---
title: "From Noise to Signal — Zve 2026"
description: "Dynamic topic models usually discard low-density documents as noise, though some of them may be early hints of topics that have not yet formed."
authors: ["Evangelia Zve", "Gauvain Bourgne", "Benjamin Icard", "Jean-Gabriel Ganascia"]
year: 2026
venue: "Proceedings of the Fifteenth Language Resources and Evaluation Conference"
source: "https://aclanthology.org/2026.lrec-1.596/"
doi: "10.63317/5c6zvq4nbjdq"
licence: "CC BY-NC 4.0 (ELRA, LREC proceedings)"
licence_url: "https://creativecommons.org/licenses/by-nc/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/zve-2026-noise-signal-when-outliers-seed/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/zve-2026-noise-signal-when-outliers-seed/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2026](https://md.research.elamur.ai/free/en/papers/2026/page/5/) / From Noise to Signal: When Outliers Seed New Topics

# From Noise to Signal: When Outliers Seed New Topics

**Authors:** Evangelia Zve, Gauvain Bourgne, Benjamin Icard, Jean-Gabriel Ganascia · **Year:** 2026 · **Venue:** Proceedings of the Fifteenth Language Resources and Evaluation Conference · **Licence:** [CC BY-NC 4.0 (ELRA, LREC proceedings)](https://creativecommons.org/licenses/by-nc/4.0/)

**Links:** [Publisher's page](https://aclanthology.org/2026.lrec-1.596/) · [PDF](https://aclanthology.org/2026.lrec-1.596.pdf) · [DOI 10.63317/5c6zvq4nbjdq](https://doi.org/10.63317/5c6zvq4nbjdq) · [OpenAlex](https://openalex.org/W7160026432)

**Citations:** 0 citations (openalex-snapshot · 2026-10-14)

## Summary

Dynamic topic models usually discard low-density documents as noise, though some of them may be early hints of topics that have not yet formed. The authors propose a document-level temporal taxonomy, inspired by Allen's interval algebra. It tracks three moments for each article: when it first appears, when its eventual topic is created, and when the article first joins that topic. Ordering these events gives seven mutually exclusive trajectory types. The key type is the *anticipatory outlier* (T OA), a document that shows up before the cluster it later enters.

To test the taxonomy, they built HydroNewsFr, a French corpus about the hydrogen economy with continuous daily coverage. Articles were embedded with eleven language models and clustered cumulatively in daily windows using density-based algorithms (HDBSCAN, OPTICS) after UMAP reduction. Clusters were then aligned across time and each document was labeled retrospectively.

Agreement across models singles out a small, high-consensus set of anticipatory outliers, which makes those labels more trustworthy. Qualitative case studies show these trajectories in concrete topic developments. The work links weak-signal detection with temporal topic modeling.

## Key points

- **HydroNewsFr corpus:** 1,616 deduplicated French articles covering 81 consecutive days (20 March–8 June 2025).
  - Sources were X posts containing *hydrogène* (1,533 posts, 726 unique articles) and the GNews library (891 articles).
  - Each document is a headline plus description, averaging about 280 characters.
- **Seven taxonomy cases:**
  - T OA first and T OA late (anticipatory outliers)
  - T OD late (drift outlier that reinforces an existing topic)
  - T first and T late (non-outlier integrations)
  - O recent and O old (never integrated)
  - The cutoff θ delay separating O recent from O old is the 90th percentile of observed integration delays.
- **Embedding models:** eleven models chosen via MTEB, including bge-m3, LaBSE, multilingual-e5-large, sentence-camembert-base, Solon-embeddings-large-0.1, mistral-embed, gemini and text-embedding-3-small.
- **Share of anticipatory outliers:** at θ align = 0.30, T OA make up roughly 0.11–0.46 of all outlier documents, with averages near 0.27–0.33, across UMAP dimensionalities of 2D–40D.
- **Label stability:** Fleiss' κ across UMAP dimensionalities has mean values of about 0.30–0.47, depending on setup and κ variant. These values barely change when θ align varies from 0.2 to 0.7.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

Fields: [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/)

## Cite

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@inproceedings{zve2026noise,
  title = {{From Noise to Signal: When Outliers Seed New Topics}},
  author = {Zve, Evangelia and Bourgne, Gauvain and Icard, Benjamin and Ganascia, Jean-Gabriel},
  booktitle = {Proceedings of the Fifteenth Language Resources and Evaluation Conference},
  year = {2026},
  doi = {10.63317/5c6zvq4nbjdq},
  url = {https://doi.org/10.63317/5c6zvq4nbjdq}
}
```

- [← previous From LLM Prompts to Acoustic Baselines: A Scalable Pipeline for Under-Resourced Disfluent Code-Mixed Speech](https://md.research.elamur.ai/free/en/papers/mitra-2026-llm-prompts-acoustic-baselines-scalable/)
- [next → From Propositional to Perceptual Asymmetry: Extending FPO to Asymmetric Partial Information Dialogue](https://md.research.elamur.ai/free/en/papers/zhu-2026-propositional-perceptual-asymmetry-extending-fpo/)
- [↑ back to papers of 2026](https://md.research.elamur.ai/free/en/papers/2026/page/5/)

## References

25 references, 5 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. James Allan. 2002. Introduction to topic detection and tracking. In James Allan, editor, Topic Detection and Tracking: Event-based Information Organization, pages 1–16. Springer, Boston, MA. [find](https://scholar.google.com/scholar?q=James+Allan%2E+2002%2E+Introduction+to+topic+detection+and+tracking%2E+In+James+Allan%2C+editor%2C+Topic+Detection+and+Tracking%3A+Event%2Dbased+Information+Organization%2C+pages+1%E2%80%9316%2E+Springer%2C+Boston%2C+MA%2E)
2. James F Allen. 1983. Maintaining knowledge about temporal intervals. Communications of the ACM, 26(11):832–843. [find](https://scholar.google.com/scholar?q=James+F+Allen%2E+1983%2E+Maintaining+knowledge+about+temporal+intervals%2E+Communications+of+the+ACM%2C+26%2811%29%3A832%E2%80%93843%2E)
3. Mihael Ankerst, Markus M Breunig, Hans-Peter Kriegel, and Jörg Sander. 1999. Optics: Ordering points to identify the clustering structure. ACM Sigmod record, 28(2):49–60. [find](https://scholar.google.com/scholar?q=Mihael+Ankerst%2C+Markus+M+Breunig%2C+Hans%2DPeter+Kriegel%2C+and+J%C3%B6rg+Sander%2E+1999%2E+Optics%3A+Ordering+points+to+identify+the+clustering+structure%2E+ACM+Sigmod+record%2C+28%282%29%3A49%E2%80%9360%2E)
4. H. Igor Ansoff. 1980. Strategic issue management. Strategic Management Journal, 1(2):131–148. [find](https://scholar.google.com/scholar?q=H%2E+Igor+Ansoff%2E+1980%2E+Strategic+issue+management%2E+Strategic+Management+Journal%2C+1%282%29%3A131%E2%80%93148%2E)
5. David M Blei and John D Lafferty. 2006. Dynamic topic models. In Proceedings of the 23rd International Conference on Machine Learning, pages 113–120. ACM. [find](https://scholar.google.com/scholar?q=David+M+Blei+and+John+D+Lafferty%2E+2006%2E+Dynamic+topic+models%2E+In+Proceedings+of+the+23rd+International+Conference+on+Machine+Learning%2C+pages+113%E2%80%93120%2E+ACM%2E) [our card](https://md.research.elamur.ai/free/en/papers/blei-2006-dynamic-topic-models/)
6. David M Blei, Andrew Y Ng, and Michael I Jordan. 2003. Latent dirichlet allocation. Journal of machine Learning research, 3(Jan):993–1022. [find](https://scholar.google.com/scholar?q=David+M+Blei%2C+Andrew+Y+Ng%2C+and+Michael+I+Jordan%2E+2003%2E+Latent+dirichlet+allocation%2E+Journal+of+machine+Learning+research%2C+3%28Jan%29%3A993%E2%80%931022%2E) [our card](https://md.research.elamur.ai/free/en/papers/blei-2003-latent-dirichlet-allocation/)
7. Allaa Boutaleb, Jerome Picault, and Guillaume Grosjean. 2024. BERTrend: Neural topic modeling for emerging trends detection. In Proceedings of the Workshop on the Future of Event Detection (FuturED), pages 1–17, Miami, Florida, USA. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=Allaa+Boutaleb%2C+Jerome+Picault%2C+and+Guillaume+Grosjean%2E+2024%2E+BERTrend%3A+Neural+topic+modeling+for+emerging+trends+detection%2E+In+Proceedings+of+the+Workshop+on+the+Future+of+Event+Detection+%28FuturED%29%2C+)
8. Clément Christophe, Julien Velcin, Jairo Cugliari, Manel Boumghar, and Philippe Suignard. 2021. Monitoring geometrical properties of word embeddings for detecting the emergence of new topics. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 994–1003. [find](https://scholar.google.com/scholar?q=Cl%C3%A9ment+Christophe%2C+Julien+Velcin%2C+Jairo+Cugliari%2C+Manel+Boumghar%2C+and+Philippe+Suignard%2E+2021%2E+Monitoring+geometrical+properties+of+word+embeddings+for+detecting+the+emergence+of+new+topics%2E+In+Proce)
9. Jason Chuang, Sonal Gupta, Christopher Manning, and Jeffrey Heer. 2013. Topic model diagnostics: Assessing domain relevance via topical alignment. In International conference on machine learning, pages 612–620. PMLR. [find](https://scholar.google.com/scholar?q=Jason+Chuang%2C+Sonal+Gupta%2C+Christopher+Manning%2C+and+Jeffrey+Heer%2E+2013%2E+Topic+model+diagnostics%3A+Assessing+domain+relevance+via+topical+alignment%2E+In+International+conference+on+machine+learning%2C+page)
10. Rob Churchill and Lisa Singh. 2022. The evolution of topic modeling. ACM Computing Surveys, 54(10s):1–35. [find](https://scholar.google.com/scholar?q=Rob+Churchill+and+Lisa+Singh%2E+2022%2E+The+evolution+of+topic+modeling%2E+ACM+Computing+Surveys%2C+54%2810s%29%3A1%E2%80%9335%2E)
11. Adji B Dieng, Francisco JR Ruiz, and David M Blei. 2019. The dynamic embedded topic model. arXiv preprint arXiv:1907.05545. — [arXiv:1907.05545](https://arxiv.org/abs/1907.05545) [find](https://scholar.google.com/scholar?q=Adji+B+Dieng%2C+Francisco+JR+Ruiz%2C+and+David+M+Blei%2E+2019%2E+The+dynamic+embedded+topic+model%2E+arXiv+preprint+arXiv%3A1907%2E05545%2E+%E2%80%94+%5BarXiv%3A1907%2E05545%5D%28https%3A%2F%2Farxiv%2Eorg%2Fabs%2F1907%2E05545%29)
12. Maarten Grootendorst. 2022. Bertopic: Neural topic modeling with a class-based tf-idf procedure. arXiv preprint arXiv:2203.05794. [find](https://scholar.google.com/scholar?q=Maarten+Grootendorst%2E+2022%2E+Bertopic%3A+Neural+topic+modeling+with+a+class%2Dbased+tf%2Didf+procedure%2E+arXiv+preprint+arXiv%3A2203%2E05794%2E) [our card](https://md.research.elamur.ai/free/en/papers/grootendorst-2022-bertopic-neural-topic-modeling-class/)
13. David Hall, Daniel Jurafsky, and Christopher D. Manning. 2008. Studying the history of ideas using topic models. In Proceedings of the conference on empirical methods in natural language processing, pages 363–371. Association for Computational Linguistics. [find](https://scholar.google.com/scholar?q=David+Hall%2C+Daniel+Jurafsky%2C+and+Christopher+D%2E+Manning%2E+2008%2E+Studying+the+history+of+ideas+using+topic+models%2E+In+Proceedings+of+the+conference+on+empirical+methods+in+natural+language+processing%2C+p)
14. John A Hartigan and Manchek A Wong. 1979. Algorithm as 136: A k-means clustering algorithm. Journal of the Royal Statistical Society: Series C (Applied Statistics), 28(1):100–108. [find](https://scholar.google.com/scholar?q=John+A+Hartigan+and+Manchek+A+Wong%2E+1979%2E+Algorithm+as+136%3A+A+k%2Dmeans+clustering+algorithm%2E+Journal+of+the+Royal+Statistical+Society%3A+Series+C+%28Applied+Statistics%29%2C+28%281%29%3A100%E2%80%93108%2E)
15. Elina Hiltunen. 2008. The future sign and its three dimensions. Futures, 40(3):247–260. [find](https://scholar.google.com/scholar?q=Elina+Hiltunen%2E+2008%2E+The+future+sign+and+its+three+dimensions%2E+Futures%2C+40%283%29%3A247%E2%80%93260%2E)
16. Sarah Kaplan and Keyvan Vakili. 2015. The double-edged sword of recombination in breakthrough innovation. Strategic Management Journal, 36(10):1435–1457. [find](https://scholar.google.com/scholar?q=Sarah+Kaplan+and+Keyvan+Vakili%2E+2015%2E+The+double%2Dedged+sword+of+recombination+in+breakthrough+innovation%2E+Strategic+Management+Journal%2C+36%2810%29%3A1435%E2%80%931457%2E)
17. Harold W Kuhn. 1955. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83–97. [find](https://scholar.google.com/scholar?q=Harold+W+Kuhn%2E+1955%2E+The+hungarian+method+for+the+assignment+problem%2E+Naval+research+logistics+quarterly%2C+2%281%2D2%29%3A83%E2%80%9397%2E)
18. Leland McInnes, John Healy, Steve Astels, et al. 2017. hdbscan: Hierarchical density based clustering. J. Open Source Softw., 2(11):205. [find](https://scholar.google.com/scholar?q=Leland+McInnes%2C+John+Healy%2C+Steve+Astels%2C+et+al%2E+2017%2E+hdbscan%3A+Hierarchical+density+based+clustering%2E+J%2E+Open+Source+Softw%2E%2C+2%2811%29%3A205%2E) [our card](https://md.research.elamur.ai/free/en/papers/mcinnes-2017-hdbscan-hierarchical-density-clustering/)
19. Leland McInnes, John Healy, Nathaniel Saul, and Lukas Großberger. 2018. Umap: Uniform manifold approximation and projection. Journal of Open Source Software, 3(29):861. [find](https://scholar.google.com/scholar?q=Leland+McInnes%2C+John+Healy%2C+Nathaniel+Saul%2C+and+Lukas+Gro%C3%9Fberger%2E+2018%2E+Umap%3A+Uniform+manifold+approximation+and+projection%2E+Journal+of+Open+Source+Software%2C+3%2829%29%3A861%2E)
20. Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. 2023. Mteb: Massive text embedding benchmark. In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics, pages 2014–2037. [find](https://scholar.google.com/scholar?q=Niklas+Muennighoff%2C+Nouamane+Tazi%2C+Loic+Magne%2C+and+Nils+Reimers%2E+2023%2E+Mteb%3A+Massive+text+embedding+benchmark%2E+In+Proceedings+of+the+17th+Conference+of+the+European+Chapter+of+the+Association+for+Comp) [our card](https://md.research.elamur.ai/free/en/papers/muennighoff-2022-mteb-massive-text-embedding-benchmark/)
21. Hamed Rahimi, Hubert Naacke, Camelia Constantin, and Bernd Amann. 2024. Antm: aligned neural topic models for exploring evolving topics. In Transactions on Large-Scale Data-and Knowledge-Centered Systems LVI: Special Issue on Data Management-Principles, Technologies, and Applications, pages 76–97. S… [find](https://scholar.google.com/scholar?q=Hamed+Rahimi%2C+Hubert+Naacke%2C+Camelia+Constantin%2C+and+Bernd+Amann%2E+2024%2E+Antm%3A+aligned+neural+topic+models+for+exploring+evolving+topics%2E+In+Transactions+on+Large%2DScale+Data%2Dand+Knowledge%2DCentered+Syst)
22. Ketan Rajshekhar Shahapure and Charles Nicholas. 2020. Cluster quality analysis using silhouette score. In 2020 IEEE 7th international conference on data science and advanced analytics (DSAA), pages 747–748. IEEE. [find](https://scholar.google.com/scholar?q=Ketan+Rajshekhar+Shahapure+and+Charles+Nicholas%2E+2020%2E+Cluster+quality+analysis+using+silhouette+score%2E+In+2020+IEEE+7th+international+conference+on+data+science+and+advanced+analytics+%28DSAA%29%2C+pages+7)
23. Carmen K Vaca, Amin Mantrach, Alejandro Jaimes, and Marco Saerens. 2014. A time-based collective factorization for topic discovery and monitoring in news. In Proceedings of the 23rd international conference on World wide web, pages 527–538. [find](https://scholar.google.com/scholar?q=Carmen+K+Vaca%2C+Amin+Mantrach%2C+Alejandro+Jaimes%2C+and+Marco+Saerens%2E+2014%2E+A+time%2Dbased+collective+factorization+for+topic+discovery+and+monitoring+in+news%2E+In+Proceedings+of+the+23rd+international+conf)
24. Xuerui Wang and Andrew McCallum. 2006. Topics over time: a non-markov continuous-time model of topical trends. In Proceedings of the 12th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 424–433. ACM. [find](https://scholar.google.com/scholar?q=Xuerui+Wang+and+Andrew+McCallum%2E+2006%2E+Topics+over+time%3A+a+non%2Dmarkov+continuous%2Dtime+model+of+topical+trends%2E+In+Proceedings+of+the+12th+ACM+SIGKDD+international+conference+on+Knowledge+discovery+and)
25. Evangelia Zve, Benjamin Icard, Alice Breton, Lila Sainero, Gauvain Bourgne, and Jean-Gabriel Ganascia. 2025. From outliers to topics in language models: Anticipating trends in news corpora. In Proceedings of the 8th International Conference on Natural Language and Speech Processing (ICNLSP-2025), pa… [find](https://scholar.google.com/scholar?q=Evangelia+Zve%2C+Benjamin+Icard%2C+Alice+Breton%2C+Lila+Sainero%2C+Gauvain+Bourgne%2C+and+Jean%2DGabriel+Ganascia%2E+2025%2E+From+outliers+to+topics+in+language+models%3A+Anticipating+trends+in+news+corpora%2E+In+Proceed)