Pith. sign in

REVIEW 4 major objections 6 minor 67 references

Paths to Causality: Finding Informative Subgraphs Within Knowledge Graphs for Knowledge-Based Causal Discovery

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that ranking metapath subgraphs from knowledge graphs with a learning-to-rank model and placing the top subgraph in the prompt materially improves zero-shot LLM causal discovery, by up to 44.4 F1 points.

desk verdict Plausible incremental method for selecting KG subgraphs via learning-to-rank, but the missing train/test split makes the headline F1 gains hard to trust. read the letter →

arxiv 2506.08771 v1 pith:OILIXY35 submitted 2025-06-10 cs.AI cs.CLcs.IRcs.LG

Yuni Susanti , Michael Färber This is my paper
classification cs.AIcs.CLcs.IRcs.LG
keywords causaldiscoveryknowledgegraphslargelanguagemodelsmetapathlearningtorankzero-shotpromptingpairwiserelationclassificationbiomedicalNLP
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that knowledge-based causal discovery—deciding whether one variable causes another from names and textual context alone—becomes substantially more reliable when an LLM's prompt includes a carefully chosen subgraph from a knowledge graph, instead of no graph context. Its central claim is that the choice of subgraph matters: a dedicated learning-to-rank model trained on LLM-generated relevance labels selects the informative paths, and adding just the top-ranked path to zero-shot prompts outperforms no-subgraph, random-subgraph, similarity-based, and GPT-based ranking baselines. If true, this would give practitioners a cheap, data-free route to causal inference in domains like biomedicine where observational data is scarce, by leveraging structured knowledge already present in knowledge graphs.

What carries the argument

The central object is the metapath-based subgraph: a path through a heterogeneous knowledge graph connecting the two variables, with node types and relation labels intact, for example [drug–gene–disease] for raloxifene–melanoma. The machinery that carries the argument is the SubgraphRanker, a learning-to-rank model trained on relevance labels produced by Eq. (1), which rewards a subgraph when the LLM's answer matches the ground-truth causal label and penalizes when it does not, scaled by the LLM's output probability. At test time the ranker's top subgraph is verbalized into a prompt for zero-shot causal classification, which is the step credited with the F1 improvements. The paper compares pointwise (RMSE), pairwise (RankNet), and listwise (ListNet) objectives with cross-encoder and XGBoost architectures.

What would settle it

Check whether the variable pairs used to train the SubgraphRanker are disjoint from the pairs used in the final zero-shot causal-discovery evaluation; if they overlap, retrain the ranker on one half and evaluate on the other, and compare F1. A drop toward the random-subgraph baseline would show that the gains depended on the ranker having seen the answer.

Watch

Extended reading notes

Core claim

The core discovery is that informative subgraphs for causal inference can be identified and ranked automatically, and that using the top-ranked subgraph in a zero-shot prompt improves F1 scores beyond the no-subgraph baseline across three biomedical datasets and one open-domain dataset, for three different 7–8B LLMs, with the largest gain being 44.4 F1 points on COMAGC with llama-3-8b. The paper also reports that when subgraphs are taken from the Hetionet biomedical KG and Wikidata, the method beats statistical causal discovery baselines (PC, Exact Search, DirectLiNGAM) on the SACHS protein data, by up to 33.42 F1 points and a 17-point Hamming distance reduction. The mechanism is a two-stage distillation: an LLM scores each metapath by whether it steers the model to the human-labeled causal answer, and a SubgraphRanker (cross-encoder or XGBoost) learns to reproduce that ranking, so at inference time only the top path goes into the prompt.

Load-bearing premise

The SubgraphRanker is trained on relevance labels that are defined using the ground-truth causal label, and the paper never states where the ranker's training data ends and the final causal-discovery evaluation begins, so reported F1 gains could be inflated if the evaluation pairs overlap with the training pairs.

Editorial extensions

If this is right

  • Even randomly selected subgraphs improve F1 over the no-subgraph baseline by an average of 12.5 points, so graph context itself carries signal beyond the ranking.
  • The SubgraphRanker beats a similarity-based ranker by an average of 7.74 F1 points, indicating that causal informativeness is not captured by embedding similarity.
  • Among the learning-to-rank objectives, the pairwise RankNet variant performs best in most experiments, followed by pointwise and then listwise.
  • Using only the top-1 subgraph in the prompt yields the best average results across LLMs, so adding more paths can hurt rather than help.
  • The approach generalizes across knowledge graphs: improvements reach 44.4 F1 points with Hetionet on biomedical data and 22.37 points with Wikidata on open-domain data.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper does not state the train/test split between SubgraphRanker training and the final causal-discovery evaluation; if the evaluation pairs overlap with the ranker's training pairs, the reported F1 gains could be inflated by the ranker having seen the answers.
  • A natural extension is full causal graph discovery: apply the pairwise method to all variable pairs and then orient edges, a direction the paper itself flags as future work.
  • Because the relevance labels in Eq. (1) use ground-truth causal labels, the pipeline is not truly label-free; a testable variant would replace the ground-truth supervision with self-consistency scores across different subgraphs to isolate the contribution of subgraph selection itself.
  • A cheaper alternative to training a ranker would be to prompt an LLM to score subgraphs directly at inference time; the paper's GPT-based baseline suggests this is less effective, but the comparison is confounded by using different LLMs for ranking versus final classification.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes a pipeline for knowledge-based causal discovery over entity pairs: (1) Subgraph Relevancy Estimation uses mistral-7b to score metapath-based subgraphs from Hetionet or Wikidata by whether the LLM's prediction with that subgraph matches the ground-truth relation label (Eq. 1); (2) a Subgraph Ranker is trained with pointwise, pairwise, or listwise learning-to-rank objectives and with cross-encoder or XGBoost architectures; and (3) the top-ranked subgraph is inserted into a zero-shot prompt for causal/non-causal classification. The authors report F1 gains over a no-subgraph baseline of up to 44.4 points across GENEC, ADE, COMAGC, and SEMEVAL, under three 7-8B LLMs and two KGs, and they additionally compare against statistical causal discovery on SACHS. The code and data are released on GitHub.

Significance. The contribution is potentially useful: it is, to my knowledge, the first dedicated learning-to-rank treatment of subgraph selection for KG-grounded causal discovery, and the evaluation is broad, covering multiple datasets, LLMs, KGs, and ranker architectures, with code and data released. The paper also includes useful baseline comparisons (random, similarity, and GPT-based rankers) and an appendix with implementation details. However, the validity of the headline empirical claim depends on a train/test separation that the manuscript never states; without it, the reported gains may be an artifact of label-informed subgraph selection. The SACHS comparison also needs clarification before the claims can be accepted.

major comments (4)
  1. [§5.1, §4.2, Appendix D, Figure 2] The manuscript never defines a train/test split between the Subgraph Ranker training data and the final evaluation. Eq. (1) scores each subgraph using the ground-truth label and the LLM's prediction probability; Section 5.1 says the ranker is trained on this dataset; Appendix D repeats this without mentioning any split; Figure 2 labels both the ranker input and the final discovery stage as "test data." If the same pairs are used in both stages, then the top-k subgraph is selected by a model trained on answer-key information, and the Subgraph Ranker rows in Tables 1 and 2 are inflated relative to the no-subgraph, random, similarity, and GPT rankers. This is load-bearing for the central claim in Section 6 ("improvements of up to 44.4 points in F1 scores"), so the authors must specify the exact split and report results on truly held-out pairs.
  2. [§6.1, Table 3] The comparison to statistical causal discovery is not a controlled comparison. PC, Exact Search, and DirectLiNGAM estimate a full causal graph from the SACHS observational data, while the proposed method produces pairwise causal/non-causal predictions from metadata and KG paths. The paper does not explain how these pairwise outputs are converted to an adjacency matrix over the 11 SACHS variables, how Hamming distance is computed for the LLM-based method, or whether the ground-truth edges used for scoring are the same for both families. Since some statistical scores are quoted from [52], the experimental settings across rows may differ. The claim of "up to a 33.42-point improvement in F1" is therefore not supported by the reported protocol and needs either a precise task alignment or a revised claim.
  3. [§5.3, Appendix A] All main results are computed on a filtered subset of each dataset ("we filter the datasets ... to include only the instances that contain subgraph in the KGs"), but the paper never reports how many instances remain after filtering or the class balance of the retained subsets. Because the no-subgraph baseline is evaluated on the same filtered subset, the reported absolute F1 values cannot be compared to any published results on the full datasets, and the conclusions may not transfer to instances with no KG connectivity. The authors should report retained counts per dataset and per KG, and should consider reporting performance on the full datasets as well.
  4. [§4.2, Eq. (1), §6] Even if a proper split is introduced, the labels used to train the ranker are defined by whether mistral-7b's prediction with a subgraph matches the ground truth. For the mistral-7b rows in Table 1, the final evaluation then measures how well the ranker can select a subgraph under which mistral-7b tends to produce the ground-truth label, rather than how informative the subgraph is in general. The paper should acknowledge this and report at least one experiment in which the relevance labels are generated by a different LLM (or from intrinsic subgraph features) to separate selection quality from answer-key recovery.
minor comments (6)
  1. [Abstract / Contributions] The abstract states "up to 44.4 points" while the contribution list states "up to a 44.8-point increase"; Section 6 reports 44.4. The numbers should be reconciled.
  2. [Figure 2] Figure 2 labels both the Subgraph Relevancy Estimation input and the final causal discovery stage as "test data"; the caption should clarify whether these are the same pairs or different splits.
  3. [Eq. (4)] Eq. (4) uses softmax(y) where y denotes ground-truth relevance scores, but those scores are arbitrary real values from Eq. (1); the authors should define how they are normalized to a distribution.
  4. [§6, item 1] The phrase "metadata alone in prompts is not sufficient" is imprecise, since the no-subgraph condition still includes the textual context in the prompt (Eq. 8).
  5. [Table 3] Table 3 uses "NHD" and "HD" without defining them in the table caption; the authors should define normalized Hamming distance and state how it is computed for each method.
  6. [§4.3.2, Appendix B] There are small typographical errors, including "Transfomer" in Figure 4, "GDBT" instead of "GBDT" in Section 4.3.2, and "allShortestPaths" in Appendix B, which should be formatted as code.

Circularity Check

1 steps flagged · score 4.0 of 10

SubgraphRanker is trained on ground-truth-derived relevance labels and the paper never states a train/test split, so the headline F1 gains may be label-informed retrieval rather than independent causal discovery.

  1. fitted input called prediction [Section 4.2 (Eq. 1); Section 5.1; Appendix D.1; Figure 2]
    "A subgraph is considered as relevant if the LLM predicts the relation correctly, according to the human-labeled ground truth relation. The relevance score s_i ... s_i = (1+p('causal/non-causal')) if y_LLM=y_truth; 1-p(...) if y_LLM≠y_truth (Eq. 1). ... we train SubgraphRanker (§4.3) model using this dataset ... We provided up to 10 subgraphs per pair for subgraph ranker training, excluding pairs without any metapath-based subgraphs in our selected KGs."

    Eq. (1) defines a subgraph's relevance score using the ground-truth relation label y_truth and the LLM's probability of predicting that label. The SubgraphRanker is fitted to these scores (Section 5.1, Appendix D.1). The paper then uses the ranker to select the top-1 subgraph for the same 'test data' on which the zero-shot causal F1 is reported (Figure 2), without stating any train/test split between the pairs used to fit the ranker and the pairs scored in Tables 1-2. If those sets overlap, the selected subgraph for a test pair is by construction the one already known to make the LLM answer correctly with highest confidence; the final F1 gain (e.g., 44.4 points on COMAGC with llama-3-8b) then measures label-informed retrieval rather than the value of informative subgraph selection.

full rationale

The main empirical claim is that LLM+KG with SubgraphRanker improves zero-shot causal classification by up to 44.4 F1. This pipeline is not circular by construction if the SubgraphRanker is trained on a disjoint train split and applied to a held-out test split, because mimicking Eq. (1) relevance labels is a legitimate supervised retrieval objective. However, the paper never states such a split. Section 5.1 says the ranked dataset is generated 'per variable pair' and the ranker is trained 'using this dataset'; Appendix D.1 says 'we provided up to 10 subgraphs per pair for subgraph ranker training' with no separate evaluation split; and Figure 2 routes 'test data' through the ranker into the final zero-shot stage. If the final evaluation pairs are the same pairs used to fit the ranker, then for each test pair the top-1 subgraph is chosen using Eq. (1), which already encodes the ground-truth relation label and the LLM's probability of predicting it, so the reported F1 is partly a retrieval artifact. I therefore flag this as a fitted-input-called-prediction concern. I do not see other circular steps: the cited prior work [50] supplies a dataset and a related method but is not used to forbid alternatives, and the general benefit of subgraphs is independently supported by the random-subgraph baseline. Score 4 reflects that the central subgraph-helping claim has independent content, but the headline SubgraphRanker improvement is not verified as split-clean.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new entities, forces, or conserved quantities; it combines existing KG structures, LLMs, and ranking models. The free parameters are evaluation-tuned choices (k, hop limits, n-gram size) rather than fitted physical constants.

free parameters (3)
  • k (top subgraphs per prompt) = 1
    The number of subgraphs included in the prompt is chosen as k=1 because it 'yielded the best average results across all LLMs' (Appendix E), i.e., tuned on evaluation data.
  • Hop limit for subgraph extraction = 4 hops for Hetionet, 2 for Wikidata
    Maximum path length for metapath extraction is set without analysis of sensitivity; it determines the candidate subgraph set (Appendix B).
  • N-gram embedding dimension and n = n=2, dim=128
    Hyperparameters for the XGBoost ranker (Appendix D.2) selected by the authors without ablations.
assumptions (5)
  • domain assumption Knowledge graphs (Hetionet, Wikidata) contain reliable and sufficient causal structure for the variable pairs
    The entire method depends on the KG containing informative metapaths that are predictive of the ground-truth causal label; this is assumed in Section 4.
  • domain assumption LLM log-probability scores are usable as relevance labels for ranking subgraphs
    Equation (1) converts LLM prediction confidence into subgraph relevance scores; the validity of this calibration is not evaluated.
  • domain assumption Ground-truth causal labels in GENEC, ADE, COMAGC, SEMEVAL are correct and applicable to the filtered subset
    Dataset filtering (Appendix A) and label quality are taken for granted.
  • domain assumption Metapath node-type sequences carry causal signal (drug-gene-disease paths are more informative)
    Section 4 argues this by example; no formal justification is given.
  • standard math Standard transformer and XGBoost learning-to-rank machinery behaves as described
    The ranker objectives (RankNet, ListNet, RMSE) are standard (Section 4.3.1), but the claimed ranking performance rests on their implementation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Paths to Causality: Finding Informative Subgraphs Within Knowledge Graphs for Knowledge-Based Causal Discovery." pith.science (2026). https://pith.science/paper/OILIXY35

@misc{pith2026250608771,
  author       = {Pith},
  title        = {Pith review of: Paths to Causality: Finding Informative Subgraphs Within Knowledge Graphs for Knowledge-Based Causal Discovery},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OILIXY35}},
  note         = {Machine review of arXiv:2506.08771}
}
read the original abstract

Inferring causal relationships between variable pairs is crucial for understanding multivariate interactions in complex systems. Knowledge-based causal discovery -- which involves inferring causal relationships by reasoning over the metadata of variables (e.g., names or textual context) -- offers a compelling alternative to traditional methods that rely on observational data. However, existing methods using Large Language Models (LLMs) often produce unstable and inconsistent results, compromising their reliability for causal inference. To address this, we introduce a novel approach that integrates Knowledge Graphs (KGs) with LLMs to enhance knowledge-based causal discovery. Our approach identifies informative metapath-based subgraphs within KGs and further refines the selection of these subgraphs using Learning-to-Rank-based models. The top-ranked subgraphs are then incorporated into zero-shot prompts, improving the effectiveness of LLMs in inferring the causal relationship. Extensive experiments on biomedical and open-domain datasets demonstrate that our method outperforms most baselines by up to 44.4 points in F1 scores, evaluated across diverse LLMs and KGs. Our code and datasets are available on GitHub: https://github.com/susantiyuni/path-to-causality

Figures

Figures reproduced from arXiv: 2506.08771 by the authors.

Figure 1
Figure 1. 1-hop connections between (raloxifene, melanoma) in Hetionet [18]. Relation types are omitted for readability. treatment. Thus, inferring causal relationships between pairs of variables, such as (smoking, lung cancer), is a fundamental step in causal discovery, as it enhances our understanding of the complex interactions among interconnected variables. Consequently, un￾covering causal information through the analysi… view at source ↗
Figure 2
Figure 2. Overview of the proposed approach, composed of three main modules: (1) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Metapaths connecting raloxifene and melanoma: (1) [raloxifene-ERBB2-melanoma] with [drug-gene-disease] node sequence, and (2) [raloxifene-breast cancer-ERBB2-melanoma] with [drug-disease-gene-disease] node type sequence. task as classification task, classifying the relation as causal or non￾causal. We evaluate our approach on dataset D = {X, Y}, where X is a set of data instance and Y = {𝑐𝑎𝑢𝑠𝑎𝑙, 𝑛𝑜𝑛-𝑐𝑎𝑢𝑠𝑎𝑙 } is a se… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Subgraph example. Different node colors represent [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

67 extracted references · 32 canonical work pages

  1. [52]

    Masayuki Takayama, Tadahisa Okuda, Thong Pham, Tatsuyoshi Ikenoue, Shingo Fukuma, Shohei Shimizu, and Akiyoshi Sannai. 2024. Integrating Large Language Models in Causal Discovery: A Statistical Causal Approach. arXiv:2402.01454 [cs.LG] https://arxiv.org/abs/2402.01454

    arXiv 2024

  2. [1]

    Castro, and Daniel C

    Ahmed Abdulaal, adamos hadjivasiliou, Nina Montana-Brown, Tiantian He, Ayodeji Ijishakin, Ivana Drobnjak, Daniel C. Castro, and Daniel C. Alexander. 2024. Causal Modelling Agents: Causal Graph Discovery through Synergising Metadata- and Data-driven Reasoning. InICLR 2024 poster. International Conference on Learning Representations, Kigali, Rwanda, 123–130

    work page 2024

  3. [2]

    Jinheon Baek, Alham Fikri Aji, and Amir Saffari. 2023. Knowledge-Augmented Language Model Prompting for Zero-Shot Knowledge Graph Question Answer- ing. InProceedings of MATCHING 2023, Estevam Hruschka, Tom Mitchell, Sajjadur Rahman, Dunja Mladenić, and Marko Grobelnik (Eds.). Association for Computa- tional Linguistics, Toronto, ON, Canada, 70–98. doi:10....

    work page doi:10.18653/v1/2023.matching- 2023

  4. [3]

    Jinheon Baek, Soyeong Jeong, Minki Kang, Jong Park, and Sung Hwang. 2023. Knowledge-Augmented Language Model Verification. InProceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational Linguistics, Singapore, 1720–1736. doi:10.18653/v1/2023.emnlp-main.107

    doi:10.18653/v1/2023.emnlp-main.107 2023

  5. [4]

    Taiyu Ban, Lyvzhou Chen, Xiangyu Wang, and Huanhuan Chen. 2023. From Query Tools to Causal Architects: Harnessing Large Language Models for Ad- vanced Causal Discovery from Data. arXiv:2306.16902 [cs.AI] https://arxiv.org/ abs/2306.16902

    arXiv 2023

  6. [5]

    Christopher J. C. Burges. 2010.From RankNet to LambdaRank to LambdaMART: An Overview. Technical Report MSR-TR-2010-82. Microsoft Research

    work page 2010

  7. [6]

    2007.Learning to Rank: From Pairwise Approach to Listwise Approach

    Zhe Cao, Tao Qin, Tie-Yan Liu, Ming-Feng Tsai, and Hang Li. 2007.Learning to Rank: From Pairwise Approach to Listwise Approach. Technical Report MSR-TR- 2007-40. Microsoft Research. 9 pages

    work page 2007

  8. [7]

    Heng Chang, Jiangnan Ye, Alejo Lopez-Avila, Jinhua Du, and Jia Li. 2024. Path- based Explanation for Knowledge Graph Completion. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD ’24). ACM, 231–242. doi:10.1145/3637528.3671683

    arXiv 2024

Show all 67 references
  1. [8]

    Pulkit Chatwal, Amit Agarwal, and Ankush Mittal. 2025. Enhancing Causal Relationship Detection Using Prompt Engineering and Large Language Models. InProceedings of the Joint Workshop of the 9th Financial Technology and Natural Language Processing (FinNLP), the 6th Financial Na...

    2025

  2. [9]

    Tianqi Chen and Carlos Guestrin. 2016. XGBoost: A Scalable Tree Boosting System. InProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining(San Francisco, California, USA)(KDD ’16). Association for Computing Machinery, New York, NY, US...

    2016

  3. [10]

    Philipp Christmann, Rishiraj Saha Roy, and Gerhard Weikum. 2022. Beyond NED: Fast and Effective Search Space Reduction for Complex Question Answering over Knowledge Bases. InProceedings of the Fifteenth ACM International Conference on Web Search and Data Mining(Virtual Event, ...

    2022 doi

  4. [11]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2018. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. CoRRabs/1810.04805 (2018). arXiv:1810.04805 http://arxiv.org/abs/1810.04805

    2018 arXiv

  5. [12]

    Keith, Emaad Manzoor, Reid Pryzant, Dhanya Sridhar, Zach Wood-Doughty, Jacob Eisenstein, Justin Grimmer, Roi Reichart, Margaret E

    Amir Feder, Katherine A. Keith, Emaad Manzoor, Reid Pryzant, Dhanya Sridhar, Zach Wood-Doughty, Jacob Eisenstein, Justin Grimmer, Roi Reichart, Margaret E. Roberts, Brandon M. Stewart, Victor Veitch, and Diyi Yang. 2022. Causal Infer- ence in Natural Language Processing: Estim...

    2022 doi

  6. [13]

    Jinglong Gao, Xiao Ding, Bing Qin, and Ting Liu. 2023. Is ChatGPT a Good Causal Reasoner? A Comprehensive Evaluation. InFindings of the Association for Computational Linguistics: EMNLP 2023, Houda Bouamor, Juan Pino, and Kalika Bali (Eds.). Association for Computational Lingui...

    2023 doi

  7. [14]

    Luyu Gao, Zhuyun Dai, and Jamie Callan. 2021. Rethink Training of BERT Rerankers in Multi-stage Retrieval Pipeline. InAdvances in Information Retrieval: 43rd European Conference on IR Research, ECIR 2021, Virtual Event, March 28 – April 1, 2021, Proceedings, Part II. Springer-...

    2021 doi

  8. [15]

    Andrea Treviño Gavito, Diego Klabjan, and Jean Utke. 2023. Gradient-Boosted Based Structured and Unstructured Learning. arXiv:2302.14299 [cs.LG] https: //arxiv.org/abs/2302.14299

    2023 arXiv

  9. [16]

    Harsha Gurulingappa, Abdul Mateen Rajput, Angus Roberts, Juliane Fluck, Martin Hofmann-Apitius, and Luca Toldo. 2012. Development of a benchmark corpus to support the automatic extraction of drug-related adverse effects from medical case reports.Journal of Biomedical Informati...

    2012 doi

  10. [17]

    Iris Hendrickx, Su Nam Kim, Zornitsa Kozareva, Preslav Nakov, Diarmuid Ó Séaghdha, Sebastian Padó, Marco Pennacchiotti, Lorenza Romano, and Stan Szpakowicz. 2010. SemEval-2010 Task 8: Multi-Way Classification of Semantic Relations between Pairs of Nominals. InProceedings of th...

    2010

  11. [18]

    Daniel Scott Himmelstein, Antoine Lizee, Christine Hessler, Leo Brueggeman, Sabrina L Chen, Dexter Hadley, Ari Green, Pouya Khankhanian, and Sergio E Baranzini. 2017. Systematic integration of biomedical knowledge prioritizes drugs for repurposing.eLife6 (sep 2017), e26726. do...

    2017 doi

  12. [19]

    Ruixin Hong, Hongming Zhang, Hong Zhao, Dong Yu, and Changshui Zhang. 2023. Faithful Question Answering with Monte-Carlo Planning. arXiv:2305.02556 [cs.CL] https://arxiv.org/abs/2305.02556

    2023 arXiv

  13. [20]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, De- vendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thoma...

    2023 arXiv

  14. [21]

    Xinke Jiang, Ruizhe Zhang, Yongxin Xu, Rihong Qiu, Yue Fang, Zhiyuan Wang, Jinyi Tang, Hongxin Ding, Xu Chu, Junfeng Zhao, and Yasha Wang. 2024. HyKGE: A Hypothesis Knowledge Graph Enhanced Framework for Accurate and Reliable Medical LLMs Responses. arXiv:2312.15883 https://ar...

    2024 arXiv

  15. [22]

    Thomas Jiralerspong, Xiaoyin Chen, Yash More, Vedant Shah, and Yoshua Ben- gio. 2024. Efficient Causal Graph Discovery Using Large Language Models. arXiv:2402.01207 [cs.LG] https://arxiv.org/abs/2402.01207

    2024 arXiv

  16. [23]

    Minki Kang, Jin Myung Kwak, Jinheon Baek, and Sung Ju Hwang. 2023. Knowl- edge Graph-Augmented Language Models for Knowledge-Grounded Dialogue Generation. arXiv:2305.18846 [cs.CL] https://arxiv.org/abs/2305.18846

    2023 arXiv

  17. [24]

    Vivek Khetan, Md Imbesat Rizvi, Jessica Huber, Paige Bartusiak, Bogdan Sacaleanu, and Andrew Fano. 2022. MIMICause: Representation and automatic extraction of causal relation types from clinical notes. InFindings of the Associ- ation for Computational Linguistics: ACL 2022. As...

    2022 doi

  18. [25]

    Emre Kıcıman, Robert Ness, Amit Sharma, and Chenhao Tan. 2023. Causal Reasoning and Large Language Models: Opening a New Frontier for Causality

    2023

  19. [26]

    Hee-Jin Lee, Sang-Hyung Shim, Mi-Ryoung Song, Hyunju Lee, and Jong C. Park

  20. [27]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. InProceedings o...

    2020

  21. [28]

    Chen Li, Weiqi Wang, Jingcheng Hu, Yixuan Wei, Nanning Zheng, Han Hu, Zheng Zhang, and Houwen Peng. 2024. Common 7B Language Models Already Possess Strong Math Capabilities. arXiv:2403.04706 https://arxiv.org/abs/2403.04706

    2024 arXiv

  22. [29]

    Tie-Yan Liu. 2009. Learning to Rank for Information Retrieval.Found. Trends Inf. Retr.3, 3 (mar 2009), 225–331. doi:10.1561/1500000016

    2009 doi

  23. [30]

    Linhao Luo, Yuan-Fang Li, Gholamreza Haffari, and Shirui Pan. 2024. Reasoning on Graphs: Faithful and Interpretable Large Language Model Reasoning. In International Conference on Learning Representations (ICLR)

    2024

  24. [31]

    Meta. [n. d.]. Introducing Meta Llama 3: The most capable openly available LLM to date. https://ai.meta.com/blog/meta-llama-3/. [Accessed 29-07-2024]

    2024

  25. [32]

    Rodrigo Nogueira and Kyunghyun Cho. 2019. Passage Re-ranking with BERT. ArXivabs/1901.04085 (2019)

    2019 arXiv

  26. [33]

    Ayush Noori, Michelle M Li, Amelia L M Tan, and Marinka Zitnik. 2023. Metap- aths: similarity search in heterogeneous knowledge graphs via meta-paths.Bioin- formatics39, 5 (05 2023), btad297. https://doi.org/10.1093/bioinformatics/btad297

    2023 doi

  27. [34]

    Shirui Pan, Linhao Luo, Yufei Wang, Chen Chen, Jiapu Wang, and Xindong Wu

  28. [35]

    Fabio Petroni, Patrick S. H. Lewis, Aleksandra Piktus, Tim Rocktäschel, Yuxiang Wu, Alexander H. Miller, and Sebastian Riedel. 2020. How Context Affects KDD ’25, August 3–7, 2025, Toronto, ON, Canada Yuni Susanti and Michael Färber Language Models’ Factual Predictions.. InAKBC...

    2020

  29. [36]

    Przemyslaw Pobrotyn, Tomasz Bartczak, Mikolaj Synowiec, Radoslaw Bialo- brzeski, and Jaroslaw Bojar. 2020. Context-Aware Learning to Rank with Self- Attention. InProceedings of ACM SIGIR Workshop on eCommerce (SIGI ReCom (China (Virtual Event))

    2020

  30. [37]

    Zhen Qin, Le Yan, Honglei Zhuang, Yi Tay, Rama Kumar Pasumarthi, Xuanhui Wang, Mike Bendersky, and Marc Najork. 2021. Are Neural Rankers still Out- performed by Gradient Boosted Decision Trees?. InInternational Conference on Learning Representations (ICLR)

    2021

  31. [38]

    Priyanka Ranade and Anupam Joshi. 2024. FABULA: Intelligence Report Genera- tion Using Retrieval-Augmented Narrative Construction. InProceedings of the 2023 IEEE/ACM International Conference on Advances in Social Networks Analy- sis and Mining(Kusadasi, Turkiye)(ASONAM ’23). A...

    2024

  32. [39]

    Nils Reimers and Iryna Gurevych. 2019. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. InProceedings of the 2019 Conference on Em- pirical Methods in Natural Language Processing. Association for Computational Linguistics. http://arxiv.org/abs/1908.10084

    2019 arXiv

  33. [40]

    Zhong-Hao Ren, Zhu-Hong You, Quan Zou, Chang-Qing Yu, Yan-Fang Ma, Yong- Jian Guan, Hai-Ru You, Xin-Fei Wang, and Jie Pan. 2023. DeepMPF: deep learning framework for predicting drug–target interactions based on multi-modal repre- sentation with meta-path semantic analysis.Jour...

    2023 doi

  34. [41]

    Laria Reynolds and Kyle McDonell. 2021. Prompt Programming for Large Lan- guage Models: Beyond the Few-Shot Paradigm. InExtended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems(Yokohama, Japan)(CHI EA ’21). Association for Computing Machinery, New Yo...

    2021

  35. [42]

    Lauffenburger, and Garry P

    Karen Sachs, Omar Perez, Dana Pe’er, Douglas A. Lauffenburger, and Garry P. Nolan. 2005. Causal Protein-Signaling Networks Derived from Multiparameter Single-Cell Data.Science308, 5721 (2005), 523–529. doi:10.1126/science.1105809 arXiv:https://www.science.org/doi/pdf/10.1126/s...

    2005 doi

  36. [43]

    Weizhou Shen, Yingqi Gao, Canbin Huang, Fanqi Wan, Xiaojun Quan, and Wei Bi. 2023. Retrieval-Generation Alignment for End-to-End Task-Oriented Dialogue System. InProceedings of the 2023 Conference on Empirical Meth- ods in Natural Language Processing, Houda Bouamor, Juan Pino,...

    2023 doi

  37. [44]

    Hoyer, and Kenneth Bollen

    Shohei Shimizu, Takanori Inazumi, Yasuhiro Sogawa, Aapo Hyvärinen, Yoshi- nobu Kawahara, Takashi Washio, Patrik O. Hoyer, and Kenneth Bollen. 2011. DirectLiNGAM: A Direct Method for Learning a Linear Non-Gaussian Structural Equation Model.J. Mach. Learn. Res.12, null (July 201...

    2011

  38. [45]

    Peter Spirtes, Clark Glymour, and Scheines R. 2000. Constructing bayesian net- works models of gene expression networks from microarray data. InProceedings of the Atlantic Symposium on Computational Biology(North Carolina)

    2000

  39. [46]

    Spirtes, C

    P. Spirtes, C. Glymour, and R. Scheines. 2001.Causation, Prediction, and Search (2nd ed.). MIT press

    2001

  40. [47]

    2001.Causation, Prediction, and Search

    Peter Spirtes, Clark Glymour, and Richard Scheines. 2001.Causation, Prediction, and Search. The MIT Press. doi:10.7551/mitpress/1754.001.0001

    2001 doi

  41. [48]

    Weiwei Sun, Lingyong Yan, Xinyu Ma, Shuaiqiang Wang, Pengjie Ren, Zhumin Chen, Dawei Yin, and Zhaochun Ren. 2023. Is ChatGPT Good at Search? In- vestigating Large Language Models as Re-Ranking Agents. InProceedings of the 2023 Conference on Empirical Methods in Natural Languag...

    2023 doi

  42. [49]

    Yu, and Tianyi Wu

    Yizhou Sun, Jiawei Han, Xifeng Yan, Philip S. Yu, and Tianyi Wu. 2011. PathSim: meta path-based top-K similarity search in heterogeneous information networks. Proc. VLDB Endow.4, 11 (aug 2011), 992–1003. doi:10.14778/3402707.3402736

    2011

  43. [50]

    Yuni Susanti and Michael Färber. 2024. Knowledge Graph Structure as Prompt: Improving Small Language Models Capabilities for Knowledge-based Causal Discovery. InProceedings of the ISWC’24(Baltimore, USA)

    2024

  44. [51]

    Yuni Susanti and Kanji Uchino. 2024. Causal-Evidence Graph for Causal Relation Classification. InProceedings of the 39th ACM/SIGAPP Symposium on Applied Computing(Avila, Spain)(SAC ’24). Association for Computing Machinery, New York, NY, USA, 714–722. doi:10.1145/3605098.3635894

    2024

  45. [53]

    Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupati- raju, Shreya Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juli- ette Love, Pouya Tafti, Léonard Hussenot, Pier Giuseppe Sessa, Aakanksha Chowdhery, Adam Roberts, Aditya Barua, Alex Botev, ...

    2024 arXiv

  46. [54]

    Hiren Thakkar and A Manimaran. 2023. Comprehensive Examination of Instruction-Based Language Models: A Comparative Analysis of Mistral-7B and Llama-2-7B. In2023 International Conference on Emerging Research in Computa- tional Science (ICERCS). 1–6. doi:10.1109/ICERCS57948.2023...

    2023

  47. [55]

    Ruibo Tu, Chao Ma, and Cheng Zhang. 2023. Causal-Discovery Performance of ChatGPT in the context of Neuropathic Pain Diagnosis. arXiv:2301.13819 [cs.CL]

    2023 arXiv

  48. [56]

    Denny Vrandečić and Markus Krötzsch. 2014. Wikidata: a free collaborative knowledgebase.Commun. ACM57, 10 (sep 2014), 78–85. https://doi.org/10. 1145/2629489

    2014

  49. [58]

    Keheng Wang, Feiyu Duan, Sirui Wang, Peiguang Li, Yunsen Xian, Chuantao Yin, Wenge Rong, and Zhang Xiong. 2023. Knowledge-Driven CoT: Explor- ing Faithful Reasoning in LLMs for Knowledge-intensive Question Answering. arXiv:2308.13259 [cs.CL] https://arxiv.org/abs/2308.13259

    2023 arXiv

  50. [59]

    Xintao Wang, Qianwen Yang, Yongting Qiu, Jiaqing Liang, Qianyu He, Zhouhong Gu, Yanghua Xiao, and Wei Wang. 2023. KnowledGPT: Enhancing Large Language Models with Retrieval and Storage Access on Knowledge Bases. arXiv:2308.11761 [cs.CL] https://arxiv.org/abs/2308.11761

    2023 arXiv

  51. [60]

    Moritz Willig, Matej Zečević, Devendra Singh Dhami, and Kristian Kersting. 2022. Can Foundation Models Talk Causality? arXiv:2206.10591 [cs.AI]

    2022 arXiv

  52. [61]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mari...

    2020

  53. [62]

    Wenjie Yao, Weizhong Zhao, Xingpeng Jiang, Xianjun Shen, and Tingting He

  54. [63]

    Changhe Yuan and Brandon Malone. 2013. Learning optimal bayesian networks: a shortest path perspective.J. Artif. Int. Res.48, 1 (Oct. 2013), 23–65

    2013

  55. [64]

    Cheng Zhang, Stefan Bauer, Paul Bennett, Jiangfeng Gao, Wenbo Gong, Agrin Hilmkil, Joel Jennings, Chao Ma, Tom Minka, Nick Pawlowski, and James Vaughan. 2023. Understanding Causality with Large Language Models: Fea- sibility and Opportunities. arXiv:2304.05524 [cs.LG]

    2023 arXiv

  56. [65]

    Shichang Zhang, Jiani Zhang, Xiang Song, Soji Adeshina, Da Zheng, Christos Faloutsos, and Yizhou Sun. 2023. PaGE-Link: Path-based Graph Neural Network Explanation for Heterogeneous Link Prediction. arXiv:2302.12465 [cs.LG] https: //arxiv.org/abs/2302.12465

    2023 arXiv

  57. [2013]

    BMC Bioinformatics14, 1 (Nov 2013), 323

    CoMAGC: a corpus with multi-faceted annotations of gene-cancer relations. BMC Bioinformatics14, 1 (Nov 2013), 323. https://doi.org/10.1186/1471-2105-14- 323

    2013 doi

  58. [2022]

    InIEEE International Conference on Bioinformatics and Biomedicine, BIBM 2022, Las Vegas, NV, USA, December 6-8, 2022, Donald A

    MPGNN-DSA: A Meta-path-based Graph Neural Network for drug-side effect association prediction. InIEEE International Conference on Bioinformatics and Biomedicine, BIBM 2022, Las Vegas, NV, USA, December 6-8, 2022, Donald A. Adjeroh, Qi Long, Xinghua Mindy Shi, Fei Guo, Xiaohua ...

    2022

  59. [2024]

    IEEE Transactions on Knowledge and Data Engineering36, 7 (July 2024), 3580–3599

    Unifying Large Language Models and Knowledge Graphs: A Roadmap. IEEE Transactions on Knowledge and Data Engineering36, 7 (July 2024), 3580–3599. doi:10.1109/tkde.2024.3352100

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.