TL;DR
This paper demonstrates that leveraging Chinese verb translations can significantly improve the automatic classification of English verbs, even outperforming English-only features. The key insight is that syntactic differences across languages can reveal underlying semantic properties that are otherwise hard to detect, and that this crosslinguistic information is most powerful when combined.
What Problem Does This Paper Address?
Automatically classifying verbs into semantic categories is crucial for building comprehensive natural language processing (NLP) systems. These categories help understand verb meaning and usage, but manually creating such lexicons is expensive and time-consuming. Existing corpus-based methods rely on the assumption that a verb's meaning is reflected in its syntactic usage. However, this mapping can be noisy, and some semantic distinctions might not be overtly expressed in a single language's syntax. The paper tackles this by exploring how information from a different language, specifically Chinese, can augment English verb classification.
Methods
The authors build upon previous work (Merlo & Stevenson, 2001) that uses statistical features derived from argument structure to classify English verbs. They extend this by:
- Feature Extraction: They extract 5 core English features related to animacy, transitivity, and causativity from a large English corpus (British National Corpus).
- Crosslinguistic Feature Extraction: For Chinese, they develop 28 features, including:
- POS-tags: Using the CKIP tagger to capture subcategorization and stative/active distinctions.
- Passive Particles: Identifying overt particles that signal passive constructions.
- Periphrastic (Causative) Particles: Identifying external particles indicating a causal agent.
- Morpheme Information: (Found to be less useful and not heavily discussed).
- Translation Strategy: Crucially, they do not rely on parallel corpora. Instead, they use a monolingual Chinese corpus (MNews) to find translations for English verbs and extract Chinese features from these translations.
- Machine Learning: They employ the C5.0 decision tree algorithm for classification, using 10-fold cross-validation repeated 50 times for robust evaluation. They experiment with English-only features, Chinese-only features, and combinations of both.
Datasets
- English Data: British National Corpus (BNC, 100M words), POS-tagged and chunked.
- Chinese Data: Mandarin Chinese News Text (MNews, People's Daily and Xinhua newswire sections, ~165M characters), tagged with the CKIP tagger.
- Verb Classes: Three semantic classes of English verbs were used: manner of motion, change of state, and creation/transformation.
Key Insights and Ideas
- Syntactic Differences Reveal Semantics: Languages express semantic properties differently. What's subtle in English syntax might be overt in Chinese, providing complementary information.
- Monolingual Transfer is Sufficient: You don't need perfectly aligned parallel corpora to benefit from crosslinguistic transfer. Using monolingual data to find translations and extract features is effective.
- Multilingual Combination is Powerful: Combining features from different languages consistently outperforms monolingual features, suggesting that diverse syntactic views of the same underlying semantics enrich the classification process.
- Chinese Features Shine: Chinese-only features alone often outperform the best English-only features for classifying English verbs, highlighting the value of the crosslinguistic information.
Key Results
- 3-Way Classification (Manner of Motion, Change of State, Creation/Transformation):
- English-only (best features): 67.6% accuracy.
- Chinese-only (best features): 82.1% accuracy.
- Multilingual (best combination): 85.2% accuracy (outperforming both monolingual sets).
- Baseline accuracy: 33.3%.
- 2-Way Classifications: In most pairwise comparisons, multilingual combinations achieved the highest accuracy, often exceeding 90%. For example, in the manner of motion vs. creation/transformation task, the multilingual approach reached 93.7% accuracy.
- Chinese Features Outperform English: Across the board, the best Chinese-only features consistently achieved higher accuracy than the best English-only features.
Limitations and Future Work
- Sense Ambiguity: The authors acknowledge that their corpora are not sense-tagged, and they assume statistical features reflect the predominant sense. This is a common limitation in unsupervised methods.
- Feature Engineering: While they found effective features, the process of manually identifying them for specific classes is labor-intensive.
- Future Work: The paper mentions ongoing work in other languages (Italian and German) that confirms similar crosslinguistic transfer benefits. They also note that future research could explore more general feature spaces that don't require manual feature selection for specific classes.
Summary updated