Computer network activity summarizer and chat interface
Various disclosed embodiments are directed to deriving, via a language model, a summary of data by converting or encoding table data into one or more natural language sentences, which are then used as input to the language model for generating the summary. One or more embodiments are additionally or alternatively directed to deriving, via a language model, a response to a user question or command via a chat interface by providing the language model with the generated summary as input. In this way, for example, the language model can use the summary as a prompt or other target context for providing a response.
1 . A system comprising:
at least one computer processor; and
one or more computer storage media storing computer-useable instructions that, when used by the at least one computer processor, cause the at least one computer processor to perform operations comprising:
accessing, from the one or more computer storage media, user activity data in one or more database tables, the user activity data including statistics indicating a quantity of different types of computer user input, each database table, of the one or more database tables, having one or more columns and one or more rows, a column, of the one or more columns, having a column name indicated in natural language, a row, of the one or more rows, having one or more values that are a part of the statistics indicating the quantity of different types of computer user input;
generating, via computer natural language processing, a machine-readable tag, the machine-readable tag providing additional contextual text that supplements the column name to indicate an entity type of the column;
for the row, generating a natural language sentence by combining the column name and the machine-readable tag of the column with the one or more values that are the part of the statistics indicating the quantity of different types of computer user input; and
at least partially responsive to the generating of the natural language sentence, encoding the natural-language sentence into a numerical tensor representation and providing the numerical tensor representation as input to a transformer-based language model comprising an attention mechanism, wherein an output layer of the language model generates an embedding vector corresponding to a semantic summary associated with the row based on the attention mechanism.
2 . The system of claim 1 , wherein the operations further comprising:
receiving a user question or command via a chat interface;
in response to receiving of the user question or command, providing the language model or a second language model with the summary and the user question or command as input; and
based on the providing, receiving an output comprising a response to the user question or command.
3 . The system of claim 2 , wherein the providing the summary as input includes calling the second language model with a query in response to the receiving of the user question or command, and wherein the second model executes the query and generates the response based on retrieving information from the summary.
4 . The system of claim 2 , wherein the operations further comprising:
based on the one or more tables, generating, via a machine learning model, a prediction indicative of at least one of: an indication of an account or lead that should be sent a digital marketing advertisement, a particular campaign to use for the account or lead, and a particular channel and time to provide the digital marketing advertisement or the particular campaign, and wherein the response to the user question or command is further based on the prediction.
5 . The system of claim 1 , wherein the one or more tables includes a first table and a second table, the first table describes lead data of multiple leads, a lead being a member of an account, and wherein second table describes one or more accounts, an account being a particular organization.
6 . The system of claim 5 , wherein the operations further comprising stitching the first table and the second table together such that the first table and the second table are merged.
7 . The system of claim 1 , wherein the one or more tables includes at least one of: a role of each lead, user activity data of each lead, opportunity stage and type of each lead, an advertising campaign that was used for each lead, account name, account industry, revenue amount of each account, and number of employees of each account.
8 . The system of claim 1 , wherein the operations further comprising preprocessing the user activity data, and wherein the preprocessing includes anonymizing all personally identifiable information in the user activity data.
9 . The system of claim 1 , wherein the at least one summary includes a lead summary that summarizes user activity data of a first lead and an account summary that summarizes user activity data of the first lead and at least a second lead for an account.
10 . A computer-implemented method comprising:
accessing, from one or more computer storage media, data from one or more tables, the data includes at least one of lead data of a lead and account data of an account, each table, of the one or more tables having one or more columns and one or more rows, each column, of the one or more columns, having a column name, each row, of the one or more rows, having one or more values corresponding to statistics of computer user input of the lead or the account;
generating, for each column name and via computer natural language processing, a machine-readable tag that indicates a semantic meaning of a respective column name of the one or more tables;
generating one or more natural language sentences by combining at least the computer-readable tag with the one or more values corresponding to the statistics of computer user input of the lead or account for each row of the one or more tables; and
encoding the one or more natural language sentences into a numerical tensor representation and providing the numerical tensor representation as input into a language model comprising an attention mechanism, wherein an output layer of the language model generates an embedding vector representing at least one summary associated with each row based on the attention mechanism.
11 . The computer-implemented method of claim 10 , further comprising receiving, via a chat interface, a user question or command associated with the at least one summary, wherein the providing the numerical tensor representation as input includes calling the language model with a query in response to the receiving of the user question, and wherein the language model executes the query and generates an answer based on retrieving information from the at least one summary.
12 . The computer-implemented method of claim 11 , further comprising:
receiving, via a chat interface, a user question or command associated with the at least one summary;
generating, via a machine learning model, a prediction indicative of at least one of: an indication of the account or the lead that should be sent a digital marketing advertisement, a particular campaign to use for the account or the lead, and a particular channel and time to provide the digital marketing advertisement or the particular campaign, and wherein a response to the user question or command is further based on the prediction.
13 . The computer-implemented method of claim 10 , wherein the lead is a member of the account, and wherein the account is a particular organization.
14 . The computer-implemented method of claim 10 , wherein the one or more tables include a first table and a second table, and wherein the computer-implemented method further comprises stitching the first table and the second table together such that the first table and the second table are merged.
15 . The computer-implemented method of claim 10 , wherein the one or more tables includes at least one of: a role of each lead, user activity data of each lead, opportunity stage and type of each lead, an advertising campaign that was used for each lead, account name, account industry, revenue amount of each account, or a number of employees of each account.
16 . The computer-implemented method of claim 10 , wherein the operations further comprising preprocessing the lead data and the account data, and wherein the preprocessing includes anonymizing all personally identifiable information.
17 . The computer-implemented method of claim 10 , wherein the at least one summary summarizes user activity data of a first lead and at least a second lead for an account.
18 . A computer system comprising:
a user activity extractor means for accessing, from the one or more computer storage media, data from one or more tables, the data includes at least one of lead data of a lead and account data of an account, each table, of the one or more tables having one or more columns and one or more rows, each column, of the one or more columns, having a column name, each row, of the one or more rows, having one or more values corresponding to statistics of computer user input of the lead or the account;
a verbalization component means for generating, for each column name and via computer natural language processing, a machine-readable tag that indicates a semantic meaning of a respective column name of the one or more tables;
wherein the verbalization component means further for generating one or more natural language sentences by combining at least the computer-readable tag with the one or more values corresponding to the statistics of computer user input of the lead or account for each row of the one or more tables; and
a summarization component means for encoding the one or more natural language sentences into a numerical tensor representation and providing the numerical tensor representation as input into a language model comprising an attention mechanism, wherein an output layer of the language model generates an embedding vector representing at least one summary associated with each row based on the attention mechanism.
19 . The system of claim 18 , further comprising:
a chat interface component means for receiving a user question or command via a chat interface;
wherein the chat interface component means is further for providing the first language model or a second language model with the summary and the user question or command as input in response to receiving of the user question or command; and
wherein the chat interface component means is further for receiving an output comprising a response to the user question or command based on the providing.