Path dropout for natural language processing
Techniques are provided for improved training of a machine-learning model that includes multiple layers and is configured to process textual language input. The machine-learning model includes one or more blocks in which each block includes a multi-head self-attention network, a first connection for providing input to the multi-head self-attention network, and a second (residual) connection for providing the input to a normalization layer, bypassing the multi-head self-attention network. During training, the second connection is dropped out according to a dropout parameter. Additionally, or alternatively, an attention weight matrix is used for dropout by blocking diagonal entries in the attention weight matrix. As a result, the machine-learning model increasingly focuses on contextual information, which provides more accurate language processing results.
1. A computer-implemented method for training a machine learning model to process audio or textual language input, the method comprising:
accessing a machine-learning model, the machine-learning model including one or more blocks in which each block includes a multi-head self-attention network, a first connection for providing input to the multi-head self-attention network, and a second connection for providing the input to a normalization layer, bypassing the multi-head self-attention network;
accessing a training data set;
identifying a dropout parameter, wherein the dropout parameter is a dropout rate;
applying the dropout parameter to the second connection, wherein the second connection is a residual connection, and wherein applying the dropout parameter comprises dropping out the residual connection according to the dropout rate;
training the machine-learning model using the training data set to generate a trained machine-learning model, wherein the second connection is dropped out according to the dropout parameter;
inputting an utterance to the trained machine-learning model to identify an entity; and
based on the identified entity, preparing and providing a response to the utterance.
2. The method of claim 1 , wherein:
the machine learning model includes a plurality of attention heads, each attention head providing an output that is concatenated and multiplied by an attention weight matrix; and
the method further comprising dropping out diagonal entries in the attention weight matrix.
3. The method of claim 1 , wherein the dropout parameter is a hyperparameter of the machine-learning model, the method further comprising:
performing hypertuning to identify the dropout parameter.
4. The method of claim 1 , wherein, when the second connection is dropped out, the machine-learning model is caused to learn contextual information through multi-headed self-attention.
5. The method of claim 4 , wherein:
the training data set comprises one or more utterances, each utterance comprising one or more words that are labeled as an entity and one or more words that are not labeled as entities; and
the contextual information is determined based on the one or more words that are not labeled as entities.
6. The method of claim 1 , wherein providing the response to the utterance comprises providing audio output, via a speaker, responsive to the utterance.
7. The method of claim 1 , wherein providing the response to the utterance comprises providing text output, via a display, responsive to the utterance.
8. A system comprising:
one or more processors; and
a memory coupled to the one or more processors, the memory storing a plurality of instructions executable by the one or more processors, the plurality of instructions comprising instructions that when executed by the one or more processors cause the one or more processors to perform operations comprising:
accessing a machine-learning model, the machine-learning model including one or more blocks in which each block includes a multi-head self-attention network, a first connection for providing input to the multi-head self-attention network, and a second connection for providing the input to a normalization layer, bypassing the multi-head self-attention network;
accessing a training data set;
identifying a dropout parameter, wherein the dropout parameter is a dropout rate;
applying the dropout parameter to the second connection, wherein the second connection is a residual connection, and wherein applying the dropout parameter comprises dropping out the residual connection according to the dropout rate;
training the machine-learning model using the training data set to generate a trained machine-learning model, wherein the second connection is dropped out according to the dropout parameter;
inputting an utterance to the trained machine-learning model to identify an entity; and
based on the identified entity, preparing and providing a response to the utterance.
9. The system of claim 8 , wherein:
the machine learning model includes a plurality of attention heads, each attention head providing an output that is concatenated and multiplied by an attention weight matrix; and
the operations further comprise dropping out diagonal entries in the attention weight matrix.
10. The system of claim 8 , wherein:
the dropout parameter is a hyperparameter of the machine-learning model, and
the operations further comprise performing hypertuning to identify the dropout parameter.
11. The system of claim 8 , wherein:
when the second connection is dropped out, the machine-learning model is caused to learn contextual information through multi-headed self-attention.
12. The system of claim 11 , wherein:
the training data set comprises one or more utterances, each utterance comprising one or more words that are labeled as an entity and one or more words that are not labeled as entities; and
the contextual information is determined based on the one or more words that are not labeled as entities.
13. The system of claim 11 , wherein providing the response to the utterance comprises providing audio output, via a speaker, responsive to the utterance.
14. The system of claim 11 , wherein providing the response to the utterance comprises providing text output, via a display, responsive to the utterance.
15. A non-transitory computer-readable memory storing a plurality of instructions executable by one or more processors to cause the one or more processors to perform operations comprising:
accessing a machine-learning model, the machine-learning model including one or more blocks in which each block includes a multi-head self-attention network, a first connection for providing input to the multi-head self-attention network, and a second connection for providing the input to a normalization layer, bypassing the multi-head self-attention network;
accessing a training data set;
identifying a dropout parameter, wherein the dropout parameter is a dropout rate;
applying the dropout parameter to the second connection, wherein the second connection is a residual connection, and wherein applying the dropout parameter comprises dropping out the residual connection according to the dropout rate;
training the machine-learning model using the training data set to generate a trained machine-learning model, wherein the second connection is dropped out according to the dropout parameter;
inputting an utterance to the trained machine-learning model to identify an entity; and
based on the identified entity, preparing and providing a response to the utterance.
16. The non-transitory computer-readable memory of claim 15 , wherein:
the machine learning model includes a plurality of attention heads, each attention head providing an output that is concatenated and multiplied by an attention weight matrix; and
the operations further comprise dropping out diagonal entries in the attention weight matrix.
17. The non-transitory computer-readable memory of claim 15 , wherein:
the dropout parameter is a hyperparameter of the machine-learning model, and
the operations further comprise performing hypertuning to identify the dropout parameter.
18. The non-transitory computer-readable memory of claim 15 , wherein:
when the second connection is dropped out, the machine-learning model is caused to learn contextual information through multi-headed self-attention.
19. The non-transitory computer-readable memory of claim 15 , wherein providing the response to the utterance comprises providing audio output, via a speaker, responsive to the utterance.
20. The non-transitory computer-readable memory of claim 15 , wherein providing the response to the utterance comprises providing text output, via a display, responsive to the utterance.