[Paper Review] The Parallel Meaning Bank: Towards a Multilingual Corpus of Translations Annotated with Compositional Meaning Representations

Lasha Abzianidze, Johannes Bjerva+6 more|University of Groningen research database (University of Groningen / Centre for Information Technology)|Feb 13, 2017

Natural Language Processing Techniques18 references14 citations

TL;DR

The Parallel Meaning Bank (PMB) introduces a multilingual corpus of over 11 million words in English, German, Dutch, and Italian, annotated with shared, compositional meaning representations via cross-lingual projection. By leveraging automatically generated, semi-supervised semantic annotations for English and projecting them onto aligned translations using syntactic and semantic alignment, the PMB enables cross-lingual compositional semantics and supports the development of multilingual semantic parsers with minimal manual effort.

ABSTRACT

The Parallel Meaning Bank is a corpus of translations annotated with shared, formal meaning representations comprising over 11 million words divided over four languages (English, German, Italian, and Dutch). Our approach is based on cross-lingual projection: automatically produced (and manually corrected) semantic annotations for English sentences are mapped onto their word-aligned translations, assuming that the translations are meaning-preserving. The semantic annotation consists of five main steps: (i) segmentation of the text in sentences and lexical items; (ii) syntactic parsing with Combinatory Categorial Grammar; (iii) universal semantic tagging; (iv) symbolization; and (v) compositional semantic analysis based on Discourse Representation Theory. These steps are performed using statistical models trained in a semi-supervised manner. The employed annotation models are all language-neutral. Our first results are promising.

Motivation & Objective

• To develop a multilingual corpus with shared, formal meaning representations across multiple languages to support cross-lingual compositional semantics.
• To reduce the cost and effort of manual semantic annotation by projecting high-quality English semantic annotations onto translated texts.
• To enable systematic study of meaning shifts in translation and support the development of multilingual semantic parsers.
• To create a scalable, language-neutral framework for semantic annotation that can be adapted to new languages with minimal retraining.

Proposed method

• The PMB uses a five-step annotation pipeline: sentence and lexical segmentation, Combinatory Categorial Grammar (CCG) parsing, universal semantic tagging, symbolization, and compositional semantic analysis using Discourse Representation Theory (DRT).
• Semantic annotations are first generated for English using semi-supervised statistical models trained on existing tools and human-corrected data.
• Cross-lingual projection is achieved by aligning sentences and words between English and its translations using heuristic sentence alignment and GIZA++ for word alignment.
• The same DRT-based meaning representation is projected to other languages under the assumption that translations preserve meaning, enabling shared formal semantics across languages.
• Manual corrections—called 'Bits of Wisdom'—are used to resolve conflicts and improve model performance, distinguishing gold, silver, and bronze annotation standards.
• Models are retrained iteratively with human-corrected data to improve accuracy and adapt to evolving annotation guidelines.

Experimental results

Research questions

• RQ1 Can compositional meaning representations from English be reliably projected onto translated sentences in other languages using alignment and semantic parsing?
• RQ2 To what extent does cross-lingual projection reduce the need for manual annotation in multilingual semantic annotation compared to building resources from scratch?
• RQ3 How effective is the use of language-neutral semantic tagging and symbolic formalisms like CCG and DRT in enabling cross-lingual semantic consistency?
• RQ4 What is the impact of human-annotated 'Bits of Wisdom' on model convergence and annotation quality across multiple languages?
• RQ5 Can the same semantic annotation pipeline be effectively applied to typologically diverse languages such as Italian, in addition to Germanic languages?

Key findings

• The PMB corpus contains over 11.3 million tokens across 285,154 documents in four languages, with 5% of documents existing in all four languages.
• The method achieved promising results in Dutch, demonstrating the feasibility of cross-lingual projection for compositional semantics.
• Gold-standard annotations were established on 6,810 English, 4,757 German, 2,843 Italian, and 945 Dutch documents, ensuring high-quality reference data.
• The use of 'Bits of Wisdom' significantly improved model performance by focusing human annotation on conflicting or ambiguous cases.
• The pipeline reduced annotation costs by reusing English semantic annotations and minimizing manual effort for other languages.
• The corpus is publicly accessible via a web interface, with stable releases available for download, supporting reproducibility and further research.