The second idea from our Baltic DB&IS 2026 paper: an LLM writes a short document-level context for every chunk before it is embedded, and how this compares with late chunking, propositions and other ways to keep a chunk's context.

One idea from our Baltic DB&IS 2026 paper on a production RAG agent platform: storing each chunk under its own vector and under the questions it answers.

How to spend limited compute in an NLP shared task: a three-phase, 10-step workflow from my doctoral thesis.

In this post I’ll focus on Demo papers and tools that could be relevant in research and development. GenGO: ACL Paper Explorer with Semantic Features Motivation The researchers aimed to create a ...

This post focuses on ACL’2024 papers, that are studying information extraction tasks. Hypergraph based Understanding for Document Semantic Entity Recognition Motivation The researchers aimed to i...

When Phrases Meet Probabilities: Enabling Open Relation Extraction with Cooperating Large Language Models Motivation Existing clustering-based open relation extraction methods use pre-trained lang...

This year, I attended ACL 2024 in Bangkok. In this series of posts, I’ll share my personal highlights from the conference, focusing primarily on papers and ideas with practical applications. Promp...

DarkBERT: A Language Model for the Dark Side of the Internet Motivation Most of pretrained models are trained on the “Surface Web” data, i.e., the content that is readily available and indexed in ...

DIONYSUS: A Pre-trained Model for Low-Resource Dialogue Summarization Motivation The main challenge of dialogue summarization is to select the most relevant information, that could be spread acro...

Improving the Robustness of Summarization Systems with Dial Augmentation Motivation A robust summarization system should be able to capture the gist of the document, regardless of the specific word...