TL;DR
This paper presents a system for automatically extracting paraphrastic phrases from medium-sized corpora by leveraging a semantic network and a corpus. The approach aims to reduce the manual effort required for building resources for Natural Language Processing (NLP) tasks like Information Extraction (IE).
What Problem Does This Paper Address?
Developing NLP systems, particularly for Information Extraction (IE), often requires extensive manual creation of linguistic resources like extraction patterns. This process is time-consuming and limits the portability of IE systems to new domains or corpora. Existing methods that rely on large web redundancy are not suitable for smaller, more focused corpora. The authors propose a method to automatically acquire paraphrastic phrases, which are semantically equivalent expressions, to alleviate this burden.
Methods
The core of the proposed approach is a knowledge-rich method for paraphrase acquisition. It involves:
- Semantic Network Integration: Utilizing a large, knowledge-rich semantic network (like WordNet) to infer semantic similarity between words. This goes beyond simple hyponymy to include relationships like meronymy and stereotypical associations.
- Corpus-Based Validation: Using the target corpus to discover actual expressions and filter out irrelevant ones suggested by the semantic network.
- Weakly Supervised Learning: The system is initiated with a single "seed pattern" (a predicative linguistic structure) and a representative corpus.
- Semantic and Syntactic Expansion: The system automatically acquires paraphrases by:
- Comparing words from the seed pattern with words in the corpus using semantic similarity measures derived from the semantic network.
- Performing shallow syntactic analysis to ensure the found word pairs correspond to a paraphrastic structure.
- Employing meta-graphs to model and recognize syntactic variations (e.g., passive voice, insertion of modifiers) that can alter the surface form of paraphrases.
- Manual Validation: The acquired patterns are then manually validated by an end-user, though the authors emphasize this process is rapid.
Datasets
- Training Corpus: 200 texts from a French financial corpus about companies buying other companies.
- Test Corpus: 100 texts from the same French financial corpus.
- Semantic Network: A knowledge-rich semantic net encoding a large variety of semantic relationships, described as being particularly rich for French.
Key Insights and Ideas
- Bridging Semantic Networks and Corpora: The paper highlights the power of combining structured knowledge from semantic networks with the real-world expressions found in a corpus. This synergy allows for both broad semantic coverage and practical linguistic validation.
- Weak Supervision for Efficiency: Requiring only a single seed pattern significantly reduces the initial manual effort compared to traditional methods.
- Handling Syntactic Variations: The use of meta-graphs to model syntactic transformations is crucial for capturing paraphrases that might otherwise be missed due to structural differences.
- The Importance of Semantic Distance: The paper emphasizes that semantic relatedness is more than just hyponymy, advocating for measures that incorporate a wider range of semantic relationships.
Key Results
- The system successfully acquired paraphrastic phrases from a French financial corpus. For the example "cession de société" (company transfer), it discovered expressions like "reprise des activités," "rachat d'activité," and "acquérir des magasins."
- The acquired patterns include both nominal and verbal phrases.
- An evaluation on a financial corpus showed that the system using automatically acquired resources achieved approximately 10% lower performance (measured by precision, recall, and F-score) compared to a system with manually defined resources. However, the authors argue that the time saved in resource development makes this approach highly usable.
- The manual validation of acquired patterns was found to be very rapid, with only a small percentage (estimated 25%) needing rejection.
Limitations and Future Work
- Performance Gap: The automatically acquired resources are about 10% less efficient than manually defined ones. This decrease can vary depending on the specific slot being analyzed.
- Errors: Errors can arise from missing relations in the semantic net or from syntactic transformations not registered by the system.
- Future Work: The paper suggests that the performance gap could be reduced by improving the semantic network and the syntactic variation management module. The authors also mention the potential for co-training algorithms to further enhance the acquisition process.
Summary updated