acceptodds
Under review as a conference paper at ICLR 2027

Disambiguating Semantic Units for EEG-to-Text Reconstruction

Abstract

EEG-to-text decoding aims to recover linguistic semantics from non-invasive neural signals. However, reliable decoding requires not only fluent generation, but also identifiable semantic evidence and its compositional structure. We reveal a previously underexplored failure mode in fine-grained EEG semantic decoding: semantic ambiguity arises at two distinct scales. Within a sample, shared sentence context dominates query-specific information, causing different semantic queries to converge toward similar representations; across samples, highly central semantic candidates repeatedly attract unrelated EEG queries, inducing systematic overmatching. These observations suggest that reliable semantic decoding requires jointly resolving context-induced convergence and hubness before composing the recovered evidence into language. Based on this finding, we propose Neural conteXtual Unified Semantics (NEXUS), a unified framework that disentangles shared contextual semantics from query-specific evidence, calibrates biased semantic matching from both candidate and query perspectives, and recomposes the recovered semantic units using delexicalized structural priors. Experiments on heterogeneous EEG datasets spanning natural reading and imagined across languages speech demonstrate that NEXUS substantially reduces within-sample representation convergence and cross-sample overmatching, leading to more discriminative semantic-unit identification and more faithful EEG-to-text reconstruction.

Primary area:
applications to neuroscience & cognitive science
Keywords:
Brain–Computer Interface, EEG-to-Text Decoding, Semantic Alignment, Representation Collapse

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

This paper proposes NEXUS, an EEG-to-text framework which first retrieves semantic units and similar sentences from EEG, and then uses a frozen LLM to reconstruct text. The authors identify two sources of semantic ambiguity in unit retrieval: within and cross-sample ambiguity. They address them by separating shared contextual semantics from query-specific evidence and calibrating biased semantic matching.

Critical strengths
  1. The paper addresses EEG-to-text decoding, an important and challenging problem for non-invasive BCI.
  2. The experiments cover three datasets that span two languages (English and Chinese) and two paradigms (natural reading and imagined speech).
Critical weaknesses
  1. The writing of this paper is very poor and difficult to follow.
  2. Standard metrics such as BLEU and ROUGE are missing, which makes comparison with prior work difficult. The experiment does not compare with latest published EEG-to-text works, such as [1]. In addition, the paper shows no examples of decoded sentences. Together, these limitations make the reported decoding performance less convincing.
  3. The hubness calibration is similar to QB-Norm [2], which makes the contribution seem incremental. Please specify the differences. Besides, Its gain in retrieval accuracy is also marginal (Table 2(b)).
  4. The inference gallery appears to be dataset-wide. Could the retrieved sentences include the ground truth?

[1] Assembling the Mind’s Mosaic: Towards EEG Semantic Intent Decoding.

[2] Cross Modal Retrieval with Querybank Normalisation.

Questions for authors

Please see weaknesses.

3 comments

We thank the reviewer for the constructive feedback and for recognizing the importance of EEG-to-text decoding and the diversity of our experimental settings. We have carefully revisited the concerns and made substantive revisions, particularly to the inference gallery construction and evaluation protocol. We address the main points below.

W1. Presentation and clarity We acknowledge that the original presentation could be clearer. We will further reorganize the methodology to explicitly distinguish the three components of NEXUS: (1) within-sample disambiguation through shared-residual decomposition, (2) cross-sample disambiguation through centrality-aware Unit learning, and (3) compositional reconstruction using delexicalized structural priors. We will also make the training and inference protocols more explicit to clarify the purpose and interaction of each component.

W2. Evaluation metrics and comparison with recent EEG-to-text methods Regarding BLEU and ROUGE, we respectfully argue that lexical-overlap metrics are not well suited as primary measures of EEG-to-text semantic reconstruction.

Unlike conventional machine translation, EEG-to-text decoding attempts to reconstruct linguistic meaning from noisy, indirect neural measurements. EEG signals do not provide a deterministic encoding of the exact lexical realization of a sentence. Consequently, semantically faithful reconstructions may differ substantially from the reference in word choice, syntax, or expression. BLEU and ROUGE predominantly measure surface-level n-gram overlap. As a result, they can penalize semantically equivalent paraphrases while rewarding frequent lexical patterns that do not necessarily reflect the target-specific information encoded in EEG signals. This concern is especially relevant to our framework, which explicitly separates fine-grained semantic evidence from lexicalized sentence structures.

The limitations of lexical-overlap evaluation in EEG-to-text decoding have also been recognized in recent work, such as Escaping the BLEU Trap (Wang et al., 2026), which discusses how high-frequency lexical patterns can inflate conventional scores without establishing genuine signal-grounded semantic recovery.

To ensure comparability with existing methods, we already report BERTScore F1 under a consistent evaluation protocol. BERTScore evaluates contextual semantic similarity rather than relying solely on exact lexical matches, making it more closely aligned with our objective of semantic reconstruction. Importantly, NEXUS and the baselines are evaluated using the same BERTScore configuration and aggregation procedure within each dataset.

As reported in the revised Table 1, NEXUS achieves BERTScore F1 values of 0.8457, 0.8483, and 0.5882 on ZuCo1, ZuCo2, and Chisco, respectively, outperforming the evaluated baselines. We additionally report RSRA50 and Groundedness to assess target-specific semantic recovery and unsupported generated content, since semantic similarity alone cannot establish EEG grounding. We would also like to clarify that the reviewer's suggested work [1], Assembling the Mind's Mosaic: Towards EEG Semantic Intent Decoding, is already included as SID (Li et al., 2026) in our comparisons on all three datasets. Our evaluation uses a common scoring protocol, and the SID implementation employs the same text embedding model and generation backbone as NEXUS. We will explicitly identify SID by its full paper title to avoid further ambiguity.

Therefore, our metric selection is motivated by the need to distinguish semantic fidelity from superficial lexical similarity, rather than by the omission of conventional evaluation or relevant prior work.

W3. Distinction from QB-Norm and effectiveness of hubness calibration We appreciate the reviewer's observation that QB-Norm [2] also addresses hubness in cross-modal retrieval. While both approaches share the motivation of reducing hub-dominated matching, NEXUS differs from QB-Norm in its optimization mechanism, calibration scope, and integration with semantic-unit learning.

(1) Post-hoc similarity normalization vs. training-time semantic optimization. QB-Norm is a retrieval-time normalization method that adjusts query-gallery similarity scores using a querybank, without modifying the learned representations or requiring model retraining.

In contrast, our NBI and WTI modules operate during training and directly influence the optimization of EEG Unit representations. Specifically, NEXUS minimizes a combined objective involving semantic-unit alignment and hubness-aware supervision. At inference, we use raw cosine similarity without any memory-dependent score correction, as explicitly described in Appendix D.3.

Thus, our method aims to improve the learned semantic matching behavior rather than applying a post-hoc correction to retrieval scores.

(2) Candidate-side correction vs. dual-sided calibration.

QB-Norm primarily addresses the tendency of certain gallery candidates to become hubs. NEXUS additionally considers the centrality of EEG queries themselves.

Our candidate-side NBI estimates the background affinity of semantic-key prototypes using a memory bank of training EEG queries. Unlike direct score normalization, this estimate is used to construct adaptive supervision targets over non-ground-truth neighbors, influencing how the Unit decoder learns to distinguish semantically ambiguous candidates.

Meanwhile, query-side WTI estimates the intra-modal centrality of EEG queries and adaptively reweights their semantic alignment losses. This addresses query-dependent susceptibility to ambiguous matching, which is not explicitly modeled by candidate-side normalization alone.

The two components therefore provide complementary training signals from both the candidate and query perspectives, rather than merely extending a querybank-based similarity correction.

(3) Interpreting the reported retrieval gains.

We agree that the improvement in Unit retrieval ranking is limited, and we would like to clarify the primary objective of hubness calibration.

As shown in the revised Table 2(b), NBI+WTI reduces CR10 from 93.53% to 82.56% and increases the effective number of retrieved candidates from 4.90 to 8.82 compared with anchor-only training. These changes indicate substantially less concentration of erroneous matches among a small number of hub candidates.

At the same time, Unit MRR changes from 0.06585 to 0.06272, revealing a modest trade-off between hubness reduction and local retrieval ranking. We do not claim that hubness calibration universally improves MRR; rather, its purpose is to mitigate systematic cross-sample overmatching that conventional alignment objectives do not explicitly address. Furthermore, the overall novelty of NEXUS does not rest solely on hubness calibration. Our framework identifies and jointly addresses two distinct sources of semantic ambiguity: within-sample context-induced query convergence and cross-sample hubness-induced overmatching.

As shown in Table 2(a), shared-residual supervision reduces query cosine similarity from 0.99996 to 0.30178 and increases effective rank from 1.00044 to 2.66065, demonstrating substantially improved within-sample semantic differentiation. Taken together, these components constitute a unified semantic disambiguation framework for EEG-to-text reconstruction, rather than a direct application of QB-Norm.

W4. Inference gallery construction and additional verification against data leakage

We thank the reviewer for raising this important concern. We acknowledge that the original submission used a dataset-wide inference gallery, which could include held-out target sentences and contextualized Unit representations. Although the original pipeline incorporated safeguards against direct copying, we agree that the gallery construction required a stricter evaluation protocol.

We have therefore made substantive change to address this concern.

In the revised manuscript, both the sentence retrieval gallery and semantic-key prototype gallery are constructed exclusively from the training partition.

Specifically, the sentence gallery contains only training text embeddings, while semantic-key prototypes are constructed solely from training Unit annotations. Validation and test sentences, together with their contextualized Unit targets, are explicitly excluded from gallery construction.

The revised sentence/Unit gallery sizes are: • ZuCo1: 788 / 4,161 • ZuCo2: 515 / 2,675 • Chisco: 5,581 / 5,578

Furthermore, training, validation, and test partitions use disjoint text IDs. Ground-truth semantic keys that are absent from the training gallery are treated as unretrieved during evaluation, rather than being added from the test set.

These revisions remove the original concern that a ground-truth test sentence or its test-derived contextualized Unit representation could be retrieved directly from the inference gallery. The results in the revised manuscript are reported under this stricter protocol, as detailed in Appendix E.1 and F.1.

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

NEXUS reconstructs text from EEG by retrieving sentences and semantic Units. Shared–residual supervision addresses convergence among Unit queries, while candidate- and query-side calibration targets retrieval hubness. A frozen LLM combines retrieved Units with delexicalized structural templates. Experiments on ZuCo1, ZuCo2, and Chisco report improved reconstruction scores and targeted ablations. The evaluation holds out text IDs from downstream training but uses dataset-wide text galleries at inference; it is not held-out-subject evaluation.

Critical strengths
  1. The distinction between within-sample query convergence and cross-sample hubness is a useful way to analyze semantic-unit retrieval. The shared–residual formulation has a clear motivation, and the residual-loss ablation improves Unit retrieval as well as representation diversity. The anchor-only calibration control also helps distinguish further training from the added hubness objectives (Table 2).

  2. The appendix provides detailed losses, prompts, and evaluation rules. Checkpoints are selected using validation scores, training prototype banks and query memories are training-only, and all five generated candidates are averaged rather than choosing the best against the reference. The template ablation keeps retrieval evidence and initial generation seeds matched. These are valuable controls, and the reported RSRA50 values are consistent with the supplied calibration formula and Table 6.

Critical weaknesses

1. The evidence for EEG-grounded reconstruction is not yet sufficient. Appendix E.1 builds inference galleries from all supplied text targets, including held-out sentence contexts. Retrieved sentences also determine Unit verdicts in the first LLM call. This is a legitimate known-corpus retrieval setting, but removing verbatim sentences from the second call does not remove their influence on content selection. Table 6’s random-reference checks keep predictions fixed; they do not replace shuffled-EEG and sentence-only reconstruction controls. There are also unresolved comparison boundaries: Appendix B.1 does not specify whether the Chisco-pretrained front-end excluded the current validation/test trials, and Table 5 gives only NEXUS’s splits. The cross-method reference unions in Appendix F.2 do not establish that the methods were evaluated on identical EEG/text IDs. These details are needed to interpret Table 1. None of them, by itself, proves test leakage.

2. The fidelity metrics and component evidence do not adequately support the compositional claim. Groundedness explicitly omits recall and gives 100 to a nonempty subset of reference content. With a prompt encouraging short outputs and omission of most Units, its improvement needs a content-coverage and output-length analysis. More importantly, the English content metrics ignore order and Structural Alignment collapses content words to coarse tags. For example, “The dog chased the cat” and “The cat chased the dog” have identical content multisets and structural sequences, despite reversing the relation (Appendices F.2–F.4). BERT-F1 is useful complementary evidence, but direct role/polarity checks are still missing. Templates improve Structural Alignment by only 1.30/0.55 points while slightly reducing Groundedness, and the no-template condition retains sentence-derived roles and verdicts. The hubness evidence is a historical source-validation diagnostic: adding NBI/WTI raises MRR from 0.1360 to 0.1377 but slightly lowers R@1 relative to anchor-only. These results support some benefits, not yet a reliable link from both disambiguation mechanisms to faithful final text. Final-output ablations and paired uncertainty across subjects and training runs would strengthen that link.

The retrieval formulation is promising, but I lean toward rejection until the evaluation boundaries and the evidence for faithful composition are better established.

Questions for authors
  1. Did the Chisco front-end’s 39-way pretraining exclude every validation/test trial and text ID used in the current evaluation? Which data were used to estimate the global RMS, and can you provide the corresponding split manifests?

  2. Were all methods evaluated on the same EEG/text IDs and given the same retrieval galleries? Please explain the larger cross-method reference unions in Appendix F.2 and provide a comparison restricted to common test trials.

  3. What happens to Unit retrieval and final reconstruction when EEG trials are shuffled while galleries and prompts remain fixed? How does NEXUS compare with sentence-retrieval-only reconstruction using the same LLM? A train-only-gallery comparison, with Unit coverage reported, would also clarify the scope of the result.

  4. Do the improvements remain when reporting recovered content, output length, and agent–patient or polarity accuracy? The current English metrics cannot distinguish some role-reversed sentences, so these checks would materially affect the compositional interpretation.

  5. Are final-text results available with residual supervision and NBI/WTI separately disabled, keeping the generation procedure fixed? Please provide paired uncertainty for these comparisons and the template gains, and distinguish test-set evidence from the historical validation diagnostic in Table 2(b).

3 comments

We thank the reviewer for the detailed comments. We have revised the pretraining and retrieval protocols, standardized the evaluation samples across methods, and added EEG-shuffling, relation-sensitive diagnostics, and final-text ablations. All references below refer to the revised manuscript.

Q1. Pretraining boundaries and normalization

We replaced the shared Chisco-pretrained front-end with dataset-specific EEG front-ends pretrained independently on each dataset's training partition.

Parameter optimization uses training trials only. Validation data are used for checkpoint selection, and test data are reserved for evaluation. Training, validation, and test partitions have disjoint text IDs, so validation/test sentence identities are excluded from the pretraining optimization data.

Normalization statistics are also estimated exclusively from training data. Each EEG trial is temporally centered within each channel. The dataset-specific global RMS is then computed from centered training signals and applied unchanged to validation and test trials.

Appendices B.1 and F.1 specify the pretraining, normalization, and partitioning procedures.

Q2. Identical evaluation samples across methods

All methods are evaluated on exactly the same held-out EEG trial IDs and corresponding text IDs within each dataset. For ZuCo1 and ZuCo2, we follow GLIM's training/validation/test assignments and apply sample-retention criteria consistently across methods. NEXUS and SID use the same fixed sentence-disjoint partition on Chisco. Table 1 therefore already provides the requested common-test-trial comparison.

The offline reference pools are used to compute scoring statistics, including document frequencies. They are separate from the training-only inference galleries. Sample-level comparability is established by the shared EEG/text IDs.

All methods use the same within-dataset scoring implementation and aggregation protocol: candidate scores are averaged within each trial, followed by equal averaging across trials. All generated candidates contribute to evaluation, with no reference-based best-candidate selection.

Q3. Training-only galleries, EEG dependence, and the sentence-level baseline

Training-only retrieval. We rebuilt both the sentence gallery and the semantic-key prototype gallery exclusively from training data. Validation/test sentence contexts and their contextualized Unit targets are excluded. The sentence/Unit gallery sizes are 788/4,161 for ZuCo1, 515/2,675 for ZuCo2, and 5,581/5,578 for Chisco.

Ground-truth Unit keys absent from the training gallery remain in the evaluation denominator and receive zero reciprocal rank and recall. This evaluates retrieval over the complete set of annotated target Units, including those outside the training gallery's coverage (Appendix E.1).

The first LLM call analyzes retrieved training evidence, and the final generation call receives Unit records, role information, and delexicalized templates. Retrieved sentence text is explicitly excluded from the final generation input (Appendices E.2–E.3).

EEG-shuffling control. We added an experiment that reassigns EEG inputs to different text identities within each subject and reading task. We then rerun retrieval, structural analysis, and generation, keeping model weights, training-only galleries, retrieval settings, and prompts fixed.

On ZuCo1, shuffling reduces RSRA50 from 59.25 to 55.04, Groundedness from 79.63 to 78.65, and BERT-F1 from 0.84570 to 0.84482. Structural Alignment changes from 42.34 to 43.26. The 4.21-point RSRA50 decrease directly links target-specific reconstruction performance to the correct EEG–text correspondence (Appendix G.3, Table 8).

Why we selected GLIM instead of an additional sentence-only generator. The proposed sentence-only system reconstructs text from sentence-level evidence without explicit Unit retrieval. We selected GLIM as an established, more comprehensive implementation of the sentence-level semantic-generation route. GLIM learns EEG representations aligned with the latent semantic space of a language model and uses those representations for text generation, going beyond a basic retrieve-and-rewrite procedure.

Our evaluation therefore uses a purpose-built sentence-level decoding method as the external comparator and targeted NEXUS ablations to evaluate the proposed disambiguation objectives. This motivated our choice of GLIM and component ablations rather than constructing an additional sentence-only generator.

The revised manuscript also includes a fixed-reconstruction comparison using identical retrieved sentences, prompts, and LLM settings with different Unit sources, as detailed in Q5.

Q4. Semantic fidelity, role binding, and polarity

Groundedness measures unsupported generated content. We report it together with RSRA50 for target specificity and BERT-F1 for contextual semantic similarity. To evaluate semantic composition directly, we added participant-role and polarity diagnostics on the final Chisco outputs (Section 4.3).

The participant-role diagnostic checks whether the direction of an action between person-referent participants is reversed. NEXUS reduces this reversal rate from 2.67% for SID to 1.35%, directly examining the “who acts on whom” relationship.

For polarity, sentences are categorized as positive, neutral, or negative, and direct positive–negative reversals are counted. NEXUS reduces the reversal rate from 43.65% for SID to 10.65%.

These evaluations directly measure relational and polarity errors that content-token overlap and coarse structural tags do not distinguish. They provide final-output evidence of improved semantic composition.

The template ablation separately evaluates explicit structural conditioning. EEG weights, retrieved Units, retrieved sentences, role assignments, and consistency verdicts are shared between conditions; the explicit template fields and corresponding instructions are removed in the ablated condition. Generation settings and initial seeds are matched (Appendix G.2). This isolates the template's incremental role while the relation-sensitive diagnostics evaluate the semantic relationships in the final reconstructions.

Q5. Final-text ablations and paired statistical evidence

We added final-text component ablations on ZuCo1, separately removing residual supervision and the combined NBI+WTI objective (Table 3):

SettingRSRA₅₀ ↑Ground. ↑Struct. ↑
w/o Residual58.233175.220540.8666
w/o NBI+WTI58.694080.456242.8916
Full NEXUS59.250479.631742.3431

Relative to removing residual supervision, the full model improves RSRA50 by 1.02 points, Groundedness by 4.41 points, and Structural Alignment by 1.48 points. Adding NBI+WTI improves RSRA50 by 0.56 points. These results quantify the contributions of the disambiguation objectives to final text reconstruction.

The revised Table 2(b) replaces the historical Chisco source-validation diagnostic with a ZuCo1 analysis. Relative to anchor-only training, NBI+WTI reduces CR10 from 93.53% to 82.56% and increases NeffN_{\mathrm{eff}} from 4.90 to 8.82; the corresponding Unit MRR values are 0.06585 and 0.06272. Table 2 measures retrieval concentration, while Table 3 traces the objectives through final reconstruction.

We also added a controlled Unit-source comparison with the downstream reconstruction pipeline fixed (Appendix G.4, Table 9). SID-predicted and NEXUS-predicted Units are evaluated on the same ZuCo1 EEG trials using the same NEXUS-retrieved Top-3 training sentences, reconstruction prompts, Qwen3.5-9B model, and generation settings.

Under this shared pipeline, NEXUS Units improve RSRA50 from 51.11 to 59.25 and BERT-F1 from 0.82862 to 0.84570. The paired confidence intervals for both differences are above zero. This establishes a contribution from the recovered Unit evidence under identical downstream processing.

Table 1 now reports 95% confidence intervals for the main reconstruction results, and Appendix G.4 provides paired uncertainty for the fixed-pipeline comparison. Together with the final-text ablations, these additions connect semantic disambiguation and recovered Unit evidence to measurable reconstruction outcomes.

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

This paper presents NEXUS for recovering sentence-level meaning and fine-grained semantic Units from EEG. It aims to distinguish semantic components within each sample and reduce cross-sample overmatching toward highly central candidates, so that recovered content can be organized into coherent sentences. Experiments on ZuCo1, ZuCo2, and Chisco report improvements in Unit retrieval and text reconstruction across natural reading and imagined speech.

Critical strengths
  1. The paper identifies a concrete and practically relevant failure mode in fine-grained EEG semantic decoding. It distinguishes within-sample query convergence from cross-sample hubness, providing a clear motivation for why existing semantic-unit decoders may retrieve redundant or biased semantic evidence.

  2. The proposed method is well aligned with the identified failure modes. Shared–residual decomposition and balanced Query–Unit assignment target within-sample ambiguity, while NBI and WTI explicitly address candidate-side and query-side hubness.

  3. The evaluation covers heterogeneous EEG-language settings. NEXUS is evaluated on ZuCo1, ZuCo2, and Chisco, spanning natural reading and imagined speech across English and Chinese settings.

Critical weaknesses
  1. The contribution of the proposed disambiguation mechanisms to the final reconstruction gains is only indirectly established. The residual-loss and hubness analyses mainly report Unit-level retrieval and concentration metrics, but do not trace these components through the full reconstruction pipeline, so their contribution to the gains in Table 1 remains unclear.

  2. The evidence for compositional fidelity is still limited. The template ablation improves coarse Structural Alignment only modestly, while RSRA50 and Groundedness are largely unchanged, and the current metric does not directly evaluate role-binding errors such as agent–patient reversals.

  3. The final improvements are partially confounded by a substantially stronger downstream retrieval-and-LLM pipeline. NEXUS adds sentence retrieval, Unit validation, role assignment, structural abstraction, and LLM-based generation, making it difficult to isolate how much of the improvement comes from EEG-side semantic disambiguation itself.

  4. The robustness and comparability of the main results are not fully established. Table 1 reports only point estimates without uncertainty or significance tests, and it is not fully clear whether all methods are evaluated on identical EEG trials and text IDs; this matters especially when some ZuCo gains are very small.

Questions for authors
  1. How much do the residual-loss and hubness-calibration components improve final text reconstruction rather than only Unit-level retrieval? Can the authors report matched end-to-end ablations under the same gallery and generation settings?

  2. Can the authors directly evaluate relation-sensitive compositional errors in the final outputs? In particular, how often does NEXUS preserve or reverse agent–predicate–patient relations when lexical content is similar?

  3. How much of the improvement over SID comes from EEG-side semantic disambiguation versus the stronger downstream retrieval-and-LLM reconstruction pipeline? A controlled comparison with the reconstruction pipeline fixed would clarify this attribution.

2 comments

We thank the reviewer for the constructive comments. We have added final-text component ablations, direct semantic-role and polarity evaluations, and a controlled comparison with the reconstruction pipeline fixed. These experiments directly address the contribution of EEG-side decoding to final reconstruction. All references below refer to the revised manuscript.

W1 / Q1. Disambiguation improves final text reconstruction

We have added final-text ablations on ZuCo1 that separately remove residual supervision and the combined NBI+WTI objective (Table 3).

SettingRSRA₅₀ ↑Ground. ↑Struct. ↑
w/o Residual58.233175.220540.8666
w/o NBI+WTI58.694080.456242.8916
Full NEXUS59.250479.631742.3431

Compared with removing residual supervision, the full model improves RSRA50 by 1.02 points, Groundedness by 4.41 points, and Structural Alignment by 1.48 points. Residual supervision therefore benefits the final reconstructed text across all three reported measures.

NBI+WTI improves RSRA50 by 0.56 points relative to its ablated variant, demonstrating its contribution to target-specific reconstruction.

These final-output ablations directly connect the proposed disambiguation objectives to reconstruction performance. The fixed-pipeline experiment below further evaluates the contribution of decoded Unit evidence under identical downstream processing.

W2 / Q2. Direct evaluation of semantic-role and polarity errors

We have added relation-sensitive evaluations of the final Chisco outputs (Section 4.3).

The participant-role diagnostic checks whether the direction of an action between person-referent participants is reversed. NEXUS reduces the reversal rate from 2.67% for SID to 1.35%. This directly evaluates the “who acts on whom” relationship raised by the reviewer.

We additionally evaluate semantic polarity by categorizing sentences as positive, neutral, or negative and counting direct positive–negative reversals. NEXUS reduces the polarity reversal rate from 43.65% for SID to 10.65%.

These results provide direct evidence of fewer role-binding and polarity errors in the reconstructed sentences. They complement Structural Alignment by evaluating the semantic relationships expressed in the final outputs.

The template ablation in Table 4 measures the incremental effect of explicit template conditioning, while the new role and polarity diagnostics evaluate the final system’s relational fidelity. Together, these analyses assess both the structural guidance mechanism and the resulting semantic composition.

W3 / Q3. Controlled comparison with the reconstruction pipeline fixed

We implemented the requested controlled comparison in Appendix G.4 and Table 9. On the same ZuCo1 test EEG trials, SID-predicted and NEXUS-predicted Units are supplied to the same downstream reconstruction pipeline.

Both conditions use the same NEXUS-retrieved Top-3 training sentences, reconstruction prompts, Qwen3.5-9B model, and generation settings. Unit validation, role assignment, structural abstraction, and final generation follow the same procedure. The experimental variable is the source of predicted Unit evidence: SID or NEXUS. All five generated candidates are retained and evaluated using the same averaging protocol.

MetricSID UnitsNEXUS Units
RSRA₅₀ ↑51.1159.25
Groundedness ↑73.6979.63
Structural Alignment ↑49.1642.34
BERT-F1 ↑0.828620.84570

Under this shared reconstruction pipeline, NEXUS Units improve RSRA50 by 8.14 points, Groundedness by 5.94 points, and BERT-F1 by 0.01708. The paired confidence intervals for the RSRA50 and BERT-F1 differences are both above zero (Appendix G.4).

The gains in target specificity and semantic similarity persist when SID receives exactly the same sentence evidence and downstream LLM processing. This directly demonstrates the contribution of the decoded Unit evidence to the reconstruction advantage.

Table 3 and Table 9 provide complementary evidence: the former quantifies the effects of residual supervision and hubness calibration on final text, while the latter establishes the benefit of NEXUS’s Unit evidence with the reconstruction procedure fixed.

W4. Identical evaluation samples and uncertainty estimates

All methods are evaluated on identical held-out EEG trial IDs and corresponding text IDs within each dataset. For ZuCo1 and ZuCo2, we follow GLIM’s training, validation, and test assignments and apply sample-retention criteria consistently across methods. NEXUS and SID share the same fixed sentence-disjoint test cohort on Chisco.

All comparisons use the same within-dataset scoring implementation and aggregation protocol: candidate scores are averaged within each trial, followed by equal averaging across trials. All candidates contribute to evaluation.

Revised Table 1 now reports 95% confidence intervals for the main reconstruction results. Appendix G.4 additionally reports paired uncertainty for the fixed-pipeline Unit-source comparison, with positive intervals for the RSRA50 and BERT-F1 differences. The revised evaluation therefore provides common-sample comparisons and uncertainty estimates for both the main results and the controlled reconstruction experiment.

Together, these additions directly connect semantic disambiguation to final-text performance, evaluate relation-sensitive reconstruction errors, and establish the contribution of EEG-derived Unit evidence under a shared downstream pipeline.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.