acceptodds
Under review as a conference paper at ICLR 2027

Confidence-Conditioned Slot Composition for Structured Retrieval under Noisy Grounding

Abstract

Reliability-weighted fusion has been widely used, from products of experts (Hinton, 2002) and gated multimodal units (Arevalo et al., 2017) to trusted multi-view classification (Han et al., 2021). Yet, how per-slot confidence should enter a structured retrieval query, and whether it should be used during training, is rarely examined. In this work, we systematically study CCSC (Confidence-Conditioned Slot Composition) for intent-to-knowledge-graph retrieval, comparing eight alternative composers over 37 matched runs on five benchmarks built from BINS, CoDEx-M, WN18RR, and FB15k-237. Synthetic confidence that perfectly separates correct from wrong slots isolates how confidence is used from how well it is estimated. Our central finding is that a simple design, CCSC, which adds slot confidence as a log-bias to query-side attention logits, consistently improves retrieval: it achieves higher mean reciprocal rank (MRR) than role-wise matching, pooling, SumMLP, and learned scalar weighting on all five benchmarks, and than a model trained without confidence on all five when both receive the same confidence. Ablations trace the gain to two factors: (1) training with confidence in the attention logits, since a confidence-only variant already recovers most of the gain, and (2) learned content scores, which help further when correct and wrong slots are mixed. Notably, we find that a model trained with confidence should also receive it at deployment, as removing it reverses its advantage over the confidence-free model on four benchmarks. We release code and checkpoints to facilitate future research.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
structured retrieval, knowledge graph retrieval, slot attention, slot composition, confidence-aware learning, contrastive learning, noisy grounding, shared projection

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_LBxyOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies downstream structured query composition for intent-to-knowledge-graph retrieval. Given grounded action, object, and target slots with confidence scores supplied by an upstream system, the model retrieves a complete triple from a fixed candidate catalog. The proposed CCSC method combines MLP-based slot content scores with log-confidence biases, applies softmax to obtain slot weights, and uses role-aligned concatenation and a shared projection to construct retrieval representations. Training uses a contrastive retrieval objective with entropy regularization.

Experiments cover two BINS-derived panels and three known-fact catalogs constructed from existing knowledge graphs. Grounding errors are simulated by replacing or removing slots. CCSC outperforms the evaluated confidence-free composition baselines, although its additional gains over GateOnly, which receives the same confidence signal, are small and condition-dependent. The paper also examines confidence use during training and inference, temperature sensitivity, and confidence transformations.

Critical strengths
  1. Clear task formulation and method specification. The paper explicitly separates upstream grounding from downstream retrieval and defines the correspondence between semantic roles and triple positions. The method is relatively simple to implement.

  2. Informative controlled comparisons. GateOnly, NoConf, and the training/inference confidence grid help distinguish the benefit of supplied confidence from the additional contribution of learned content weighting.

  3. Candid reporting of limitations and unfavorable results. The authors disclose the gold-dependent construction of confidence and the use of candidate facts seen during training. They also report conditions under which GateOnly or separate projections perform better.

Critical weaknesses
  1. Limited methodological novelty and insufficient added value beyond a combination of simple components. The method primarily combines per-slot MLP scoring, confidence weighting, softmax, ordered concatenation, shared projection, and contrastive learning. Suppressing unreliable inputs based on their confidence is a well-established idea, and the log-confidence bias is equivalent to normalized multiplicative weighting. Simplicity is not itself a weakness, but the paper does not establish a sufficiently novel mechanism, explanation, or substantial practical benefit to justify this combination as a research contribution. In particular, under the same confidence signal, the MRR differences between CCSC and GateOnly range from only −0.0002 to +0.0052, leaving the necessity of the full learned architecture insufficiently established.

  2. The comparisons do not adequately establish competitiveness. Table 1 mainly compares elementary composition strategies, including role-wise cosine matching, average pooling, SumMLP, and global scalar weighting, rather than establishing competitiveness against relevant existing retrieval approaches. Moreover, these baselines do not use the external confidence supplied to CCSC, so the reported improvements partly reflect access to additional reliability information. GateOnly is an important control, but strong simple alternatives directly matched to the setting—such as confidence-weighted role-wise matching and candidate filtering using high-confidence slots—are still missing.

  3. The evaluation is highly controlled and remains substantially disconnected from the practical motivation. Correct slots always receive confidence 1, whereas incorrect slots receive values in [0.3, 0.6], perfectly separating correct from incorrect inputs. The paper does not evaluate actual grounder outputs or systematically examine partially informative confidence, overlapping confidence distributions, or confidently incorrect slots. The evaluation primarily measures recovery from artificially corrupted fields within known candidate catalogs. Although this is consistent with the stated experimental task, it provides insufficient evidence that the method addresses reliability challenges in practical intent-to-KG retrieval.

Questions for authors

What is the main methodological novelty of CCSC beyond applying existing confidence weighting and gating mechanisms to intent-to-knowledge-graph retrieval? Since GateOnly achieves most of the performance gains, what evidence shows the added value of the full CCSC design?

2 comments
AuthorsOpenReview ↗

What CCSC brings to the table

Thank you for asking us to clarify the value of the complete design.

Q1. What is the main methodological novelty of CCSC beyond applying existing confidence weighting and gating mechanisms to intent-to-knowledge-graph retrieval?

CCSC combines learned content and supplied reliability in a role-aligned representation before a shared query–candidate projection. We compare it with GateOnly, which uses confidence-only weights and a learned projection, and with direct matching on the same inputs. NoConf retains the content scorer and shared projection but is trained without confidence; ON/OFF supplies/omits it at evaluation.

On CoDEx-M with 50% of present positions selected for within-role confidence reassignment, CCSC exceeds GateOnly by 0.0064±0.00140.0064 \pm 0.0014 MRR and NoConf receiving the same evaluation confidence by 0.0040±0.00040.0040 \pm 0.0004 (paired mean ± SD over three backbone means). Both differences are positive in all three means. Under submitted ties, it also exceeds squared-confidence cosine by 0.0190 and match-priority by 0.0117. These matched-input comparisons show additional retrieval utility of the fitted CCSC procedure in this condition.

At this level, one CoDEx-M relation-slot realization has correctness AUROC 0.7502. CCSC's own MRR decreases from 0.7303 to 0.7195. Reassignment changes only evaluation confidence, retaining the observations, masks, candidates and fitted models on previously evaluated panels.

Appendix B.1 gives the normalized-product equivalence. In Figure 3, squaring confidence preserves its ordering and improves synthetic Brier/ECE by construction, yet lowers fixed-CCSC mean MRR on all five panels. We propose revising the abstract's attribution of gains to match the comparisons of fitted procedures under common evaluation signals in Section 3.3.

Q2. Since GateOnly achieves most of the performance gains, what evidence shows the added value of the full CCSC design?

The submitted mean CCSC−GateOnly differences range from -0.0002 to +0.0052 MRR; Table 7 details their dependence on input correctness, including cases favoring GateOnly.

Under the original confidence and submitted tie rules, the table reports CCSC minus comparator (paired mean ± SD over three graph backbone means or five BINS repeats conditional on one backbone):

PanelSquared cosineMatch-priority
BINS curated−0.0070±0.0060-0.0070 \pm 0.0060−0.0080±0.0061-0.0080 \pm 0.0061
BINS narrow+0.0015±0.0104+0.0015 \pm 0.0104+0.0015±0.0104+0.0015 \pm 0.0104
CoDEx-M−0.0045±0.0029-0.0045 \pm 0.0029−0.0249±0.0030-0.0249 \pm 0.0030
WN18RR+0.0155±0.0009+0.0155 \pm 0.0009−0.0117±0.0018-0.0117 \pm 0.0018
FB15k-237−0.0115±0.0027-0.0115 \pm 0.0027−0.0335±0.0021-0.0335 \pm 0.0021

Squared cosine weights role-wise similarity by squared confidence. Match-priority places candidates matching all confidence-1 roles first, then uses squared cosine within each group. These full-catalog measurements are target-rank-equivalent to filtering with correct confidence-1 roles, identical scores/ties and full-catalog fallback for no selected role. Match-priority also leads on all three graphs under expected tie credit (CCSC differences -0.0321/-0.0156/-0.0412).

With naturally predicted roles on narrow BINS using frozen RoBERTa, the AP gains are 0.0191±0.01200.0191 \pm 0.0120 over GateOnly and 0.0063±0.02100.0063 \pm 0.0210 over NoConf-ON. GateOnly differences are positive in all five fitted states; NoConf differences have three positive and two negative states. NoConf-ON leads in the other three curated/narrow × TF–IDF/RoBERTa conditions. Curated/RoBERTa also favors linear matching by 0.0284 AP. These evaluations use 512 all-role-present development queries per condition from the historical grounder-training pool, with states sharing predictor outputs and queries; they are not independent generalization tests.

At 50% reassignment, mean MRR favors CCSC over GateOnly on the three graph panels and GateOnly on both BINS panels. In mean MRR, NoConf-ON leads by 0.0133/0.0215 on curated/narrow BINS and about 0.0005 on WN18RR/FB15k-237; CoDEx-M favors CCSC. These differences limit a general preference for fitting with confidence. At 75%, all five panels favor NoConf-OFF over CCSC-ON; at 100%, all favor NoConf-ON. Even the natural query-averaged gain can depend on aggregation: curated/TF–IDF CCSC−GateOnly changes from +0.0140 AP to -0.0019 under equal group weights.

In a separate single-state synthetic curated-BINS test with the object role missing, GateOnly had higher MRR@10 and lower latency. The comparisons do not establish every component's necessity or deployment superiority.

At 50% reassignment on CoDEx-M, the fitted CCSC procedure achieves higher mean MRR than GateOnly and NoConf-ON under the same evaluation confidence. We hope this clarifies the added value of the complete design in this tested setting.

Thank you for the additional experiments. They show conditional gains, but do not resolve my concerns about methodological novelty or establish sufficient added value over simpler alternatives. Several broader weaknesses in my review also remain insufficiently addressed. I therefore maintain my score of 2.

Reviewer_yT3POpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposed a method called Confidence-Conditioned Slot Composition(CCSC) for structured retrieval under noisy grounding.Specifically, the authors represent the user intent using multiple slots, with each slot assigned a confidence score. They then combine the content relevance and confidence of each slot to determine the weight of different slots, furtherly constructing the final retrieval query representation. And then they use this representation to retrieve the corresponding triple from a knowledge graph. Additionally, they test this method on five benchmarks under different settings, which shows that the confidence information can improve the retrieval performance. And plenty of ablation experiment also support this idea.

Critical strengths
  1. The limitations, including synthetic confidence and the known candidate bank, are clearly stated. The paper also reports unfavorable results. This helps readers understand the conditions under which the findings apply.​
  2. The experiments give a fairly detailed picture of incorrect and missing slots, and of confidence use during training and testing. Table 3 is particularly useful because it compares different training procedures under the same test-time confidence conditions.
Critical weaknesses
  1. GateOnly already achieves MRR close to that of CCSC. The mean MRR difference in Table 2, measured as CCSC minus GateOnly, ranges from −0.0002 to +0.0052. GateOnly also uses confidence and learns the projection. I would like to see comparisons with confidence-weighted cosine matching and candidate filtering based on high-confidence slots. These are direct ways to use the available information. Without these comparisons, I am still not sure how much the learned composer adds.​
  2. The findings are still fairly limited in scope. The experiments use three grounded slots and a known candidate bank, with confidence that perfectly separates correct and incorrect slots in the main setting. The comparisons of confidence use during training and testing are useful, but it is unclear how much they tell us about retrieval when confidence is only partly reliable.
Questions for authors

The following additional evaluations could affect my assessment.​

  1. Could the authors add confidence-weighted cosine matching and high-confidence slot filtering baselines, using the same supplied confidence? In the current setup, what fraction of test queries have a unique target when matching only the slots with confidence 1?​
  2. Could the key comparisons in Tables 2 and 3 be repeated when the confidence distributions for correct and incorrect slots overlap but remain informative? The existing transformations and full permutation do not answer this intermediate case.
2 comments
AuthorsOpenReview ↗

Putting confidence to the test that matching and overlap

Thank you for these concrete comparison requests. We report the matched results below.

Q1. Could the authors add confidence-weighted cosine matching and high-confidence slot filtering baselines, using the same supplied confidence? In the current setup, what fraction of test queries have a unique target when matching only the slots with confidence 1?

We compare linear and squared-confidence role-wise cosine on the same observations, confidence and candidate order as CCSC. Priority puts candidates matching every present confidence-1 role first, then ranks within each group by squared-confidence cosine. The table gives CCSC minus comparator in mean MRR under the submitted tie conventions.

PanelUnique targetLinearSquaredPriority
BINS curated46.32%-0.0054-0.0070-0.0080
BINS narrow58.69%+0.0023+0.0015+0.0015
CoDEx-M57.82%+0.0131-0.0045-0.0249
WN18RR72.83%+0.0556+0.0155-0.0117
FB15k-23760.63%+0.0043-0.0115-0.0335

“Unique target” means exactly one candidate matches all selected roles; no-selected-role queries remain in the denominator. The narrow BINS squared/Priority differences have paired SD 0.0104.

We report Priority using full-catalog ranking. In the original synthetic setting, confidence-1 roles are correct and the target is in the catalog. Restricting candidates to the matching group gives the same target rank with the same within-group scores and tie rule: every candidate above or tied with the target belongs to that group. With no selected role, both retain the full catalog and use squared-confidence scoring. This equivalence does not extend to erroneous confidence-1 roles that exclude the target.

Two additional input-slot rules use selected-role average cosine or ID agreement, giving zero scores for empty selection. Their panel MRRs coincide. CCSC leads on the three graphs and trails on BINS, where tied-first credit gives both rules MRR 1.0000. Expected reciprocal rank under uniformly random tie orders, applied to both methods, favors CCSC on curated/narrow BINS by 0.0909/0.0803. Priority still leads on CoDEx-M/WN18RR/FB15k-237 under expected credit (CCSC differences -0.0321/-0.0156/-0.0412). The tie analysis therefore does not overturn the strong matching comparator's graph advantages.

Q2. Could the key comparisons in Tables 2 and 3 be repeated when the confidence distributions for correct and incorrect slots overlap but remain informative? The existing transformations and full permutation do not answer this intermediate case.

We selected 50% of present positions within each role for confidence reassignment, preserving marginal values while changing their association with correctness. In one CoDEx-M relation-slot realization, 735 wrong slots receive confidence 1 and 735 correct slots receive lower values; correctness AUROC is 0.7502. The selection fraction is neither AUROC nor the fraction of values actually changed. Each model keeps its fitted parameters; only evaluation confidence changes.

The table shows paired mean MRR differences ± SD across three graph backbone means or five BINS repeats conditional on one backbone. ON supplies the same overlapping confidence to both methods. The last column is both models OFF, reproduced from submitted Table 3.

PanelCCSC−GateOnly, ONCCSC−NoConf, ONCCSC−NoConf, both OFF
BINS curated−0.0024±0.0029-0.0024 \pm 0.0029−0.0133±0.0022-0.0133 \pm 0.0022−0.0303±0.0054-0.0303 \pm 0.0054
BINS narrow−0.0059±0.0033-0.0059 \pm 0.0033−0.0215±0.0032-0.0215 \pm 0.0032−0.0356±0.0043-0.0356 \pm 0.0043
CoDEx-M+0.0064±0.0014+0.0064 \pm 0.0014+0.0040±0.0004+0.0040 \pm 0.0004+0.0030±0.0003+0.0030 \pm 0.0003
WN18RR+0.0009±0.0012+0.0009 \pm 0.0012−0.0005±0.0005-0.0005 \pm 0.0005−0.0013±0.0010-0.0013 \pm 0.0010
FB15k-237+0.0053±0.0013+0.0053 \pm 0.0013−0.0005±0.0006-0.0005 \pm 0.0006−0.0030±0.0009-0.0030 \pm 0.0009

CCSC was trained with confidence and NoConf without it. The common-OFF results are unchanged because neither model uses the reassigned confidence. We compare fitted procedures under common evaluation conditions; this does not isolate the training-time effect of confidence.

At this overlap level, CCSC exceeds GateOnly in mean MRR on the three graphs, while both BINS panels favor GateOnly. Against NoConf-ON, CCSC has higher mean MRR on CoDEx-M, larger deficits on BINS, and small negative differences on the other two graphs. On CoDEx-M, the +0.0064 GateOnly contrast and the +0.0040 NoConf-ON contrast are positive in all three backbone means.

We also compare CCSC-ON with NoConf evaluated without confidence: at 75% reassignment, CoDEx-M CCSC-ON is 0.0018 above NoConf-ON but 0.0062 below NoConf-OFF. At that level all five panels favor NoConf-OFF over CCSC-ON; at 100%, all five favor NoConf-ON. We propose adding the fixed-model 50% reassignment results alongside Table 3 in Section 3.3, including cases favoring NoConf-ON and CCSC's gains on CoDEx-M.

Thank you for the additional experiments and explanations. I now have a better understanding of this paper; however, I still believe that the novelty and completeness of this paper do not meet the standard for a score of 3, and I will maintain my score at 2.

Reviewer_i3rjOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies how to compose three role slots (action, object, target) into a query that retrieves a triple from a fixed candidate set. Each slot has a frozen TransH embedding and a confidence score. CCSC adds log-confidence to per-slot attention logits before weighting and concatenating the role vectors. Slot corruption is synthetic, and the confidence is also synthetic and perfectly separates correct from wrong slots. Under this setup, CCSC slightly outperforms composers that do not use confidence on BINS and three KG-derived catalogs.

Critical strengths
  1. Transparent experiments. Runs are matched, and post-hoc choices and limitations are reported openly.
Critical weaknesses
  1. Simple, but not effective beyond the signal it is given. The method amounts to multiplicative confidence weighting (Appendix B.1). That would be fine if the learned composer added something, but it does not: in Table 1 only CCSC receives the confidence that marks the wrong slots, and GateOnly, which gets the same confidence without the learned content scores, performs almost the same (Table 2).
  2. Artificial setting. Correct slots always receive full confidence and wrong slots a lower, gold-derived value, so confidence perfectly separates them, and a simple threshold would already remove every wrong slot. Slots come directly from gold triples with random corruption rather than from a real grounder, so systematic errors never appear. The candidate sets are either tiny (BINS) or made of triples seen in training (KG catalogs).
  3. Cherry-picked evaluation. FB15k-237 was adopted as a benchmark only after its results were observed (Section 3.1 and Limitations), and the later controls and stress tests are post-hoc analyses on panels whose results were already known. Choosing benchmarks after seeing results undermines the claim that CCSC "consistently improves retrieval".
  4. Expected findings. The second main finding, that removing confidence at test time hurts a model trained to rely on it, would hold for any model that loses an input it was trained on.
  5. Defensive writing. Much of the text hedges and pre-empts objections ("Mean orderings are descriptive, not claims of statistical significance on every panel", "this is a post-hoc description, not a rule for selecting future tasks") instead of stating what was found, so the contribution is hard to identify.
Questions for authors
  1. How does CCSC perform with realistic confidence, or with synthetic confidence whose correct and wrong distributions overlap?
  2. How does it compare with simple baselines such as confidence thresholding or confidence-weighted cosine?
  3. Which other benchmarks were run but not reported, and how was FB15k-237 selected?
1 comment
AuthorsOpenReview ↗

A closer look at confidence, baselines, and benchmark choices

Thank you for your questions and sharp critique.

Q1. How does CCSC perform with realistic confidence, or with synthetic confidence whose correct and wrong distributions overlap?

We evaluated fixed models with partially reassigned confidence and with naturally predicted roles. NoConf-ON/OFF denotes the model trained without confidence, evaluated with/without it. At 50% within-role reassignment on CoDEx-M, CCSC exceeds GateOnly by 0.0064±0.00140.0064 \pm 0.0014 MRR and NoConf receiving the same evaluation confidence by 0.0040±0.00040.0040 \pm 0.0004 (paired mean ± SD across three backbone means; all three differences are positive for each comparison). Reassignment preserves marginal confidence values while changing their association with correctness. In one CoDEx-M relation-slot realization, 735 of 2,083 wrong slots receive confidence 1 and correctness AUROC is 0.7502. The 50% selection fraction is not AUROC or the fraction of values actually changed.

On narrow BINS with frozen RoBERTa-predicted roles, CCSC/GateOnly/NoConf-ON obtain 0.4610/0.4419/0.4547 average precision (AP). CCSC−GateOnly is positive in all five fitted states; CCSC−NoConf-ON has three positive and two negative states (SD 0.0210). Across the four curated/narrow × TF–IDF/RoBERTa conditions, CCSC exceeds GateOnly by 0.0022–0.0191 mean AP, but NoConf-ON leads in the other three conditions. Each development condition uses 512 all-role-present queries from the historical grounder-training pool. Fitted states share predictor outputs and queries; independent generalization remains untested.

Table 3 compares both fitted procedures under the same evaluation confidence, not just CCSC with its signal removed. CCSC leads on all five means with original confidence ON; both-OFF evaluation reverses four orderings.

Q2. How does it compare with simple baselines such as confidence thresholding or confidence-weighted cosine?

Squared cosine weights role-wise cosine by squared confidence. Priority ranks candidates matching every confidence-1 role first, then uses squared cosine within each group. Entries are CCSC minus comparator under submitted ties (paired mean ± SD over three graph backbone means or five BINS repeats conditional on one backbone).

PanelSquared cosinePriority
BINS curated−0.0070±0.0060-0.0070 \pm 0.0060−0.0080±0.0061-0.0080 \pm 0.0061
BINS narrow+0.0015±0.0104+0.0015 \pm 0.0104+0.0015±0.0104+0.0015 \pm 0.0104
CoDEx-M−0.0045±0.0029-0.0045 \pm 0.0029−0.0249±0.0030-0.0249 \pm 0.0030
WN18RR+0.0155±0.0009+0.0155 \pm 0.0009−0.0117±0.0018-0.0117 \pm 0.0018
FB15k-237−0.0115±0.0027-0.0115 \pm 0.0027−0.0335±0.0021-0.0335 \pm 0.0021

Linear confidence-weighted cosine favors CCSC on four panels; curated BINS is the exception. Two further rules select confidence-1 input roles and score the full catalog by average cosine or ID agreement, using zero scores for empty selection. Their panel MRRs coincide. CCSC leads on the three graphs by 0.0571–0.0704 MRR and trails on BINS, whose tied-first credit gives both rules MRR 1.0000. Expected reciprocal rank over uniformly random tie orders, applied to both methods, favors CCSC on BINS by 0.0909/0.0803. Priority still leads on all three graphs under that metric.

At 50% reassignment on CoDEx-M, CCSC−squared cosine becomes +0.0190 and CCSC−Priority +0.0117 MRR under submitted ties, while CCSC itself falls from 0.7303 to 0.7195. At 50%, CCSC exceeds GateOnly in mean MRR on the three graphs; both BINS panels favor GateOnly. In mean MRR, NoConf-ON leads by 0.0133/0.0215 on curated/narrow BINS and about 0.0005 on WN18RR/FB15k-237; CCSC leads by 0.0040 on CoDEx-M. At 75%, all five panels favor NoConf-OFF; at 100%, all five favor NoConf-ON.

Q3. Which other benchmarks were run but not reported, and how was FB15k-237 selected?

In our documented submission lineage (August 11–September 27, 2026), the submitted whole-triple graph protocol was evaluated on CoDEx-M, WN18RR, and FB15k-237. All three are reported in the paper.

Earlier development included BINS grounding-feasibility studies that stopped before downstream CCSC evaluation and a 2,048-candidate CoDEx-M catalog. Masked-component studies on the three graphs addressed different prediction tasks and did not meet their predefined continuation or method-order reproducibility criteria.

FB15k-237 extended the CoDEx-M/WN18RR protocol. We specified its candidate catalog, methods, learning-rate search, and confirmation procedure before execution. We decided on manuscript inclusion after inspecting the results, as disclosed in §3.1.

We evaluated squared-confidence-weighted cosine ("Squared cosine" in Q2) before submission but did not report it, leaving the same-confidence comparison incomplete. Baseline/CCSC MRR is 0.734792/0.730303 on CoDEx-M, 0.830576/0.846046 on WN18RR, and 0.773666/0.762209 on FB15k-237. Q2 reports the paired differences. We propose adding this baseline to the main table and revising the abstract and conclusion to reflect conditional advantages and counterexamples.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of manuscript revisions

We revised the abstract and conclusion to clarify where CCSC and simpler methods each perform better. Table 1 (p. 5) now reports squared-confidence-weighted cosine and match-priority ranking alongside CCSC, all using the same confidence inputs. Table 3B (p. 6) reports comparisons across all five panels under partial within-role confidence reassignment, with the fitted models held fixed (Section 3.3, p. 7).

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.