TL;DR
This paper introduces a novel method for automatically detecting "false friends" – words that look or sound similar across languages but have different meanings – using word embeddings. Their approach achieves high accuracy and coverage for Spanish and Portuguese, offering a valuable tool for language learners and NLP applications by leveraging readily available monolingual corpora and a small bilingual lexicon.
What Problem Does This Paper Address?
Language learners often struggle with "false friends," words that appear identical or very similar in two languages but carry distinct meanings. This can lead to significant confusion and errors. While traditional false friends are well-known, the paper also highlights less obvious types, like words with partial meaning differences or those used in different syntactic structures. Existing methods for automatically detecting these linguistic pitfalls often rely on large, expensive bilingual corpora or struggle with less common words. This paper aims to bridge this gap by proposing a high-coverage, automatic method that can be applied to language pairs with limited bilingual resources, focusing initially on the closely related Spanish and Portuguese.
Methods
The core of the proposed method lies in leveraging word embeddings, specifically the word2vec model, to represent words in a vector space. The authors' key innovation is to use a technique developed by Mikolov et al. (2013b) to learn a linear transformation between the vector spaces of two languages. This transformation allows them to project word vectors from one language into the other's vector space.
The process involves:
- Building Language-Specific Vector Spaces: Using
word2vecon large, untagged monolingual Wikipedia corpora for Spanish and Portuguese. - Learning a Linear Transformation: This is achieved by using a small bilingual lexicon (initially derived from WordNet, but adaptable to other sources like machine translation outputs). The transformation aims to map word vectors from the source language space to the target language space such that cognates (words with similar meanings) remain close.
- Classifying False Friends vs. Cognates: For a candidate word pair, the source word's vector is transformed into the target language's space. If the transformed vector is close to the target word's vector, they are classified as cognates. If they are distant, they are classified as false friends.
- Supervised Classification: To refine this, a Support Vector Machine (SVM) classifier is trained using three features:
- Cosine distance between the transformed source vector and the target vector.
- The number of word vectors in the target space closer to the target vector than the transformed source vector.
- The sum of distances between the target vector and the transformed source vectors of the five nearest neighbors of the source word in its original space.
Datasets
- Monolingual Corpora: Spanish and Portuguese Wikipedia dumps were used to train the
word2vecmodels. - Bilingual Lexicon: Initially, a lexicon was built using Open Multilingual WordNet, specifically aligning Spanish and Portuguese synsets. The paper notes that this resource can be substituted with other bilingual dictionaries or even machine translation outputs.
- Evaluation Dataset: The authors used a dataset of 710 Spanish-Portuguese word pairs, comprising 338 cognates and 372 false friends, previously compiled by Sepúlveda and Aluísio (2011). This dataset was annotated by two people and includes pairs with identical spellings and those differing by a few characters.
Key Insights and Ideas
- Word Embeddings as a Bridge: The core idea that word embeddings can capture semantic relationships and that a learned linear transformation can align these spaces across languages is powerful. This allows for cross-lingual comparison without requiring massive parallel corpora.
- Small Lexicons Suffice: The research demonstrates that a relatively small bilingual lexicon (even fewer than a thousand entries) can be sufficient to learn a good transformation, making the method applicable to low-resource language pairs.
- Beyond Traditional False Friends: While the paper focuses on traditional false friends for simplicity, the underlying method has the potential to be extended to detect more subtle linguistic differences.
- Feature Engineering Matters: The inclusion of multiple features for the SVM classifier, beyond just the direct vector distance, significantly improves performance, suggesting that capturing contextual and neighborly relationships is crucial.
Key Results
The proposed method achieved state-of-the-art results for the Spanish-Portuguese pair.
- Accuracy: The method achieved an accuracy of 77.28% with a high coverage of 97.91%.
- Comparison: It outperformed a WordNet baseline (68.18% accuracy, 55.38% coverage) and Sepúlveda and Aluísio's methods (63.52% and 76.37% accuracy, respectively). It was also competitive with Apertium (77.75% accuracy, 66.01% coverage).
- Frequency Enhancement: Including word frequency information as an additional feature boosted accuracy to 79.42% while maintaining high coverage.
- Vector Configuration: Experiments showed that higher vector dimensions did not consistently improve accuracy, and in some cases, worsened performance. Using Spanish as the source language yielded slightly better results, possibly due to the larger corpus size. Including two-word phrases did not improve results.
Limitations and Future Work
- Limitations:
- The paper acknowledges that the accuracy is measured only on covered pairs, meaning words not present in the training corpora or the bilingual lexicon are excluded.
- The method does not explicitly deal with determining if two words are "similar" in the first place, which is a prerequisite for identifying them as potential false friends or cognates.
- The current focus is on traditional false friends; more nuanced types are not fully addressed.
- Future Work:
- Experimenting with more advanced word vector representations and linear transformation techniques (e.g., those from Artetxe et al., 2017, 2018).
- Developing fine-grained classifications to distinguish between different types of false friends and cognates.
- Applying the method to other language pairs, especially those with scarce bilingual resources.
- Exploring the integration of other information sources, such as word frequencies, to further enhance classifier performance.
Summary updated