📄 Proceedings of Bridging Neurons and Symbols for Natural Language Processing and Knowledge Graphs Reasoning @ COLING 2025

Summary

This is the proceedings volume of a workshop held on January 20, 2025, as part of the COLING 2025 conference. The workshop's theme is combining neural and symbolic methods in natural language processing and in reasoning over knowledge graphs. The organizers start from the observation that large language models such as ChatGPT are good at sustaining conversation but weak at solving even simple logical problems. Whether these models truly reason remains an open question. Drawing on the notion of "fast" and "slow" human thinking, the organizers outline two directions. The first is to strengthen existing neural networks until they reach the level of symbolic reasoning. The second is to look for new neural implementations of symbolic logic, and this is the direction the workshop encouraged most. The volume contains 11 papers on temporal knowledge graphs, logical rule refinement, RAG with knowledge graphs, table-to-text generation, multimodal NER, in-context learning mechanisms in LLMs, and scalar implicatures. The front matter contains no experimental results. The volume is useful primarily as a snapshot of current topics in neurosymbolic NLP.

Key points

• The volume was published by the ACL (ISBN 979-8-89176-210-7). The workshop took place on January 20, 2025, as part of COLING 2025.
• The organizing committee included Kang Liu, Yangqiu Song, Zhen Han (Amazon), Rafet Sifa, Shizhu He, and Yunfei Long.
• Papers on knowledge graphs:
  • Chain of Knowledge Graph — summarization of multiple noisy documents without information loss;
  • CEGRL-TKGR — causal representation learning for temporal KGs;
  • From Chain to Tree — transforming chain-like rules into tree-like ones;
  • LAB-KG — RAG for interpreting laboratory test results.
• Papers on LLM interpretability investigate the emergence of "symbolic abstraction heads" during in-context learning. Another paper compares language model predictions with human behavior on scalar implicatures.
• Other contributions:
  • Generative FrameNet — frames generated by LLMs themselves as an interpretable knowledge repository;
  • combining OPM with deep learning for question answering tasks;
  • VCRMNER — multimodal NER based on CLIP prompts;
  • building explicit 3-D models that link language to scenes.