BabelNet Meaning Representation: A Fully Semantic Formalism to Overcome Language Barriers

Roberto Navigli, Rexhina Blloshmi, Abelardo Carlos Martinez LorenzoAAAI 2022 · DOI

Abstract

BabelNet Meaning Representation (BMR) is a novel, language-independent framework designed to represent sentence meaning through universal concepts and semantic relations. By transitioning from English-centric labels to BabelNet synsets and VerbAtlas frames, BMR establishes a true "interlingua" capable of bridging hundreds of languages and multiple modalities within a single graph structure.

Introduction: The Limitations of English-Centric Parsing

In the field of Natural Language Understanding (NLU), Semantic Parsing is the process of transforming human language into machine-readable graphs. The current gold standard, Abstract Meaning Representation (AMR), has significantly advanced the field but suffers from a fundamental "English-centric" flaw. AMR uses English words as its nodes and relies on PropBank—an English-based inventory—for its semantic relations.

An example of an English-centric Abstract Meaning Representation (AMR) graph showing lexical nodes like 'mouse' and 'person'.

This dependency creates several critical hurdles:

  1. Ambiguity: A node labeled "mouse" in AMR is lexically ambiguous, potentially referring to an animal or a computer peripheral.
  2. Language Barriers: Because AMR is tied to English syntax and lexicon, adapting it to other languages like Spanish or Chinese often results in a loss of semantic consistency.
  3. Idiomatic Expressions: Multiword expressions (e.g., "kick the bucket") are frequently represented literally in AMR, obscuring their true figurative meaning.

The BMR Framework: Toward a Universal Interlingua

BabelNet Meaning Representation (BMR) evolves the AMR structure into a language-independent graph through three primary transitions:

From Words to Synsets

Instead of using lexical words, BMR utilizes BabelNet synsets. BabelNet is a massive multilingual semantic network that groups synonymous words across 500 languages into unique concept IDs. This eliminates ambiguity; for instance, the computer peripheral "mouse" receives a specific ID that remains constant regardless of whether the input language is English, Spanish, or Chinese.

From PropBank to VerbAtlas

AMR relies on roles like :ARG0 or :ARG1, which are specific to English verbal frames. BMR replaces these with VerbAtlas, a hand-crafted inventory of cross-lingual verbal frames. This allows for the use of human-readable, universal roles such as AGENT, PATIENT, or LOCATION, facilitating a more intuitive understanding of the underlying action.

Handling Multiword Expressions (MWEs)

In traditional AMR, a term like "external hard drive" might be decomposed into three separate nodes. BMR leverages BabelNet’s extensive coverage of MWEs to represent the entire term as a single, semantically atomic node. This simplifies the graph and ensures the intended meaning is captured accurately.

Bridging Languages and Modalities

The core strength of BMR lies in its ability to act as a "language of thought." In theory, equivalent sentences in different languages should map to the exact same BMR graph, as they share the same underlying semantic concepts.

The BMR framework maps multilingual sentences to a single, unified semantic graph using BabelNet synsets and visual grounding.

Furthermore, because BabelNet synsets are often linked to images or videos, BMR provides a natural bridge for AI systems to connect text to visual or auditory data. Unlike the latent vector representations (embeddings) used in Large Language Models (LLMs), BMR graphs are human-readable, making them a prime candidate for Explainable AI (XAI).

Conclusion and Future Directions

While BMR represents a significant leap toward language independence, challenges remain. The quality of BabelNet coverage varies across its 500 supported languages, and the vast "vocabulary" of synsets may make BMR-based models harder to train than their AMR counterparts. Future work will focus on empirical validation through tasks like Machine Translation and Question Answering, as well as integrating temporal and plurality nuances into the core formalism. By moving beyond language-specific constraints, BMR paves the way for a more inclusive and semantically grounded approach to global NLU.

Summary updated

BibTeX
@inproceedings{navigli2022babelnet,
  author = {Roberto Navigli and Rexhina Blloshmi and Abelardo Carlos Martinez Lorenzo},
  title = {{BabelNet Meaning Representation: A Fully Semantic Formalism to Overcome Language Barriers}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21490},
}
Figure