Abstract
Creating a language-independent meaning representation would benefit many cross-lingual NLP tasks. We introduce the first unsupervised approach to this problem, learning clusters of semantically equivalent English and French relations between referring expressions, based on their named-entity arguments in large monolingual corpora. The clusters can be used as language-independent semantic relations, by mapping clustered expressions in different languages onto the same relation. Our approach needs no parallel text for training, but outperforms a baseline that uses machine translation on a cross-lingual question answering task. We also show how to use the semantics to improve the accuracy of machine translation, by using it in a simple reranker.
Introduction
A long-standing goal in computational linguistics is the creation of a universal, language-independent meaning representation, or "interlingua." If sentences from any language could be mapped to a common logical form, tasks such as machine translation (MT) and cross-lingual question answering would be significantly simplified.
Traditionally, MT systems rely on parallel corpora—large datasets of human-translated sentences. However, parallel data is expensive to produce and unavailable for the vast majority of language pairs. This research explores an alternative: can we learn that expressions in different languages are semantically equivalent by observing the real-world facts they describe in independent monolingual texts?
Methodology
The proposed method leverages distributional semantics, using named entities as a universal bridge between languages. The pipeline consists of four primary stages:
- Predicate Extraction: Sentences are parsed into logical triplets. English text is processed using a Combinatory Categorial Grammar (CCG) parser, while French text uses a dependency parser. For example, "Shakespeare wrote Macbeth" becomes
write(Shakespeare, Macbeth). - Entity Linking and Typing: To align data across languages, the system uses Wikipedia cross-language links to map entities (e.g., "États-Unis" to "United States") to unique identifiers. Entities are also assigned types (e.g.,
PERSON,BOOK) using Freebase. - Monolingual Clustering: Synonymous relations within a single language are grouped using the Chinese Whispers algorithm. This might group "X wrote Y" with "X is the author of Y" based on shared argument pairs.
- Cross-lingual Alignment: Finally, these monolingual clusters are aligned across languages. An English cluster is matched with a French cluster if they share a significant number of the same entity pairs, such as the same authors and their respective works.
By clustering within languages before aligning across them, the model avoids "language bias," where cultural differences in subject matter might otherwise lead the algorithm to incorrectly distinguish equivalent predicates.
Evaluation and Results
The system was evaluated on two primary tasks: cross-lingual question answering (QA) and machine translation reranking.
Cross-lingual Question Answering
In the French-to-English QA task, the unsupervised clusters achieved a 93% accuracy rate, significantly outperforming a baseline system that utilized state-of-the-art machine translation (85% accuracy). The cluster-based approach also demonstrated higher recall, identifying a larger volume of correct answers.
Translation Reranking
The semantic clusters were also used to rerank the output of a standard MT system (Moses). By checking if a suggested translation matched the semantic relation of the source sentence, the system could select more accurate candidates. Human evaluators preferred the cluster-based reranker's output in 39% of cases, compared to only 5% for the standard 1-best MT output.
Discussion and Future Work
While successful, the current model is limited to binary relations between named entities (e.g., "Person born in City"). It does not yet account for common nouns or complex sentence structures involving negation or quantifiers. Additionally, the model uses "flat" clustering, which does not capture hierarchical relationships between concepts (e.g., that a "novelist" is a type of "author").
Future research will focus on expanding these clusters to include common nouns and more sophisticated logical operators, further moving toward a comprehensive, unsupervised interlingua.
Summary updated