Systems and methods for leveraging embedded services
A method according to the present disclosure may include providing an embedded service in an application, in response to receiving an input via the embedded service, determining an applicable module of a plurality of modules based on a characteristic of at least one of the input or of the embedded service, processing the input via the applicable module, and controlling the application based on the processed input. Another method according to the present disclosure may include providing an embedded service in an application, training a monitoring model via federated learning based on data derived locally from the application, monitoring, via the trained monitoring model, a plurality of interactions with the embedded service, determining, by the trained monitoring model, that at least one of the plurality of interactions comprises an anomalous interaction, and in response to the determination, restricting further usage of the application.
1. A computer-implemented method comprising:
providing an embedded service in an application;
in response to receiving an input via the embedded service, determining an applicable module of a plurality of modules based on a characteristic of at least one of the input and of the embedded service;
processing the input via the applicable module; and
controlling the application based on the processed input,
wherein the applicable module comprises a contract module, and
wherein the input comprises a first input document and at least one second input document,
wherein processing the input comprises:
determining, by a machine learning model, embeddings representative of one or more levels of the first input document and the at least one second input document,
generating, by the machine learning model, a single embedding representative of an amount of deviation for at least one term of the first input document from at least one term of the at least one second input document based on the embeddings representative of the one or more levels of the first input document and the at least one second input document,
extracting, by the machine learning model, the at least one term of the first input document based on the single embedding exceeding a threshold,
determining, by the machine learning model, an action from a list of actions in response to the single embedding exceeding the threshold and based on the at least one term, and
presenting the action as an interactive element of the application,
wherein the at least one second input document comprises a document comprising one or more substitute terms having a relative fairness above a threshold value based on semantic features of the first input document.
2. The method of claim 1 , wherein the plurality of modules further comprises:
an abnormality detection module; and
an impersonation detection module.
3. The method of claim 2 , wherein the applicable module comprises the abnormality detection module, and wherein processing the input comprises:
determining a user associated with the input;
retrieving data regarding the determined user;
deriving, via the machine learning model, a behavior pattern from the retrieved data; and
determining, based on a comparison of the input to the derived behavior pattern, whether the input comprises an abnormality from the derived behavior pattern.
4. The method of claim 3 , wherein controlling the application comprises:
in response to determining that the input comprises the abnormality, restricting further access for the associated user to the application.
5. The method of claim 4 , wherein the at least one term comprises:
a term determined to be different from a previous version of the contract; or
a term having a relative fairness below a threshold value.
6. The method of claim 2 , wherein the applicable module comprises the impersonation detection module, and wherein processing the input comprises:
building, by the machine learning model, a security profile based on the input;
enriching, by the machine learning model, the security profile with data derived from the application; and
determining whether the enriched security profile is fraudulent.
7. The method of claim 1 , further comprising training the machine learning model by:
obtaining a training data set comprising a plurality of contracts;
training a first portion of the machine learning model, according to paragraph-level text of the plurality of contracts, to extract at least one paragraph-level portion from input documents;
training a second portion of the machine learning model, according to sentence-level text of the plurality of contracts, to extract at least one sentence-level portion from the input documents; and
training a third portion of the machine learning model, according to token-level text of the plurality of contracts, to extract at least one token-level portion from the input documents.
8. The method of claim 1 , wherein the first input document comprises a contract, and wherein the at least one second input document comprises a previous version of the contract.
9. A non-transitory computer readable medium storing program instructions that, when executed by a processor, a computer system including the processor is configured to perform operations comprising:
presenting an application on a graphical user interface (GUI) of a user device;
generating, via a service embedded in the application, a first interactive element on the GUI;
in response to receiving an interaction with the first interactive element, processing the interaction via at least one of a plurality of modules identified based on characteristics of the interaction; and
generating, based on the processed interaction, a second interactive element on the GUI,
wherein the second interactive element controls interactions with the application corresponding to inputs from the user device,
wherein the plurality of modules comprises an impersonation detection module, and wherein processing the interaction comprises:
building, by a machine learning model, a security profile for a user of the user device based on the inputs and the interaction of the user with the first interactive element within a same session as the inputs,
enriching, by the machine learning model, the security profile with user device data and with behavioral data derived from the application, and
determining whether the enriched security profile is fraudulent based on comparing a characteristic of the interaction and the security profile.
10. The non-transitory computer readable medium of claim 9 , wherein the plurality of modules comprise:
an abnormality detection module; and
a contract module.
11. The non-transitory computer readable medium of claim 10 , wherein the at least one module comprises the abnormality detection module, and wherein processing the interaction comprises:
determining the user associated with the input;
retrieving data regarding the determined user;
deriving, via the machine learning model, a behavior pattern from the retrieved data; and
determining, based on a comparison of the input to the derived behavior pattern, whether the input comprises an abnormality from the derived behavior pattern.
12. The non-transitory computer readable medium of claim 11 , wherein the second interactive element comprises, in response to determining that the input comprises the abnormality, a restriction on the user device for the embedded service.
13. The non-transitory computer readable medium of claim 10 , wherein:
the at least one module comprises the contract module,
the interaction comprises a contract,
processing the interaction comprises:
determining, by the machine learning model, embeddings representative of one or more levels of the contract;
generating, by the machine learning model, a single embedding representative of an amount of deviation for at least one term of the contract from at least one term of at least one second input document based on the embeddings representative of the one or more levels of the contract,
extracting, by the machine learning model, the at least one term of the contract based on the single embedding exceeding a threshold; and
determining, by the machine learning model, an action from a list of actions based on the at least one term, and
wherein the second interactive element comprises a selectable portion of the GUI to initiate execution of the determined action,
wherein the at least one second input document comprises a document comprising one or more substitute terms having a relative fairness above a threshold value based on semantic features of the contract.
14. The non-transitory computer readable medium of claim 13 , wherein the at least one term comprises:
a term determined to be different from a previous version of the contract; or
a term having a relative fairness below a threshold value.
15. The non-transitory computer readable medium of claim 13 , wherein the at least one second input document comprises a previous version of the contract.
16. The non-transitory computer readable medium of claim 9 , wherein the second interactive element comprises a prompt for verifying information as additional input from the user device, the prompt being displayed at the second interactive element on the GUI.
17. A computer-implemented system comprising:
a processor; and
a non-transitory computer readable medium stored thereon instructions that are executable by the processor to perform operations comprising:
an embedded service in an application;
receiving an input via the embedded service;
generating a profile of a user associated with the input, the profile being based on the input and interactions of the user within a same session;
determining an applicable module of a plurality of modules based on a characteristic of at least one of the input, of the embedded service, or of the user;
determining an action from a list of actions based on the characteristic; and
presenting the action via the embedded service,
wherein the plurality of modules comprises a contract module and an impersonation detection module, and
wherein the contract module determining the action from the list of actions based on the characteristic comprises:
determining, by a machine learning model, embeddings representative of one or more levels of an input document,
generating, by the machine learning model, a single embedding representative of an amount of deviation for at least one term of the input document from at least one term of at least one second document based on the embeddings representative of the one or more levels of the input document; and
extracting, by the machine learning model, the at least one term of the input document based on the single embedding exceeding a threshold,
wherein the action is in response to the single embedding representative of the deviation exceeding the threshold and based on the at least one term,
wherein the at least one second document comprises a document comprising one or more substitute terms having a relative fairness above a threshold value based on semantic features of the input document,
wherein the impersonation detection module determining the action from the list of actions based on the characteristic comprises:
enriching, by the machine learning model, the profile with user device data and with behavioral data derived from the application, and
determining whether the enriched profile is fraudulent based on the characteristic of the input and the profile.
18. The computer-implemented system of claim 17 , wherein the plurality of modules comprise:
an abnormality detection module.
19. The computer-implemented system of claim 18 , wherein:
the applicable module comprises the contract module,
the input comprises a contract,
the characteristic comprises at least one term of the contract determined, by the machine learning model of the contract module, to have a relative fairness below a threshold value; and
the determined responsive action comprises striking the at least one term.
20. The computer-implemented system of claim 17 , wherein the characteristic comprises at least one of:
a field of the embedded service;
a type of the input;
an interaction history of the user;
a sensitivity of the embedded service; or
a complexity of the input.