TL;DR
This paper presents a web service for automatically extracting Multi-Word Expressions (MWEs) and building lexicons, integrated into a larger platform for linguistic resource creation. The tool is designed to be computationally light, language-independent, and produce MWE lexicons in XML format, achieving approximately 80% precision.
What Problem Does This Paper Address?
Multi-word expressions (MWEs) pose significant challenges in various Natural Language Processing (NLP) tasks, including Machine Translation, Information Retrieval, and text mining. Their non-literal and often unpredictable meanings make them difficult to handle with traditional word-by-word processing. While methods for MWE acquisition exist, readily available and customizable tools, especially for languages other than English, are scarce. This paper addresses this gap by developing a practical, web-service-based tool for automatic MWE lexicon building.
Methods
The core of the MWE acquisition tool involves a multi-step process:
- Collocation Extraction (Window-based): This initial step identifies candidate MWEs by looking at sequences of tokens within a defined window, using only Part-of-Speech (POS) tags. It's language-independent and doesn't require seed words.
- Pre-filtering and Ranking: To reduce computational load and noise, collocations are filtered based on frequency. Two methods are proposed: AverageFrequency PreFilter and MaxFrequency PreFilter. The AverageFrequency PreFilter is used in experiments.
- Pattern Extraction: For retained collocations, the algorithm extracts all intervening patterns between the first and last words, including their POS tags and lemmas.
- Pattern-based Collocation Filtering and MWE Selection (SigmaPatternExtraction): This is a key novel step. It analyzes the distribution of patterns associated with a collocation. If the pattern distribution shows significant outliers (indicating a strong, consistent pattern for the MWE), the collocation is retained. This is quantified using mean and standard deviation of pattern frequencies. A simpler baseline method, FirstPatternExtraction, is also described.
- Lexicon Building: Selected MWEs are compiled into an XML-encoded lexicon conforming to the Lexical Markup Framework (LMF) standard.
Datasets
- Corpus: The experiments were conducted on a domain-specific corpus of Italian from the ENVIRONMENT domain, automatically produced through focused monolingual crawling. The corpus size is approximately 37 million word tokens.
- Gold Standard: A reference resource was created by semi-manually collecting relevant nominal Italian MWEs from authoritative web glossaries and thesauri. Frequencies of these MWEs were computed in the corpus.
Key Insights and Ideas
- Language Independence: The core extraction algorithm relies on POS tags and positional information, making it adaptable to different languages without requiring language-specific lexicons or rules.
- Computational Efficiency: The tool is designed to be "computationally light," avoiding overly intensive methods to facilitate its use in a distributed web service environment.
- Pattern Distribution for MWE Identification: The SigmaPatternExtraction algorithm's novel approach of analyzing the variance in pattern distributions is crucial. A high variance suggests a stable, genuine MWE, distinguishing it from random word co-occurrences.
- Web Service Deployment: Integrating the tool as a web service within a distributed platform is a significant contribution, making linguistic resources more accessible and enabling workflow chaining.
- Precision vs. Recall Trade-off: The paper highlights that while pre-filtering reduces recall, it significantly increases precision, which is often a more critical factor for practical applications.
Key Results
- The tool achieved approximately 80% precision in evaluation experiments.
- The AverageFrequency PreFilter reduced the number of collocations from over 2 million to around 260,000, while maintaining a significantly higher density of eligible pairs relative to the gold standard.
- The SIGMA extraction method (using SigmaPatternExtraction) generally outperformed the simpler FIRST extraction method in terms of precision and recall, especially in the REDUCED evaluation setting.
- In a manual evaluation of the top 1000 highest frequency results, the initial precision was 40%, which improved to 78% after adding genuine MWEs found during the evaluation to the gold standard. This suggests that even without seed words, high precision can be achieved.
Limitations and Future Work
- Gold Standard Incompleteness: The authors acknowledge that the gold standard is not exhaustive, which can affect precision evaluation.
- Stemming Simplification: The reduction algorithm for evaluation uses a naive stemming method, which might slightly underestimate precision.
- Future Work:
- Adding further filtering/cleaning of extracted MWEs, including merging substring patterns.
- Extracting MWEs with POS patterns to learn "free slots."
- Implementing language-specific fine-tuning through post-filtering and post-editing.
- Automatic conversion to RDF format for publishable Linked Open Data.
- Conducting task-based evaluations (e.g., machine translation, syntactic parsing) to assess the impact of the acquired MWEs.
- Extending manual evaluation for a more accurate precision estimate.
- Improving user feedback integration for configurable parameters and output formats.
Summary updated