cat articles/weekly-ai-news

Starting Weekly AI News: automated summaries with clustering and GPT

AI-related topics continue to be numerous, and I thought it would be useful to know roughly what became a topic each week. So I started a weekly newsletter on Substack. The content is created fully automatically. For example, the AI news summary for the week going back from July 28, 2023 looked like this:

I would not call it perfect, but I think it gathers reasonably notable topics in a decent way. If you are interested, please subscribe or read it through a feed reader.

Substack signup form

The newsletter title is a tribute to Weekly Kaggle News.

That would be only publicity, so I will also write about the internal implementation. It has not changed drastically from the material I linked before, but roughly:

  • Generate features, or sentence vectors, from title + summary using multilingual-e5-small.
    • I use small so that it runs on the low-spec VPS environment. Subjectively, small did not feel much less accurate.
  • Add a standardized article timestamp vector to the 384 dimensions from e5-small, making a 385-dimensional representation.
  • Run KMeans without dimensionality reduction. The number of clusters is total article count divided by 8, chosen roughly. With about 250 target articles, this gives around 30 clusters.
    • Reducing dimensions with UMAP or PCA did not produce very good results.
  • Look at overall distances, extract only articles near each cluster center, and use clusters where at least N articles remain.

This extracts clusters that look meaningful as groups of articles from the week. Then I generate titles and summaries for those clusters with gpt-3.5-turbo. It is basically ordinary BERTopic-like clustering plus GPT-based topic representation. In other words, a topic model implementation.

Recent BERTopic implementations on GitHub also seem to include OpenAI and LLM-based features, such as creating sentence vectors with OpenAI embeddings, or ada-v2, in addition to sentence-transformers, and creating topic representations with ChatGPT or GPT-4. By default, it seems to include c-TF-IDF keyword extraction in the prompt for generation. If you want to try this quickly with a library, BERTopic may be a good option.

Incidentally, newsletter article creation is fully automatic, but Substack itself does not seem to have a mechanism that lets me send the newsletter by calling an API. The final delivery flow is manual through the Web UI, which is unfortunate for my own workload.

cat related_articles/weekly-ai-news.yaml

  1. Launching AI News and how I used OpenAI behind itHow AI News collects AI, data science, and machine learning topics and uses OpenAI APIs to classify and summarize them in three lines.
  2. Releasing a Japanese StaticEmbedding Model for Practical 100x Faster Text Embeddingsstatic-embedding-japanese is a fast non-Transformer embedding model for Japanese and English text, evaluated on JMTEB.
  3. Measuring speed, data size, and accuracy for vector search algorithms and quantization parametersA benchmark of FAISS vector search settings, including IVF, HNSW, and product quantization, with a focus on recall@1, @3, and @5 for RAG systems where top-N retrieval quality matters.