Introduction
In Information Extraction (IE), knowledge is typically structured as triples: (subject, relation, object). While this format is efficient for machines, human language is inherently varied. A single relation can be expressed through numerous paraphrases—for instance, "is the vocalist of," "performed," and "recorded" all describe a similar relationship between an artist and a song. Conversely, words like "shot" are ambiguous, changing meaning based on whether the arguments are (Director, Movie) or (Assassin, Politician).
POLY is a massive linguistic resource designed to address these challenges. It organizes over 1 million relational phrases into synonymous clusters across 61 languages. By combining fine-grained semantic typing with a novel multilingual bridging technique, POLY achieves high precision and recall, significantly outperforming previous state-of-the-art tools.
Methodology: Semantic Typing and Multilingual Bridging
The construction of POLY involves a two-stage pipeline: relational phrase typing and clustering.
Relational Phrase Typing
To handle ambiguity, POLY "canonicalizes" arguments using Knowledge Bases like YAGO3 and WordNet. Instead of treating phrases as raw text, it assigns fine-grained types to the arguments. For example, it distinguishes "wrote" in the context of <author> and <paper> from other uses. This process utilizes a hierarchy of 734 WordNet types to ensure semantic consistency across the dataset.
Multilingual Bridging
The core innovation of POLY is the use of translation as a semantic filter. The authors hypothesize that if two distinct English phrases frequently translate to the same phrase in a foreign language, they are likely synonyms. This "multilingual bridging" acts as a natural disambiguator, as different senses of an English word often translate to different words in other languages.
As shown above, the Spanish phrase "fue filmado por" acts as a bridge between "was shot by" and "was filmed by." The system calculates probabilistic similarity through these foreign bridges and uses graph clustering (modularity-based partitioning) to group synonymous phrases into equivalence classes.
Key Results and Insights
POLY's scale and accuracy represent a significant leap forward in relational paraphrase mining.
- Scale: The resource contains approximately 1 million relational phrases organized into 160,000 clusters.
- Precision: Human evaluation yielded a precision of 0.91 for top-ranked synonyms and 0.83 for random samples.
- Performance: POLY achieved an F1-score of 0.58, significantly higher than previous benchmarks like PATTY (0.42) and DEFIE (0.44).
The use of semantic type signatures (e.g., <location> ... <location>) ensures that clusters remain coherent and contextually relevant. This approach also proved highly effective in downstream tasks; for example, in Question Answering (QA) on the WebQuestions dataset, POLY increased the number of correctly answered questions from 365 to 530.
Conclusion and Future Directions
By leveraging the natural disambiguation provided by translation, POLY provides a high-coverage, high-precision repository for natural language processing. While the system is currently sensitive to machine translation errors and language imbalances, future work aims to process all 61 languages jointly to further refine the clusters and improve global coverage.
Summary updated