Patent Grant
10721190
An embodiment of the present subject matter relates generally to generating recommended messages and, more specifically, to a sequence to sequence to classification model for generating recommended messages.
Digital messaging is quickly becoming the most common way to quickly communicate with friends and colleagues. For example, messaging applications are commonly used to send and receive messages between individual users and groups of users. While these messaging applications are useful, they still require users to take the time to type responses. Professionals are busy people and do not always have time to craft these responses. There are current systems that provide users with suggested responses; however, these systems have limitations. Accordingly, improvements are needed.
In the drawings, which are not necessarily drawn to scale, like numerals may describe similar components in different views. Like numerals having different letter suffixes may represent different instances of similar components. Some embodiments are illustrated by way of example, and not limitation, in the figures of the accompanying drawings in which:
In the following description, for purposes of explanation, various details are set forth in order to provide a thorough understanding of some example embodiments. It will be apparent, however, to one skilled in the art, that the present subject matter may be practiced without these specific details, or with slight alterations.
Reference in the specification to “one embodiment” or “an embodiment” means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the present subject matter. Thus, the appearances of the phrase “in one embodiment” or “in an embodiment” appearing in various places throughout the specification are not necessarily all referring to the same embodiment.
For purposes of explanation, specific configurations and details are set forth in order to provide a thorough understanding of the present subject matter. However, it will be apparent to one of ordinary skill in the art that embodiments of the subject matter described may be practiced without the specific details presented herein, or in various combinations, as described herein. Furthermore, well-known features may be omitted or simplified in order not to obscure the described embodiments. Various examples may be given throughout this description. These are merely descriptions of specific embodiments. The scope or meaning of the claims is not limited to the examples given.
Disclosed are systems, methods, and non-transitory computer-readable media for using a sequence to sequence to classification model to generate recommended response messages. Sequence to sequence models provide the benefit of being trained based on an entire sequence of data, rather than just snippets of the data. For example, a sequence to sequence model can be trained based on the full sequence of words in a text, rather than just a piece of text, such as individual words or set of characters from the text. Once trained, the sequence to sequence model includes an encoder that generates an embedding vector representing a given input text. The encoder generates the embedding vector based on the entirety of the input text (e.g., sequence of words in the input text), rather than just snippets of the text. The sequence to sequence model also includes a decoder that generates a likely response to the text based on the embedding vector generated by the encoder. The decoder can generate a response without having to select from a predefined set of candidate responses, and act as a weakly supervision to enforce the encoder to be optimized toward the embedding space for a proper reply.
While a sequence to sequence model provides the benefit of analyzing an entire text and not being limited to a set of responses to select from, it also has some shortcoming. Specifically, a sequence to sequence model takes significant time and resources to train and change, if needed. Further, the outputs provided may include spelling and/or grammatical errors.
In contrast, a text classification model is trained based on smaller portions of a text, such as a word or set of characters, rather than the entirety of a text. Once trained, the text classification model, like a sequence to sequence model, generates an embedding vector from an input text, however the generated embedding vector is a combination of embedding vectors or values mapped to smaller portions of the input text, such as individual words, rather than an entire sequence of the words in the text. The text classification model determines likely response to an input text using categorization labels that correspond to the candidate responses. That is, the text classification module assigns values to each categorization label that indicate the likelihood that the corresponding candidate response will be used to respond to the input text.
While a text classification model isn't trained based on the rich data that a sequence to sequence model is and is limited in the responses it generates, it does provide other benefits. For example, the responses generated by the text classification model are not as likely to include spelling or grammatical errors. Further, the model is much easier to train and modify if needed.
A sequence to sequence to classification model provides the benefits of both a sequence to sequence model and a text classification model. That is, the sequence to sequence to classification model uses an encoder from a sequence to sequence model to generate an embedding vector based on an entirety of an input text, rather than portions of the text. The generated embedding vector is used as an input in a text classification model that determines probability values for a set of predefined responses. Each probability values indicates a determined likelihood that a given response will be used in response to the input message. By selecting from a set of predetermined responses, both spelling and grammatical errors are greatly minimized. Another advantage of the sequence to sequence to classification model is that is can be easily modified because the text classification portion of the model can be modified, retrained, etc., separately from the sequence to sequence portion of the model.
As shown, multiple devices (i.e., client device 102, client device 104, and a messaging system 106) are connected to a communication network 108 and configured to communicate with each other through use of the communication network 108. The communication network 108 is any type of network, including a local area network (LAN), such as an intranet, a wide area network (WAN), such as the internet, or any combination thereof. Further, the communication network 108 may be a public network, a private network, or a combination thereof. The communication network 108 is implemented using any number of communications links associated with one or more service providers, including one or more wired communication links, one or more wireless communication links, or any combination thereof. Additionally, the communication network 108 is configured to support the transmission of data formatted using any number of protocols.
Multiple computing devices can be connected to the communication network 108. A computing device is any type of general computing device capable of network communication with other computing devices. For example, a computing device can be a personal computing device such as a desktop or workstation, a business server, or a portable computing device, such as a laptop, smart phone, or a tablet personal computer (PC). A computing device can include some or all of the features, components, and peripherals of the machine 900 shown in
To facilitate communication with other computing devices, a computing device includes a communication interface configured to receive a communication, such as a request, data, etc., from another computing device in network communication with the computing device and pass the communication along to an appropriate module running on the computing device. The communication interface also sends a communication to another computing device in network communication with the computing device.
In the system 100, users interact with the messaging system 106 to establish and participate in communication sessions with other users. For example, users use the client devices 102 and 104 that are connected to the communication network 108 by direct and/or indirect communication to communicate with and utilize the functionality of the messaging system 106. Although the shown system 100 includes only two client devices 102, 104, this is only for ease of explanation and is not meant to be limiting. One skilled in the art would appreciate that the system 100 can include any number of client devices 102, 104. Further, the messaging system 106 may concurrently accept connections from and interact with any number of client devices 102, 104. The messaging system 106 supports connections from a variety of different types of client devices 102, 104, such as desktop computers; mobile computers; mobile communications devices, e.g., mobile phones, smart phones, tablets; smart televisions; set-top boxes; and/or any other network enabled computing devices. Hence, the client devices 102 and 104 may be of varying type, capabilities, operating systems, etc.
A user interacts with the messaging system 106 via a client-side application installed on the client devices 102 and 104. In some embodiments, the client-side application includes a messaging system specific component. For example, the component may be a stand-alone application, one or more application plug-ins, and/or a browser extension. However, the users may also interact with the messaging system 106 via a third-party application, such as a web browser, that resides on the client devices 102 and 104 and is configured to communicate with the messaging system 106. In either case, the client-side application presents a user interface (UI) for the user to interact with the messaging system 106. For example, the user interacts with the messaging system 106 via a client-side application integrated with the file system or via a webpage displayed using a web browser application.
The messaging system 106 is one or more computing devices configured to facilitate and manage communication sessions between various client devices 102, 104. The messaging system 106 can be a standalone system or integrated into other systems or services, such as being integrated into an online service, such as a social networking service, new service, etc. In either case, the messaging system 106 facilitates a communication session between client devices 102 and 104, where a user participant using one client device 102 can send and receive messages with a user participant using the other client device 104.
The messaging system 106 enables a user to initiate a communication session by providing a messaging interface where the user can select other users to include in the communication session, draft messages to be transmitted to the selected other users as part of a communication session, and read messages received from the other users as part of the communication sessions. Messages transmitted by a user's client device 102 as part of a communication session are received by the messaging system 106, which forwards the message to the recipient user's client device 104. The messaging system 106 can also store the received messages along with metadata describing the messages, such as the time the messages were sent, the originating user of the message, the recipient of the message, etc.
The messaging system 106 includes a recommended response generation system 110 that enables the messaging system 106 to generate and provides users with recommended responses based on historical message data and user profile data. The historical message data includes messages that were transmitted between client devices 102, 104 as part of previous communication sessions, as well as an order in which the messages were sent. The recommended response generation system 110 uses the historical message data to generate a sequence to sequence to classification model, which is used to determine recommended responses based on a received message. For example, upon receiving a message from client device 102 that is intended for client device 104 as part of a communication session, the recommended response generation system 110 uses the received message as input in the sequence to sequence to classification model, which results in a set of candidate responses to the message. The recommended response generation system 110 then selects a set of recommended responses from the set of candidate responses and the messaging system 106 transmits the set of recommended responses along with the received message to client device 104 as part of the communication session.
The user of client device 104 can use the provided recommended responses to quickly and easily provide a response to the received message. For example, the user of client device 104 can provide a single input (e.g., tapping one of the recommended responses) to provide a response to the user of client device 102, rather than typing out a reply message letter by letter and then selecting to send the message. Upon receiving a user selection of one of the recommended responses from client device 104, the messaging system 106 transmits the recommended response to client device 102 as part of the communication session. Additionally, the messaging system 106 may generate and provide client device 102 with a set of recommended responses.
A sequence to sequence to classification model provides the benefits of both a sequence to sequence model and a text classification model. That is, the sequence to sequence to classification model uses an encoder from a sequence to sequence model to generate an embedding vector based on an entirety of an input text, rather than portions of the text. The generated embedding vector is used as an input in a text classification model that determines probability values for a set of predefined response. Each probability value indicates a determined likelihood that a given response will be used in response to the input message. By selecting from a set of predetermined responses, both spelling and grammatical errors are greatly minimized. Another advantage of the sequence to sequence to classification model is that is can be easily modified because the text classification portion of the model can be modified, retrained, etc., separately from the sequence to sequence portion of the model.
As shown, the messaging system 106 includes an interface module 202, a receiving module 204, a storing module 206, a recommended response generation system 110, an output module 208, and a data storage 210. The interface module 202 provides a messaging interface that enables users to initiate and participate in communication sessions with other users. For example, the messaging interface includes user interface elements (e.g., buttons, scrollbars, text fields, etc.) that enable a user to select users and draft messages to initiate and participate in a communication session. Further, the messaging interface presents the users with a listing of available contacts to include in a communication session. The messaging interface also presents the user with a listing of existing communication sessions, which a user can select from to read the previous messages transmitted as part of the communication session as well as to draft and send new messages as part of the communication session.
The receiving module 204 receives messages that are being transmitted as part of a communication session. The messages are received from the client device 102, 104 of an originating user and intended for one or more other client devices 102, 104 of recipient users in the communication session. For example, a user may use the client device 102 to generate and transmit a message to the client device 104 of another user as part of a communication session. The message is initially received by the receiving module 204 of the messaging system 106. The received messages may include metadata, such as a timestamp indicating the time at which the message was transmitted, identifiers identifying the source and/or destination client devices 102, 104, an identifier identifying the communication session, etc.
The storing module 206 stores message data consisting of the received messages along with associated metadata in the data storage 210. In some embodiments, the storing module 206 anonymizes the message data to protect the privacy of the users. For example, the storing module 206 removes names and other personal information from the message data. The storing module 206 may also store the message data for a limited period of time, after which the message data is deleted. In some embodiments, a user is allowed to opt in or opt out of having their message data stored by the storing module. Accordingly, users that do not want to have their message data stored can opt out, resulting in the storing module 206 not storing their message data
The recommended response generation system 110 generates recommended responses based on a received message and the historical message data stored in the data storage 210. As explained above, the historical message data stored in the data storage 210 includes messages that were transmitted between client devices 102, 104 as part of previous communication sessions, as well as metadata describing the messages, such as the order in which the messages were sent, the originating and destination client devices, etc. The recommended response generation system 110 uses the historical message data to generate a sequence to sequence to classification model, which is used to determine recommended responses based on a received message. For example, upon receiving a message from client device 102 that is intended for client device 104 as part of a communication session, the recommended response generation system 110 uses the received message as input in the sequence to sequence to classification model, which results in a set of candidate responses to the message. The recommended response generation system 110 then selects a set of recommended responses from the set of candidate response, which can be provided to a user. The functionality of the recommended response generation system 110 is discussed in greater detail below in relation to
The output module 208 transmits received messages and a set of recommended responses to a recipient user's client device (e.g., client device 104) as part of a communication session. The recipient user can use their client device (e.g., client device 104) to select a recommended response to quickly and easily respond to the received message. For example, the user of client device 104 can provide a single input (e.g., tapping one of the recommended responses) to provide a response to the user of client device 102, rather than typing out a reply message letter by letter and then selecting to send the message. As described above, the messaging system 106, upon receiving a user selection of one of the recommended responses from client device 104, transmits the recommended response to client device 102 as part of the communication session. Additionally, the recommendation generation module may generate and provide client device 102 with a set of recommended responses.
As shown, the recommended response generation system 110 includes a model generation module 302, a candidate grouping module 304, a candidate response determination module 306, a recommended response selection module 308, and a de-normalization module 310.
The model generation module 302 generates a sequence to sequence to classification model based on historical message data stored in the data storage 210. The sequence to sequence generation model includes aspects of both a sequence to sequence model and a text classification model. Specifically, the sequence to sequence to classification model includes an encoder that is generated as would be with a sequence to sequence model. That is, the encoder is trained based on the entirety of a text, including the sequence of words in the text, rather than being trained on smaller portions of the text, such as individual words. As a result, the generated encoder determines an embedding vector for an input text that represents the entirety of the input text, rather than an embedding vector that is a combination of embeddings mapped to smaller portions of the input text.
The sequence to sequence to classification model uses the embedding vectors generated by the encoder as an input into a text classification portion of the model. That is, the embedding vector may be either a portion or the whole of a feature vector that is used as input in the text classification portion of the model. The text classification portion of the model determines probability values for a set of categorization labels that correspond to potential responses to the input message. The probability value assigned to a categorization label indicates a likelihood that that the corresponding response will be used to respond to the input message.
The text classification portion of the sequence to sequence to classification model is trained based on the embedding vectors generated by the encoder, as well as any other desired data. For example, the classification model may also be trained on other data such as user features, environmental features, other features, etc.
Once generated, the sequence to sequence to classification model receives, as input, a message transmitted as part of a communication session, and returns a set of candidate responses that are likely to be used in response to the message. That is, the encoder portion of the model analyzes the text of the input message to generate an embedding vector representing the input text. The classification portion of the model then uses the embedding vector as at least a portion of an input feature vector used to determine a set of classification labels that correspond to responses that are most likely to be used to respond to the input message.
The model generation module 302 generates the sequence to sequence to classification model using historical message data stored in the data storage 210. Specifically, the model generation module 302 uses previous messages transmitted during previous communication sessions to train the sequence to sequence to classification model. In some embodiments, the model generation module 302 normalizes each of the previous messages by, for example, replacing personalized terms (e.g., names, addresses, etc.) with placeholders, removing or normalizing punctuation, etc. Normalizing the data in this manner allows the model generation module 302 to identify multiple instances of a unique message. For example, the message “Hi Bob” and “Hi Frank” would both be normalized to “Hi [First Name],” which the model generation module 302 will identify as being two instances of the same message.
After the messages have been normalized, the model generation module 302 also identifies a set of the normalized messages that occurred at least a threshold number of times. This provides the model generation module 302 with commonly used responses, which the model generation module 302 designates as a set of available candidate responses that the sequence to sequence to classification model chooses from to provide recommended responses to users. That is, the model generation module 302 designates the commonly used responses as classification labels for classifying input messages. The model generation module 302 stores the classification labels and their corresponding embeddings in the data storage 210.
The resulting sequence to sequence to classification model receives a message transmitted as part of a communication session as input and returns a set of candidate responses based on a provided message input. The set of candidate responses is a subset of the full set of responses that are assigned as classification labels.
The candidate grouping module 304 groups similar candidate responses into candidate groups. Multiple candidate responses may have similar meanings. For example, the terms ‘yes,’ ‘yeah,’ ‘yup,’ ‘sure,’ etc., all have essentially the same meaning. Providing a user with multiple recommended responses with similar meanings may be redundant and provide the user with limited response options. Accordingly, it is preferable to provide the user with recommended responses that have a variety of meanings. For example, providing the user with recommended responses ‘yes,’ ‘maybe,’ and ‘no’ provides the user with three options that each have a distinct meaning from the others, whereas the recommended responses ‘yes,’‘yup,’ and ‘sure’ provide the user with three options that all have a single meaning.
To this end, the candidate grouping module 304 groups each of the candidate response from the available candidate responses into a candidate group based on the meaning of the candidate response. As a result, each candidate response is included in a candidate group with a similar meaning.
The candidate response determination module 306 determines a set of candidate response based on a received message. The received message is a message received as part of a communication session. To accomplish this, the candidate response determination module 306 uses the received message as input in the sequence to sequence to classification model generated by the model generation module 302. As explained above, sequence to sequence to classification model returns a set of candidate responses that are commonly used in response to a received message. The set of candidate responses are a subset of the available candidate responses.
To generate the set of candidate response, the candidate response determination module 306 uses the received message as input into the sequence to sequence portion of the sequence to sequence to classification model. That is, the candidate response determination module 306 uses the received message as input in the sequence to sequence generated encoder. The encoder uses the input message to generate an embedding vector that represents the input message.
The candidate response determination module 306 uses the resulting embedding vector to generate a feature vector that is used as input into the text classification portion of the sequence to sequence to classification model. The feature vector may include other features in addition to the embedding vector, such as environmental features, user features, etc. The text classification portion of the sequence to sequence to classification model then returns the set of the candidate responses based on the feature vector. The text classification portion of the model may be any type of suitable text classification model, such as a logistic regression model.
In some embodiments, the set of candidate responses returned by the sequence to sequence to classification model are ranked based on how likely or commonly each candidate response is used in response to the input message. For example, the sequence to sequence to classification model may assign values indicating a percentage value for how likely the candidate response is to be used in response to the input message. The values may be assigned using a multinomial system, in which the sum of all the values assigned to the candidate responses is equal to one. Alternatively, the values may be assigned using a multiclass system, in which the sum of all the values assigned to the candidate responses does not have to equal one.
The recommended response selection module 308 selects a set of recommended responses from the set of candidate responses based on a set of candidate selection rules. The set of candidate selection rules dictate how the set of recommended responses should be selected from the set of candidate responses. For example, the set of candidate selection rules may dictate the number of candidate responses that should be selected as recommended responses, such as 3 candidate responses. As another example, the set of candidate selection rules may dictate that the top rated candidate selection rules be selected. As another example, the set of candidate selection rules may dictate a maximum limit on the number of candidate response that can be selected from a candidate group. For example, the candidate selection rules may dictate that only one candidate response can be selected from each candidate group. This type of candidate selection rule ensures that a user is provided with recommended response with a variety of meanings. In some embodiments, the candidate selection rules can dictate a minimum number of candidate groups from which a candidate response should be picked, such as at least 2 candidate groups. As a result, the recommended responses provided to the user will include at least 2 responses with different meanings. The set of candidate selection rules can be derived from user behavior and feedback. For example, user behavior and feedback may be analyzed to determined user preferences, such as preferring emoticons over text or a mixture of text and emoticons. Accordingly, a candidate selection rule can dictate selecting a set of recommended responses based on the learned user preference.
The de-normalization module 310 de-normalizes recommended responses prior to the recommended response being provided to the user. De-normalizing a recommended response includes replacing a placeholder included in the recommended response with personalized data. For example, a placeholder such as [first name] can be replaced with the recipient user's first name. As another example a placeholder such as [date] can be replaced with the current date. As a result, the recommended messages provided to the user will be personalized based on the communication session and current context (e.g., date, time, etc.).
As shown, the model generation module 302 includes a message data gathering module 402, a normalization module 404, an available candidate determination module 406, and a model training module 408. The message data gathering module 402 gathers historical message data from the data storage 210. The historical message data includes messages that were transmitted as part of previous communication sessions, as well as an order in which the messages were transmitted.
The normalization module 404 normalizes the messages from the historical message data. Normalizing the messages may include multiple steps. For example, normalizing the message data includes replacing personalized data with placeholders. Personalized data is data that is unique to the context of the communication session. Examples of personalized data are names, dates, email addresses, addresses, phone numbers, URLs, etc. The normalization module 404 identifies these types of personalized terms on the messages and replaces the terms with placeholders indicating the type of term that was replaced. For example, a user's first name is replaced with a placeholder such as [first name]. Likewise, a date is replaced with a placeholder such as [current date].
In addition to replacing personalized terms with placeholders, the normalization module 404 further normalizes the messages to remove unnecessary characters, spacing, etc. For example, the normalization module 404 removes additional white spaces between words, additional or repeat punctuation, etc.
The available candidate determination module 406 determines the set of available candidate responses that are commonly used by users in communication sessions. To accomplish this, the available candidate determination module 406 determines a number of occurrences of each unique normalized message generated by the normalization module 404. The number of occurrences indicates a number of times that the normalized message was used in communication sessions. The available candidate determination module 406 then compares the number of occurrences that each unique normalized message occurred to a predetermined threshold number of occurrences, such as 1000 occurrences. The available candidate determination module 406 adds each normalized message that occurred at least the threshold number of times to the set of available candidate responses. As a result, the set of available candidate responses only includes responses that were used often or commonly during communication sessions, and messages that were not used commonly are not included in the set of available candidate responses. The set of available candidate responses is designated as classification labels for the text classification portion of the sequence to sequence to classification model.
The model training module 408 generations the sequence to sequence to classification model based on the historical message data gathered by the message data gathering module 402. The sequence to sequence generation model includes aspects of both a sequence to sequence model and a text classification model. Specifically, the sequence to sequence to classification model includes an encoder that is generated as would be with a sequence to sequence model. That is, the model training module 408 trains the encoder based on analyzing an entirety of a text, including the sequence of words in the text, rather than being trained on smaller portions of the text, such as individual words. As a result, the generated encoder determines an embedding vector based on an input text that represents the entirety of the input text. This contrasts with an encoder used in a standard text classification model that generates an embedding vector based on smaller portions of the input text.
The sequence to sequence to classification model uses the embedding vectors generated by the encoder as an input into a text classification portion of the model. That is, the embedding vector may be a portion or the whole of a feature vector that is used as input in the text classification portion of the model. The text classification portion of the model determines probability values for a set of categorization labels that correspond to potential responses to the input message. The probability value assigned to a categorization label indicates a likelihood that that the corresponding response will be used to respond to the input message.
The model training module 408 trains the text classification portion of the sequence to sequence to classification model based on the embedding vectors generated by the encoder, as well as any other desired data. For example, the classification model may also be trained on other data such as user features, environmental features, other features, etc.
Once generated, the sequence to sequence to classification model receives as input, a message transmitted as part of a communication session, and returns a set of candidate responses that are likely to be used in response to the message. That is, the encoder portion of the model analyzes the text of the input message to generate an embedding vector representing the input text. The classification portion of the model then uses the embedding vector as at least a portion of an input feature vector used to determine a set of classification labels that correspond to responses that are most likely to be used to respond to the input message.
At operation 502, the receiving module 204 receives a message as part of a communication session. The messages are received from the client device 102, 104 of an originating user and intended for one or more other client devices 102, 104 of recipient users in the communication session. For example, a user may use the client device 102 to generate and transmit a message to the client device 104 of another user as part of a communication session. The message is initially received by the receiving module 204 of the messaging system 106. The received messages may include metadata, such as a timestamp indicating the time at which the message was transmitted, identifiers identifying the source and/or destination client devices, an identifier identifying the communication session, etc.
At operation 504, the candidate response determination module 306 determines, using the message as input in a sequence to sequence to classification model, a set of candidate responses for replying to the message. To accomplish this, the candidate response determination module 306 uses the received message as input in the sequence to sequence to classification model generated by the model generation module 302. As explained above, a sequence to sequence to classification model returns a set of candidate responses that are commonly used in response to a received message. The set of candidate responses are a subset of the available candidate responses.
To generate the set of candidate response, the candidate response determination module 306 uses the received message as input into the sequence to sequence portion of the sequence to sequence to classification model. That is, the candidate response determination module 306 uses the received message as input in the sequence to sequence generated encoder. The encoder uses the input message to generate an embedding vector that represents the input message.
The candidate response determination module 306 uses the resulting embedding vector to generate a feature vector that is used as input into the text classification portion of the sequence to sequence to classification model. The feature vector may include other features in addition to the embedding vector, such as environmental features, user features, etc. The text classification portion of the sequence to sequence to classification model then returns the set of the candidate responses based on the feature vector. The text classification portion of the model may be any type of suitable text classification model, such as a logistic regression model.
At operation 506, the recommended response selection module 308 determines a set of recommended responses from the set of candidate responses. In some embodiments, the recommended response selection module 308 determines the set of recommended responses based on a set of candidate selection rules that dictate how the set of recommended responses should be selected from the set of candidate responses. For example, the set of candidate selection rules may dictate the number of candidate responses that should be selected as recommended responses, such as 3 candidate responses. As another example, the set of candidate selection rules may dictate that the top rated candidate selection rules be selected. As another example, the set of candidate selection rules may dictate a maximum limit on the number of candidate responses that can be selected from a candidate group. For example, the candidate selection rules may dictate that only one candidate response can be selected from each candidate group. This type of candidate selection rule ensures that a user is provided with recommended responses with a variety of meanings. In some embodiments, the candidate selection rules can dictate a minimum number of candidate groups from which a candidate response should be picked, such as at least 2 candidate groups. As a result, the recommended responses provided to the user will include at least 2 responses with different meanings.
At operation 508, the de-normalization module 310 de-normalizes the set of recommended responses. De-normalizing a recommended response includes replacing a placeholder included in the recommended response with personalized data. For example, a placeholder such as [first name] can be replaced with the recipient user's first name. As another example a placeholder such as [date] can be replaced with the current date. As a result, the recommended messages provided to the user will be personalized based on the communication session and current context (e.g., date, time, etc.).
At operation 510, the output module 208 transmits the message and the set of recommended responses to a destination client device. The recipient user can use their client device (e.g., client device 104) to select a recommend response to quickly and easily respond to the received message. For example, the user of client device 104 can provide a single input (e.g., tapping one of the recommended responses) to provide a response to the user of client device 102, rather than typing out a reply message letter by letter and then selecting to send the message.
At operation 602, the recommended response generation system 110 access a text transmitted as part of a communication session. The text may be any type of text consisting of sets of characters. For example, the text may be a message transmitted from one user to another user as part of a communication session.
At operation 604, the recommended response generation system 110 generates an embedding vector using a sequence to sequence encoder. A sequence to sequence encoder is an encoder that was trained based on analyzing a totality of training texts, as opposed to an encoder trained on portions of training texts. As a result, the embedding vector generated by the encoder represents the entirety of the input text, rather than being based on a combination of portions of the input text.
At operation 606, the recommended response generation system 110 generates a feature vector based on the embedding vector. The feature vector may include a portion or all of the embedding vector generated by the sequence to sequence encoder, as well as any additional feature data. For example, the feature vector may include other feature data such as user feature data, environmental feature data, etc.
At operation 608, the recommended response generation system 110 determines a set of candidate responses by using the feature vector as input in a text classification system portion of a sequence to sequence to classification model. The text classification portion of the model is trained based on embedding vectors generated by the sequence to sequence encoder, as well as other feature data. The text classification portion of the model may be any type of text classification model, such as a logistic regression model. The text classification portion of the model assigns probability values to classification labels that each correspond to a potential known response. The assigned probability values assigned to each categorization label indicate a determined likelihood that the corresponding response will be used in response to the input message. The recommended response generation system 110 determines as subset of the categorization labels that are assigned the highest probability values as the set of candidate responses to the message.
Software Architecture
In the example architecture of
The operating system 802 may manage hardware resources and provide common services. The operating system 802 may include, for example, a kernel 822, services 824, and drivers 826. The kernel 822 may act as an abstraction layer between the hardware and the other software layers. For example, the kernel 822 may be responsible for memory management, processor management (e.g., scheduling), component management, networking, security settings, and so on. The services 824 may provide other common services for the other software layers. The drivers 826 are responsible for controlling or interfacing with the underlying hardware. For instance, the drivers 826 include display drivers, camera drivers, Bluetooth® drivers, flash memory drivers, serial communication drivers (e.g., Universal Serial Bus (USB) drivers), Wi-Fi® drivers, audio drivers, power management drivers, and so forth, depending on the hardware configuration.
The libraries 820 provide a common infrastructure that is used by the applications 816 and/or other components and/or layers. The libraries 820 provide functionality that allows other software components to perform tasks in an easier fashion than to interface directly with the underlying operating system 802 functionality (e.g., kernel 822, services 824 and/or drivers 826). The libraries 820 may include system libraries 844 (e.g., C standard library) that may provide functions such as memory allocation functions, string manipulation functions, mathematical functions, and the like. In addition, the libraries 820 may include API libraries 846 such as media libraries (e.g., libraries to support presentation and manipulation of various media format such as MPEG4, H.264, MP3, AAC, AMR, JPG, PNG), graphics libraries (e.g., an OpenGL framework that may be used to render 2D and 3D in a graphic content on a display), database libraries (e.g., SQLite that may provide various relational database functions), web libraries (e.g., WebKit that may provide web browsing functionality), and the like. The libraries 820 may also include a wide variety of other libraries 848 to provide many other APIs to the applications 816 and other software components/modules.
The frameworks/middleware 818 (also sometimes referred to as middleware) provide a higher-level common infrastructure that may be used by the applications 816 and/or other software components/modules. For example, the frameworks/middleware 818 may provide various graphic user interface (GUI) functions, high-level resource management, high-level location services, and so forth. The frameworks/middleware 818 may provide a broad spectrum of other APIs that may be used by the applications 816 and/or other software components/modules, some of which may be specific to a particular operating system 802 or platform.
The applications 816 include built-in applications 838 and/or third-party applications 840. Examples of representative built-in applications 838 may include, but are not limited to, a contacts application, a browser application, a book reader application, a location application, a media application, a messaging application, and/or a game application. Third-party applications 840 may include an application developed using the ANDROID™ or IOS™ software development kit (SDK) by an entity other than the vendor of the particular platform, and may be mobile software running on a mobile operating system such as IOS™, ANDROID™, WINDOWS® Phone, or other mobile operating systems. The third-party applications 840 may invoke the API calls 808 provided by the mobile operating system (such as operating system 802) to facilitate functionality described herein.
The applications 816 may use built in operating system functions (e.g., kernel 822, services 824 and/or drivers 826), libraries 820, and frameworks/middleware 818 to create user interfaces to interact with users of the system. Alternatively, or additionally, in some systems, interactions with a user may occur through a presentation layer, such as presentation layer 814. In these systems, the application/component “logic” can be separated from the aspects of the application/component that interact with a user.
The machine 900 may include processors 904, memory/storage 906, and I/O components 918, which may be configured to communicate with each other such as via a bus 902. The memory/storage 906 may include a memory 914, such as a main memory, or other memory storage, and a storage unit 916, both accessible to the processors 904 such as via the bus 902. The storage unit 916 and memory 914 store the instructions 910 embodying any one or more of the methodologies or functions described herein. The instructions 910 may also reside, completely or partially, within the memory 914, within the storage unit 916, within at least one of the processors 904 (e.g., within the processor's cache memory), or any suitable combination thereof, during execution thereof by the machine 900. Accordingly, the memory 914, the storage unit 916, and the memory of processors 904 are examples of machine-readable media.
The I/O components 918 may include a wide variety of components to receive input, provide output, produce output, transmit information, exchange information, capture measurements, and so on. The specific I/O components 918 that are included in a particular machine 900 will depend on the type of machine. For example, portable machines such as mobile phones will likely include a touch input device or other such input mechanisms, while a headless server machine will likely not include such a touch input device. It will be appreciated that the I/O components 918 may include many other components that are not shown in
In further example embodiments, the I/O components 918 may include biometric components 930, motion components 934, environmental components 936, or position components 938 among a wide array of other components. For example, the biometric components 930 may include components to detect expressions (e.g., hand expressions, facial expressions, vocal expressions, body gestures, or eye tracking), measure biosignals (e.g., blood pressure, heart rate, body temperature, perspiration, or brain waves), identify a person (e.g., voice identification, retinal identification, facial identification, fingerprint identification, or electroencephalogram based identification), and the like. The motion components 934 may include acceleration sensor components (e.g., accelerometer), gravitation sensor components, rotation sensor components (e.g., gyroscope), and so forth. The environmental components 936 may include, for example, illumination sensor components (e.g., photometer), temperature sensor components (e.g., one or more thermometer that detect ambient temperature), humidity sensor components, pressure sensor components (e.g., barometer), acoustic sensor components (e.g., one or more microphones that detect background noise), proximity sensor components (e.g., infrared sensors that detect nearby objects), gas sensors (e.g., gas detection sensors to detect concentrations of hazardous gases for safety or to measure pollutants in the atmosphere), or other components that may provide indications, measurements, or signals corresponding to a surrounding physical environment. The position components 938 may include location sensor components (e.g., a GPS receiver component), altitude sensor components (e.g., altimeters or barometers that detect air pressure from which altitude may be derived), orientation sensor components (e.g., magnetometers), and the like.
Communication may be implemented using a wide variety of technologies. The I/O components 918 may include communication components 940 operable to couple the machine 900 to a network 932 or devices 920 via coupling 924 and coupling 922, respectively. For example, the communication components 940 may include a network interface component or other suitable device to interface with the network 932. In further examples, communication components 940 may include wired communication components, wireless communication components, cellular communication components, near field communication (NFC) components, Bluetooth® components (e.g., Bluetooth® Low Energy), Wi-Fi® components, and other communication components to provide communication via other modalities. The devices 920 may be another machine or any of a wide variety of peripheral devices (e.g., a peripheral device coupled via a USB).
Moreover, the communication components 940 may detect identifiers or include components operable to detect identifiers. For example, the communication components 940 may include radio frequency identification (RFID) tag reader components, NFC smart tag detection components, optical reader components (e.g., an optical sensor to detect one-dimensional bar codes such as Universal Product Code (UPC) bar code, multi-dimensional bar codes such as Quick Response (QR) code, Aztec code, Data Matrix, Dataglyph, MaxiCode, PDF417, Ultra Code, UCC RSS-2D bar code, and other optical codes), or acoustic detection components (e.g., microphones to identify tagged audio signals). In addition, a variety of information may be derived via the communication components 940, such as, location via Internet Protocol (IP) geo-location, location via Wi-Fi® signal triangulation, location via detecting a NFC beacon signal that may indicate a particular location, and so forth.
Glossary
“CARRIER SIGNAL” in this context refers to any intangible medium that is capable of storing, encoding, or carrying instructions 910 for execution by the machine 900, and includes digital or analog communications signals or other intangible medium to facilitate communication of such instructions 910. Instructions 910 may be transmitted or received over the network 932 using a transmission medium via a network interface device and using any one of a number of well-known transfer protocols.
“CLIENT DEVICE” in this context refers to any machine 900 that interfaces to a communications network 932 to obtain resources from one or more server systems or other client devices. A client device 102, 104 may be, but is not limited to, a mobile phone, desktop computer, laptop, PDAs, smart phones, tablets, ultra books, netbooks, laptops, multi-processor systems, microprocessor-based or programmable consumer electronics, game consoles, STBs, or any other communication device that a user may use to access a network 932.
“COMMUNICATIONS NETWORK” in this context refers to one or more portions of a network 932 that may be an ad hoc network, an intranet, an extranet, a virtual private network (VPN), a local area network (LAN), a wireless LAN (WLAN), a wide area network (WAN), a wireless WAN (WWAN), a metropolitan area network (MAN), the Internet, a portion of the Internet, a portion of the Public Switched Telephone Network (PSTN), a plain old telephone service (POTS) network, a cellular telephone network, a wireless network, a Wi-Fi® network, another type of network, or a combination of two or more such networks. For example, a network 932 or a portion of a network 932 may include a wireless or cellular network and the coupling may be a Code Division Multiple Access (CDMA) connection, a Global System for Mobile communications (GSM) connection, or other type of cellular or wireless coupling. In this example, the coupling may implement any of a variety of types of data transfer technology, such as Single Carrier Radio Transmission Technology (1×RTT), Evolution-Data Optimized (EVDO) technology, General Packet Radio Service (GPRS) technology, Enhanced Data rates for GSM Evolution (EDGE) technology, third Generation Partnership Project (3GPP) including 3G, fourth generation wireless (4G) networks, Universal Mobile Telecommunications System (UMTS), High Speed Packet Access (HSPA), Worldwide Interoperability for Microwave Access (WiMAX), Long Term Evolution (LTE) standard, others defined by various standard setting organizations, other long range protocols, or other data transfer technology.
“MACHINE-READABLE MEDIUM” in this context refers to a component, device or other tangible media able to store instructions 910 and data temporarily or permanently and may include, but is not be limited to, random-access memory (RAM), read-only memory (ROM), buffer memory, flash memory, optical media, magnetic media, cache memory, other types of storage (e.g., erasable programmable read-only memory (EEPROM)), and/or any suitable combination thereof. The term “machine-readable medium” should be taken to include a single medium or multiple media (e.g., a centralized or distributed database, or associated caches and servers) able to store instructions 910. The term “machine-readable medium” shall also be taken to include any medium, or combination of multiple media, that is capable of storing instructions 910 (e.g., code) for execution by a machine 900, such that the instructions 910, when executed by one or more processors 904 of the machine 900, cause the machine 900 to perform any one or more of the methodologies described herein. Accordingly, a “machine-readable medium” refers to a single storage apparatus or device, as well as “cloud-based” storage systems or storage networks that include multiple storage apparatus or devices. The term “machine-readable medium” excludes signals per se.
“COMPONENT” in this context refers to a device, physical entity, or logic having boundaries defined by function or subroutine calls, branch points, APIs, or other technologies that provide for the partitioning or modularization of particular processing or control functions. Components may be combined via their interfaces with other components to carry out a machine process. A component may be a packaged functional hardware unit designed for use with other components and a part of a program that usually performs a particular function of related functions. Components may constitute either software components (e.g., code embodied on a machine-readable medium) or hardware components. A “hardware component” is a tangible unit capable of performing certain operations and may be configured or arranged in a certain physical manner. In various example embodiments, one or more computer systems (e.g., a standalone computer system, a client computer system, or a server computer system) or one or more hardware components of a computer system (e.g., a processor or a group of processors 904) may be configured by software (e.g., an application 816 or application portion) as a hardware component that operates to perform certain operations as described herein. A hardware component may also be implemented mechanically, electronically, or any suitable combination thereof. For example, a hardware component may include dedicated circuitry or logic that is permanently configured to perform certain operations. A hardware component may be a special-purpose processor, such as a field-programmable gate array (FPGA) or an application specific integrated circuit (ASIC). A hardware component may also include programmable logic or circuitry that is temporarily configured by software to perform certain operations. For example, a hardware component may include software executed by a general-purpose processor 904 or other programmable processor 904. Once configured by such software, hardware components become specific machines 900 (or specific components of a machine 900) uniquely tailored to perform the configured functions and are no longer general-purpose processors 904. It will be appreciated that the decision to implement a hardware component mechanically, in dedicated and permanently configured circuitry, or in temporarily configured circuitry (e.g., configured by software), may be driven by cost and time considerations. Accordingly, the phrase “hardware component” (or “hardware-implemented component”) should be understood to encompass a tangible entity, be that an entity that is physically constructed, permanently configured (e.g., hardwired), or temporarily configured (e.g., programmed) to operate in a certain manner or to perform certain operations described herein. Considering embodiments in which hardware components are temporarily configured (e.g., programmed), each of the hardware components need not be configured or instantiated at any one instance in time. For example, where a hardware component comprises a general-purpose processor 904 configured by software to become a special-purpose processor, the general-purpose processor 904 may be configured as respectively different special-purpose processors (e.g., comprising different hardware components) at different times. Software accordingly configures a particular processor or processors 904, for example, to constitute a particular hardware component at one instance of time and to constitute a different hardware component at a different instance of time. Hardware components can provide information to, and receive information from, other hardware components. Accordingly, the described hardware components may be regarded as being communicatively coupled. Where multiple hardware components exist contemporaneously, communications may be achieved through signal transmission (e.g., over appropriate circuits and buses 902) between or among two or more of the hardware components. In embodiments in which multiple hardware components are configured or instantiated at different times, communications between such hardware components may be achieved, for example, through the storage and retrieval of information in memory structures to which the multiple hardware components have access. For example, one hardware component may perform an operation and store the output of that operation in a memory device to which it is communicatively coupled. A further hardware component may then, at a later time, access the memory device to retrieve and process the stored output. Hardware components may also initiate communications with input or output devices, and can operate on a resource (e.g., a collection of information). The various operations of example methods described herein may be performed, at least partially, by one or more processors 904 that are temporarily configured (e.g., by software) or permanently configured to perform the relevant operations. Whether temporarily or permanently configured, such processors 904 may constitute processor-implemented components that operate to perform one or more operations or functions described herein. As used herein, “processor-implemented component” refers to a hardware component implemented using one or more processors 904. Similarly, the methods described herein may be at least partially processor-implemented, with a particular processor or processors 904 being an example of hardware. For example, at least some of the operations of a method may be performed by one or more processors 904 or processor-implemented components. Moreover, the one or more processors 904 may also operate to support performance of the relevant operations in a “cloud computing” environment or as a “software as a service” (SaaS). For example, at least some of the operations may be performed by a group of computers (as examples of machines 900 including processors 904), with these operations being accessible via a network 932 (e.g., the Internet) and via one or more appropriate interfaces (e.g., an API). The performance of certain of the operations may be distributed among the processors 904, not only residing within a single machine 900, but deployed across a number of machines 900. In some example embodiments, the processors 904 or processor-implemented components may be located in a single geographic location (e.g., within a home environment, an office environment, or a server farm). In other example embodiments, the processors 904 or processor-implemented components may be distributed across a number of geographic locations.
“PROCESSOR” in this context refers to any circuit or virtual circuit (a physical circuit emulated by logic executing on an actual processor) that manipulates data values according to control signals (e.g., “commands,” “op codes,” “machine code,” etc.) and which produces corresponding output signals that are applied to operate a machine 900. A processor 904 may be, for example, a central processing unit (CPU), a reduced instruction set computing (RISC) processor, a complex instruction set computing (CISC) processor, a graphics processing unit (GPU), a digital signal processor (DSP), an ASIC, a radio-frequency integrated circuit (RFIC) or any combination thereof. A processor may further be a multi-core processor having two or more independent processors 904 (sometimes referred to as “cores”) that may execute instructions 910 contemporaneously.
| Number | Name | Date | Kind |
|---|---|---|---|
| 7783486 | Rosser | Aug 2010 | B2 |
| 8204751 | Di Fabbrizio | Jun 2012 | B1 |
| 9576050 | Walters | Feb 2017 | B1 |
| 9715496 | Sapoznik | Jul 2017 | B1 |
| 9824692 | Khoury | Nov 2017 | B1 |
| 9852177 | Cheung | Dec 2017 | B1 |
| 9860200 | Braun | Jan 2018 | B1 |
| 9899020 | Huang | Feb 2018 | B2 |
| 10091140 | Galley | Oct 2018 | B2 |
| 10109275 | Henry | Oct 2018 | B2 |
| 10169315 | Heckel | Jan 2019 | B1 |
| 10210244 | Branavan | Feb 2019 | B1 |
| 10268671 | Kaiser | Apr 2019 | B2 |
| 10339919 | Raux | Jul 2019 | B1 |
| 10354182 | Chang | Jul 2019 | B2 |
| 10366341 | Weber | Jul 2019 | B2 |
| 10387571 | Baldwin | Aug 2019 | B2 |
| 10417266 | Patel | Sep 2019 | B2 |
| 10475438 | Chicote | Nov 2019 | B1 |
| 20070070979 | Kim | Mar 2007 | A1 |
| 20130226562 | Arnon | Aug 2013 | A1 |
| 20130325992 | McGann et al. | Dec 2013 | A1 |
| 20150213372 | Shah | Jul 2015 | A1 |
| 20160282956 | Ouyang | Sep 2016 | A1 |
| 20160337301 | Rollins | Nov 2016 | A1 |
| 20170032035 | Gao | Feb 2017 | A1 |
| 20170092264 | Hakkani-Tur | Mar 2017 | A1 |
| 20170098153 | Mao | Apr 2017 | A1 |
| 20170118336 | Tapuhi | Apr 2017 | A1 |
| 20170124432 | Chen et al. | May 2017 | A1 |
| 20170132019 | Karashchuk | May 2017 | A1 |
| 20170180294 | Milligan | Jun 2017 | A1 |
| 20170195269 | Miklos | Jul 2017 | A1 |
| 20170201471 | Miklos | Jul 2017 | A1 |
| 20180018971 | Park | Jan 2018 | A1 |
| 20180024991 | Baldwin | Jan 2018 | A1 |
| 20180052829 | Lee | Feb 2018 | A1 |
| 20180083901 | McGregor, Jr. | Mar 2018 | A1 |
| 20180144234 | Devesa | May 2018 | A1 |
| 20180144248 | Lu | May 2018 | A1 |
| 20180157664 | Howell | Jun 2018 | A1 |
| 20180157992 | Susskind | Jun 2018 | A1 |
| 20180189271 | Noh | Jul 2018 | A1 |
| 20180189559 | Lee | Jul 2018 | A1 |
| 20180210874 | Fuxman | Jul 2018 | A1 |
| 20180218382 | Ye | Aug 2018 | A1 |
| 20180239770 | Ghotbi et al. | Aug 2018 | A1 |
| 20180268317 | Dharwadker | Sep 2018 | A1 |
| 20180285459 | Soni | Oct 2018 | A1 |
| 20180329982 | Patel | Nov 2018 | A1 |
| 20180336193 | Liu | Nov 2018 | A1 |
| 20180341839 | Malak | Nov 2018 | A1 |
| 20180367480 | Housman | Dec 2018 | A1 |
| 20180367484 | Rodriguez | Dec 2018 | A1 |
| 20190012374 | Petroni | Jan 2019 | A1 |
| 20190042748 | Shabtai | Feb 2019 | A1 |
| 20190073351 | Zhang | Mar 2019 | A1 |
| 20190073580 | Dzhulgakov | Mar 2019 | A1 |
| 20190114362 | Subbian | Apr 2019 | A1 |
| 20190124019 | Leon | Apr 2019 | A1 |
| 20190129947 | Shin | May 2019 | A1 |
| 20190130249 | Bradbury | May 2019 | A1 |
| 20190220513 | Xin | Jul 2019 | A1 |
| 20190251165 | Bachrach | Aug 2019 | A1 |
| 20190258854 | Hosabettu | Aug 2019 | A1 |
| 20190266482 | Erez | Aug 2019 | A1 |
| 20190286943 | Leskovec | Sep 2019 | A1 |
| 20190325293 | Wang | Oct 2019 | A1 |
| 20190333020 | Zhao | Oct 2019 | A1 |
| 20190333400 | Saini | Oct 2019 | A1 |
| 20190341021 | Shang | Nov 2019 | A1 |
| Entry |
|---|
| Brownlee—What Are Word Embeddings for Text—Machine Learning Mastery—Oct. 2017. |
| Chollet—A ten minute introduction to sequence to sequence learning in Keras—the Keras Blog—Sep. 2017. |
| Dw et al.—Apply word embeddings to entire document to get a feature vector—Cross Validated—Oct. 2016. |
| Zia et al.—What is embedding in deep neural architectures—Quora—Dec. 2016. |
| lenz et al.—Why are word embedding actually vectors—Stack Overflow—Oct. 2017. |
| “Non Final Office Action Issued in U.S. Patent Application No. 16/353,397”, dated Mar. 4, 2020, 13 Pages. |
| Kannan, et al., “Smart Reply: Automated Response Suggestion for Email”, In Proceedings of the 22nd ACM SIGKDDR International Conference on Knowledge Discovery and Data Mining, Aug. 13, 2016, pp. 955-964. |
| Number | Date | Country | |
|---|---|---|---|
| 20200044990 A1 | Feb 2020 | US |
