Reviewer has chosen to be Anonymous

Overall Impression: Weak

Content:
Technical Quality of the paper: Weak
Originality of the paper: Yes
Adequacy of the bibliography: Yes, but see detailed comments

Presentation:
Adequacy of the abstract: Yes
Introduction: background and motivation: Limited
Organization of the paper: Needs improvement
Level of English: Satisfactory
Overall presentation: Average

Detailed Comments:

Summary:
This paper proposes an approach for entity relationship extraction/classification for agglutinative languages with a morphology-based method. To this end, a syntax tree is treated as a graph from which a Tsetlin Machine (TM) learns propositional rules jointly with binarized node embeddings of a Relational GCN. Thereby, the neural representation is to be integrated with the symbolic representation of morphological information in a TM for the purpose of relation classification. To evaluate their approach, the authors also propose their own dataset in the Manipuri language.

Strengths:
- The idea to integrate morphological information into a relation extraction approach might not be novel, but the concrete approach proposed by the authors is interesting.
- Topic and approach both fit well within the scope of the NAI journal.
- The submission not only proposes an approach but also a dataset in a low-resource agglutinative language.

Weaknesses:
- The approach is designed for agglutinative languages in plural but only evaluated or practically applied to a single language and an unclear dataset, which limits the evidence for the generalizability of the proposed approach. Since the approach is morphology-aware, generalizability most likely strongly depends on the availability of a corresponding inventories or parsers.
- To what extend the proposed syntax tree really qualifies as a graph can only be determined once sufficient details are provided (see detailed comments).
- The related work section focuses on morphological parsing/approaches and neuro-symbolic approaches, but not on relation extraction at all. It should, however, detail the state of the art on relation extraction as the main topic and task of the submission.
- There is a severe lack of details on the proposed approach that makes it hard to understand what exactly is being proposed and entirely impossible to potentially reproduce the approach (see detailed comments).
- The same is true for the design and creation of the proposed dataset, where all the details on how the dataset was created are missing. The only description provided is what the final datasets consists of, but not how it was derived.
- All variables in represented equations need to be introduced and explained, which is currently not the case starting from page 3. The level of details should be sufficient, e.g. nonlinear activation function is not sufficiently detailed. As regards Equation 3, I am quite at a loss what it represents.
- It is extremely rare and hard to believe that accuracy and Micro-F1 score are exactly identical in a relation extraction task, even if the proposed task seems to be more of a relation classification than actual relation extraction task. Equally the results of the baseline models are hard to believe and truly require the code and the dataset so that these can be checked. First, it is unclear why exactly these models where chosen as baseline models and second, what exactly did XLM-R predict and were the models equally applied in a multi-class classification mode as it was done for the proposed approach? If a multi-class classification approach, for both the proposed approach and the baseline models none should be included. Where the models fine-tuned or a relation classification fine-tuned version used for the experiment or just used as is? The whole argument is that this TM-based approach outperforms neural approaches, however, to make this argument valid, the neural approach needs to be applied correctly and as intended, not just using comparatively small pre-trained models without fine-tuning as baseline.

Detailed comments:
One of the main issues of this submission is the lack of details of the proposed approach, the results, and in general in the paper.

While the nature of the syntax tree aka graph is somewhat specified, the method of how it is derived as well as a lot of other details are rather unclear. Where does the Manipuri morpheme inventory come from and how is it exactly applied? Does a suffix inventory truly suffice to obtain lexical stems? Which tools, parsers or algorithms were applied to lemmatize and derive the POS and NER tags? Table 2 is even more surprising as it suggests that specific suffixes can directly be mapped to semantic relations, not just linguistic relations. This is not only surprising but very hard to believe as this would be highly specific to Manipuri. This would definitely require further evidence and also still details on how this table and idea was practically applied.

If the root note is the lexical stem, are NER tags also assigned based on the stems or based on the original form? Because stemming and lemmatization can destroy the surface form and recognition of named entities. What do E and R represent in G? How are the fastText embeddings used? Which R-GCN is being used exactly? What exactly is the meaning of all the different edges in the edge schema?

In the Neuro-Symbolic Relation Classification Approach, the assumption is that entities resp. entity candidates are already given. Where do these entity mentions come from? How exactly are GCN embeddings binarized for input to the Tsetlin Machine? What exactly do all the elements in the clause in Equation 5 mean? The same is true for all the clauses in the Results section, which should clearly be separate from a discussion section. Without a thorough introduction on what these literals in the clauses mean, it is difficult to understand the clauses and their relevance to the mentioned relations.

Overall evaluation:
What is described as a graph in this paper is essentially a syntax tree, which is not the same as a knowledge graph. Syntax trees can be considered symbolic to some degree, however, only the way that this morphological information is integrated as structured knowledge with GCN node embeddings make it a viable contribution to neuro-symbolic AI.

The main and very convincing argument on why considering morphology in a relation extraction approach is unfortunately repeated several times, without, however, providing a sufficiently detailed account on the approach and how this objective is exactly achieved.

There are two reasons for suggesting a major revision. One is the severity of the details being ommitted from a scientific perspective as well as from the perspective of fully understanding the proposed approach. The second reason is the time that will probably be needed to revise the paper thoroughly and introduce experiments on a second agglutinative language for validity.

Minor comments:
Usually abstracts have no paragraphs but represent one long abstract.
The framework in the abstract becomes a model that is not more closely described in the next sentence. What kind of model is it?
On p. 2 there is a missing Section ref: "analyzed in
Section ."
Review of Literature => this is usually called Related Work
There seems to be a color mismatch between Figure 1 and 2, both of which seem to represent one of the proposed graphs. In Figure 1, -da is represented in green, the color assigned to e supposedly entities (2) in Figure 2. I am also not sure what r is as blue nodes in Figure 1 are named entities, but then there is a node called ENTITIES in Figure 1 as well, which is confusing as this would mix the T-Box and the A-Box of an ontology directly resp. introduce types randomly as nodes.
R-GCN (neural baseline) => all other baselines are neural baselines as well

Reviewer has chosen not to be Anonymous

Overall Impression: Weak

Content:
Technical Quality of the paper: Weak
Originality of the paper: Yes
Adequacy of the bibliography: Yes, but see detailed comments

Presentation:
Adequacy of the abstract: Yes
Introduction: background and motivation: Good
Organization of the paper: Needs improvement
Level of English: Satisfactory
Overall presentation: Average

Detailed Comments:

The paper proposes a parser-independent, morphology-aware framework for relation extraction in low-resource agglutinative languages, using Manipuri (Meiteilon) as the case study. Each sentence is converted into a heterogeneous graph in which lexical roots, grammatical affixes, and entity mentions are explicit nodes connected by typed edges (HAS_AFFIX, NEXT_TOKEN, SPAN_OF, SELF_LOOP), node representations are learned with a two-layer R-GCN, and relation classification is performed by a Tsetlin Machine (TM) over binary symbolic features plus binarized R-GCN embeddings. A new dataset (MERED) is introduced, and the system reports 91.72% accuracy / 91.16% Macro-F1, ahead of BiLSTM, CNN–BiLSTM, dependency-GCN, and XLM-R baselines.

I think the core idea is genuinely appealing and a good fit for a neurosymbolic venue: promoting role-marking affixes to first-class graph nodes is a clean way to inject morphology into message passing, and pairing an R-GCN encoder with a Tsetlin Machine gives a credible interpretability story.

The motivation and linguistic framing are well written. My concerns are concentrated in the experimental section. As part of the review I downloaded and inspected the released dataset and code repository (as of June 16, 2026), and I was not able to reconcile the reported experiments with those artifacts on several points. I describe each observation factually below; in several places there may be a straightforward explanation (for example, a more complete dataset or additional scripts that were not yet added to the public release), and if so, I would ask the authors to provide the exact artifacts so the results can be verified.

Strengths
-------------
1. Well-motivated, well-scoped problem. Morphology-as-structure for agglutinative, low-resource RE is clearly argued, the parser-independence motivation is convincing.
2. Conceptually elegant architecture. Affixes as graph nodes + relation-specific R-GCN message passing + a symbolic TM classifier is a natural, defensible design.
3. Interpretability angle. Translating embeddings into propositional clauses fits the venue, and the neurosymbolic "bridge" analysis (Table 7) is a nice idea.
4. New resource + public code. A Manipuri RE dataset and a public repository are valuable for an under-resourced language, provided the released artifacts can be matched to the reported results.

Major points (please address)
-----------------------------
M1: I could not reconcile the released dataset with the reported results.
Using the provided Kaggle and GitHub links, I could not reproduce the 495-sample MERED dataset referred throughout the article. Both released JSON files only contain 25 sentences and 43 relations (although the github version in MERED.json seems to feature some duplicates), which is an order of magnitude less than the 495 relations refered in Tables 8-13 and in the confusion matrix. Both files also mark samples with "annotation_status": "auto + heuristic" metadata, which seems to contrast the manual annotation claim made in paper (page 5). Releasing the used dataset as stated in page 11-12 (Data availability) would resolve this confusion.

M2: Released code is incomplete as published
The entry-point scripts import modules that are not present in the repository - rgcn_encoder.py (the R-GCN definition), tm_features.py (candidate-pair features), and evaluation.py (metrics). The TM training script named in the README (tm_train_encode_positive.py) and the default MERED data file used by every script (MERED_reduced_relations_v2.json) are also absent. As a result the pipeline cannot currently be run end-to-end from the public repository. For a paper whose contributions include interpretability and reproducibility, completing the release (including a runnable example and the exact data) would substantially strengthen it.

M3: The abstract-listed score has no explicit held-out test-split
While the article does not address the train/test splitting of MERED dataset used for experimentation at all, upon investigating the attached training and evaluation scripts, the TM seems (as per pipeline_eval_tm_with_encode_positive_new.py) to be applied to all pairs from the full dataset, with no train/test partition in the evaluation path. Please report a disjoint, reproducible test split, so the reported metrics are clearly on unseen data.

M4: The abstract-listed score is on a positive-only subset; end-to-end RE is not reported.
The strongest result is computed on the 495 positive instances after candidate pruning ("to mitigate the dominance of the NONE class"). This gives a performance metric of relation classification given that relation exists, rather than that of an end-to-end relationship extraction. Please add the realistic end-to-end numbers, including the NONE class.

M5: Dataset documentation
As a stated contribution, MERED should come with corpus statistics (sentences/documents, entities, per-class relation counts on the full corpus), annotation details (number of annotators, guidelines, inter-annotator agreement), and the train/dev/test split used. None of these are currently given, and with so few instances per class, confidence intervals or cross-validation would help establish that the results are stable.

M6: The baselines and ablations may benefit from further clarification
Some reported numbers are unexpectedly low or unstable, especially the neural R-GCN + MLP ablation with Macro-F1 0.16 and the XLM-R baseline with near-zero scores for many classes. These results may be correct in this low-resource setting, but the paper should at least briefly explain their training setup. Without this, it is difficult to know whether the large gains come from the proposed neuro-symbolic design, from the positive-only setup, from feature engineering/candidate filtering, or from under-optimized baselines.

Minor / presentation
--------------------
- Broken cross-reference: "…further analyzed in Section ." (empty reference).
- Introduction and Review of Literature overlap substantially; please consolidate.
- Bibliography: several author fields look mis-parsed where "et al." became a first name ("A. e. a. Conneau", "X. e. a. He", "J. e. a. Devlin", "D. e. a. Zeng", "J. e. a. Nivre"); duplicate entries appear — Comrie 1989a/1989b (same book), Haspelmath 2002/2010 (same Understanding Morphology), Schlichtkrull 2018a/2018b (same ESWC paper), Wu et al. 2021a/2021b (identical survey), and multiple Zeng entries. Please de-duplicate and correct.
- Segmentation is not evaluated although the pipeline depends on the greedy suffix segmenter; it might be worth reporting its accuracy/coverage or discussing its possible error contribution.
- Transferability claims to Uyghur/Bodo/Santali/Kokborok are asserted but not demonstrated; I would suggest to slightly soften them (or add at least one transfer experiment).

Recommendation rationale
------------------------------------
I recommend Major revisions. The core idea is original, interesting, and well suited to the venue, and the manuscript has a plausible contribution if the empirical claims can be verified. However, the present version has serious reproducibility and evaluation gaps: the released dataset/code do not appear to match the reported experiments, the train/test split is unclear, and the headline score is reported on a positive-only subset rather than a full end-to-end RE setting.

I would be happy to reassess the paper if the authors provide the exact dataset and splits used for the reported results, code, or a more complete description of the experiments that produced the numbers.