双向循环网络测验

当您参加测验时，就会出现答案和解释。

1. 双向 RNN 如何处理输入序列？
   • Processes only the first half forward
   • Processes it in forward and backward directions and combines the contextual states
   • Uses future tokens as training labels
   • Runs the same hidden state twice in the same direction
2. 为什么来自两个方向的上下文有助于标记诸如“bank”之类的词？
   • Words before and after it can clarify the intended meaning
   • The word has only one possible meaning
   • The model needs no surrounding text
   • The label is stored in the sequence position
3. 两个方向的隐藏状态通常如何在某个位置进行组合？
   • Average all positions in the sequence first
   • Discard the backward state
   • Use only the final forward state for every position
   • Concatenate them or otherwise combine them for the output
4. 为什么双向 RNN 可能不适合流式生成？
   • It cannot process tokens in order
   • It requires class labels at inference
   • It needs future sequence context that is not yet available
   • 它仅适用于图像
5. 离线序列标记与双向性结合的一个优势是什么？
   • Each position can use context on both sides
   • The model needs no sequence boundaries
   • It removes the need for recurrent parameters
   • It guarantees correct token labels
6. Do forward and backward passes share the same recurrent parameters in the guide’s described architecture?
   • Yes, by definition every bidirectional RNN shares weights
   • No; it uses separate recurrent layers for the two directions
   • Only when the sequence has even length
   • They have no parameters
7. Which task is a natural fit when the entire sequence is available?
   • Left-to-right text generation before later tokens exist
   • Live word prediction with zero lookahead
   • Part-of-speech tagging or named-entity recognition
   • Predicting the next unseen sensor reading in a stream
8. What does bidirectional context not guarantee?
   • Access to following input tokens in offline use
   • A combined representation at each position
   • Two traversal directions
   • That the model will correctly interpret every token