Comparative Evaluation of Machine Learning and Deep Learning Models for Targeted Metaphor Detection

<title>Abstract</title> <p>Metaphor detection is an important natural language processing task because figurative language often conveys meanings that differ from literal word usage. This paper evaluates multiple machine learning and deep learning approaches for targeted metaphor detection, where the goal is to classify whether a predefined candidate word is used metaphorically or literally within a given textual context. Using a labeled dataset of 1,870 training samples and 800 test samples covering seven target words---road, candle, light, spice, ride, train, and boat---we compare TF-IDF-based logistic regression, LDA combined with Sentence-BERT embeddings and random forests, LSTM-based neural models, BERT-enhanced LSTM representations, XGBoost with Word2Vec embeddings, and a majority-vote consensus model. The LDA-SBERT-Random Forest model achieved the strongest overall performance, with 84\% accuracy and a weighted F1-score of 0.82. The consensus model achieved 82\% accuracy and a weighted F1-score of 0.80, suggesting that ensemble voting provided balanced predictions but did not outperform the best individual model. These results indicate that hybrid feature representations combining contextual sentence embeddings with topic-level information can be effective for metaphor detection in small labeled datasets.</p>