Conceptual
Deprecated

Superseded as the default sentence encoder by BERT-style bidirectional transformer encoders, which condition every position on both directions in one parallel layer instead of two serial passes.

Learn more →
Why study this historical topic?

Superseded as the default sentence encoder by BERT-style bidirectional transformer encoders, which condition every position on both directions in one parallel layer instead of two serial passes.

Bidirectional LSTMs

Two independent LSTMs read the sequence forward and backward, and their hidden states are concatenated per position so every output sees full left and right context. This requires the whole sequence up front, which rules it out for streaming or causal prediction.