Methods and apparatus for natural language understanding in conversational systems using machine learning processes
This application relates to apparatus and methods for natural language understanding in conversational systems using machine learning processes. In some examples, a computing device receives a request that identifies textual data. The computing device applies a natural language model to the textual data to generate first embeddings. In some examples, the natural language model is trained on retail data, such as item descriptions and chat session data. The computing device also applies a dependency based model to the textual data to generate second embeddings. Further, the computing device concatenates the first and second embeddings, and applies an intent and entity classifier to the concatenated embeddings to determine entities, and an intent, for the request. The computing device may generate a response to the request based on the determined intent and entities.
1. A system, comprising:
a database;
a processor communicatively coupled to the database; and non-transitory memory storing instructions that, when executed, cause the processor to:
receive input data comprising a plurality of characters;
generate word embeddings based on the plurality of characters;
train a Bidirectional Encoder Representation from Transformers (BERT) model based on catalog data identifying one or more attributes of a plurality of items and chat session data;
apply the BERT model to the word embeddings to generate first output embeddings and a pooled output;
train a dependency embedding generation model to generate dependency based embeddings, wherein the dependency embedding generation model comprises a syntactic dependency embedding model to generate syntactic dependency-based word embeddings, a transformer encoder comprising a single layer transformer block to generate encoded data from the syntactic dependency-based word embeddings, and a linear network to generate the dependency based embeddings from the encoded data;
generate concatenated embeddings by stepwise concatenation of the first output embeddings and the dependency based embeddings;
apply a first linear layer to the concatenated embeddings to generate second output embeddings;
apply a second linear layer to the concatenated embeddings and the pooled output to predict an intent class; and
store the second output embeddings and the intent class in the database.
2. The system of claim 1 , wherein the plurality of characters comprise a previous context and a current context.
3. The system of claim 2 , wherein the previous context is based on a previous chat session, and the current context is based on an inquiry.
4. The system of claim 1 , wherein the processor reads the instructions to apply a neural network to previously generated intent data to generate third output embeddings, where the linear layer is applied to the first output embeddings and the third output embeddings to generate the second output embeddings.
5. The system of claim 1 , wherein the processor reads the instructions to:
normalize the concatenated embeddings, and
generate the second output embeddings based on the normalized concatenated embeddings.
6. The system of claim 1 , wherein the processor reads the instructions to tokenize the input data into a plurality of tokens, wherein the BERT model is applied to the tokenized input data.
7. The system of claim 1 , wherein the second output embeddings characterize a named entity of the plurality of characters.
8. The system of claim 1 , wherein the plurality of characters identify a first item title, and the second output embeddings characterize a second item title that is shorter than the first item title.
9. The system of claim 1 , wherein the input data is received in a request from a second computing device, and wherein the computing device is further configured to generate a response to the request based on the output values.
10. A method, comprising:
receiving input data comprising a plurality of characters;
generating word embeddings based on the plurality of characters;
training a Bidirectional Encoder Representation from Transformers (BERT) model based on catalog data identifying one or more attributes of a plurality of items and chat session data;
applying the BERT model to the word embeddings to generate first output embeddings and a pooled output;
training a dependency embedding generation model to generate dependency based embeddings, wherein the dependency embedding generation model comprises a syntactic dependency embedding model to generate syntactic dependency-based word embeddings, a transformer encoder comprising a single layer transformer block to generate encoded data from the syntactic dependency-based word embeddings, and a linear network to generate the dependency based embeddings from the encoded data;
generating concatenated embeddings by stepwise concatenation of the first output embeddings and the dependency based embeddings;
applying a first linear layer to the concatenated embeddings to generate second output embeddings;
applying a second linear layer to the concatenated embeddings and the pooled output to predict an intent class; and
storing the second output embeddings and the intent class in a database.
11. The method of claim 10 , further comprising applying a neural network to previously generated intent data to generate third output embeddings, where the linear layer is applied to the first output embeddings and the third output embeddings to generate the second output embeddings.
12. The method of claim 11 , further comprising concatenating the concatenated embeddings and the third output embeddings to generate second concatenated embeddings, wherein the linear layer is applied to the second concatenated embeddings.
13. The method of claim 10 , further comprising tokenizing the input data into a plurality of tokens, wherein the BERT model is applied to the tokenized input data.
14. The method of claim 10 , wherein the second output embeddings characterize a named entity of the plurality of characters.
15. The method of claim 10 , wherein the plurality of characters identify a first item title, and the second output embeddings characterize a second item title that is shorter than the first item title.
16. A non-transitory computer readable medium having instructions stored thereon, wherein the instructions, when executed by at least one processor, cause a device to perform operations comprising:
receiving input data comprising a plurality of characters;
generating word embeddings based on the plurality of characters;
training a Bidirectional Encoder Representation from Transformers (BERT) model based on catalog data identifying one or more attributes of a plurality of items and chat session data;
applying the BERT model to the word embeddings to generate first output embeddings and a pooled output;
training a dependency embedding generation model to generate dependency based embeddings, wherein the dependency embedding generation model comprises a syntactic dependency embedding model to generate syntactic dependency-based word embeddings, a transformer encoder comprising a single layer transformer block to generate encoded data from the syntactic dependency-based word embeddings, and a linear network to generate the dependency based embeddings from the encoded data;
generating concatenated embeddings by stepwise concatenation of the first output embeddings and the dependency based embeddings;
applying a first linear layer to the concatenated embeddings to generate second output embeddings;
applying a second linear layer to the concatenated embeddings and the pooled output to predict an intent class; and
storing the second output embeddings and the intent class in a database.