Pith. sign in

REVIEW 3 major objections 5 minor 21 references

Combining semantic topic vectors with graph-based structural vectors produces sharply separated, interpretable clusters for unlabeled legal corpora.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

Concatenating Top2Vec and Node2Vec embeddings, where the Node2Vec graph encodes Top2Vec's own topic labels, yields compact clusters, but the gain is largely circular.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A cleanly-written pipeline with a circular graph construction: Node2Vec re-encodes Top2Vec's own topic labels, so the headline internal-metric gains are not evidence of synergy. the 3 major comments →

arxiv 2509.00990 v1 pith:OCUVPBVG submitted 2025-08-31 stat.ML cs.CLcs.LG

Hybrid Topic-Semantic Labeling and Graph Embeddings for Unsupervised Legal Document Clustering

Deepak Bastola , Woohyeok Choi This is my paper
classification stat.ML cs.CLcs.LG
keywords unsupervised clusteringlegal document clusteringTop2VecNode2Vecbipartite document–topic graphKMeansUMAPtopic modeling
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper tries to establish that legal documents can be clustered into coherent, well-separated topics without any labels by combining two embedding views: semantic vectors from topic modeling and structural vectors from a graph of document–topic relationships. The hybrid pipeline reportedly outclasses text-only and graph-only baselines on internal clustering metrics, with Silhouette 0.927, Davies-Bouldin 0.111, and Calinski-Harabasz 29,186 on the ACORD contract dataset. The reason to care is that legal corpora are large, jargon-heavy, and rarely labeled, so an unsupervised method that yields interpretable, tight clusters could support e-discovery, compliance checks, and legal research. The paper also shows that most of the quality is retained on 70–80% of the data, a 'representational saturation' that reduces computation, and that external alignment with ground-truth categories is modest, so the clusters are clean but not the same as human labels.

Core claim

Central claim: concatenating Top2Vec semantic embeddings with Node2Vec embeddings of a bipartite document–topic graph, then clustering with KMeans, gives legal document clusters much more compact and well separated than LDA, NMF, or TF-IDF. On ACORD the authors report Silhouette 0.927, DBI 0.111, CHS 29,186; the ablation shows the jump from semantic-only (0.685 Silhouette) to hybrid (0.927) is driven by Node2Vec. Because each document node connects to its dominant Top2Vec topic node, random walks through the shared hub pull co-topic documents together, so the graph step acts as a structural refiner. The same clusters align only modestly with ground-truth labels (NMI 0.153, ARI 0.051), making

What carries the argument

The load-bearing object is the bipartite document–topic graph G=(V,E), with document nodes and topic nodes; an edge connects a document to the single topic Top2Vec assigned it as dominant. Node2Vec's biased random walks on this graph are the mechanism: a walk frequently goes document → topic hub → another document of the same topic, so documents sharing a topic end up near each other in the graph embedding. Concatenating those vectors with Top2Vec's semantic vectors and clustering with KMeans yields the reported separations. The implicit assumption is that the graph encodes topic co-membership as structure rather than duplicating the semantic signal.

Load-bearing premise

The load-bearing premise is that a graph whose edges are just each document's dominant Top2Vec topic assignment carries structural information beyond the semantic embedding that produced those assignments; if it does not, Node2Vec merely re-labels Top2Vec's output, and the internal-metric gains are built into the graph construction.

What would settle it

Run the same pipeline with a placebo graph in which each document is linked to a random or shuffled topic node; if Silhouette and CHS stay near 0.927, the gains are an artifact of edge construction. Alternatively, replace the topic-assignment edges with edges from an independent source (e.g., contract cross-references or shared clauses); if clustering quality then still beats text-only, the graph contributes real structural signal.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • Because the hybrid pipeline needs no labels, it can be applied directly to large legal corpora for exploratory topic discovery, e-discovery triage, and compliance review.
  • The reported representational saturation—cluster quality at 70–80% of the data matches 100%—means practitioners can subsample the corpus and cut computation with little loss in cluster quality.
  • The modest NMI (0.153) and ARI (0.051) imply the clusters should be treated as internally consistent themes, not as a substitute for human legal categories; they are a scaffolding for review, not a final judgment.
  • The ablation's large internal-metric jump when Node2Vec is added indicates the graph step compacts clusters, which suggests the pipeline could improve other unsupervised tasks that benefit from tight groupings.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • My reading: because the bipartite graph's edges come directly from Top2Vec's dominant-topic assignments, the 'structural' signal is a deterministic re-encoding of the semantic topic model; the synergy may be a smoothing of the same signal rather than evidence that network structure adds independent information. A graph built from an independent source (citations, shared parties, cross-referenced c
  • My extension: the huge gains in Silhouette/CHS may be partly tautological, since random walks through topic hubs are trained to make co-topic documents neighbors. External metrics (NMI/ARI) should be the primary yardstick in future comparisons, and their modest values here are the more honest measure of label recovery.
  • My suggestion: a direct comparison against Top2Vec-only with a tuned KMeans (rather than default settings) would clarify how much of the improvement comes from graph embedding versus from the final clustering step.
  • My testable prediction: swapping Top2Vec for a domain-specific legal embedder should improve external alignment more than adding Node2Vec, because legal vocabulary is the bottleneck.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes an unsupervised legal document clustering pipeline. Top2Vec produces semantic document embeddings and topic assignments; a bipartite document–topic graph is built by connecting each document to its dominant Top2Vec topic; Node2Vec learns node embeddings on this graph; the Top2Vec and Node2Vec document embeddings are concatenated and clustered with KMeans; UMAP is used for visualization. The authors claim the hybrid 'Top2Vec+Node2Vec' approach outperforms LDA, NMF, and TF-IDF baselines on internal metrics and external NMI/ARI, and they report an ablation study and a sensitivity analysis on the ACORD and CUAD datasets.

Significance. If the central claim were supported, combining semantic topic embeddings with graph-based structural information for legal text would be a useful contribution. The paper includes several strengths: a sensitivity analysis over hyperparameters, experiments on two legal corpora, qualitative UMAP inspections, and an explicit ablation study. However, the core evaluation is compromised by construction: the graph used for Node2Vec is defined entirely by Top2Vec's topic assignments, so the reported internal-metric improvements largely re-encode the same partition rather than demonstrate synergy between semantic and structural information. The absence of a graph-only baseline and the low external metrics (NMI 0.153, ARI 0.051) leave the central claim unsupported. The manuscript is not suitable for publication in its current form.

major comments (3)
  1. [Sections 1.2–1.3, Table 3] The bipartite graph G=(V,E) is constructed by connecting each document to the topic node it was 'predominantly assigned to by Top2Vec' (Section 1.2). Thus the edge set is a deterministic function of the topic partition. Node2Vec random walks through shared topic hubs pull documents of the same topic together in the embedding space, as the paper itself states in Section 1.3. The large gains in Silhouette (0.927), DBI (0.111), and CHS (29,186) in Tables 1 and 3 are therefore expected by construction: the Node2Vec component re-encodes the Top2Vec topic labels. The ablation study (Table 3) omits a Node2Vec-only (graph-only) baseline and any independent source of graph structure, so the abstract's claim that the hybrid improves over 'text-only or graph-only embeddings' is never directly tested. The 'synergistic' interpretation in Section 2.4 is not supported by the experimental design.
  2. [Section 2.2, Table 1] The external metrics are weak: NMI=0.153 and ARI=0.051, only marginally above NMF (NMI=0.143, ARI=0.045) and far from the values one would expect for 'superior alignment with ground-truth labels.' Given that the internal metrics are inflated by the circular graph construction, the external metrics are the only non-circular evidence of clustering quality, and they provide little support for the paper's central claim. The text should be substantially revised to acknowledge this, and the analysis should focus on why the hybrid method does not align better with ground truth.
  3. [Section 2.3.1 vs. Appendix A] The sensitivity analysis in Section 2.3.1 and Figure 7 identifies k=25 as the optimal number of clusters for the hybrid pipeline. However, Appendix A states 'The final model used K = 5, which aligned with the known number of categories in the ground-truth data.' This is a direct contradiction. The reported benchmark results in Table 1 and the visualizations need to specify which k was used. If k=5 was used for all reported results, the model was not evaluated at its purported optimal configuration; if k=25 was used, Appendix A and the 'known categories' explanation are incorrect. Either way, the validity of the quantitative comparison is compromised.
minor comments (5)
  1. [Tables 1 and 3] There are small numerical inconsistencies between the two tables: TF-IDF CHS is 20 in Table 1 but 20.4 in Table 3, and the hybrid DBI is 0.111 in Table 1 but 0.110 in Table 3. Please ensure consistency.
  2. [Appendix A and Table 2] The reported Node2Vec hyperparameters differ: Appendix A says walk length 30 and num walks 200, while Table 2 uses walk length 20 and num walks 20 as the central configuration. Clarify which configuration produced the headline results, and align Appendix A with the actual experimental settings.
  3. [Section 2.3.2, Figure 8] The text says 'Both metrics show improved cluster quality as the number of topics increases,' and Figure 8's x-axis is labeled 'number of topics,' but the discussion is about data fractions (50%–100%). Clarify the axes and explain the relationship between data fraction, number of topics, and the reported 'representational saturation.'
  4. [Abstract and Introduction] The abstract mentions 'combining unsupervised topic and graph embeddings with a supervised model,' but no supervised model is used anywhere in the paper. Either remove this reference or add a supervised evaluation component.
  5. [Table 3] Typo: 'T op2V ec' should be 'Top2Vec'.

Circularity Check

3 steps flagged

Graph edges are defined by Top2Vec's own topic assignments, so Node2Vec re-encodes the semantic partition and the claimed hybrid synergy is definitional rather than empirical.

specific steps
  1. self definitional [Section 1.2 (p.5) and Section 1.3 (pp.5–6)]
    "An edge ( d, t) ∈ E exists if document d ∈ VD is assigned to topic t ∈ VT. In our implementation, each document node is connected to the specific topic node representing the topic it was predominantly assigned to by Top2Vec. This graph structure explicitly encodes the community information within the corpus: documents sharing a common topic are all linked to the same topic node, effectively forming implicit connections between these documents via their shared thematic anchor."

    The bipartite graph is not an independent structural signal; every edge is a direct function of Top2Vec's fitted topic assignment. Node2Vec then learns embeddings that place documents sharing a topic node close together (the paper states this explicitly in Section 1.3). Therefore the 'structural' information added to the semantic embedding is the same partition that Top2Vec already produced. The hybrid's cluster geometry is by construction aligned with the input topic labels, so internal metrics such as Silhouette/DBI measure the graph's re-encoding of the topic assignment rather than a synergy between text and network structure.

  2. fitted input called prediction [Section 2.4 Ablation Study (p.14), paragraph interpreting Table 3]
    "Notably, the most significant performance leap occurs in the internal metrics (Silhouette, CHS, DBI) when adding the Node2Vec component. This provides a clear mechanistic explanation: Top2Vec first identifies semantically coherent groups, and the Node2Vec component then acts as a structural refiner. By learning from the document-topic graph, it actively compacts the intra-cluster density and sharpens inter-cluster boundaries."

    The ablation labels the Node2Vec component a 'structural refiner' and attributes the internal-metric leap to it, but the graph that Node2Vec learns from is constructed from the same Top2Vec topic assignments. Compacting intra-cluster density is the direct objective of embedding a graph in which same-topic documents are connected through a shared hub; the improvement is therefore a fitted consequence of the input, not an independent structural signal. No graph-only (Node2Vec without Top2Vec) baseline appears in Table 3, so the claimed 'synergistic' contribution is never tested against a graph that carries information beyond the topic labels.

  3. renaming known result [Section 2.1 (p.9), discussion of Figure 5]
    "Subsequently, Node2Vec’s graph embedding process, by reinforcing connections between documents sharing the same topic via the bipartite document-topic graph, pulls these documents closer in the embedding space, leading to the observed high intra-cluster cohesion and inter-cluster separation."

    This is a renaming of Top2Vec's topic co-membership as 'graph-based structural information.' The graph has no edges except those defined by topic assignment, so the 'reinforcing connections' are literally the topic labels. The observed intra-cluster cohesion is therefore not a new property discovered by the graph embedding; it is the original Top2Vec partition re-expressed in Node2Vec coordinates and then reported as evidence of hybrid synergy.

full rationale

The central claim of the paper—that combining Top2Vec semantic embeddings with Node2Vec structural embeddings is 'synergistic' and that the hybrid 'consistently and significantly outperforms all baseline models across every internal metric'—reduces to the way the graph is built. Section 1.2 defines the bipartite graph's edges solely by each document's dominant Top2Vec topic assignment. Section 1.3 then states explicitly that Node2Vec maps documents sharing a topic node to proximate regions. Thus the Node2Vec component is a re-encoder of the same topic partition, and KMeans on the concatenated embeddings will yield clusters aligned with that partition. The large internal-metric gains (Silhouette 0.927, DBI 0.111, CHS 29,186) are therefore a by-construction consequence of embedding a graph derived from the very labels being clustered, not evidence of an independent structural signal. Table 3 includes no Node2Vec-only/graph-only baseline and no graph built from an independent source, so the claimed 'graph-only' advantage is untested. This is not a case of a few harmless self-citations; the load-bearing conclusion of synergy is forced by the graph construction. Accordingly, the circularity score is high. The paper does contain a self-contained pipeline and some legitimate comparisons (e.g., Top2Vec vs TF-IDF/LDA/NMF), but its headline claim of structural synergy is definitional, not empirical.

Axiom & Free-Parameter Ledger

7 free parameters · 6 axioms · 0 invented entities

The central claim relies on several model hyperparameters and on the assumption that the document-topic graph carries independent structural information. The ledger lists the free parameters and assumptions the authors used but did not derive. No new physical or model entities are introduced; the document-topic bipartite graph is a construct from existing data.

free parameters (7)
  • KMeans cluster count k = 5 (final); sensitivity analysis reports optimal k=25
    Chosen to match the five ground-truth categories (Appendix A), conflicting with the reported optimal k=25 (Section 2.3.1); directly determines the output clusters.
  • Node2Vec embedding dimensionality m = 64
    Sensitivity analysis shows CHS changes from 21,450 (m=32) to 29,186 (m=64) to 34,912 (m=128); the reported headline numbers use m=64.
  • Node2Vec walk length and number of walks = walk length 30, num walks 200 (Appendix A); sensitivity table uses walk length 20 and num walks 20 as base
    Random walk parameters control how strongly graph co-membership is pushed into the embeddings.
  • Node2Vec p, q = p=q=1
    Default settings; no tuning reported.
  • Top2Vec embedding size = 300
    Default Top2Vec embedding size.
  • UMAP n_neighbors and min_dist = not specified; 'selected configuration'
    The text reports effects of n_neighbors=15 and min_dist=0.3 but does not give the final values used in figures.
  • KMeans restarts/initialization = k-means++, 10 restarts
    Affects cluster stability; not part of central claim but listed for completeness.
axioms (6)
  • domain assumption Top2Vec's automatic topic discovery produces meaningful legal topic assignments.
    The entire graph and downstream clustering inherit Top2Vec's topic quality; the conclusion section acknowledges the pipeline's efficacy is contingent on initial topic quality.
  • standard math Random walks on the bipartite document-topic graph produce embeddings in which documents sharing a topic are close.
    This is the standard Node2Vec behavior on a graph that is a star per topic; it is the mechanism that makes the circularity possible.
  • domain assumption Concatenating Top2Vec and Node2Vec embeddings preserves or improves cluster structure relative to either alone.
    The paper does not derive conditions under which concatenation helps; it is asserted and tested on two datasets.
  • domain assumption ACORD and CUAD labels are reliable ground truth for external metrics.
    NMI/ARI are computed against these expert labels; the paper does not audit label quality or class distribution beyond a stratified sampling note.
  • standard math UMAP projections faithfully reflect cluster structure.
    The paper itself cautions that global distances and relative sizes in UMAP plots should not be over-interpreted (Section 1.4), yet qualitative claims rely on the plots.
  • domain assumption Silhouette, DBI, and CHS are appropriate for comparing embeddings of different dimensionalities and algorithms.
    Plug-in metrics like these are known to favor certain geometries (e.g., compact ball-like clusters), which Node2Vec on a star graph will trivially produce; no correction is applied.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Hybrid Topic-Semantic Labeling and Graph Embeddings for Unsupervised Legal Document Clustering." pith.science (2026). https://pith.science/paper/OCUVPBVG

@misc{pith2026250900990,
  author       = {Pith},
  title        = {Pith review of: Hybrid Topic-Semantic Labeling and Graph Embeddings for Unsupervised Legal Document Clustering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/OCUVPBVG}},
  note         = {Machine review of arXiv:2509.00990}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Legal documents pose unique challenges for text classification due to their domain-specific language and often limited labeled data. This paper proposes a hybrid approach for classifying legal texts by combining unsupervised topic and graph embeddings with a supervised model. We employ Top2Vec to learn semantic document embeddings and automatically discover latent topics, and Node2Vec to capture structural relationships via a bipartite graph of legal documents. The embeddings are combined and clustered using KMeans, yielding coherent groupings of documents. Our computations on a legal document dataset demonstrate that the combined Top2Vec+Node2Vec approach improves clustering quality over text-only or graph-only embeddings. We conduct a sensitivity analysis of hyperparameters, such as the number of clusters and the dimensionality of the embeddings, and demonstrate that our method achieves competitive performance against baseline Latent Dirichlet Allocation (LDA) and Non-Negative Matrix Factorization (NMF) models. Key findings indicate that while the pipeline presents an innovative approach to unsupervised legal document analysis by combining semantic topic modeling with graph embedding techniques, its efficacy is contingent upon the quality of initial topic generation and the representational power of the chosen embedding models for specialized legal language. Strategic recommendations include the exploration of domain-specific embeddings, more comprehensive hyperparameter tuning for Node2Vec, dynamic determination of cluster numbers, and robust human-in-the-loop validation processes to enhance legal relevance and trustworthiness. The pipeline demonstrates potential for exploratory legal data analysis and as a precursor to supervised learning tasks but requires further refinement and domain-specific adaptation for practical legal applications.

Figures

Figures reproduced from arXiv: 2509.00990 by Deepak Bastola, Woohyeok Choi.

Figure 1
Figure 1. Figure 1: Pipeline comparison: hybrid Top2Vec + Node2Vec (semantic + graph [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: ACORD bipartite graph linking each legal document (gray) to its [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Bar chart showing the number of documents assigned to each topic [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: UMAP projections for NMF and LDA, showing each cluster’s docu [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: UMAP projections comparing optimal results from hybrid ‘Top2Vec+Node2Vec’. This capability extends to other complex legal corpora [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: UMAP projection of CUAD legal-clause clusters using the hybrid [PITH_FULL_IMAGE:figures/full_fig_p010_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Sensitivity analysis across cluster counts (5–50) showing Silhouette, [PITH_FULL_IMAGE:figures/full_fig_p012_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Davies–Bouldin (lower is better) and Silhouette (higher is better) [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

21 extracted references · 9 canonical work pages · 2 internal anchors

  1. [1]

    Top2vec: Distributed representations of topics

    Dimo Angelov. Top2vec: Distributed representations of topics. arXiv preprint arXiv:2008.09470, 2020. URL https://arxiv.org/abs/2008.09470

    Pith/arXiv arXiv 2008

  2. [2]

    Ariai and G

    F. Ariai and G. Demartini. Natural language processing for the legal domain: A survey of tasks, datasets, models, and challenges. arXiv preprint arXiv:2410.21306, 2024. URL https://arxiv.org/abs/2410.21306

    arXiv 2024

  3. [3]

    Methods for Computing Legal Document Similarity: A Comparative Study

    Paheli Bhattacharya, Kripabandhu Ghosh, Arindam Pal, and Saptarshi Ghosh. Methods for computing legal document similarity: A comparative study. arXiv preprint arXiv:2004.12307, 2020. URL https://arxiv.org/abs/2004.12307

    work page internal anchor Pith review Pith/arXiv arXiv 2004

  4. [4]

    Legal case document similarity: You need both network and text

    Paheli Bhattacharya, Kripabandhu Ghosh, Arindam Pal, and Saptarshi Ghosh. Legal case document similarity: You need both network and text. Information Processing & Management, 59 0 (6): 0 103069, 2022. doi:10.1016/j.ipm.2022.103069. URL https://www.sciencedirect.com/science/article/pii/S0306457322001716

    work page arXiv 2022

  5. [5]

    Latent dirichlet allocation

    David M Blei, Andrew Y Ng, and Michael I Jordan. Latent dirichlet allocation. Journal of Machine Learning Research, 3 0 (Jan): 0 993--1022, 2003. URL https://www.jmlr.org/papers/volume3/blei03a/blei03a.pdf

    work page 2003

  6. [6]

    Large-scale multi-label text classification on EU legislation

    Ilias Chalkidis, Emmanouil Fergadiotis, Prodromos Malakasiotis, and Ion Androutsopoulos. Large-scale multi-label text classification on EU legislation. In Anna Korhonen, David Traum, and Llu \'i s M \`a rquez, editors, Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, Florence, Italy, July 2019. Association for Compu...

    doi:10.18653/v1/p19-1636 2019

  7. [7]

    Legal-bert: The muppets straight out of law school

    Ilias Chalkidis, Ion Androutsopoulos, and Nikolaos Aletras. Legal-bert: The muppets straight out of law school. arXiv preprint arXiv:2010.02559, 2020. URL https://arxiv.org/abs/2010.02559

    Pith/arXiv arXiv 2010

  8. [8]

    Unveiling Themes in Judicial Proceedings: A Cross-Country Study Using Topic Modeling on Legal Documents from India and the UK

    Kavya Didwania, Durga Toshniwal, and Ankit Agarwal. Unveiling themes in judicial proceedings: A cross-country study using topic modeling on legal documents from india and the uk. arXiv preprint arXiv:2406.00040, 2024. URL https://arxiv.org/abs/2406.00040

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  9. [9]

    node2vec: Scalable feature learning for networks

    Aditya Grover and Jure Leskovec. node2vec: Scalable feature learning for networks. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2016. URL https://doi.org/10.1145/2939672.2939754

    arXiv 2016

  10. [10]

    Cuad: An expert-annotated nlp dataset for legal contract review

    Dan Hendrycks, Collin Burns, Ananya Chen, and Spencer Ball. Cuad: An expert-annotated nlp dataset for legal contract review. arXiv preprint arXiv:2103.06268, 2021. URL https://arxiv.org/abs/2103.06268

    Pith/arXiv arXiv 2021

  11. [11]

    Learning the parts of objects by non-negative matrix factorization

    Daniel D Lee and H Sebastian Seung. Learning the parts of objects by non-negative matrix factorization. Nature, 401: 0 788--791, 1999. doi:10.1038/44565. URL https://doi.org/10.1038/44565

    doi:10.1038/44565 1999

  12. [12]

    Exploratory analysis of legal case citation data using node embedding

    Swati Lodha and Rakesh Wagh. Exploratory analysis of legal case citation data using node embedding. ICIC Express Letters, 13 0 (10): 0 917--922, 2019. URL http://www.icicel.org/ell/contents/2019/10/el-13-10-02.pdf

    work page 2019

  13. [13]

    MacQueen

    J. MacQueen. Some methods for classification and analysis of multivariate observations. In Proceedings of the fifth Berkeley symposium on mathematical statistics and probability, volume 1, pages 281--297. University of California Press, 1967. URL https://projecteuclid.org/euclid.bsmsp/1200512992

    arXiv 1967

  14. [14]

    Normalized mutual information to evaluate overlapping community finding algorithms

    Aaron F McDaid, Derek Greene, and Neil Hurley. Normalized mutual information to evaluate overlapping community finding algorithms. arXiv preprint arXiv:1110.2515, 2011. URL https://arxiv.org/abs/1110.2515

    Pith/arXiv arXiv 2011

  15. [15]

    Umap: Uniform manifold approximation and projection for dimension reduction

    Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. URL https://arxiv.org/abs/1802.03426

    Pith/arXiv arXiv 2018

  16. [16]

    An analysis of topic modelling for legislative texts

    James O’Neill, Ciarán Robin, Liam O’Brien, and Paul Buitelaar. An analysis of topic modelling for legislative texts. Technical report, University College Cork, 2016. URL https://cora.ucc.ie/handle/10468/6774

    work page 2016

  17. [17]

    Silhouettes: a graphical aid to the interpretation and validation of cluster analysis

    Peter J Rousseeuw. Silhouettes: a graphical aid to the interpretation and validation of cluster analysis. Journal of computational and applied mathematics, 20: 0 53--65, 1987. URL https://www.sciencedirect.com/science/article/pii/0377042787901257

    arXiv 1987

  18. [18]

    Estimating the dimension of a model

    Gideon Schwarz. Estimating the dimension of a model. The annals of statistics, 6 0 (2): 0 461--464, 1978. URL https://projecteuclid.org/euclid.aos/1176344136

    arXiv 1978

  19. [19]

    Email thread identification using latent dirichlet allocation and non-negative matrix factorization based clustering techniques

    Aakanksha Sharaff and Naresh Kumar Nagwani. Email thread identification using latent dirichlet allocation and non-negative matrix factorization based clustering techniques. Journal of Information Science, 42 0 (5): 0 662--678, 2016. doi:10.1177/0165551515587854. URL https://doi.org/10.1177/0165551515587854

    work page doi:10.1177/0165551515587854 2016

  20. [20]

    Acord: An expert-annotated retrieval dataset for legal contract drafting

    SH Wang, M Zubkov, K Fan, S Harrell, Y Sun, et al. Acord: An expert-annotated retrieval dataset for legal contract drafting. arXiv preprint arXiv:2501.06582, 2025. URL https://arxiv.org/abs/2501.06582

    work page arXiv 2025

  21. [21]

    When does pretraining help? assessing self-supervised learning for law and the casehold dataset of 53,000+ legal holdings

    Lucy Zheng, Nikhil Guha, Brian Anderson, and Peter Henderson. When does pretraining help? assessing self-supervised learning for law and the casehold dataset of 53,000+ legal holdings. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 2021. doi:10.1145/3462757.3466088. URL https://dl.acm.org/doi/abs/10.1145/3462757.3466088

    work page arXiv 2021

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.