面向自然语言理解的语义表示方法综述

Summary

Semantic representations used in natural language understanding are fragmented. Query graphs and SparQL in knowledge-base QA, SQL in table QA, and frame semantics or AMR in sentence-level parsing are each studied in isolation. As a result, they are rarely compared, hard to choose between for a given task, and show large performance gaps. This survey, by researchers at UCAS and the CAS Institute of Automation, places them in one framework.

It first sorts task-driven representations by methodology: logical, cognitive or distributional. It then links each to a theory of meaning and regroups them by how they treat the world–language relation. This yields two classes: External World Semantics (EWS) and Language Interior Semantics (LIS). For EWS, the authors review research trends and how new representations are designed and compared. For LIS, they outline the debate over whether systems such as pretrained language models really capture meaning.

They conclude that hybrid methods combining both classes are especially promising. The framework gives a principled basis for choosing representations and for rethinking symbolic–neural integration in the LLM era.

Key points

• Literature collection:
  • A "semantic parsing" search on Google Scholar (2021 onward), routed through the Semantic Scholar API, returned 2,258 papers.
  • Adding keyword-filtered ACL Anthology papers from the last three years gave 2,887.
  • The authors manually annotated 373 of these (all 2023 papers plus highly cited earlier ones) and kept 213 relevant works.
• Task-driven families:
  • Logical: programming languages (Java, Python, SQL, SparQL), first-order logic with λ-calculus, and graph-based forms (MRS, AMR, UMR, KB query graphs).
  • Cognitive: FrameNet-style frames and Construction Grammar (SBCG).
  • Distributional: LSI/LDA, Word2Vec, GloVe, ELMo, BERT and GPT.
• Placement on the world–language axis:
  • Referential theory, truth-conditional semantics, predicate logic and programming languages sit near the world (EWS) end.
  • Relational semantics (e.g., WordNet), semantic field theory, conceptual role semantics and PLMs sit near the language (LIS) end.
  • Frames, construction grammar, sense theory and descriptivism fall in between.
• EWS parsing history and current topics:
  • Parsing evolved from CCG-driven lexical mappings, to statistical models, to end-to-end neural parsing.
  • Current hot topics are domain applications (e.g., Coq proofs, biomedical summarisation using AMR), combination with LLMs, compositional generalisation, multilinguality and evaluation metrics.
• Publication details: 中文信息学报, Vol. 38, No. 12, pp. 1–17. The PDF header dates the issue to December 2024, not 2025 as given in the request; please verify the citation.