{"id":"d1183df6-6c39-435f-85dc-44e2df05d176","arxiv_id":"2508.08128","paper_version":3,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"FuzzyVis composes ontology concepts into fuzzy queries and finds similar concepts via membership-vector cosine similarity, without requiring formal syntax.","lead":"FuzzyVis is a prototype system that lets people explore large ontologies by building visual queries from concepts with AND, OR, and NOT, and matching them with fuzzy membership vectors. It aims to make ontology search easier for non-experts who find SPARQL and other formal query languages too difficult.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The α-embedding construction assumes a tree taxonomy, yet the HPO usage scenario uses an ontology the paper itself shows is a DAG, leaving the central demonstration ill-defined.","rationale":"The reader correctly identified the α-embedding's tree assumption and arbitrary α as the weakest point. My stress-test sharpens this: the paper's own example contradicts the tree assumption for HPO, so the specific reported case study is not well-defined. This is load-bearing because the central claim of supporting subtle information needs is supported mainly by that scenario, not by a user study or benchmark. However, the underlying compositional fuzzy-embedding idea is coherent for true tree taxonomies, and the DAG issue is fixable by specifying a deterministic handling of multiple parents and LCAs (or by restricting the demonstration to a tree subset). Therefore the appropriate verdict remains conditional acceptance rather than rejection: the system and math are plausible, but the paper must disambiguate the construction and provide a sensitivity analysis/evaluation. I do not see evidence of fraud or gross error; the concern is about reproducibility and unstated assumptions.","tokens_in":18750,"tokens_out":5158,"duration_ms":64928,"concrete_test":"Reproduce Q1 from Section 5.4 on HPO using two explicit DAG disambiguations: (a) resolve each multi-parent node to one parent uniformly at random, and (b) take the minimum distance over all lowest common ancestors / aggregate fuzzy union over all parents. Compare the top-10 primitive concepts retrieved under each variant and against the results stated in Section 5.4. If the top-k lists change materially, the case study is not reproducible as written; if they are stable, quantify the frequency of multi-parent nodes in HPO to confirm the tree assumption is a harmless simplification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.1 defines α-embeddings under the assumption that \"the ontology O forms a tree-structured taxonomy\" and explicitly lists HPO as an example. However, Section 5.3.1 describes HPO's Cardiac valve calcification as \"a child of both Cardiovascular calcification and Abnormal heart morphology\", so HPO has multiple inheritance and is not a tree. In a DAG the key definitions are underspecified: the graph distance d(C,C') to \"their lowest common ancestor\" is not unique for concept pairs with multiple LCAs, and the recursive fuzzy-union aggregation for internal nodes is ambiguous when a node has several parents (aggregate over all parents, or pick one?). The usage scenario (Section 5.4) nevertheless applies α-embeddings to HPO and reports concrete top-k outputs (Pseudobulbar paralysis, Abnormal esophagus physiology). Since the embedding vectors are not uniquely defined for HPO, those results are not reproducible, and the only demonstration of the central claim rests on an ill-defined construction. This is not merely a missing evaluation: the algorithm as described cannot be run on the demonstrated ontology without additional assumptions.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents FuzzyVis, a proof-of-concept system that combines fuzzy ontology embeddings with a visual query builder for ontology exploration. Primitive concepts are encoded as membership vectors over a finite domain (Eq. 1), and composite queries are formed by element-wise fuzzy conjunction, disjunction, and negation. A synthetic 'α-embedding' construction is proposed for hierarchical ontologies (Section 4.1), using distance-based membership decay from randomly sampled leaves. The front end supports treemap/network visualization, search, highlighting, and drag-and-drop query composition. The central demonstration is a narrative usage scenario on the Human Phenotype Ontology (HPO) that reports top-k results for a composite query such as Slurred speech ⊓ Dysphagia ⊓ ¬Abnormality of the immune system.","tokens_in":19023,"tokens_out":6761,"duration_ms":79587,"significance":"The underlying idea—compositional membership-vector embeddings with on-the-fly fuzzy operations—is coherent for tree-structured taxonomies and could offer a lightweight, syntax-free alternative to formal ontology querying. The interface design is described in substantial detail, and the mathematical formulation of Eq. (1) is clear. However, the paper provides no user study, no comparison against baselines (e.g., keyword search, SPARQL, or other ontology embeddings), no sensitivity analysis for α or sample size, and no reproducibility controls for the randomized α-embedding. More seriously, the only demonstration applies the α-embedding to HPO, which is a DAG, even though Section 4.1 assumes a tree. The central effectiveness claim therefore rests on an ill-defined construction and an illustrative narrative rather than systematic evidence.","major_comments":[{"comment":"Section 4.1 defines α-embeddings under the assumption that 'the ontology O forms a tree-structured taxonomy' and uses 'the graph distance d(C,C′) ... to their lowest common ancestor,' which is unique only in a tree. Section 5.3.1, however, states that HPO's Cardiac valve calcification is a child of both Cardiovascular calcification and Abnormal heart morphology, so HPO is a DAG with multiple inheritance. In a DAG, the lowest common ancestor is not unique and the distance is undefined; the recursive fuzzy-union step also needs a consistent convention for nodes with multiple parents. Section 5.4 nevertheless applies α-embeddings to HPO and reports concrete top-k results (Pseudobulbar paralysis, Abnormal esophagus physiology). These results are not reproducible from the described algorithm. The paper must either extend the construction to DAGs (e.g., define distance via min/max over common","section":"§4.1, §5.3.1, §5.4"},{"comment":"The central claim that FuzzyVis 'supports subtle information needs and helps users uncover relevant concepts' is supported only by a narrative usage scenario. No user study, task-completion metrics, comparison with baseline search/query tools, or ablation of the fuzzy operators is reported. Thus, the paper does not provide evidence that the proposed querying model outperforms keyword search or SPARQL for non-experts, nor that the retrieved concepts are semantically appropriate beyond the authors' own interpretation. I recommend adding a systematic evaluation—even a small user study or a benchmark with ground-truth query relevance—or explicitly reframing the contributions as an illustrative prototype and tempering the effectiveness claims.","section":"§5.4, §6"},{"comment":"The α-embedding construction has at least three free choices that are not analyzed: the decay parameter α (set to 0.25 in §5.4 without sensitivity analysis), the number of sampled domain elements |ΔI| (not reported for the HPO scenario), and the random choice of leaf concepts for each domain element (no seed mentioned). Cosine similarity rankings can change with α and with the random sample, so the specific top-k outputs in §5.4 may not be stable. The manuscript should report these parameters, include a sensitivity analysis, and either fix a random seed or average over repeated samples.","section":"§4.1"}],"minor_comments":[{"comment":"Typo: 'treemps' should be 'treemaps'.","section":"§5.3"},{"comment":"'air way' should be 'airway'.","section":"§1, Example 1"},{"comment":"The phrase 'a typical ontology views' is grammatically awkward; consider rephrasing.","section":"§1"},{"comment":"The statement 'We assume ... tree-structured taxonomy' is immediately followed by 'e.g., HPO [31]' as an example. Since Section 5.3.1 later shows HPO has multiple inheritance, this example should be corrected or qualified.","section":"§4.1"},{"comment":"The claim that the generated interpretation 'satisfies all taxonomic axioms' covers subsumption but does not address disjointness or other non-taxonomic axioms mentioned in Section 3.1. The phrase 'semantically grounded' is therefore stronger than what the α-embedding actually guarantees.","section":"§4"}],"recommendation":"major_revision","confidential_remarks":"The paper reads as a system/prototype description rather than a completed HCI evaluation. The DAG-versus-tree inconsistency is fixable—either by redefining the embedding for DAGs or by changing the demonstration—but it is currently load-bearing, as the only illustrative results are not reproducible as stated. The lack of user evaluation and parameter sensitivity analysis further weakens the central claims. If the authors add a focused evaluation and address the DAG ambiguity, the paper could become publishable; in its current form, the evidence is too thin for a journal-level acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take: the core of the paper is the fuzzy membership vector representation and the compositional query evaluation via element-wise t-norms. That part is new relative to the ontology-embedding literature they cite, and it is a real idea: it lets you build new concepts on the fly with conjunction, disjunction, and negation, and get top-k matches without retraining. The visual interface is described in enough detail that a reader could rebuild it, and the system description (vector DB, Flask, query resolver) is straightforward. Credit where due: this is a competent prototype write-up.\n\nThe soft spots are not small. The α-embedding construction in Section 4.1 explicitly assumes a tree-structured taxonomy, but the usage scenario in Section 5.4 uses HPO, which the paper itself shows is a DAG (Cardiac valve calcification has two parents). The definitions of graph distance and the recursive aggregation are underspecified for a DAG: 'lowest common ancestor' is not unique, and how to handle a node with multiple parents is not stated. That means the reported query results in the scenario are not actually reproducible from the algorithm as written. That is a load-bearing flaw in the demonstration, not just a missing evaluation.\n\nBeyond that, there's no user study, no comparison to keyword search or SPARQL, and no sensitivity analysis for α or the choice of t-norm. The 'effectiveness' claims rest entirely on one narrative example. The formal core is coherent—the equations make sense—but the paper's only evidence that the approach helps real users is anecdotal. So I'd say: the idea is worth publishing in some form, but this draft has a definitional gap where its main example lives.\n\nFor peer review: yes, I would send this to a serious referee. The novelty of the compositional fuzzy embedding is enough to warrant careful review, and the incompleteness on DAGs is exactly the kind of thing a referee should catch. The authors need to either extend the construction to DAGs or re-demonstrate on a tree-structured ontology, and they need some kind of evaluation beyond a walkthrough. A desk reject would be unfair; this is a promising draft with a specific, identifiable fix.","headline":"FuzzyVis has a genuinely neat compositional embedding idea, but its only worked example applies that idea to HPO—which is a DAG—while the construction is defined only for trees, and there is no evaluation beyond a narrative scenario.","tokens_in":19485,"tokens_out":3541,"would_cite":false,"duration_ms":40205,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper introduces fuzzy membership vectors that let users compose ontology concepts with AND, OR, NOT and retrieve semantically similar concepts without formal query syntax.","keywords":["fuzzy ontology embeddings","ontology exploration","visual query building","fuzzy logic","concept similarity","ontology visualization","alpha-embeddings","Human Phenotype Ontology"],"falsifier":"Take HPO, build an alpha-embedding with alpha=0.25, run the query Slurred speech ⊓ Dysphagia ⊓ ¬Abnormality of the immune system, and check whether the top-k list consistently contains known related phenotypes such as Pseudobulbar paralysis across repeated random leaf samples and across alpha values such as 0.1, 0.25, and 0.5; if rankings change drastically or unrelated concepts rank highly, the semantic-similarity claim fails.","tokens_in":18666,"feed_emoji":"🧩","tokens_out":3823,"duration_ms":48793,"temperature":0.7,"pith_summary":"The paper tries to establish that ontology exploration can be made both expressive and beginner-friendly by turning each concept into a fuzzy membership vector and letting users build composite concepts with conjunction, disjunction, and negation. If true, non-experts can issue complex semantic queries by dragging and combining familiar concepts instead of learning SPARQL or memorizing ontology structure. The central mechanism is a vector embedding defined by membership degrees over a fixed domain, with composite queries computed on the fly by element-wise fuzzy operations and answered by cosine similarity against stored primitive-concept embeddings. The paper also proposes a synthetic construction, called alpha-embeddings, for tree-structured ontologies, using graph distance and a decay parameter. A usage scenario in the Human Phenotype Ontology shows a physician finding relevant phenotypes by combining speech and swallowing symptoms while excluding immune abnormalities.","feed_headline":"Drag, combine, search: fuzzy vectors make ontologies queryable","feed_subtitle":"No formal syntax needed—composite queries return semantically similar concepts via membership vectors.","key_machinery":"The $\\alpha$-embedding construction: a synthetic fuzzy interpretation over a tree-structured taxonomy built by random leaf sampling, with membership decaying as $alpha^{{d(C,C')}}$ along graph distance and internal concepts aggregated through fuzzy union. It supplies the fixed domain and membership values that define Eq. 1, guarantees fuzzy satisfaction of taxonomic subsumption axioms, and supports on-the-fly composite query embeddings through element-wise fuzzy operators. Cosine similarity over these vectors is the query-answering mechanism.","core_discovery":"The core claim is that ontology concepts can be represented as membership vectors over a fixed domain, and that user-defined composite concepts built with logical operators can be matched against those vectors to retrieve the most semantically similar primitive concepts. For a fuzzy interpretation I with domain elements x_1,...,x_d, each concept C is embedded as v_C = [mu^I_C(x_1),...,mu^I_C(x_d)] in [0,1]^d. A query Q such as Slurred speech ⊓ Dysphagia ⊓ ¬Abnormality of the immune system gets its embedding by applying the chosen t-norm, t-conorm, and negation element-wise to the primitive embeddings; the system then returns the primitive concepts with highest cosine similarity. This composi","pith_inferences":["A testable extension is to sweep alpha and repeat random leaf sampling on HPO, measuring whether the top-k results for a fixed composite query remain stable; large rank changes would indicate the semantic-similarity claim depends strongly on a parameter chosen without sensitivity analysis.","The same compositional fuzzy-vector scheme could be extended beyond tree-structured taxonomies to roles and non-taxonomic axioms, provided a fuzzy reasoner supplies an interpretation over a fixed domain; the paper lists this as future work.","The membership vectors may double as interpretable concept profiles, allowing users to inspect which domain elements drive a similarity score, which could support explainable ontology exploration.","A direct comparison of retrieval precision and user effort against keyword search and SPARQL baselines would test whether the visual fuzzy-query approach actually reduces vocabulary mismatch for non-experts; the paper demonstrates the scenario but does not measure it."],"forward_implications":["Users can express vague or partially specified information needs as combinations of known concepts and still obtain ranked, semantically similar primitive concepts even when the composite concept does not exist in the ontology.","Composite query embeddings are computed at query time with vectorized fuzzy operations, so new concepts do not require retraining or re-embedding, unlike static ontology embedding methods.","Precomputing primitive-concept embeddings and storing them in a vector database enables near-instant top-k retrieval over large ontologies.","For tree-structured ontologies, alpha-embeddings automatically satisfy subsumption axioms and give users a tunable parameter, alpha, that controls smoothness and generalization of the semantic space.","The visual interface allows users to build these queries by dragging and dropping concepts and nesting queries, lowering the syntactic barrier compared to SPARQL and similar formal languages."],"supporting_citations":[{"why":"Supplies the Human Phenotype Ontology used as the target domain in the usage scenario and the specific primitive concepts in the example queries.","marker":"[31]"},{"why":"Provides the foundational fuzzy description logic semantics of membership degrees and graded concept membership on which the embeddings rest.","marker":"[57]"},{"why":"Supplies the continuous t-norms, t-conorms, and negation operators used to interpret conjunction, disjunction, and negation in the composite query embeddings.","marker":"[30]"},{"why":"Offers the FALCON fuzzy reasoner that can produce fuzzy interpretations whose sample size determines the embedding dimension.","marker":"[59]"},{"why":"Represents the translation-based embedding family that the paper contrasts with its compositional approach.","marker":"[12]"},{"why":"Frames the survey of ontology embedding methods and their limitation of static concept sets, motivating the paper's on-the-fly composition. ","marker":"[19]"},{"why":"Names the Chroma vector database used in the FuzzyVis back-end to store primitive embeddings and perform fast top-k retrieval.","marker":"[20]"},{"why":"Provides the OwlReady2 library used to parse and store uploaded ontologies in the back-end.","marker":"[42]"},{"why":"Illustrates a geometric embedding approach that can approximate conjunction but does not support negation or query-time disjunction, highlighting the gap the fuzzy method fills.","marker":"[41]"}],"fun_headline_variants":["Visual query builder uses fuzzy embeddings to explore ontologies","Fuzzy vectors and visual logic make ontology search intuitive","Build composite concepts visually, match them with fuzzy embeddings","No formal syntax: fuzzy ontology embeddings enable expressive queries","Explore ontologies by combining concepts with fuzzy logic and visuals"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The load-bearing premise is that for a tree-structured ontology, sampling leaves uniformly at random and assigning membership by graph-distance decay $\\alpha$^d produces vectors whose cosine similarity reflects true semantic relevance, with the decay parameter $\\alpha$=0.25 chosen in the demonstration without a reported sensitivity analysis.","fun_headline_variants_meta":{"raw":{"variants":["Visual query builder uses fuzzy embeddings to explore ontologies","Fuzzy vectors and visual logic make ontology search intuitive","Build composite concepts visually, match them with fuzzy embeddings","No formal syntax: fuzzy ontology embeddings enable expressive queries","Explore ontologies by combining concepts with fuzzy logic and visuals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000217,"raw_usage":{"total_tokens":1293,"prompt_tokens":785,"completion_tokens":508,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":529,"completion_tokens_details":{"reasoning_tokens":431}},"tokens_in":529,"tokens_out":508,"duration_ms":5577,"temperature":1.0,"reasoning_tokens":431,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T21:36:46.010070+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take HPO, build an alpha-embedding with alpha=0.25, run the query Slurred speech ⊓ Dysphagia ⊓ ¬Abnormality of the immune system, and check whether the top-k list consistently contains known related phenotypes such as Pseudobulbar paralysis across repeated random leaf samples and across alpha values such as 0.1, 0.25, and 0.5; if rankings change drastically or unrelated concepts rank highly, the semantic-similarity claim fails.","supporting_citations":[{"cited_title":"The human phenotype ontology in 2024: phenotypes around the world","cited_arxiv_id":null,"evidence_quote":"Supplies the Human Phenotype Ontology used as the target domain in the usage scenario and the specific primitive concepts in the example queries."},{"cited_title":"A fuzzy description logic","cited_arxiv_id":null,"evidence_quote":"Provides the foundational fuzzy description logic semantics of membership degrees and graded concept membership on which the embeddings rest."},{"cited_title":"Fuzzy description logics and t-norm based fuzzy logics","cited_arxiv_id":null,"evidence_quote":"Supplies the continuous t-norms, t-conorms, and negation operators used to interpret conjunction, disjunction, and negation in the composite query embeddings."},{"cited_title":"FALCON: Scalable Reasoning over Inconsistent ALC Ontologies","cited_arxiv_id":"2208.07628","evidence_quote":"Offers the FALCON fuzzy reasoner that can produce fuzzy interpretations whose sample size determines the embedding dimension."},{"cited_title":"Translating embeddings for modeling multi- relational data","cited_arxiv_id":null,"evidence_quote":"Represents the translation-based embedding family that the paper contrasts with its compositional approach."},{"cited_title":"Ontology embedding: a survey of methods, applications and resources","cited_arxiv_id":null,"evidence_quote":"Frames the survey of ontology embedding methods and their limitation of static concept sets, motivating the paper's on-the-fly composition."},{"cited_title":"Chroma (chromadb): an open-source embed- ding/vector database, 2025","cited_arxiv_id":null,"evidence_quote":"Names the Chroma vector database used in the FuzzyVis back-end to store primitive embeddings and perform fast top-k retrieval."},{"cited_title":"Owlready: Ontology-oriented programming in python with automatic classification and high level constructs for biomedical ontologies","cited_arxiv_id":null,"evidence_quote":"Provides the OwlReady2 library used to parse and store uploaded ontologies in the back-end."},{"cited_title":"EL Embeddings: Geometric construction of models for the Description Logic EL ++","cited_arxiv_id":"1902.10499","evidence_quote":"Illustrates a geometric embedding approach that can approximate conjunction but does not support negation or query-time disjunction, highlighting the gap the fuzzy method fills."}],"review_version":1}