Pith. sign in

REVIEW 3 major objections 4 minor 37 references

Meta Reasoning over Knowledge Graphs

T0 review · 3 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Task-specific meta-encoding improves few-shot reasoning over knowledge graphs beyond MAML.

desk verdict A plausible meta-encoder for few-shot KG reasoning, but the evaluation is curated toward positive-transfer tasks, so the general claims are not yet earned. read the letter →

arxiv 1908.04877 v1 pith:XMKXEN2L submitted 2019-08-13 cs.CL

Hong Wang , Wenhan Xiong , Mo Yu , Xiaoxiao Guo , Shiyu Chang , William Yang Wang This is my paper · ORCID
classification cs.CL
keywords few-shotlearningmeta-learningknowledgegraphreasoningmulti-hopbasecompletiontask-specificinitializationMAMLencoders
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 tackles few-shot knowledge base completion, where a new relation must be reasoned over from only a handful of example triples. It argues that meta-learning can do much better if the model is initialized not with one global starting point, but with a task-specific one: an encoder reads the local graph neighbors (or reasoning paths) of the support triples and produces the parameters that will be fine-tuned. On two benchmarks, FB15K-237 and NELL, this task-specific initialization gives a stronger starting point than MAML and leads to higher Hits@1, Hits@10, and MRR after fine-tuning. A sympathetic reader would take the intended contribution to be that task-identity information, expressed through the graph structure around a few examples, is what makes few-shot KG reasoning work.

What carries the argument

The load-bearing object is the meta-encoder, an auxiliary network that maps the few support triples of a new relation into a task representation used as the initial query embedding of the reasoning agent. Two instantiations are defined: the neighbor encoder averages learned feature vectors over each entity's local neighborhood and takes their difference as the relation vector, borrowing the translation idea from TransE; the path encoder runs an LSTM over each successful reasoning path between the support entities and averages the resulting path embeddings. The meta-encoder makes the initial model depend on the actual content of the new task, so different relations start from different points, but a newly seen relation can still be encoded without any pre-existing embedding. This is what carries the claimed improvement over MAML, which starts every task from the same parameters.

What would settle it

Construct a set of few-shot relations whose support triples connect entities with no shared neighbors and no connecting paths in the background graph. If the neighbor and path encoders give initial predictions no better than a randomly initialized reasoner on these relations, then the advantage of the meta-encoder is not coming from the task-specific information, and the paper's central claim is refuted.

Watch

Extended reading notes

Core claim

The central claim is that encoding task-specific meta-information into the initial parameters of a multi-hop reasoning agent improves few-shot knowledge graph reasoning. For each relation, treated as a task, two encoders are proposed: a neighbor encoder that represents the relation as the TransE-style difference between the average neighbor embeddings of the start and end entities ($R_r = NE_{e_t} - NE_{e_s}$), and a path encoder that averages LSTM embeddings of reasoning paths connecting the two entities. Both produce a task representation that replaces the query-relation embedding in a MINERVA-style policy, and the whole system is trained with a MAML-like two-loop gradient procedure, using a first-order update. Experiments on 5-shot tasks show that the neighbor encoder is best on FB15K-237 while the path encoder is best on the sparser NELL, and that both give a much better initial point than MAML, Transfer, and Random baselines before any fine-tuning.

Load-bearing premise

The method assumes that the handful of support triples for a new relation reveal enough of its meaning from the local graph around the example entities or from a connecting reasoning path; if neither is informative (as on very sparse graphs), the claimed better initialization disappears.

Editorial extensions

If this is right

  • Unseen relations can be handled at meta-test time without learning an embedding for the new relation; the meta-encoder's output on the support triples supplies the task identity.
  • Dense and sparse graphs benefit from different kinds of meta-information: the neighbor encoder wins on FB15K-237, while the path encoder wins on NELL, where the median entity degree is 1.
  • The advantage is visible immediately at initialization, before any fine-tuning, and the gap persists after fine-tuning on the 5-shot tasks.
  • Even a single support triple gives a meaningful task representation: the ablation shows the encoder with 1 sample beats a randomly initialized model by a large margin on FB15K-237.

Reading between the lines

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

  • An untested hybrid that selects or combines neighbor and path encoders based on local graph density could outperform either alone, since the two succeed in complementary regimes.
  • The same meta-encoder trick should transfer to other few-shot structured prediction settings where task identity is unavailable at test time, such as relation extraction or link prediction over temporal or heterogeneous graphs.
  • The TransE-style subtraction assumes the difference of neighbor embeddings behaves like a translation vector for the relation; on graphs with more varied relational structure, composition operations from other embedding models (rotation, complex product) might be swapped in and tested.
  • The performance plateau for MAML beyond 10 shots in Figure 2 hints that the initialization, not the fine-tuning budget, is the main bottleneck; this could be tested by comparing adaptation curves across few-shot sizes.
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

3 major / 4 minor

Summary. The paper proposes Meta Reasoning over Knowledge Graphs, a meta-learning framework for few-shot multi-hop reasoning. Each query relation defines a task; the model uses a MINERVA-style policy to predict an end entity and a reasoning path from a start entity. To avoid MAML's single shared initialization, the authors introduce a meta-encoder that maps task-specific meta-information (neighbor entity embeddings or reasoning paths) from the support set into a task-dependent initialization for the policy. They present two encoders: a neighbor encoder using the difference of aggregated neighbor embeddings (TransE-style) and a path encoder using an LSTM over successful reasoning paths. The method is evaluated on FB15K-237 and NELL under a 5-shot setting, reporting final and initial MRR/Hits@k against Random, Transfer, MAML, and MAML-Mask baselines, plus ablations and few-shot-size analysis.

Significance. If the empirical claims are robust, the paper makes a useful contribution by showing that task-specific meta-information can be encoded into a MAML-style initialization for KG reasoning, improving both the starting point and final few-shot performance. The two encoders are simple and intuitive, and the ablation (Table 3) and few-shot-size study (Figure 2) provide useful analysis. However, the evaluation is restricted to hand-picked positive-transfer tasks, which biases the comparison; the central claim of a general improvement over MAML therefore needs additional validation on representative task splits. The paper is clearly written and the algorithm is reproducible from the description, though no code or machine-checked proofs are provided.

major comments (3)
  1. [Section 5.1] The meta-dev/meta-test tasks are selected for positive transfer (gains of at least 0.04 on FB15K-237 and 0.02 on NELL), leaving only 5/10 and 4/7 relations for evaluation. Since the paper's central claim is that the meta-encoder provides a generally better starting point than MAML, benchmarking only on tasks known a priori to benefit from multi-task training biases the comparison in favor of the proposed method; the large reported gains (e.g., Initial MRR 0.056 vs 0.010 on FB15K-237, Table 2) may not hold for typical long-tail relations that dominate KG completion. Please evaluate on all held-out relations or a random split, or justify why the selected subset is representative.
  2. [Table 2 and Figure 2] No error bars, confidence intervals, or significance tests are reported anywhere in the experimental section. With only 5 (FB15K-237) and 4 (NELL) meta-test tasks, the observed differences between methods (e.g., Neighbor MRR 0.080 vs Path 0.060 on FB15K-237) may be within noise. Report variance over multiple runs or bootstrap over tasks, and state whether the improvements over the baselines are statistically reliable.
  3. [Section 4.2.2 and Table 2] The claimed advantage is not uniform across the two encoders: on FB15K-237 the neighbor encoder is best (MRR 0.080) while on NELL the path encoder is best (MRR 0.137), and the paper does not provide a principled criterion for choosing the encoder for a new task. Since the meta-information sources are only reliable in different regimes (neighbors require sufficient degree, paths require connectivity between support entities), the central claim of a better initial point should be qualified by the conditions under which each encoder is applicable; otherwise the method may default to MAML-like behavior in sparse or poorly connected graphs.
minor comments (4)
  1. [Section 5.3] The sentence "We also note that path encoder outperforms other baselines on FB15K-237" is ambiguous: Table 2 shows Neighbor MRR 0.080 vs Path MRR 0.060, so "other baselines" must mean the non-encoder baselines (Random, Transfer, MAML, MAML-Mask), not the neighbor encoder. Please rephrase to avoid the apparent contradiction.
  2. [Section 5.1] The description that thresholds 0.04 and 0.02 are "carefully chosen" so that enough tasks with positive transfer remain suggests threshold tuning on the evaluation tasks; please specify how the thresholds were selected and whether meta-dev tasks were involved, to rule out selection bias in the threshold choice.
  3. [Section 4.2.2] The notation "Cri,ei" should be typeset as C_{r_i,e_i}, and the dimensions of W_c and b_c should be defined for completeness.
  4. [Introduction] In the first section, "since the presentation of the task is available" should read "representation of the task".

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the meta-encoder conditions on the support set and is evaluated on held-out queries, so the claimed gains are empirical rather than built into the equations.

full rationale

The derivation chain is self-contained against external benchmarks. The task representation is computed from the few-shot support triples (neighbor difference or encoded paths) and then used as initialization, but all reported metrics are on held-out meta-test queries with unseen end entities; the support labels are not re-served as test predictions. This is the standard meta-learning evaluation protocol, not a fitted-input-called-prediction cycle. The neighbor-encoder component is borrowed from Xiong et al. (2018), an independently published prior work by overlapping authors, but it is used here as an architectural building block rather than as the justification for the central claim; the central claim is supported by the new FB15K-237 and NELL experiments. The explicit task-selection procedure in Section 5.1, which keeps relations with positive transfer, could bias the evaluation toward tasks favorable to multi-task methods, but it does not force the meta-encoder's advantage over MAML and is therefore an experimental-design concern, not circularity. No equation reduces to its own input by construction, and no fitted parameter is renamed as a prediction. The 'better initial point' claim is an empirical comparison in Table 2 between a support-conditioned initial function and an unconditioned MAML initializer, which is exactly the proposed method rather than a tautology.

Assumptions & free parameters 1 free parameters · 5 assumptions · 0 invented entities

The method's central claim rests on the availability and representativeness of task-specific graph structure (neighbors or paths) for unseen relations, and on an evaluation set chosen for positive transfer.

free parameters (1)
  • Positive-transfer selection thresholds = 0.04 (FB15K-237), 0.02 (NELL)
    Chosen in Section 5.1 to include enough tasks with positive transfer; directly shapes the meta-dev/meta-test evaluation sets and can inflate apparent transfer gains.
assumptions (5)
  • domain assumption TransE-style translational property holds on neighbor embeddings: Rr = NE_et - NE_es represents the query relation.
    Section 4.2.2, neighbor encoder. If neighbor embeddings do not encode relational translation, the task initialization is meaningless.
  • domain assumption The support set (few-shot examples) is representative enough to compute a valid task representation via neighbor or path encoding.
    Sections 4.2.2 and 5.4; with 1-shot on sparse graphs the method degrades.
  • domain assumption The background graph contains successful reasoning paths between start and end entities of support triples.
    Path encoder Section 4.2.2 requires paths; on sparse graphs paths may not exist.
  • ad hoc to paper Meta-training tasks share structure with meta-test tasks (positive transfer).
    Section 5.1 selects tasks with positive transfer thresholds, presupposing transfer is beneficial for the selected evaluation.
  • domain assumption The MINERVA LSTM policy is an appropriate reasoning model and is differentiable through the action distribution.
    Section 4.2.1, borrowed from Das et al. 2018.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Meta Reasoning over Knowledge Graphs." pith.science (2026). https://pith.science/paper/XMKXEN2L

@misc{pith2026190804877,
  author       = {Pith},
  title        = {Pith review of: Meta Reasoning over Knowledge Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XMKXEN2L}},
  note         = {Machine review of arXiv:1908.04877}
}
read the original abstract

The ability to reason over learned knowledge is an innate ability for humans and humans can easily master new reasoning rules with only a few demonstrations. While most existing studies on knowledge graph (KG) reasoning assume enough training examples, we study the challenging and practical problem of few-shot knowledge graph reasoning under the paradigm of meta-learning. We propose a new meta learning framework that effectively utilizes the task-specific meta information such as local graph neighbors and reasoning paths in KGs. Specifically, we design a meta-encoder that encodes the meta information into task-specific initialization parameters for different tasks. This allows our reasoning module to have diverse starting points when learning to reason over different relations, which is expected to better fit the target task. On two few-shot knowledge base completion benchmarks, we show that the augmented task-specific meta-encoder yields much better initial point than MAML and outperforms several few-shot learning baselines.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references · 22 canonical work pages

  1. [1]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    S \" o ren Auer, Christian Bizer, Georgi Kobilarov, Jens Lehmann, Richard Cyganiak, and Zachary G. Ives. 2007. https://doi.org/10.1007/978-3-540-76298-0\_52 Dbpedia: A nucleus for a web of open data . In The Semantic Web, 6th International Semantic Web Conference, 2nd Asian Semantic Web Conference, ISWC 2007 + ASWC 2007, Busan, Korea, November 11-15, 2007...

    doi:10.1007/978-3-540-76298-0 2007

  4. [4]

    Bollacker, Colin Evans, Praveen Paritosh, Tim Sturge, and Jamie Taylor

    Kurt D. Bollacker, Colin Evans, Praveen Paritosh, Tim Sturge, and Jamie Taylor. 2008. https://doi.org/10.1145/1376616.1376746 Freebase: a collaboratively created graph database for structuring human knowledge . In Proceedings of the ACM SIGMOD International Conference on Management of Data, SIGMOD 2008, Vancouver, BC, Canada, June 10-12, 2008 , pages 1247...

    arXiv 2008

  5. [5]

    Antoine Bordes, Nicolas Usunier, Sumit Chopra, and Jason Weston. 2015. http://arxiv.org/abs/1506.02075 Large-scale simple question answering with memory networks . CoRR, abs/1506.02075

    arXiv 2015

  6. [6]

    Antoine Bordes, Nicolas Usunier, Alberto Garc \' a - Dur \' a n, Jason Weston, and Oksana Yakhnenko. 2013. http://papers.nips.cc/paper/5071-translating-embeddings-for-modeling-multi-relational-data Translating embeddings for modeling multi-relational data . In Advances in Neural Information Processing Systems 26: 27th Annual Conference on Neural Informati...

    work page 2013

  7. [7]

    Wei - Yu Chen, Yen - Cheng Liu, Zsolt Kira, Yu - Chiang Frank Wang, and Jia - Bin Huang. 2019. http://arxiv.org/abs/1904.04232 A closer look at few-shot classification . CoRR, abs/1904.04232

    arXiv 2019

  8. [8]

    Wenhu Chen, Wenhan Xiong, Xifeng Yan, and William Yang Wang. 2018. https://aclanthology.info/papers/N18-1165/n18-1165 Variational knowledge graph reasoning . In Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, NAACL-HLT 2018, New Orleans, Louisiana, USA, June 1-...

    work page 2018

Show all 37 references
  1. [9]

    Rajarshi Das, Shehzaad Dhuliawala, Manzil Zaheer, Luke Vilnis, Ishan Durugkar, Akshay Krishnamurthy, Alex Smola, and Andrew McCallum. 2018. https://openreview.net/forum?id=Syg-YfWCW Go for a walk and arrive at the answer: Reasoning over paths in knowledge bases using reinforce...

    2018

  2. [10]

    Chelsea Finn, Pieter Abbeel, and Sergey Levine. 2017. http://proceedings.mlr.press/v70/finn17a.html Model-agnostic meta-learning for fast adaptation of deep networks . In Proceedings of the 34th International Conference on Machine Learning, ICML 2017, Sydney, NSW, Australia, 6...

    2017

  3. [11]

    Mitchell

    Matt Gardner, Partha Pratim Talukdar, Bryan Kisiel, and Tom M. Mitchell. 2013. http://aclweb.org/anthology/D/D13/D13-1080.pdf Improving learning and inference in a large knowledge-base using latent syntactic cues . In Proceedings of the 2013 Conference on Empirical Methods in ...

    2013

  4. [12]

    Mitchell

    Matt Gardner, Partha Pratim Talukdar, Jayant Krishnamurthy, and Tom M. Mitchell. 2014. http://aclweb.org/anthology/D/D14/D14-1044.pdf Incorporating vector space similarity in random walk inference over knowledge bases . In Proceedings of the 2014 Conference on Empirical Method...

    2014

  5. [13]

    Jiatao Gu, Yong Wang, Yun Chen, Victor O. K. Li, and Kyunghyun Cho. 2018. https://aclanthology.info/papers/D18-1398/d18-1398 Meta-learning for low-resource neural machine translation . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, B...

    2018

  6. [14]

    Sepp Hochreiter and J \" u rgen Schmidhuber. 1997. https://doi.org/10.1162/neco.1997.9.8.1735 Long short-term memory . Neural Computation, 9(8):1735--1780

    1997 doi

  7. [15]

    Po - Sen Huang, Chenglong Wang, Rishabh Singh, Wen - tau Yih, and Xiaodong He. 2018. https://aclanthology.info/papers/N18-2115/n18-2115 Natural language to structured query generation via meta-learning . In Proceedings of the 2018 Conference of the North American Chapter of th...

    2018

  8. [16]

    Ni Lao and William W. Cohen. 2010. https://doi.org/10.1007/s10994-010-5205-8 Relational retrieval using a combination of path-constrained random walks . Machine Learning, 81(1):53--67

    2010 doi

  9. [17]

    Xi Victoria Lin, Richard Socher, and Caiming Xiong. 2018. https://aclanthology.info/papers/D18-1362/d18-1362 Multi-hop knowledge graph reasoning with reward shaping . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brussels, Belgium, ...

    2018

  10. [18]

    Nikhil Mishra, Mostafa Rohaninejad, Xi Chen, and Pieter Abbeel. 2018. https://openreview.net/forum?id=B1DmUzWAW A simple neural attentive meta-learner . In 6th International Conference on Learning Representations, ICLR 2018, Vancouver, BC, Canada, April 30 - May 3, 2018, Confe...

    2018

  11. [19]

    Mitchell, William W

    Tom M. Mitchell, William W. Cohen, Estevam R. Hruschka Jr., Partha P. Talukdar, Bo Yang, Justin Betteridge, Andrew Carlson, Bhavana Dalvi Mishra, Matt Gardner, Bryan Kisiel, Jayant Krishnamurthy, Ni Lao, Kathryn Mazaitis, Thahir Mohamed, Ndapandula Nakashole, Emmanouil A. Plat...

    2018 doi

  12. [20]

    Arvind Neelakantan, Benjamin Roth, and Andrew McCallum. 2015. http://www.aaai.org/ocs/index.php/SSS/SSS15/paper/view/10254 Compositional vector space models for knowledge base inference . In 2015 AAAI Spring Symposia, Stanford University, Palo Alto, California, USA, March 22-2...

    2015

  13. [21]

    Alex Nichol, Joshua Achiam, and John Schulman. 2018. http://arxiv.org/abs/1803.02999 On first-order meta-learning algorithms . CoRR, abs/1803.02999

    2018 arXiv

  14. [22]

    Maximilian Nickel, Volker Tresp, and Hans - Peter Kriegel. 2011. https://icml.cc/2011/papers/438\_icmlpaper.pdf A three-way model for collective learning on multi-relational data . In Proceedings of the 28th International Conference on Machine Learning, ICML 2011, Bellevue, Wa...

    2011

  15. [23]

    Sachin Ravi and Hugo Larochelle. 2017. https://openreview.net/forum?id=rJY0-Kcll Optimization as a model for few-shot learning . In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings . OpenReview.net

    2017

  16. [24]

    Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell

    Andrei A. Rusu, Dushyant Rao, Jakub Sygnowski, Oriol Vinyals, Razvan Pascanu, Simon Osindero, and Raia Hadsell. 2018. http://arxiv.org/abs/1807.05960 Meta-learning with latent embedding optimization . CoRR, abs/1807.05960

    2018 arXiv

  17. [25]

    Yelong Shen, Jianshu Chen, Po - Sen Huang, Yuqing Guo, and Jianfeng Gao. 2018. http://papers.nips.cc/paper/7912-m-walk-learning-to-walk-over-graphs-using-monte-carlo-tree-search M-walk: Learning to walk over graphs using monte carlo tree search . In Advances in Neural Informat...

    2018

  18. [26]

    Flood Sung, Yongxin Yang, Li Zhang, Tao Xiang, Philip H. S. Torr, and Timothy M. Hospedales. 2018. https://doi.org/10.1109/CVPR.2018.00131 Learning to compare: Relation network for few-shot learning . In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 201...

    2018

  19. [27]

    Kristina Toutanova, Danqi Chen, Patrick Pantel, Hoifung Poon, Pallavi Choudhury, and Michael Gamon. 2015. http://aclweb.org/anthology/D/D15/D15-1174.pdf Representing text for joint embedding of text and knowledge bases . In Proceedings of the 2015 Conference on Empirical Metho...

    2015

  20. [28]

    Th \' e o Trouillon, Johannes Welbl, Sebastian Riedel, \' E ric Gaussier, and Guillaume Bouchard. 2016. http://jmlr.org/proceedings/papers/v48/trouillon16.html Complex embeddings for simple link prediction . In Proceedings of the 33nd International Conference on Machine Learni...

    2016

  21. [29]

    Denny Vrandecic and Markus Kr \" o tzsch. 2014. https://doi.org/10.1145/2629489 Wikidata: a free collaborative knowledgebase . Commun. ACM , 57(10):78--85

    2014 doi

  22. [30]

    William Yang Wang and William W. Cohen. 2015. http://aclweb.org/anthology/P/P15/P15-1035.pdf Joint information extraction and reasoning: A scalable statistical relational learning approach . In Proceedings of the 53rd Annual Meeting of the Association for Computational Linguis...

    2015

  23. [31]

    Jiawei Wu, Ruobing Xie, Zhiyuan Liu, and Maosong Sun. 2016. http://arxiv.org/abs/1609.07075 Knowledge representation via joint learning of sequential text and knowledge graphs

    2016 arXiv

  24. [32]

    Wenhan Xiong, Thien Hoang, and William Yang Wang. 2017. https://aclanthology.info/papers/D17-1060/d17-1060 Deeppath: A reinforcement learning method for knowledge graph reasoning . In Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, EMNLP...

    2017

  25. [33]

    Wenhan Xiong, Mo Yu, Shiyu Chang, Xiaoxiao Guo, and William Yang Wang. 2018. https://aclanthology.info/papers/D18-1223/d18-1223 One-shot relational learning for knowledge graphs . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, Brusse...

    2018

  26. [34]

    Bishan Yang, Wen - tau Yih, Xiaodong He, Jianfeng Gao, and Li Deng. 2015. http://arxiv.org/abs/1412.6575 Embedding entities and relations for learning and inference in knowledge bases . In 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA,...

    2015 arXiv

  27. [35]

    Xuchen Yao and Benjamin Van Durme. 2014. http://aclweb.org/anthology/P/P14/P14-1090.pdf Information extraction over structured data: Question answering with freebase . In Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, ACL 2014, June 22...

    2014

  28. [36]

    Wen - tau Yih, Ming - Wei Chang, Xiaodong He, and Jianfeng Gao. 2015. http://aclweb.org/anthology/P/P15/P15-1128.pdf Semantic parsing via staged query graph generation: Question answering with knowledge base . In Proceedings of the 53rd Annual Meeting of the Association for Co...

    2015

  29. [37]

    Mo Yu, Wenpeng Yin, Kazi Saidul Hasan, C \' cero Nogueira dos Santos, Bing Xiang, and Bowen Zhou. 2017. https://doi.org/10.18653/v1/P17-1053 Improved neural relation detection for knowledge base question answering . In Proceedings of the 55th Annual Meeting of the Association ...

    2017 doi

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.