{"id":"0bb9b20e-4d19-4f2b-be5b-6463b2099fb3","arxiv_id":"1909.00519","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"With margin-based losses that allow positive triples to score below a bound rather than exactly zero, TransE can encode symmetric and reflexive relations, and a new complex-space variant TransComplEx gives competitive results.","lead":"The paper argues that the well-known limitations of the TransE knowledge graph embedding model, such as an inability to encode symmetric relations, depend on the choice of loss function and can be mitigated by using margin-based losses. It also introduces TransComplEx, a complex-space translation model, and reports competitive link prediction results.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 3 proves only local pair-wise constructions; global feasibility over the full relation graph is not shown and fails for cliques with d+2 entities under condition (b).","rationale":"The reader's weakest_assumption identifies exactly the missing global consistency in Lemma 3, and my analysis confirms that this is the most load-bearing gap in the paper's central argument. The local construction for one pair is correct, but the claim that TransE can encode symmetric relation patterns under condition (b) requires a global assignment for the entire relation graph. I found a concrete counterexample: for d=2, a symmetric relation on a triangle forces three difference vectors to be ±(√3,0), which cannot sum to zero; for general d, any clique of size d+2 is impossible under the equality constraint of condition (b). Therefore the universal claim is not merely unproven—it is false for condition (b). However, the paper's broader thesis that loss functions affect which patterns can be encoded remains plausible, since conditions (c) and (d) with inequalities may allow more patterns, and the experiments provide indirect support. Thus the verdict should remain CONDITIONAL: the paper needs to either prove global feasibility under weakened claims or restrict the pattern class. I agree with the reader's assessment, and my counterexample strengthens the case for revision without overturning the core idea entirely.","tokens_in":15850,"tokens_out":8546,"duration_ms":287158,"concrete_test":"Analytical test: Fix d=2, r=(0,1), α=2, γ1=2. Under condition (b), any positive symmetric pair requires h−t ⊥ r and ‖h−t‖=√3. Take a triangle on entities a,b,c (all three pairs positive). The difference vectors a−b, b−c, c−a must each be ±(√3,0); their sum is zero, but the sum of three values from {±√3} is an odd multiple of √3, never zero—contradiction. Generalize: for dimension d, a clique of size d+2 in the symmetric relation requires d+2 points in R^{d−1} with all pairwise distances equal, impossible. This shows Lemma 3(3) is false for condition (b). Computational check: train TransE with loss (4) on this triangle graph with d=2, r initialized nonzero, and verify whether the margin constraints are satisfied at convergence.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that TransE can encode symmetric (and reflexive/transitive) relation patterns under margin-based losses rests on Lemma 3, but the proof only constructs embeddings for a single entity pair (h,t), showing that h−t must be perpendicular to r with ‖h−t‖ = sqrt(α²−1)‖r‖ when γ1 = α‖r‖ (equations 8–10). It never shows that these constraints can be satisfied simultaneously for all triples in a relation pattern, nor that negative triples can be kept at score ≥ γ2. This gap is load-bearing: for condition (b), the constraints force every linked pair's difference vector onto a sphere of fixed radius in the (d−1)-dimensional subspace orthogonal to r. In a symmetric relation forming a clique of size d+2 (e.g., a triangle in d=2, K4 in d=3), all pairwise differences would need to be equidistant with the same norm, but Euclidean space of dimension d−1 admits at most d equidistant points. Hence no embedding exists, contradicting the unqualified Lemma 3(3) claim. The same local-to-global issue affects Lemmas 4–6, where proofs again exhibit only a few vectors satisfying inequalities without a construction for the entire graph. Therefore the paper's assertion that existing limitation theories are 'inaccurate' is not established by these proofs; at best, some symmetric patterns can be encoded, not the general pattern class.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that the widely cited limitations of TransE in encoding symmetric, reflexive, and transitive relation patterns are not intrinsic to the scoring function but depend on the choice of loss function. The authors define four conditions on positive and negative triple scores: exact equality at zero (a), equality at a positive margin (b), upper-bound at a positive margin (c), and triple-specific upper-bound (d). They then state lemmas claiming that under conditions (b), (c), and (d), TransE can encode reflexive and symmetric patterns, and that limitations L4, L5, and L6 no longer hold. They also propose TransComplEx, a complex-space translation model, and provide experimental comparisons on FB15K, FB15K-237, WN18, and WN18RR using margin-based losses and relation-pattern regularizers.","tokens_in":16156,"tokens_out":8314,"duration_ms":82909,"significance":"The conceptual point is genuinely valuable: prior impossibility proofs for TransE assume that positive triples must satisfy h+r=t exactly, whereas margin-based losses only require the score to lie below or at a positive bound. If the paper's existence claims were fully established, they would correct an overstatement in the existing literature and redirect attention to the joint role of score function and loss function. The paper also contains a useful reformulation of several losses as constrained optimization problems and proposes a simple new model, TransComplEx, with competitive empirical results across four standard datasets. The experimental section is broad and reports hyperparameter ranges and multiple baselines, which is a strength.","major_comments":[{"comment":"Lemma 1(1) is false as stated for TransComplEx under condition (a). From h+r−\\bar{h}=0, the proof derives Re(r)=0 and Im(r)=−2Im(h), but this only fixes the imaginary part of every entity to be −Im(r)/2; the real parts of the entity embeddings are unconstrained. Choosing r=(0,i c) with c≠0 and h=a−i c/2 gives h+r−\\bar{h}=0, so a non-zero relation vector encodes a reflexive relation in TransComplEx. The conclusion that \"all entities will have same embedding vectors\" is not implied by the equations. This error directly affects the lemma's central claim that TransComplEx cannot infer reflexive patterns with non-zero relation vectors.","section":"Section 4.2, Supplementary Lemma 1"},{"comment":"Lemma 3(3) is not established by the given proof. The derivation shows that for a single pair (h,t), the two positive triple constraints imply u=h−t is orthogonal to r and satisfies ‖u‖^2=(α^2−1)‖r‖^2. This is a local necessary condition, not a construction of a global embedding. The proof never shows that the equalities ‖h+r−t‖=γ1 and ‖t+r−h‖=γ1 can be satisfied simultaneously for all pairs in a symmetric relation, nor that all corrupted negative triples satisfy the score constraint ≥γ2. The gap is load-bearing: for a clique of d+1 entities in dimension d, all pairwise differences must be mutually equidistant and lie in the (d−1)-dimensional subspace orthogonal to r, which is impossible (e.g., a triangle in d=2). Thus the unqualified claim that \"TransE can infer a relation pattern which is symmetric with conditions (b)\" is false for general symmetric patterns; at best a restricted family of symmetric relations is demonstrated.","section":"Section 4.2, Eqs. (8)-(10) and Supplementary Lemma 3"},{"comment":"The proofs of Lemmas 4, 5, and 6 exhibit the same local-to-global gap. For example, the proof of Lemma 6 verifies only that a single inequality ‖e2+r−s2‖>γ1 can be consistent with three equalities for a particular four-tuple of entities; it does not construct embeddings for all entities in the relation graph that simultaneously satisfy all positive and negative constraints. Similarly, the proof of Lemma 5 derives an inequality ⟨e2,(e1+e2+e3)⟩<0 as a non-contradiction condition but never shows a global assignment. These lemmas are used to conclude that limitations L4, L5, and L6 \"are not valid\" under conditions (b)-(d), so the missing global consistency argument is essential to the paper's central claim that the previous limitation theories are inaccurate.","section":"Supplementary Lemmas 4-6"},{"comment":"The supplementary statement of Lemma 3(3) says TransE can infer a relation pattern which is \"symmetric and reflexive\" under conditions (b), (c), and (d), while the main-text Lemma 3(3) claims only \"symmetric.\" The proof only treats the symmetric pair constraints and does not show reflexivity. Additionally, the supplementary proof says α≥1 whereas the main proof requires α>1; if α=1 then ‖u‖=0, which collapses the two entities and does not give a non-trivial symmetric encoding. The intended claim and its exact conditions should be stated consistently and proved completely.","section":"Supplementary Lemma 3 statement"}],"minor_comments":[{"comment":"The title contains a typo: \"Then Performance\" should read \"the Performance.\"","section":"Title"},{"comment":"There is a typographical error in the derivation following Eq. (10): \"−(|u‖2+(1−α2)‖r‖2)\" should be \"−(‖u‖2+(1−α2)‖r‖2).\"","section":"Section 4.2, Eq. (10)"},{"comment":"In the proof of Lemma 1, condition (b) is claimed to give ‖r‖=γ1 for TransComplEx; the correct residual is Re(r)+i(Im(r)+2Im(h)), so the norm is not generally ‖r‖. The existence claim may still be true with a different assignment, but the stated derivation is incorrect.","section":"Supplementary Lemma 1"},{"comment":"The subsection heading \"Dissuasion of Results\" should be \"Discussion of Results,\" and the text contains the typo \"PRTransComplEx\" for \"RPTransComplEx.\"","section":"Section 5, Discussion"},{"comment":"The references list two entries, Zhou et al. 2017a and Zhou et al. 2017b, with the same title and venue; it appears the same paper is cited twice with different year labels. The duplicate should be removed and the in-text citations unified.","section":"References"},{"comment":"The algebraic formula listed for the symmetric pattern, RS=‖Re(h)−Re(t)‖=0, appears to omit the relation vector and is not derived in the text; please provide the intermediate steps connecting fr(h,t)=fr(t,h) to that formula.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern is real and lands: the proof of Lemma 3 is local-only, and the clique argument shows the unqualified symmetric-pattern claim cannot hold. However, the paper's core idea is not dead on arrival. The theoretical section can be repaired by narrowing the claims to existence of nontrivial instances under margin conditions, by constructing global embeddings or giving a precise class of encodable symmetric patterns, and by correcting Lemma 1, which is actually false for TransComplEx. The experimental contribution is independent of these fixes and remains useful. I therefore recommend major revision rather than rejection. The novelty claim that this is the first investigation of loss-function effects on these limitations may merit additional literature checking, but that is secondary."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things worth knowing about this paper. First, the core reframing—that the loss function's upper bound on positive scores changes what TransE can encode—is genuinely useful and not something I'd seen done carefully. Second, the paper's headline claim, that proper loss selection mitigates the main limitations of TransE, is not actually proven; the key lemma only checks one pair of entities at a time.\n\nWhat the paper does well: it reformulates margin, soft margin, and ranking losses as constrained optimization problems with explicit upper/lower bounds, which makes the 'region of truth' concrete. That's a nice lens. The TransComplEx model is a simple but reasonable extension, and the experiments show it's competitive on standard benchmarks. The authors also correctly note that prior impossibility proofs (Sun et al., Kazemi and Poole) assume exact equality, condition (a), and that under a margin the residual doesn't have to be zero.\n\nThe soft spots are not minor. The proof of Lemma 3(3) constructs h and t for a single symmetric pair, with u = h−t perpendicular to r and fixed norm. It never shows these constraints can be satisfied for all triples in the relation graph. The stress-test note is right: if the symmetric relation is a clique of size d+2, all pairwise difference vectors would be equidistant in the (d−1)-dimensional subspace orthogonal to r, and Euclidean space of dimension d−1 admits at most d equidistant points. So the theorem as stated is false. Lemmas 4–6 have the same local-to-global gap. The paper also overstates by calling prior theories 'inaccurate'; they are accurate under condition (a), which is the standard setting. The experiments don't directly test pattern encoding, so they can't rescue the proof.\n\nThe framework is worth engaging with, but the current version overclaims. I'd send it to reviewers, with the expectation of major revision: either prove a global version with dimensionality constraints, or restrict the claim to patterns that don't force large cliques. The paper is for people working on expressivity of KGE models; they'll find the loss-centric view useful even after the theorems get fixed.","headline":"Useful reframing of loss functions as shaping a 'region of truth', but the central theorem that margin losses let TransE encode symmetric/reflexive patterns is only proven locally and fails for large symmetric cliques.","tokens_in":16652,"tokens_out":2999,"would_cite":false,"duration_ms":29820,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Loss function, not score function, settles TransE's limits","keywords":["knowledge graph embedding","TransE","loss function","relation patterns","symmetric relations","translation-based models","complex embeddings","link prediction"],"falsifier":"Take a synthetic knowledge graph with a single symmetric relation r and many entity pairs (a_i, b_i) with both (a_i, r, b_i) and (b_i, r, a_i) as positive triples. Train TransE with loss 6 (inside-ball, gamma1 = alpha||r||, alpha > 1). If the theorem is right, the learned relation vector should be non-zero and each pair should satisfy u_i = a_i - b_i orthogonal to r. If the optimal solution collapses to r = 0, or if the margin constraints cannot be satisfied for all pairs simultaneously, the local-to-global extension fails and the central claim is refuted.","tokens_in":15663,"feed_emoji":"🧩","tokens_out":3061,"duration_ms":29753,"temperature":0.7,"pith_summary":"This paper argues that the well-known limitations of the TransE knowledge-graph embedding model—its supposed inability to encode symmetric, reflexive, and transitive relations—are not intrinsic to the scoring function but artifacts of the loss function used in training. The authors show that when the loss imposes only an upper bound on the scores of positive triples (rather than requiring exact equality h+r=t), TransE can represent these relation patterns with a non-zero relation vector. They prove this by defining four 'conditions of truth' for a triple, and they introduce TransComplEx, a translation model in complex space that inherits these benefits and has fewer theoretical limitations. If correct, this reframes a central debate in knowledge-graph embeddings: model capacity depends as much on the loss as on the score function.","feed_headline":"Loss choice, not score function, sets TransE's limits","feed_subtitle":"New proofs show margin-based losses let translation models encode symmetric and reflexive patterns.","key_machinery":"The 'region of truth' is the central object: a hypersphere (or ball) in the score space that defines which triples count as positive. The paper distinguishes four conditions: (a) score exactly zero for positives, (b) score exactly gamma1 on the sphere boundary, (c) score at most gamma1 inside the ball, and (d) triple-specific margins. The loss functions corresponding to these conditions set different upper bounds on positive scores, and the lemmas show that limitations L1-L6 only hold under condition (a), where the region of truth is a single point. The key algebraic step for symmetric relations is deriving the necessity condition ||h + r - t|| = ||t + r - h||, which, after substituting u = h - t, yields ||u||^2 = ($alpha^{2}$ - 1)||r||^2 and forces orthogonality when $\\alpha$ > 1.","core_discovery":"The central claim is that existing proofs of TransE's limitations, which conclude that symmetric or reflexive relations force the relation vector to zero, rely on the strict condition that a positive triple must satisfy h + r = t exactly. Under less restrictive conditions—where a positive triple's score is allowed to lie on or inside a hypersphere of radius gamma1, with negative triples kept above a larger threshold gamma2—TransE can encode symmetric relations with non-zero relation vectors. In particular, Lemma 3 shows that for a symmetric relation, setting the radius to gamma1 = alpha||r|| with alpha > 1 permits a solution where the difference vector u = h - t is orthogonal to r, so entities remain distinct. The paper also proves that the other limitations (L4-L6) fail under these conditions, and it introduces TransComplEx, whose conjugate tail term allows symmetric and neither-reflexive-nor-irreflexive patterns even under the strictest condition.","pith_inferences":["The 'region of truth' perspective generalizes beyond TransE: any distance-based scoring function, including RotatE and TorusE, likely has its relation-pattern limitations modulated by the loss function, so reported 'inherent' limitations should be re-examined with loss in mind.","If loss choice is this influential, then comparisons between embedding models under a single fixed loss are incomplete; fair comparison requires tuning or reporting across the four condition types, which could change leaderboard rankings on benchmark datasets.","A testable extension: on a synthetic graph containing only one symmetric relation and many entity pairs, training TransE with condition (b) should converge to a non-zero relation vector with u orthogonal to r for every pair; observing r collapse to zero would refute the global extension of the local proof.","The paper leaves 1-N, N-1, and M-N relation handling as future work, but the same upper-bound argument suggests that these patterns can also be mitigated by loss design, potentially reducing the need for relation-specific projection matrices."],"forward_implications":["TransE trained with a loss that only upper-bounds positive scores (conditions b or c) can encode symmetric and reflexive relations, so practitioners no longer need to abandon TransE for these patterns.","The same upper-bound argument applies to TransH, TransR, and other translation-based models, meaning their known limitations are also loss-dependent, not purely architectural.","The theoretical comparison predicts that loss 6 (inside-ball) outperforms loss 4 (exact-boundary), which is confirmed by the reported experiments on FB15K and FB15K-237.","TransComplEx, by using the conjugate of the tail vector, can represent symmetric relations even under the strictest condition (a), giving it a theoretical advantage over TransE without added computational cost.","The reformulation of losses as constrained-optimization problems provides a principled way to select margins (gamma1, gamma2) for any distance-based embedding model."],"supporting_citations":[{"why":"Defines TransE's score function and the margin ranking loss, which corresponds to condition (d).","marker":"(Bordes et al., 2013)"},{"why":"Proves TransE cannot encode symmetric relations under the exact condition; this is the claim the paper overturns.","marker":"(Sun et al., 2019)"},{"why":"Proves limitations L4-L6 for TransE and variants under condition (a); these are the limitations the paper shows fail under conditions (b)-(d).","marker":"(Kazemi and Poole, 2018)"},{"why":"Claims TransE cannot encode reflexive and one-to-many/many-to-one relations, a starting point for limitation L1.","marker":"(Wang et al., 2014)"},{"why":"Provides the complex-embedding formulation that inspires TransComplEx's use of conjugate tail vectors.","marker":"(Trouillon et al., 2016)"},{"why":"Introduces the soft-margin loss with an upper bound on positive scores, which is loss 6 in this paper's experiments.","marker":"(Nayyeri et al., 2019)"},{"why":"Provides the limit-based scoring loss and baseline results used for comparison of loss 4.","marker":"(Zhou et al., 2017b)"}],"fun_headline_variants":["Loss function choice, not score, fixes TransE's limits","Margin losses let TransE model symmetric relations","New proofs: loss function sets TransE's true capabilities","Forget score function, loss determines TransE's power","TransE's limitations vanish with proper loss selection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The proofs for encoding symmetric relations construct a solution for a single pair of entities (u perpendicular to r) and assume it can be extended consistently to all entity pairs in the relation graph without violating the margin constraints, but no global construction or argument is provided.","fun_headline_variants_meta":{"raw":{"variants":["Loss function choice, not score, fixes TransE's limits","Margin losses let TransE model symmetric relations","New proofs: loss function sets TransE's true capabilities","Forget score function, loss determines TransE's power","TransE's limitations vanish with proper loss selection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000847,"raw_usage":{"total_tokens":3730,"prompt_tokens":1031,"completion_tokens":2699,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":647,"completion_tokens_details":{"reasoning_tokens":2622}},"tokens_in":647,"tokens_out":2699,"duration_ms":17107,"temperature":1.0,"reasoning_tokens":2622,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:46:29.695107+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a synthetic knowledge graph with a single symmetric relation r and many entity pairs (a_i, b_i) with both (a_i, r, b_i) and (b_i, r, a_i) as positive triples. Train TransE with loss 6 (inside-ball, gamma1 = alpha||r||, alpha > 1). If the theorem is right, the learned relation vector should be non-zero and each pair should satisfy u_i = a_i - b_i orthogonal to r. If the optimal solution collapses to r = 0, or if the margin constraints cannot be satisfied for all pairs simultaneously, the local-to-global extension fails and the central claim is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Proves limitations L4-L6 for TransE and variants under condition (a); these are the limitations the paper shows fail under conditions (b)-(d)."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Claims TransE cannot encode reflexive and one-to-many/many-to-one relations, a starting point for limitation L1."}],"review_version":1}