Abstract: Patrick Ziering, Lonneke van der Plas, Hinrich Schütze. Proceedings of the Sixth International Joint Conference on Natural Language Processing.
2013.
Introduction
Building high-quality semantic lexicons—lists of words belonging to specific categories like PROCESSES or OBJECTS—is a foundational task for information extraction. However, manual compilation is labor-intensive and costly. Researchers often turn to bootstrapping, an iterative process where a system starts with a few "seed" words, identifies linguistic patterns they appear in (e.g., "the process of [X]"), and uses those patterns to discover new terms.
A major hurdle in this approach is semantic drift. If a system incorrectly extracts an ambiguous term—such as "Iris," which could refer to a flower or a person's name—it may begin populating the lexicon with unrelated terms. Once drift occurs, the quality of the resulting lexicon degrades rapidly.
Methodology: The Multilingual Ensemble Framework
This paper introduces a multilingual approach to lexicon bootstrapping designed to "keep the system honest." By utilizing a parallel corpus (text translated across languages), the authors demonstrate that different languages can act as mutual filters to suppress errors.
Building upon the monolingual Basilisk algorithm, the authors developed a Multilingual Ensemble Framework that operates simultaneously in English and German using a parallel corpus of patent documents. The process follows three key steps:
- Iterative Induction: Independent systems in each language suggest new candidate terms based on local linguistic patterns.
- Consensus Check: Candidate terms are translated between languages using a phrase table generated by standard machine translation tools (e.g., Moses).
- Filtering: A term is only added to the final lexicon if it survives a consensus check—meaning the term (or its translation) was independently suggested by the bootstrapping process in both languages.
This method exploits the fact that languages are not isomorphic. An ambiguous term in English (e.g., "energy storage," which can be an OBJECT or a PROCESS) often maps to distinct, unambiguous terms in German (Energiespeicher vs. Energiespeicherung).
Experimental Setup
The researchers evaluated their framework using 177,317 patent documents from the European Patent Office (EPO) in German and English. They focused on two semantic categories critical for patent analysis: PROCESS (e.g., stretching, molding) and TECHNICAL QUALITY (e.g., power consumption, light reflection index). Seed sets were derived from WordNet and Wikipedia.
Key Results
The multilingual approach significantly outperformed the monolingual baseline across all test cases. By requiring cross-lingual consensus, the system effectively "triangulated" the intended meaning and pruned away the drift that typically plagues monolingual systems.
As shown in the results table, the German PROCESS lexicon saw a dramatic accuracy jump from 0.730 to 0.980. The English PROCESS lexicon also improved, though less dramatically (from 0.740 to 0.790). For TECHNICAL QUALITY, both languages achieved high precision, reaching 0.960 for German and 0.955 for English.
The researchers observed an asymmetric impact: English provided a stronger filtering effect for German than vice versa in the PROCESS category. This is attributed to the German suffix -ung, which is highly productive but ambiguous, often leading monolingual German systems toward drift. The English side acted as a stabilizer, filtering out these incorrect inductions.
Discussion and Future Work
A notable advantage of this method is that it is knowledge-lean. It does not require deep linguistic parsers or manually defined rules, relying instead on raw parallel text and standard alignment tools. Furthermore, German's linguistic structure—specifically capitalized nouns and compound words—provided "boundary information" that helped the system identify complex multi-word expressions (MWEs) in English without needing a specialized noun chunker.
The primary limitation remains that extreme ambiguity in one language can occasionally bypass the consensus check. The authors suggest that incorporating a third or fourth language into the ensemble would likely suppress semantic drift even further, as the probability of a "false" term finding consensus across multiple distinct languages is statistically negligible.
Summary updated