Dynamic intent classification based on environment variables

Abstract

To prevent intent classifiers from potentially choosing intents that are ineligible for the current input due to policies, dynamic intent classification systems and methods are provided that dynamically control the possible set of intents using environment variables (also referred to as external variables). Associations between environment variables and ineligible intents, referred to as culling rules, are used.

Description

BACKGROUND

Intelligent Virtual Assistants (IVAs) and chatterbots can be implemented in a multitude of ways, but common to all is a component for natural language understanding (NLU), which is used for the translation of user inputs into a semantic representation. Regardless of the means in which the user is interacting with the IVA (keyboard, speech, gestures, etc.), the user input is first converted into text or a numerical feature representation and fed into the NLU component for analysis. The NLU component maps user inputs, or conversational turns, to a derived semantic representation commonly known as the user intention or simply intent. In the context of natural language processing, intent is defined as an interpretation of a statement or question that allows one to formulate the ‘best’ response to the statement.

The collection of syntax, semantics, and grammar rules that defines how input language maps to an intent within the NLU component is referred to as a language model. The NLU component may perform pre-processing steps such as part-of-speech tagging, stemming, and entity identification (such as proper names or dates) before feeding the text into the language model. Each of these parsers may be trained through machine learning methods or manually constructed by human experts.

There are two primary techniques to construct the language model. The first technique uses rules-based (e.g., human-curated) methods such as regular expressions or grammars to define the structure and content of language associated to a particular intent. By applying these patterns or grammars to the input text, the correct intent can be determined. The intent associated with the matching grammar or pattern is selected in cases where there is no ambiguity. If multiple patterns match, the correct intent may be selected by a scoring mechanism applied to all patterns that match or other means such as ontological methods.

The second technique to construct a language model is through statistical means. Text that is similar to the expected input text is labeled (e.g., by humans) to be associated with particular intentions. This labeled data is then used to train a model of the distribution of input text over the known intentions using various machine learning approaches. Common approaches include support vector machines, neural networks, and decision trees.

For either language model construction technique, the language model can be constructed as a tree, where general knowledge is understood at the root and specific intents at the leaves. For example, statistical models can be trained on language related to particular domains or tasks such as looking up the weather or ordering food. These specialized models can then be leaves under a higher-level model that chooses the domain or task or learned jointly into a single hierarchical model. These trees can be multiple levels deep depending on how many sub-domain layers are modeled.

Regardless of language model design, its purpose is that of selecting the most likely intention given the user input. In this way, it performs intent classification, where each intent known by the NLU component can be considered a class label.

Current practices are to create an intent classifier for a pre-defined set of intents. Regardless of the method used to construct such a classifier, once the classifier is trained, it is static in that it will always classify a user input into one of the known classes (intentions). As language models evolve and expand, this classification task becomes more difficult. When the set of known classes grows to the thousands, it becomes increasingly difficult for the NLU component to correctly select the intent. Additionally, business rules and regulations often require that specific user behaviors or features be mapped to pre-defined intents where they can be handled differently for regulatory compliance, threat handling, or transferred to human customer service agents. Thus, static intent classifiers are time-consuming and expensive.

For example, in the insurance domain, if a user asks an IVA a policy question, the NLU component must select between all known intents, even though that particular user may not be eligible or authorized to access content returned from a large subset of the known intents. This content filtering and access control is typically handled downstream of the NLU component, after the intent is selected. Suppose that of the 2,000 known intents within the language model, only 1,500 were actually eligible for the given user due to their purchased policies, location, or time of day. The NLU component must still consider and correctly eliminate 500 of the possible intents, which, if the language model consists of patterns to be applied, can greatly increase processing time. Regardless of the model type, this consideration between ineligible classes can decrease accuracy as small sets of exceptions may be either unrepresented or poorly represented in the language model training data, leading to false positives.

SUMMARY

To prevent intent classifiers from potentially choosing intents that are ineligible for the current input due to policies, dynamic intent classification systems and methods are provided that dynamically control the possible set of intents using environment variables (also referred to as external variables). Associations between environment variables and ineligible intents, referred to as culling rules, are used.

In an implementation, a dynamic intent classification system is provided. The system includes a database configured to store a plurality of culling rules based on environment variables; and a classifier configured to determine an intent based on a user input and the plurality of culling rules, and output the intent.

Implementations may include some or all of the following features. The database and the classifier are comprised within a computing device. The database and the classifier are comprised within a chatbot. The chatbot is configured to provide a processed language output based on the intent. The classifier is a rule-based classifier, a tree-based classifier, a grammar-based classifier, or a statistically-trained classifier. The environment variables comprise at least one of time of day or input channel. The classifier is dynamically alterable using the culling rules to generate a plurality of culled intents. The classifier comprises a language model, and wherein the classifier is configured to eliminate ineligible intents using the culling rules without retraining or redeployment of the language model.

In an implementation, a method of providing a response to input data using dynamic intent classification based on environment variables is provided. The method includes maintaining a plurality of culling rules based on a plurality of environment variables, at a natural language understanding (NLU) component; receiving an input data at the NLU component; determining an intent for the input data using the culling rules, at a classifier of the NLU component; and outputting the intent.

Implementations may include some or all of the following features. Providing a processed language output, by a chatbot, based on the intent. Receiving the plurality of environment variables at the NLU component, and determining the plurality of culling rules using the plurality of environment variables. Dynamically altering the classifier using the culling rules to generate a plurality of culled intents. The classifier comprises a language model, and further comprising configuring the classifier to eliminate ineligible intents using the culling rules without retraining or redeployment of the language model. Maintaining the plurality of culling rules in storage external to the classifier. The classifier is a rule-based classifier, a tree-based classifier, a grammar-based classifier, or a statistically-trained classifier. The plurality of environment variables comprise at least one of time of day or input channel.

In an implementation, a method is provided. The method includes receiving a plurality of culling rules at a classifier, wherein the plurality of culling rules are based on a plurality of environment variables; and dynamically altering the classifier using the plurality of culling rules.

Implementations may include some or all of the following features. Maintaining the plurality of culling rules in storage external to the classifier. The classifier is a rule-based classifier, a tree-based classifier, a grammar-based classifier, or a statistically-trained classifier. The classifier comprises a language model, and further comprising configuring the classifier to eliminate ineligible intents using the culling rules without retraining or redeployment of the language model.

This summary is provided to introduce a selection of concepts in a simplified form that are further described below in the detailed description. This summary is not intended to identify key features or essential features of the claimed subject matter, nor is it intended to be used to limit the scope of the claimed subject matter.

BRIEF DESCRIPTION OF THE DRAWINGS

The foregoing summary, as well as the following detailed description of illustrative embodiments, is better understood when read in conjunction with the appended drawings. For the purpose of illustrating the embodiments, there is shown in the drawings example constructions of the embodiments; however, the embodiments are not limited to the specific methods and instrumentalities disclosed. In the drawings:

FIG. 1 is an illustration of an exemplary environment for dynamic intent classification based on environment variables;

FIG. 2 is an operational flow of an implementation of a method for providing a response to input data using dynamic intent classification based on environment variables;

FIG. 3 is an operational flow of an implementation of a rules-based method for providing a response to input data using dynamic intent classification based on environment variables;

FIG. 4 is an operational flow of an implementation of a tree-based method for providing a response to input data using dynamic intent classification based on environment variables;

FIG. 5 is an operational flow of an implementation of a grammar-based method for providing a response to input data using dynamic intent classification based on environment variables;

FIG. 6 is an operational flow of an implementation of a statistical-based method for providing a response to input data using dynamic intent classification based on environment variables; and

FIG. 7 shows an exemplary computing environment in which example embodiments and aspects may be implemented.

DETAILED DESCRIPTION

FIG. 1 is an illustration of an exemplary environment 100 for dynamic intent classification based on environment variables. The environment 100 may include a chatbot 170 and a client device 180 in communication through a network 105. The network 105 may be a variety of network types including the public switched telephone network (PSTN), a cellular telephone network, and a packet switched network (e.g., the Internet). Although only one chatbot 170 and one client device 180 are shown in FIG. 1, there is no limit to the number of chatbots and client devices that may be supported.

The chatbot 170 and the client device 180 may each be implemented using a variety of computing devices such as smartphones, desktop computers, laptop computers, tablets, set top boxes, vehicle navigation systems, and video game consoles. Other types of computing devices may be supported. A suitable computing device is illustrated in FIG. 7 as the computing device 700.

The chatbot 170 is in communication with the client device 180 and is capable of conducting a conversation with a user through the client device 180. The chatbot 170 may comprise a computer program or artificial intelligence application that conducts a conversation with the client device 180 via textual methods, in an implementation. Alternatively or additionally, the chatbot 170 may conduct a conversation with the client device via auditory methods. Depending on the implementation, the chatbot 170 may be accessed via a virtual assistant 172, a messaging app 174, or by other apps or websites.

The chatbot 170 comprises a natural language understanding (NLU) component 110. The NLU component 110 comprises a classifier 126 that determines and provides an intent 129. As described further herein, the classifier 126 outputs the intent 129. The NLU component 110 further comprises a pre-processor 123 that receives and pre-processes the user input 183 and provides it output to the classifier 126. Culling rules 121, based on or associated with environment variables 122, are also provided to the classifier 126. The environment variables 122, and/or the culling rules 121, may be stored in a database or other memory or storage. The database or other memory or storage is external to the classifier 126, and is accessed by the classifier 126 at runtime. In an implementation, the classifier may be trained by a trainer 125. The trainer 125 may receive and compile prior knowledge and performance information and train the classifier 126 accordingly.

The user input 183, such as chat data, from the client device 180 is provided to the NLU component 110. While chatbots encourage conversational interaction, the user input 183 may also comprise inputs that are keyword searches or non-conversational. Depending on the implementation, the user input 183 may be provided directly from the client device 180 (e.g., via the network 105) or from the chatbot 170 (e.g., via the network 105). The NLU component 110 processes the user input 183 and outputs processed language output 186 using the intent 129. The intent 129 is the user intention which is then used to determine how to respond to that intention. In an implementation, the processed language output 186 may comprise the input's action, object, actor, indirect object, and/or their respective modifying details.

In some implementations, the chatbot 170 comprises a training module 115. The training module 115 may comprise training data 117 and may be used to train the pre-processor 123. The training data 117 may be based on human to virtual assistant chat data. In an implementation, user inputs are captured from a commercial virtual agent chatbot and used for generating the training data 117. This data is typically short inputs ranging from two to 22 words with an average of 8.5 words. These inputs range between keyword searches and more conversational questions. The training data 117 may be weighted toward imperatives, statives, and interrogatives. In an implementation, the training data 117 is hand tagged.

The pre-processor 123 receives and processes the user input data 183 and provides its output to the classifier 126. The classifier 126 considers the output of the pre-processor in conjunction with the culling rules 121, and determines an intent 129 to return (i.e., to output as the intent 129).

The classifier 126 is a dynamic intent classifier, as the culling rules 121 comprises associations between the environment variables 122 and ineligible intents. The classifier uses the culling rules 121 to dynamically control the possible set of intents. In this manner, the classifier 126 is prevented from potentially choosing intents that are ineligible for the current input due to policies, for example.

FIG. 2 is an operational flow of an implementation of a method 200 for providing a response to input data using dynamic intent classification based on environment variables. The method 200 may be implemented using the chatbot 170 in conjunction with the environment 100, in some implementations.

At 210, environment variables 122 are received or otherwise maintained by a natural language understanding (NLU) component, such as the NLU component 110 of the chatbot 170. The environment variables 122 may be stored in storage, such as a memory device or database for example.

At 220, culling rules 121 are generated using the environment variables 122.

At 230, input data, such as the user input 183, is received and sent to the classifier 126. In some implementations, the input data is pre-processed by the pre-processor 123 prior to sending to the classifier 126.

At 240, the intent 129 of the input data is determined by the classifier 126 in conjunction with the culling rules 121.

At 250, the chatbot 170 may generate a response to the input data based on the determined intent 129 and output the response as processed language output 186.

For example, with respect to Example 1, the following culling rules could be used to prevent intents (represented by capital letters) relating to specific retirement plans from being considered based on the current user's ownership of a particular plan. Such a scenario is an example of business logic being leveraged to improve the classification accuracy.

Example 1

• • :has 401k(current user):—A, B, H
  • :has 403b(current user):—C, D, H (i.e., remove C, D, H from possible intents).
  • :has RothIRA(current user):—E, F, G, H, I

Given these culling rules, if a user were to ask an IVA for a financial company a question about a retirement plan policy, intents known to the classifier 126 that do not apply to retirement plans owned by that user will not even be considered. This can increase classification accuracy by reducing the set of intents to be considered from. It may also speed classification time if the language model is implemented in a rules-based method that requires running all rules against the input and ranking the matches.

In Example 2, the environmental knowledge of which channel the user is talking to the IVA on is used to remove any intents that do not apply to that channel from the language model:

Example 2

• • :is Web(channel):—A, B
  • :is Voice(channel):— . . .
  • :is SMS(channel):— . . .

In Example 2, note that some intents are not possible when talking to the IVA over the telephone or texting. In a banking IVA, intent A may be to open a new checking account, which is also not available over texting due to regulatory requirements of identity verification. Similarly, intent B may represent closing an account. Both of those actions are available when talking to the IVA on the company website however, and by removing them from the language model when they are not possible, misunderstandings related to account management when talking to the IVA over telephone or by text can be prevented. Thus, the knowledge of the current user environment is used to reduce the possible intents for the classifier 126 to choose between and lower possible confusion, per user input.

The embodiments disclosed and contemplated herein use a set of culling rules 121 (example of which are given in Examples 1 and 2), to dynamically modify the output of the classifier 126 itself (not invalidate the intent after it has been output by the classifier 126). This set of culling rules 121 can be modified at any time and will take effect immediately without requiring any model retraining or determination of alternative intents. The application of such culling rules 121 is dependent on the classification method used, and different approaches are described herein.

Some embodiments use intent disablement for rules-based intent classifiers. For rule-based approaches, whether manually constructed or empirically machine-learned, it is desirable to modify the model in the classifier 126 itself in order to realize performance gains in addition to increase the accuracy. Any pattern is removed that only applies to the intents that are to be culled. So those intents cannot be obtained by the classifier 126. In other words, it will be impossible for the classifier 126 to get to certain intents, because no pattern exists that will allow the classifier 126 to get to those certain intents.

FIG. 3 is an operational flow of an implementation of a rule-based method 300 for providing a response to input data using dynamic intent classification based on environment variables. The method 300 may be implemented using the chatbot 170 in conjunction with the environment 100, in some implementations.

At 310, similar to 210, environment variables 122 are received or otherwise maintained by an NLU component, such as the NLU component 110 of the chatbot 170. The environment variables 122 may be stored in storage, such as a memory device or database for example. At 320, similar to 220, culling rules 121 are generated using the environment variables 122.

At 330, the classifier 126, which in this embodiment is a rule-based classifier, is itself modified using the environment variables 122.

At 340, input data, such as the user input 183, is received and sent to the classifier 126. In some implementations, the input data is pre-processed by the pre-processor 123 prior to sending to the classifier 126.

At 350, the intent 129 of the input data is determined by the classifier 126, which has been modified using the environment variables, in conjunction with the culling rules 121.

At 360, the chatbot 170 may generate a response to the input data based on the determined intent 129 and output the response as processed language output 186.

For tree-based methods that include decision trees, for example, perform a bottom-up pruning of the tree to remove all nodes and edges that lead to a culled intent. FIG. 4 is an operational flow of an implementation of a tree-based method 400 for providing a response to input data using dynamic intent classification based on environment variables. The method 400 may be implemented using the chatbot 170 in conjunction with the environment 100, in some implementations.

At 410, input data, such as the user input 183, is received and sent to the classifier 126. In some implementations, the input data is pre-processed by the pre-processor 123 prior to sending to the classifier 126.

At 420, environment variables 122 are received or otherwise maintained by an NLU component, such as the NLU component 110 of the chatbot 170. The environment variables 122 may be stored in storage, such as a memory device or database for example.

At 430, culling rules 121 are generated using the environment variables 122.

At 440, the classifier 126, which is a tree-based classifier in this implementation, is modified using the environment variables. The modification may be performed by bottom-up pruning of the tree to remove all nodes and edges that lead to a culled intent.

At 450, the intent 129 of the input data is determined by the classifier 126 in conjunction with the culling rules 121.

At 460, the chatbot 170 may generate a response to the input data based on the determined intent 129 and output the response as processed language output 186.

At 470, the tree is restored to its original unmodified form. In this manner, the modification performed at 440 is removed.

As an example, using the culling set from Example 1, if the current user conversing with the financial IVA does not have a Roth IRA or 403B account, the leaf nodes representing the intentions C, D, E, F, G, H, and I would be pruned from the tree, and the edges and nodes leading to them would be recursively removed until the remaining decision tree had no knowledge of the culled intentions. On the next user input, the original tree is restored and the culling would be applied again based on whatever subset of culling rules corresponded to that input.

For grammar-based methods, the process is similar, with the difference being all grammars associated with the culled set of intentions would be removed. Any dependent grammars on the culled set would also be removed until the remaining grammars were only used for determining intentions other than the culled set. FIG. 5 is an operational flow of an implementation of a grammar-based method 500 for providing a response to input data using dynamic intent classification based on environment variables. The method 500 may be implemented using the chatbot 170 in conjunction with the environment 100, in some implementations.

At 510, input data, such as the user input 183, is received and sent to the classifier 126. In some implementations, the input data is pre-processed by the pre-processor 123 prior to sending to the classifier 126.

At 520, environment variables 122 are received or otherwise maintained by an NLU component, such as the NLU component 110 of the chatbot 170. The environment variables 122 may be stored in storage, such as a memory device or database for example.

At 530, culling rules 121 are generated using the environment variables 122.

At 540, the classifier 126, which is a grammar-based classifier in this implementation, is modified using the environment variables. The modification may be performed by removing grammars associated with a culled set of intentions.

At 550, the intent 129 of the input data is determined by the classifier 126 in conjunction with the culling rules 121.

At 560, the chatbot 170 may generate a response to the input data based on the determined intent 129 and output the response as processed language output 186.

At 570, the original grammar rules of the classifier are restored to the original unmodified form. In this manner, the modification performed at 540 is removed.

Thus, the language model is dynamically tailored to the business rules, regulations, or environmental variables applicable to each individual user input. Any change to the culling rules 121 have immediate effect, and no external language model refinement is needed.

Some embodiments use intent disablement for statistical intent classifiers. With statistically-trained language models such as support vector machines or neural networks, the language model cannot be dynamically altered without compromising the models performance on the remaining intentions. In addition, the classification time of such models is typically not significantly slowed by the inclusion of more classes. Therefore, focus on eliminating culled intentions from the output distribution of the model. In this way, it is guaranteed that no culled intentions are returned by the model.

FIG. 6 is an operational flow of an implementation of a statistical-based method for providing a response to input data using dynamic intent classification based on environment variables. The method 600 may be implemented using the chatbot 170 in conjunction with the environment 100, in some implementations.

At 610, environment variables 122 are received or otherwise maintained by an NLU component, such as the NLU component 110 of the chatbot 170. The environment variables 122 may be stored in storage, such as a memory device or database for example.

At 620, culling rules 121 are generated using the environment variables 122.

At 630, input data, such as the user input 183, is received and sent to the classifier 126, which in this implementation is a statistical-trained classifier. In some implementations, the input data is pre-processed by the pre-processor 123 prior to sending to the classifier 126.

At 640, the intent 129 of the input data is determined by the classifier 126 in conjunction with the culling rules 121, by eliminating culled intentions from the output distribution of the classifier 126. In an implementation, to perform the culling, apply a masking layer over the output distribution from a statistical model. This layer performs a bitwise AND operation between the intention scores and the culling mask values. This is very fast to compute and as any culled intentions will have a value of 0 in the bitmask, they will be eliminated from the model output as any value ANDed to 0 will return 0. The downstream components to the NLU component 110 will therefore never see any culled intention, and will select as the winning intent the highest scoring among the non-zero values.

At 650, the chatbot 170 may generate a response to the input data based on the determined intent 129 and output the response as processed language output 186.

FIG. 7 shows an exemplary computing environment in which example embodiments and aspects may be implemented. The computing device environment is only one example of a suitable computing environment and is not intended to suggest any limitation as to the scope of use or functionality.

Numerous other general purpose or special purpose computing devices environments or configurations may be used. Examples of well-known computing devices, environments, and/or configurations that may be suitable for use include, but are not limited to, personal computers, server computers, handheld or laptop devices, multiprocessor systems, microprocessor-based systems, network personal computers (PCs), minicomputers, mainframe computers, embedded systems, distributed computing environments that include any of the above systems or devices, and the like.

Computer-executable instructions, such as program modules, being executed by a computer may be used. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. Distributed computing environments may be used where tasks are performed by remote processing devices that are linked through a communications network or other data transmission medium. In a distributed computing environment, program modules and other data may be located in both local and remote computer storage media including memory storage devices.

With reference to FIG. 7, an exemplary system for implementing aspects described herein includes a computing device, such as computing device 700. In its most basic configuration, computing device 700 typically includes at least one processing unit 702 and memory 704. Depending on the exact configuration and type of computing device, memory 704 may be volatile (such as random access memory (RAM)), non-volatile (such as read-only memory (ROM), flash memory, etc.), or some combination of the two. This most basic configuration is illustrated in FIG. 7 by dashed line 706.

Computing device 700 may have additional features/functionality. For example, computing device 700 may include additional storage (removable and/or non-removable) including, but not limited to, magnetic or optical disks or tape. Such additional storage is illustrated in FIG. 7 by removable storage 708 and non-removable storage 710.

Computing device 700 typically includes a variety of computer readable media. Computer readable media can be any available media that can be accessed by the device 700 and includes both volatile and non-volatile media, removable and non-removable media.

Computer storage media include volatile and non-volatile, and removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Memory 704, removable storage 708, and non-removable storage 710 are all examples of computer storage media. Computer storage media include, but are not limited to, RAM, ROM, electrically erasable program read-only memory (EEPROM), flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can be accessed by computing device 700. Any such computer storage media may be part of computing device 700.

Computing device 700 may contain communication connection(s) 712 that allow the device to communicate with other devices. Computing device 700 may also have input device(s) 714 such as a keyboard, mouse, pen, voice input device, touch input device, etc. Output device(s) 716 such as a display, speakers, printer, etc. may also be included. All these devices are well known in the art and need not be discussed at length here.

It should be understood that the various techniques described herein may be implemented in connection with hardware components or software components or, where appropriate, with a combination of both. Illustrative types of hardware components that can be used include Field-programmable Gate Arrays (FPGAs), Application-specific Integrated Circuits (ASICs), Application-specific Standard Products (ASSPs), System-on-a-chip systems (SOCs), Complex Programmable Logic Devices (CPLDs), etc. The methods and apparatus of the presently disclosed subject matter, or certain aspects or portions thereof, may take the form of program code (i.e., instructions) embodied in tangible media, such as floppy diskettes, CD-ROMs, hard drives, or any other machine-readable storage medium where, when the program code is loaded into and executed by a machine, such as a computer, the machine becomes an apparatus for practicing the presently disclosed subject matter.

In an implementation, a dynamic intent classification system is provided. The system includes a database configured to store a plurality of culling rules based on environment variables; and a classifier configured to determine an intent based on a user input and the plurality of culling rules, and output the intent.

Implementations may include some or all of the following features. The database and the classifier are comprised within a computing device. The database and the classifier are comprised within a chatbot. The chatbot is configured to provide a processed language output based on the intent. The classifier is a rule-based classifier, a tree-based classifier, a grammar-based classifier, or a statistically-trained classifier. The environment variables comprise at least one of time of day or input channel. The classifier is dynamically alterable using the culling rules to generate a plurality of culled intents. The classifier comprises a language model, and wherein the classifier is configured to eliminate ineligible intents using the culling rules without retraining or redeployment of the language model.

In an implementation, a method of providing a response to input data using dynamic intent classification based on environment variables is provided. The method includes maintaining a plurality of culling rules based on a plurality of environment variables, at a natural language understanding (NLU) component; receiving an input data at the NLU component; determining an intent for the input data using the culling rules, at a classifier of the NLU component; and outputting the intent.

Implementations may include some or all of the following features. Providing a processed language output, by a chatbot, based on the intent. Receiving the plurality of environment variables at the NLU component, and determining the plurality of culling rules using the plurality of environment variables. Dynamically altering the classifier using the culling rules to generate a plurality of culled intents. The classifier comprises a language model, and further comprising configuring the classifier to eliminate ineligible intents using the culling rules without retraining or redeployment of the language model. Maintaining the plurality of culling rules in storage external to the classifier. The classifier is a rule-based classifier, a tree-based classifier, a grammar-based classifier, or a statistically-trained classifier. The plurality of environment variables comprise at least one of time of day or input channel.

In an implementation, a method is provided. The method includes receiving a plurality of culling rules at a classifier, wherein the plurality of culling rules are based on a plurality of environment variables; and dynamically altering the classifier using the plurality of culling rules.

Implementations may include some or all of the following features. Maintaining the plurality of culling rules in storage external to the classifier. The classifier is a rule-based classifier, a tree-based classifier, a grammar-based classifier, or a statistically-trained classifier. The classifier comprises a language model, and further comprising configuring the classifier to eliminate ineligible intents using the culling rules without retraining or redeployment of the language model.

Although exemplary implementations may refer to utilizing aspects of the presently disclosed subject matter in the context of one or more stand-alone computer systems, the subject matter is not so limited, but rather may be implemented in connection with any computing environment, such as a network or distributed computing environment. Still further, aspects of the presently disclosed subject matter may be implemented in or across a plurality of processing chips or devices, and storage may similarly be effected across a plurality of devices. Such devices might include personal computers, network servers, and handheld devices, for example.

Although the subject matter has been described in language specific to structural features and/or methodological acts, it is to be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or acts described above. Rather, the specific features and acts described above are disclosed as example forms of implementing the claims.

Claims

1. A computing device comprising: a processor; anda memory operably coupled to the processor, the memory having computer-executable instructions stored thereon that, when executed by the processor cause the computing device to:receive, at a natural language understanding (NLU) component, a user input;pre-process the user input to identify at least one environmental variable;provide the user input to a trained classifier of the NLU component;identify an intent subset by culling at least a portion of a plurality of stored intents based at least on the at least one environmental variable;determine, using the trained classifier, an intent associated with the user input from the intent subset; andoutput, by the NLU component and via a display of the computing device, a response to the user input based at least on the determined intent.

2. The computing device of claim 1, wherein culling at least a portion of the plurality of stored intents comprises applying a masking layer over an output distribution of the trained classifier.

3. The computing device of claim 1, wherein the computer-executable instructions are further configured to, when executed by the processor, cause the computing device to: modify the trained classifier based at least on the at least one environmental variable associated with the user input.

4. The computing device of claim 3, wherein the trained classifier comprises a grammar-based model, and wherein modifying the trained classifier comprises removing grammar associated with a culled intent set.

5. The computing device of claim 1, wherein the trained classifier comprises a tree-based model, and wherein culling at least a portion of the plurality of stored intents comprises removing at least one node or edge leading to at least one of the plurality of stored intents.

6. The computing device of claim 1, wherein the computing device comprises a chatbot trained using chat data that is configured to output processed language outputs.

7. The computing device of claim 1, wherein culling at least a portion of the plurality of stored intents based at least on the at least one environmental variable comprises: using a rules-based operation to rank the plurality of stored intents; andidentifying a predetermined top number of the plurality of stored intents as the intent subset.

8. The computing device of claim 1, wherein the plurality of stored intents is stored in a database external to the computing device.

9. A system comprising: a processor; anda memory operably coupled to the processor, the memory having computer-executable instructions stored thereon that, when executed by the processor cause the system to:receive, at a natural language understanding (NLU) component, a user input;pre-process the user input to identify at least one environmental variable;provide the user input to a trained classifier of the NLU component;identify an intent subset by culling at least a portion of a plurality of stored intents based at least on the at least one environmental variable;determine, using the trained classifier, an intent associated with the user input from the intent subset; andoutput, by the NLU component, a response to the user input based at least on the determined intent.

10. The system of claim 9, wherein culling at least a portion of the plurality of stored intents comprises applying a masking layer over an output distribution of the trained classifier.

11. The system of claim 9, wherein the computer-executable instructions are further configured to, when executed by the processor, cause the system to: modify the trained classifier based at least on the at least one environmental variable associated with the user input.

12. The system of claim 11, wherein the trained classifier comprises a grammar-based model, and wherein modifying the trained classifier comprises removing grammar associated with a culled intent set.

13. The system of claim 9, wherein the trained classifier comprises a tree-based model, and wherein culling at least a portion of the plurality of stored intents comprises removing at least one node or edge leading to at least one of the plurality of stored intents.

14. The system of claim 9, further comprising: a chatbot trained using chat data that is configured to output processed language outputs as the response.

15. The system of claim 9, wherein culling at least a portion of the plurality of stored intents based at least on the at least one environmental variable comprises: using a rules-based operation to rank the plurality of stored intents; andidentifying a predetermined top number of the plurality of stored intents as the intent subset.

16. The system of claim 9, wherein the plurality of stored intents is stored in an external database.

17. A method comprising: receiving, at a natural language understanding (NLU) component, a user input;pre-processing the user input to identify at least one environmental variable;providing the user input to a trained classifier of the NLU component;identifying an intent subset by culling at least a portion of a plurality of stored intents based at least on the at least one environmental variable;determining, using the trained classifier, an intent associated with the user input from the intent subset; andoutputting, by the NLU component, a response to the user input based at least on the determined intent.

18. The method of claim 17, wherein culling at least a portion of the plurality of stored intents comprises applying a masking layer over an output distribution of the trained classifier.

19. The method of claim 17, further comprising: modifying the trained classifier based at least on the at least one environmental variable associated with the user input.

20. The method of claim 19, wherein the trained classifier comprises a grammar-based model, and wherein modifying the trained classifier comprises removing grammar associated with a culled intent set.