Reproduce “Semantic Outlier Removal with Embedding Models and LLMs” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Eren Akbiyik, João Almeida, Rik Melis, Ritu Sriram, Viviana Petrescu, Vilhjálmur Vilhjálmsson
The SORE (Semantic Outlier Removal) system offers an efficient and cost-effective method for extracting core content from web pages, utilizing multilingual embedding models and approximate nearest-neighbor search. It achieves an F-score of 0.732, closely matching LLM performance, while significantly reducing processing costs by 25 times and latency by 12.5 times for high-volume industrial applications.
Key insights
- Combining multilingual sentence embeddings with an efficient approximate nearest-neighbor search effectively bridges the gap between LLM-level semantic quality and the cost-efficiency required for industrial-scale content extraction.
- Leveraging document metadata (title, description) as a semantic anchor provides a robust and language-agnostic method for accurately identifying the core content of a web page.
- The system's transparent logging of removal reasons and the use of iteratively refined outlier groups enable continuous improvement and greater control over the content cleaning process, distinguishing it from the 'black box' nature of LLMs.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.