TL;DR
This paper introduces a novel method for building bilingual dictionaries from comparable corpora (texts on similar topics, not direct translations) by analyzing syntactic dependencies between words and their translations. This approach significantly outperforms traditional methods that rely solely on word co-occurrence in similar contexts, especially when dealing with words that might appear in similar but non-translational contexts.
What Problem Does This Paper Address?
Building bilingual dictionaries is crucial for many Natural Language Processing (NLP) tasks, like machine translation and cross-lingual information retrieval. The ideal way to create these dictionaries is by aligning parallel corpora (texts that are direct translations of each other). However, creating parallel corpora is often expensive and time-consuming, especially for less common language pairs.
This paper focuses on extracting dictionaries from comparable corpora. These are collections of documents that cover similar topics but aren't direct translations. The challenge here is that words might appear in similar contexts even if they aren't translations of each other. For instance, "economics" and "medicine" might both appear in academic contexts, leading traditional methods to incorrectly link them. The paper aims to overcome this by leveraging deeper linguistic information than just word co-occurrence.
Methods
The core of the proposed method is dependency heterogeneity. The authors observe that a word and its translation tend to share similar syntactic relationships (modifiers and heads) within sentences, even if the surrounding words are different.
Here's a breakdown of their approach:
- Corpora Preprocessing: They start with comparable corpora (Wikipedia in this case) and use language-specific tools:
- Chinese: Morphological analyzer.
- English: POS-tagger.
- Both: A dependency parser (MaltParser) to identify syntactic relationships like subject-verb, noun-modifier, etc.
- Further cleaning: Stemming for English, stop-word removal, and filtering out dependencies involving punctuation.
- Dependency Heterogeneity Vector Calculation: For each word, they construct a vector representing the heterogeneity of its syntactic relationships. This vector has four components:
HNMODHead: Proportion of unique heads for noun modifiers.HSUBHead: Proportion of unique heads for subjects.HOBJHead: Proportion of unique heads for objects.-
HNMODMod: Proportion of unique noun modifiers. These components quantify how varied a word's syntactic partners are. -
Bilingual Dictionary Extraction: To find translation pairs, they calculate the Euclidean distance between the dependency heterogeneity vectors of candidate words from the source and target languages. A smaller distance indicates a higher probability of being a translation pair.
Datasets
- Training/Learning Data: Chinese and English Wikipedia pages linked via inter-language links. The authors mention using 1,132,492 Chinese sentences and 665,789 English sentences for learning the dependency heterogeneity vectors.
- Evaluation Data: 250 randomly selected Chinese/English single-noun pairs, sampled from the aligned titles of the Wikipedia pages. This set was divided into 5 folders for cross-validation.
Key Insights and Ideas
- Syntactic Structure Over Contextual Similarity: The most significant insight is that syntactic relationships (dependency structures) are a more robust signal for translation equivalence than mere contextual co-occurrence, especially in comparable corpora.
- Dependency Heterogeneity as a Signal: The idea that words and their translations share similar syntactic roles (being modified by similar words or acting as heads for similar words) is a powerful observation. This "dependency heterogeneity" captures a deeper semantic and structural connection.
- Robustness to Contextual Ambiguity: The method is less susceptible to false positives where words appear in similar contexts but are not translations, a common pitfall for bag-of-words approaches.
- Data-Driven Approach: The reliance on automated dependency parsers makes the approach adaptable to different language pairs without requiring manual creation of syntactic rules or templates.
Key Results
- Significant Accuracy Improvement: The proposed dependency heterogeneity approach achieved significantly higher accuracy and MMR (Mean Reciprocal Rank) scores compared to the traditional context-based (bag-of-words) approach. For example, average accuracy increased from 0.132 to 0.208 (a 57.58% improvement), and average MMR from 0.064 to 0.104 (a 62.50% improvement) for Top5 evaluation.
- Effectiveness of Vector Components: Analysis showed that both head and modifier dependencies contribute to the improved performance. Specifically, the NMOD (noun modifier) relationships (both head and modifier) were found to be particularly important.
- Qualitative Analysis: The paper highlights examples where their method successfully extracts pairs missed by the context-based approach. Conversely, it also identifies cases where corpus "dirtiness" (e.g., incorrect word combinations due to line breaks in the English corpus) negatively impacted performance, suggesting that corpus quality is still a factor.
Limitations and Future Work
- Corpus Quality: The authors acknowledge that the cleanliness of the comparable corpora is crucial. Errors in the corpus (like incorrect word combinations) can negatively affect the accuracy of dependency vectors and thus dictionary extraction. They plan to address this by cleaning the corpora in future work.
- Parsing Accuracy and Threshold Setting: The performance is dependent on the accuracy of the dependency parser and the chosen threshold
kfor sentence length. Exploring different parsers and thresholds is a future direction. - Extending to Multi-word Expressions: The current approach focuses on single nouns. Future work includes extending it to extract multi-word dictionary entries.
- Adaptation to Other Language Pairs: While the data-driven nature suggests adaptability, further experiments on other language pairs are planned.
Summary updated