Answer to question neural networks

1. A method comprising:

receiving (i) an input text passage comprising a plurality of passage tokens and (ii) an input question string comprising a respective question token at each of a plurality of time steps;

processing the input text passage using an encoder neural network, wherein the encoder neural network is configured to receive the input text passage and to process the input text passage to generate a respective encoded representation for each passage token in the input text passage;

at each time step:

processing a decoder input for the time step using a decoder neural network, wherein the decoder input comprises at least a preceding token in the input question string and a preceding internal state of the decoder neural network, and wherein the decoder neural network is configured to, for each time step, process the decoder input to at least update an internal state of the decoder neural network;

processing the respective encoded representations and a preceding output of the decoder neural network using a matching vector neural network, wherein the matching vector neural network is configured to receive the preceding output of the decoder neural network and the encoded representations and to process the preceding output and the encoded representations to generate a matching vector for the time step;

processing the respective encoded representations and the preceding output of the decoder neural network using an attention vector neural network, wherein the attention vector neural network is configured to receive the preceding output of the decoder neural network and the encoded representations and to process the preceding output and the encoded representations to generate an attention vector for the time step, wherein the attention vector includes a respective value for each passage token; and generating an answer score that indicates how well the input text passage answers a question posed by the input question string based at least on the matching vectors for the time steps.

2. The method of claim 1 , wherein the attention vector has the same dimensionality as the matching vector.

3. The method of claim 2 , wherein the attention vector neural network has the same neural network architecture as the matching vector neural network but different parameter values.

4. The method of claim 1 , wherein at each time step the decoder input further comprises the attention vector for the time step.

5. The method of claim 1 , wherein generating the answer score comprises generating a matching vector score by determining an average or a weighted average of the entries of the matching vectors for the time steps.

6. The method of claim 5 , wherein the decoder neural network is further configured to, at each time step, generate a set of question token scores that includes a respective score for each question token in a vocabulary of possible question tokens.

7. The method of claim 6 , wherein generating the answer score comprises:

determining a perplexity score from an average perplexity of the question tokens in the input question string as reflected by the question token scores generated by the decoder neural network for the question tokens in the input question string; and

combining the perplexity score and the matching vector score to generate the answer score.

8. The method of claim 1 , wherein the encoder neural network and the decoder neural network are recurrent neural networks.

9. The method of claim 1 , wherein the preceding output is an output of one of the layers of the decoder neural network at a preceding time step.

10. The method of claim 1 , further comprising:

determining, based on the answer score, that the input passage answers the question posed by the input passage string; and

in response, outputting data derived from the input passage.

11. A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to perform operations comprising:

receiving (i) an input text passage comprising a plurality of passage tokens and (ii) an input question string comprising a respective question token at each of a plurality of time steps;

processing the input text passage using an encoder neural network, wherein the encoder neural network is configured to receive the input text passage and to process the input text passage to generate a respective encoded representation for each passage token in the input text passage;

at each time step:

processing a decoder input for the time step using a decoder neural network, wherein the decoder input comprises at least a preceding token in the input question string and a preceding internal state of the decoder neural network, and wherein the decoder neural network is configured to, for each time step, process the decoder input to at least update an internal state of the decoder neural network;

processing the respective encoded representations and a preceding output of the decoder neural network using a matching vector neural network, wherein the matching vector neural network is configured to receive the preceding output of the decoder neural network and the encoded representations and to process the preceding output and the encoded representations to generate a matching vector for the time step;

processing the respective encoded representations and the preceding output of the decoder neural network using an attention vector neural network, wherein the attention vector neural network is configured to receive the preceding output of the decoder neural network and the encoded representations and to process the preceding output and the encoded representations to generate an attention vector for the time step, wherein the attention vector includes a respective value for each passage token; and

generating an answer score that indicates how well the input text passage answers a question posed by the input question string based at least on the matching vectors for the time steps.

12. The system of claim 11 , wherein the attention vector has the same dimensionality as the matching vector.

13. The system of claim 12 , wherein the attention vector neural network has the same neural network architecture as the matching vector neural network but different parameter values.

14. The system of claim 11 , wherein at each time step the decoder input further comprises the attention vector for the time step.

15. The system of claim 11 , wherein generating the answer score comprises generating a matching vector score by determining an average or a weighted average of the entries of the matching vectors for the time steps.

16. The system of claim 15 , wherein the decoder neural network is further configured to, at each time step, generate a set of question token scores that includes a respective score for each question token in a vocabulary of possible question tokens.

17. The system of claim 16 , wherein generating the answer score comprises:

determining a perplexity score from an average perplexity of the question tokens in the input question string as reflected by the question token scores generated by the decoder neural network for the question tokens in the input question string; and

combining the perplexity score and the matching vector score to generate the answer score.

18. One or more non-transitory computer readable media storing instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:

receiving (i) an input text passage comprising a plurality of passage tokens and (ii) an input question string comprising a respective question token at each of a plurality of time steps;

processing the input text passage using an encoder neural network, wherein the encoder neural network is configured to receive the input text passage and to process the input text passage to generate a respective encoded representation for each passage token in the input text passage;

at each time step:

processing a decoder input for the time step using a decoder neural network, wherein the decoder input comprises at least a preceding token in the input question string and a preceding internal state of the decoder neural network, and wherein the decoder neural network is configured to, for each time step, process the decoder input to at least update an internal state of the decoder neural network;

processing the respective encoded representations and a preceding output of the decoder neural network using a matching vector neural network, wherein the matching vector neural network is configured to receive the preceding output of the decoder neural network and the encoded representations and to process the preceding output and the encoded representations to generate a matching vector for the time step;

processing the respective encoded representations and the preceding output of the decoder neural network using an attention vector neural network, wherein the attention vector neural network is configured to receive the preceding output of the decoder neural network and the encoded representations and to process the preceding output and the encoded representations to generate an attention vector for the time step, wherein the attention vector includes a respective value for each passage token; and

generating an answer score that indicates how well the input text passage answers a question posed by the input question string based at least on the matching vectors for the time steps.