronger conclusion than the results support.

Comment 1: A few issues still need attention, especially the distinction between contextual variation and semantic ambiguity, the English-language nature of the corpus, the theory-guided nature of the architectural registers, and the relatively exploratory status of the engawa concept-alignment analysis.

Response: We appreciate the reviewer for identifying these four points. We revised the abstract, Methods, Results, Discussion, and Conclusions to distinguish contextual differentiation from semantic ambiguity. Section 2.6 now defines the structure identified by clustering as contextual differentiation. It states that stable clusters, including clusters that persist across FULL and CLEAN, do not by themselves establish distinct meanings or lexical senses. Interpreting a pattern as semantic ambiguity also requires examination of representative contexts and the architectural functions they describe. We revised the synthesis in §4.8 accordingly: the FULL/CLEAN comparison examines the persistence of contextual patterns after removing definitional sentences, while semantic interpretation remains a separate step.

We also made the corpus language and composition explicit in the abstract, Introduction, §2.2, and limitations. The analysis uses a mixed English-language corpus of explanatory sources and peer-reviewed architectural literature. We did not analyze these source types as separate conditions. The revised limitations clarify that the findings concern English-language uses of Japanese-origin terms and cannot establish Japanese-language meanings, translation equivalence, or patterns specific to specialist architectural practice.

Sections 2.1 and 2.7 now explain that the five architectural registers and their prompts were specified from prior architectural interpretation. Although the pipeline applies them after contextual clustering, they are theory-guided analytical categories, not categories discovered independently from the corpus. We also state that the resulting alignments depend on the selected prompts, thresholds, and model pathway.

Finally, we clarified that the additional engawa analysis is a focused, exploratory concept-alignment probe. Section 2.9 describes its preselected boundary-edge, connector-flow, and riverbank probes; §3.6 and the Figure 5 caption identify the analysis as exploratory; and the limitations explain why its evidence cannot be used to rank engawa against terms that did not receive comparable probes. The paper presents positive similarity as compatible with the proposed architectural reading. It does not establish that the probes are unique to engawa, outperform alternatives, or reveal an etymological origin.

Comment 2: I also recommend a final editorial check, as there are some duplicated and unfinished sentences in the revised text. Addressing these points would further strengthen the manuscript.

Response: We thank the reviewer for pointing out these editorial issues. In a separate proofreading pass, we removed a duplicated introductory passage in §2.4 while retaining its explanation of the workflow and reproducibility details. We condensed the §3.1 overview without changing the five analytical stages, reduced repeated descriptions of ma in §4.4 while retaining its representative uses, and reorganized overlapping statements in §4.8 while preserving the eight term-specific architectural readings. We also made the §3.6 heading and Figure 5 caption consistent. These editorial edits changed the order and wording of sentences, not the corpus, analytical procedure, numerical results, cited evidence, or intended architectural interpretations.