Bidirectional Language Modeling involves predicting missing or masked information conditioned on BOTH left and right context — used by BERT and RoBERTa, it enables deep understanding of sentence structure and ambiguity resolution that unidirectional (causal) models miss.

Mechanism

• Masking: Inputs are masked (MLM).
• Attention: Self-attention is unmasked (full visibility) — every token can attend to every other token.
• Prediction: The model predicts the masked token using clues from before AND after it.
• Result: "bank" could be river or finance — "The _bank_ overflowed" (right context "overflowed" disambiguates).

Why It Matters

• Understanding: Essential for tasks like Classification, NER, and QA where seeing the whole sentence is crucial.
• Representation: Produces richer contextual embeddings than unidirectional models.
• Not Generative: Cannot easily generate text (which requires left-to-right production), making it less suitable for chatbots.

Bidirectional Language Modeling is reading the whole sentence — using full context to understand meaning, primarily for understanding/discriminative tasks.