Forum Numerica - Federica GRANESE: Neural Topic Models for Evolving Text Streams

Item: Forum Numerica - Federica GRANESE: Neural Topic Models for Evolving Text Streams
Rating: 4
Author: PostSingularity Institute

Title history

Forum Numerica - Federica GRANESE: Neural Topic Models for Evolving Text Streams — observed on 20261025-1505 UTC (YouTube)

by Federica Granese · View the channel on PSI

Formal & Physical Sciences AI — Models & Research UYQArtificial intelligence (AI)UYQMMachine learning

🎙 Federica Granese 👥 158 📅 October 22, 2026 ⏱ 47 min 👁 3 📄 original study 🧭 2026-10-25

Available in: English (current) Français

Keywords

topic modelsembedded topic modeloptimal transportstick-breakingtemporal evolution

Summary

The seminar by Federica Granese addresses the challenge of discovering and tracking topics in evolving text streams. The speaker begins by defining topic models as unsupervised methods that represent topics as distributions over words, and documents as mixtures of those distributions. She reviews the Embedded Topic Model (ETM) of Dieng et al. 2020, which uses a variational autoencoder to embed both words and topics in a shared latent space. The core contribution is a streaming extension called SB-STREAM, which solves two key limitations: fixed the number of topics and naive merging of topic embeddings. The first improvement employs a stick-breaking construction on the topic proportions, allowing the model to automatically determine the effective number of topics from data, controlled by a maximum K. The second improvement replaces simple summation of embeddings with a Gaussian mapping based on optimal transport, enabling more grounded alignment of topics across time steps. Experiments on the 20 Newsgroups dataset show that SB-STREAM maintains stable topic counts and successfully discovers novel topics that were missed by the earlier model. A case study on Russia-Ukraine war articles demonstrates the practical value of tracking topic shifts in response to real-world events. The talk concludes with a discussion of open issues, such as topic granularity and evaluation metrics.

209 words

Critical Evaluation

Value of the Information & Strength of the Argument

The presentation offers a clear and valuable contribution to the field of online topic modeling. The speaker identifies two concrete limitations of existing streaming models—pre-defined topic count and arbitrary merging—and presents elegant, theoretically grounded solutions. The use of stick-breaking constructions (Kumaraswamy distribution) for adaptive topic discovery and Gaussian mapping for optimal transport-based alignment are well-motivated and clearly explained. The experimental evaluation includes both controlled synthetic-like settings (20 Newsgroups with tailored temporal splits) and a real-world case study (Russia-Ukraine articles), which strengthens the argument for practical applicability. The speaker also honestly acknowledges limitations and open questions, such as topic granularity, indicating a balanced and rigorous scientific stance.

Scientific Rigor, Source Quality, Title Accuracy

The talk demonstrates scientific rigor: prior work is explicitly cited (ETM, Gaussian mapping by Bouveyron and Corneli), and the methodology is described with sufficient detail for replication. The description provides a link to the seminar series, which is the primary external source; the talk itself does not reference external sources beyond those already mentioned, but the internal consistency is high. The title accurately mirrors the content—both the general theme (neural topic models) and the specific focus (evolving text streams) are addressed thoroughly. The presentation includes visualizations and quantitative comparisons, enhancing credibility. Audience questions are handled precisely, and the speaker clarifies confusions, further reinforcing reliability.

225 words

Title / Content Match

The title accurately reflects the content: the talk focuses on neural topic models designed for evolving text streams, covering both the problem and the proposed solutions.

Quality & Reliability

8/10

The talk presents a clear, well-structured exposition of original research, with explicit references to prior work (e.g., Embedded Topic Model) and practical implementation details. Methodology is rigorous, including experimental validation on both synthetic and real-world datasets, and the speaker engages constructively with audience questions.

Key Moments

Markers derived by PSI from the transcript: the creator did not define chapters.

Cited Sources

Forum Numerica seminar series — Official page of the seminar series where this talk is hosted.

Contribution & Novelties

This talk presents two novel contributions to online topic modeling: (1) a stick-breaking mechanism to automatically determine the number of topics from data, eliminating the need for a fixed a priori K, and (2) a Gaussian mapping approach based on optimal transport for aligning topics across timesteps, replacing the ad-hoc summation of embeddings. These innovations address critical limitations in earlier streaming topic models, leading to more stable and interpretable topic evolutions. The experimental validation, including a real-world case study on news articles, highlights the practical value of the approach.

Pour aller plus loin :

Topic model — Foundational concept used throughout the talk.
Optimal transport — Mathematical framework underpinning the alignment method.
Variational autoencoder — The neural architecture used in the Embedded Topic Model.
Stick-breaking process — The construction adapted for adaptive topic number discovery.

135 words

Radar Profile

The radar profile shows a high level of technical depth (9/10) with strong scores in quantity and quality of information (8/10 each) and a reliability score of 8/10. This indicates a dense, expert-level presentation with sound methodology, though the limited viewership and lack of external validation slightly temper the reliability.

Reliability 8/10