A developer is building a text classification model to organize legal contracts. The documents contain a vast vocabulary with specialized jargon and numerous synonyms (e.g., "attorney", "counselor", "counsel", "lawyer"). Which text representation model will best capture the semantic relationships between these terms to improve classifier performance?

In text classification, capturing semantic similarity is critical when dealing with domain-specific jargon and synonyms. Pre-trained dense word embeddings, such as Word2Vec or GloVe, represent words as dense, low-dimensional vectors in a continuous vector space, typically 100 to 300 dimensions. These models are trained on large corpora to place semantically similar words near one another in this space. For a classifier, this means that even if a training document contains "attorney" and a test document contains "lawyer," the model can utilize the proximity of their vectors to make accurate predictions. One-Hot Encoding represents words as sparse, high-dimensional vectors where only one index is 1 and all others are 0. The dot product between any two distinct word vectors is always 0, meaning it cannot capture any semantic relationship between synonyms. Bag-of-Words and Term Frequency-Inverse Document Frequency are vocabulary-dependent sparse models. They represent documents by word counts or statistical weights, treating each word as an independent feature. They do not model word semantic similarity, making them less effective at generalizing across different synonyms.