TL;DR
This paper explores how linguistic features, beyond just word usage patterns, can improve predictions of semantic change in Spanish. By combining word embeddings with features like polysemy and neological synonyms, the authors achieved promising results in a shared task, highlighting the value of expert-curated linguistic knowledge.
What Problem Does This Paper Address?
Tracking how word meanings evolve over time (semantic change) is a crucial task in computational linguistics. Traditional approaches often rely on analyzing word usage in historical text corpora using word embeddings. However, these methods can sometimes miss subtle shifts or struggle with nuanced changes. This paper addresses the question of whether incorporating explicit linguistic knowledge, such as a word's inherent complexity (polysemy) or its relationship with newer words (neological synonyms), can enhance the accuracy of predicting semantic change. The authors specifically tackle this challenge within the Spanish language context, using data from historical and modern corpora.
Methods
The CoToHiLi team's approach involved two main components:
- Word Embedding Distances: They trained word embeddings separately on an "old" corpus (1810-1906) and a "modern" corpus (1994-2020). To compare these, they aligned the embedding spaces using three methods: supervised, semi-supervised, and unsupervised alignment. The cosine distance between a word's representation in the two aligned spaces served as a proxy for its semantic shift.
- Hand-crafted Linguistic Features: They incorporated several linguistically motivated features:
- Word Polysemy: Measured by counting the number of synsets a word belongs to in WordNet. The intuition is that more polysemous words might be more prone to change.
- Number of Neological Synonyms: Identified by finding synonyms in WordNet that do not appear in the older corpus, suggesting a word might be losing ground to newer terms.
- Levenshtein Distance to Closest English Word: This feature captures potential influence from English, a significant source of borrowings and semantic shifts in Romance languages.
These features were then combined in models to predict semantic change. They explored two main strategies: manual weighting of features based on expert linguistic judgment and supervised learning to automatically learn feature weights.
Datasets
The primary dataset used was provided by the LSCDiscovery shared task for Spanish. This dataset comprised:
- Old Corpus: Texts from Project Gutenberg published between 1810 and 1906.
- Modern Corpus: Texts from the OPUS project published between 1994 and 2020.
The shared task also provided annotated examples for training supervised models, consisting of a list of 20 target words with associated semantic shift scores.
Key Insights and Ideas
- Linguistic Features Matter: The most significant insight is that incorporating linguistically informed features, beyond just word usage patterns captured by embeddings, demonstrably improves semantic change prediction.
- Polysemy as a Driver: The paper reinforces the idea that words with multiple meanings (polysemy) are more likely to undergo semantic change, a concept with historical roots in linguistic theory.
- Neologisms as Indicators: The presence of new synonyms (neologisms) for a word can signal that the original word's semantic space is shifting or being encroached upon.
- Expert Knowledge is Valuable: Manually weighting features based on linguistic expertise yielded promising results, suggesting that human linguistic intuition can effectively guide computational models.
- Cross-lingual Influence: The Levenshtein distance to English words highlights the impact of language contact on semantic evolution.
Key Results
The paper reports results for two main sub-tasks: graded discovery (ranking words by degree of change) and binary change (classifying words as changed or not).
- Graded Discovery: Their system using manual weighting of features achieved a correlation of 0.282 with the ground truth, outperforming a baseline of 0.092. A supervised learning approach also yielded positive results.
- Binary Change: The manual weighting approach achieved an F1 score of 0.636, outperforming a baseline of 0.537.
- Optional Tasks: They achieved second place in terms of recall for the "sense gain" prediction task. For "sense loss," their systems did not perform well.
Limitations and Future Work
- Small Annotated Data: The authors acknowledge that the supervised learning approach was hampered by the very small size of the annotated data (20 words), which limited the reliability of learned feature weights.
- Sense Loss Prediction: Their systems struggled to predict "sense loss," indicating a need for further development in this specific area.
- Future Directions: The authors suggest future work could involve:
- Training word embeddings with more fine-tuned parameters optimized for the task.
- Exploring more sophisticated ways to integrate linguistic features.
- Further investigating the role of features beyond word frequency.
- Expanding the dataset for supervised learning.
Summary updated