Abstract
We propose a new approach to identifying semantically similar words across languages based on the idea that two words in different languages are similar if they are likely to generate similar words as their top semantic word responses. This concept, drawn from cognitive science, detects likely human free word associations given a cue word. Our method (1) utilizes a probabilistic multilingual topic model trained on comparable data to quantify semantic word responses and (2) provides ranked lists of similar words based on response vector similarity. Evaluated on bilingual lexicon extraction (BLE), our response-based method is more robust and outperforms state-of-the-art methods that operate directly in latent cross-lingual topic spaces.
Introduction
Building bilingual dictionaries, or Bilingual Lexicon Extraction (BLE), is fundamental for machine translation and cross-lingual information retrieval. While parallel corpora (direct translations) are ideal for this task, they are scarce for most language pairs. Consequently, researchers rely on "comparable corpora"—texts discussing similar topics without being direct translations, such as Wikipedia articles in different languages.
Traditional methods using comparable corpora often map words into a shared "latent topic" space. However, these topics can be abstract and noisy. This paper introduces a more granular approach that mimics human "free word association" to identify translation pairs more accurately.
The Semantic Response Approach
The core of our method is the concept of semantic word responding. Instead of comparing words directly through abstract mathematical distributions, we ask: "If a specific word is given as a cue, what other words (in any language) are likely to be triggered?"
The process involves two main steps:
- Multilingual Topic Modeling: A Bilingual Latent Dirichlet Allocation (B-LDA) model is trained on comparable data to learn a shared set of concepts (topics) across languages.
- Generating Response Vectors: Topics serve as a bridge to calculate the probability $P(w_2|w_1)$ of one word triggering another. Each word is represented by a high-dimensional "response vector" reflecting its association strength with the entire vocabulary of both languages.
By comparing these response vectors using the Bhattacharyya coefficient, we identify semantically similar words. This approach effectively "de-noises" the semantic signal by projecting low-dimensional topic information back into the high-dimensional vocabulary space, emphasizing the most descriptive associations.
Experimental Setup
We evaluated the method on Wikipedia article pairs for three language sets: Italian-English (IT-EN-W), Spanish-English (ES-EN-W), and Dutch-English (NL-EN-W). We also tested an augmented Dutch-English set (NL-EN-W+EP) using Europarl data to assess the impact of higher-quality training material. Performance was measured against a ground-truth set of 1,000 translation pairs using Top-1 accuracy ($Acc_1$), Mean Reciprocal Rank ($MRR$), and Top-10 accuracy ($Acc_{10}$).
Results and Discussion
The proposed Response-BC method consistently outperformed standard topic-based comparison (Topic-BC) and other baselines across all datasets.
Key findings include:
- Superior Accuracy: In the IT-EN-W task, Response-BC achieved an $Acc_1$ of 62.2%, significantly higher than the 57.8% achieved by Topic-BC.
- Robustness: The method proved more robust to the noise inherent in latent topics. By using topics as a bridge rather than a final destination, the model captures specific vocabulary "signatures" that are more indicative of direct translations.
- Data Scaling: Performance improved notably with cleaner training data (NL-EN-W+EP), suggesting the method scales effectively with data quality.
Conclusion
Representing cross-lingual word similarity through semantic word responses provides a powerful alternative to direct topic-based models. While computationally more intensive due to high-dimensional vectors, the method's ability to filter noise and capture asymmetric associations leads to state-of-the-art results in bilingual lexicon extraction. Future work may involve integrating orthographic clues (cognates) to further enhance performance for related language pairs.
Summary updated