Fully-Semantic Parsing and Generation: the BabelNet Meaning Representation

Abelardo Carlos Martinez Lorenzo, Marco Maru, Roberto NavigliACL 2022 · DOI

Introduction

In the field of Natural Language Understanding (NLU), the "holy grail" is a way to represent the meaning of a sentence independently of the language it was written in. While Abstract Meaning Representation (AMR) is the current industry standard, it is deeply English-centric. AMR uses English lemmas as its nodes and relies on PropBank, a repository of English-specific predicate frames. This creates a "language gap": representing a Spanish or Chinese sentence using AMR forces a mapping back to English-like structures, making it difficult to use for low-resourced languages or as a true "interlingua" for machine translation.

The BabelNet Meaning Representation (BMR)

To address these limitations, researchers introduced BMR, a truly language-agnostic representation that departs from AMR in four fundamental ways:

  1. BabelNet Synsets: Instead of English lemmas (e.g., plane-01), BMR uses BabelNet synsets—unique identifiers representing concepts across hundreds of languages. This resolves ambiguity, such as distinguishing an airplane from a carpenter's tool.
  2. VerbAtlas Frames: BMR replaces language-specific PropBank frames with VerbAtlas, grouping verbs into semantically coherent, language-independent frames (e.g., using :agent and :theme instead of :ARG0 and :ARG1).
  3. Node Merging: BMR identifies Multiword Expressions (MWEs) and idioms (like "at the last minute") and collapses them into a single semantic node, reducing graph density by approximately 11.6%.
  4. Explicit Grammar: Unlike AMR, BMR explicitly encodes tense, aspect, and number (e.g., pluralization), which are often lost in standard AMR graphs.
BMR graph representation using BabelNet synsets and VerbAtlas frames.

Experimental Setup

The researchers developed BMR 1.0, a dataset derived from AMR 3.0 via automatic alignment and expert validation. They also created AMR+, an enhanced version of AMR 3.0 that includes BMR’s grammatical features to ensure a fair comparison. Evaluation was performed on English (EN), German (DE), Italian (IT), and Spanish (ES).

Key Results and Insights

Graph-to-Text Generation

BMR consistently outperformed both AMR and the enhanced AMR+ across all four languages. In English, BMR achieved a 50.7 BLEU score compared to AMR's 44.8. This suggests that the explicit encoding of grammar and word senses significantly aids the generation of high-quality, fluent text.

Multilingual Graph-to-Text generation performance across four languages.

Machine Translation via Interlingua

In "Text-to-Graph-to-Text" tasks (using the graph as a bridge for translation), BMR proved superior to AMR in almost every language pair. For English-to-Italian translation, BMR reached 30.9 BLEU, while standard AMR lagged at 29.0. Because the representation is language-independent, the model avoids the syntactic interference of the source language when generating the target.

The Disambiguation Trade-off

While BMR excels at generation, it presents a greater challenge for parsing (Text-to-Graph). A BMR parser must perform Word Sense Disambiguation (WSD) to select the correct BabelNet synset, a more complex task than simple lemma identification. However, the resulting semantic density and clarity provide a more robust foundation for downstream applications.

Conclusion and Future Work

BMR represents a significant step toward a universal semantic interlingua. By moving beyond English-centric labels, it enables more accurate multilingual generation and translation. Future work will focus on creating a unified multilingual model for BMR parsing and evolving the formalism to be entirely lexical-free, further bridging the gap between diverse human languages.

Summary updated

BibTeX
@inproceedings{lorenzo2022fully,
  author = {Abelardo Carlos Martinez Lorenzo and Marco Maru and Roberto Navigli},
  title = {{Fully-Semantic Parsing and Generation: the BabelNet Meaning Representation}},
  booktitle = {Annual Meeting of the Association for Computational Linguistics},
  year = {2022},
  doi = {10.18653/v1/2022.acl-long.121},
}
Figure