A web application system for text processing and web page navigation, and a method thereof

Abstract

Embodiments herein disclose a web application system comprising a system comprising a natural language processing unit (NLPU) and a graphical user interface (GUI) navigator. The user inputs their query via a conversational interface in the NLPU. An activity and context recognition engine (ACRE) identifies at least one past user activity, which is indicative of a context associated with the user. Based on the user query and the at least one past user activity, a core language engine (CLE) in the NLPU determines an intent associated with the user query, and generates a response. Based on the determined intent, the GUI navigator navigates the user to a web page having information pertaining to the determined intent associated with the user query.

Description

A WEB APPLICATION SYSTEM FOR TEXT PROCESSING AND WEB PAGE NAVIGATION, AND A METHOD THEREOFCross-Reference To Related Applications:

[0001] This application claims the benefit of and priority to Indian Provisional Application 202311045899, filed on July 7, 2023, which is hereby incorporated by reference in its entirety.Technical Field:

[0002] The present disclosure relates to the fields of computing and artificial intelligence, and more particularly relates to a web application system for text processing and web page navigation, and a method thereof.Background:

[0003] In today’s digital world, users are able to avail many services and products by accessing a web application system through their electronic device (e.g., a smartphone or a personal computer). Users can interact with a web application system through their device’s browser — by typing the website address directly or clicking on a link that leads to the desired page. Once the appropriate address is entered or the link is clicked, the application is rendered on the device’s browser.

[0004] Most web application systems are designed with multiple web pages that serve a dual purpose: presenting information and accepting user inputs. This interactive design enables users to engage with the website and perform various tasks as needed. To assist users in navigating through the web application, site maps are often provided as a visual representation of the website’s structure. However, despite their intended purpose, site maps can sometimes be challenging to use and may not effectively guide users to the correct application screen within the website. This has paved the way to leverage search-based approaches to find appropriate web pages that are relevant to the information requested or application. However, search-based approaches suffer from the limitation of requiring the user to input specific (or exact) keywords to retrieve the web pages that are associated with the requested information.

[0005] Nowadays, in order to quickly retrieve the information the user is looking out for, instead of having the user figure out the key terms to input into a search bar on the website, some web application systems utilize artificial intelligence -based chatbots for accepting a user’s input (in their own words) and providing a reply to the user’s input. In other words, the user’s query is formulated using their natural words, instead of specific terminology, and is input to the chatbot for processing. However, the reply from the chatbot may not be satisfactory to the user owing to the limitations in the input parameters considered by the chatbot for replying to the user, and the limitations in its output capabilities.Summary:

[0006] These and other problems are generally solved or circumvented, and technical advantages are generally achieved, by advantageous embodiments of the present disclosure.

[0007] A summary of certain embodiments disclosed herein is set forth below. It should be understood that these aspects are presented merely to provide the reader with a brief summary of these certain embodiments and that these aspects are not intended to limit the scope of this disclosure. Indeed, this disclosure may encompass a variety of aspects that may not be set forth below.

[0008] According to a first embodiment, a system is disclosed. The system comprises a conversational interface configured to receive a query from a user that is input via a user device. The system comprises a core language engine (CLE) configured to receive the query from the conversational interface and determine an intent associated with the user query. The system comprises a graphical user interface (GUI) navigator, configured to navigate the user to a desired web page based on the determined user query intent. The CLE determines the user query intent by converting the user query into a mathematicalvector, using a vector embedding model. The CLE converts a description of each intent, among a plurality of intents, to a mathematical vector, using the vector embedding model. The CLE compares the user query vector with each intent description vector, based on a cosine similarity, to determine the distance between the user query vector and each intent description vector. The cosine values can be indicative of those intent description vectors that are close to the user query vector. The CLE selects a subset of the intent description vectors, whose distance from the user query vector is below a threshold. The CLE feeds a first machine learning model (MLM) the following inputs: the user query; the intent and the description associated with each intent description vector in the subset of intent description vectors; and a set of instructions to be followed for determining the user query intent. The first LLM is trained to determine the user query intent, based on the inputs, and wherein the determined user query intent is one of the intents associated with the subset of intent description vectors.

[0009] According to a second embodiment, a method is disclosed. The method comprises receiving, via a conversational interface, a query from a user. The method comprises converting, using a vector embedding model, the user query into a mathematical vector. The method comprises converting a description of each intent, among a plurality of intents, to a mathematical vector using the vector embedding model. The method comprises comparing the user query vector with each intent description vector, based on a cosine similarity, to determine the distance between the user query vector and each intent description vector. The cosine values can be indicative of those intent description vectors that are close to the user query vector. The method comprises selecting a subset of the intent description vectors, wherein the subset comprises those intent description vectors that are distanced below a threshold from the user query vector. The method comprises feeding to a first machine learning model (MLM) the following inputs: the user query; the intent and the description associated with each intent description vector in the subset of intent description vectors; and a set of instructions to be followed for determining the user query intent. The first MLM is trained to determine the user query intent based on the inputs, and wherein the determined user query intent is one of the intents associated with the subset of the intent description vectors. The method comprisesnavigating, by a graphical user interface (GUI) navigator, the user to a desired web page based on the determined user query intent.Brief Description of Drawings:

[0010] The detailed description is described with reference to the accompanying figures. In the figures, the left-most digit of a reference number identifies the figure in which the reference number first appears. The same numbers are used throughout the drawings to reference like features and components.FIG. 1 illustrates the manner in which a user is able to access the web application system, according to an embodiment of the present disclosure;FIG. 2 illustrates the various components within the web application system, according to an embodiment of the present disclosure;FIG. 3A illustrates the various modules within a conversational interface of a natural language processing unit of the web application system, according to an embodiment of the present disclosure;FIG. 3B illustrates the various modules within a core language engine of the natural language processing unit of the web application system, according to an embodiment of the present disclosure;FIG. 3C illustrates the various modules within an activity intent and context recognition engine of the natural language processing unit of the web application system, according to an embodiment of the present disclosure;FIG. 4 illustrates a process flow illustrating the sequence of operations executed by the various components within the web application system, according to an embodiment of the present disclosure;FIG. 5 illustrates an example layout of the web page that is displayed by a graphical user interface navigator, according to an embodiment of the present disclosure;FIG. 6 illustrates a method performed by an activity intent and context recognition engine in the natural language processing unit, according to an embodiment of the present disclosure;FIG. 7 illustrates a method performed by a core language engine in the natural language processing unit, according to an embodiment of the present disclosure; andFIG. 8 illustrates a method performed by the graphical user interface navigator, according to an embodiment of the present disclosure.Detailed Description

[0011] Exemplary embodiments now will be described with reference to the accompanying drawings. The invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein; rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey its scope to those skilled in the art. The terminology used in the detailed description of the particular exemplary embodiments illustrated in the accompanying drawings is not intended to be limiting. In the drawings, like numbers refer to like elements. The term “exemplary embodiment” is meant to be interpreted as being an example embodiment and is not meant to be interpreted as a preferred embodiment.

[0012] The specification may refer to “an”, “one” or “some” embodiment(s) in several locations. This does not necessarily imply that each such reference is to the same embodiment(s), or that the feature only applies to a single embodiment. Single features of different embodiments may also be combined to provide other embodiments.

[0013] As used herein, the singular forms “a”, “an” and “the” are intended to include the plural forms as well, unless expressly stated otherwise. It will be further understood that the terms “includes”, “comprises”, “including” and / or “comprising” when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. Conditional language, such as among others, “can” or “may”, unless specifically stated otherwise, or otherwise understood within the context as used, is generally intended to convey that certain embodiments could include, while other embodiments may not include certain features, elements, and / or steps. Thus, such conditional language is not generally intended to imply that features, elements, and / or steps are in any way required for one or more embodiments. It will be understood that when an element is referred to as being “connected” or “coupled” to another element, it can be directly connected or coupled to the other element or intervening elements may be present. Furthermore, “connected” or “coupled” as used herein may include wirelessly connected or coupled. As used herein, the term “and / or” includes any and all combinations and arrangements of one or more of the associated listed items.

[0014] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.

[0015] The figures depict a simplified structure only showing some elements and functional entities, all being logical units whose implementation may differ from what is shown. The connections shown are logical connections; the actual physical connections may be different. In addition, all logical units described and depicted in the figures include the software and / or hardware components required for the unit to function. Further, each unit may comprise within itself one or more components, which are implicitly understood. These components may be operatively coupled to each other and be configured to communicate with each other to perform the function of the said unit.

[0016] The embodiments herein disclose a web application system (WAS) for processing the text of a user to determine the user’s intent based on the text input by the user, and navigating the user to a specific web page based on the determined intent. The WAS comprises an interface through which a user, using their natural language, can input a query (in the form of a natural language statement) to be answered by the WAS. In response to receiving the input, the WAS may analyze the user’s input and any available information relating to a previous activity by the user, for an intent of the user that is associated with the user’s input. Upon determining the user intent, the WAS may generate a desired response to the user’s input, along with navigating the user to a desired web page, wherein the desired web page may substantiate the content of the web application system’s generated response. In case the user’s input had specific named entities, the WAS identifies these named entities and pre-fills any input fields (e.g., text fields, checkboxes, or radio buttons) on the navigated-to web page, based on the specific named entities.

[0017] For instance, assume that the user was browsing web pages related to baggage limits and also booked a flight from Mumbai to New York. If the user posed a question, at a conversational interface (i.e., an interface for accepting user input) of the WAS, along the lines of “What is the baggage limit?”, the WAS may decode that the user’s intent is to determine the baggage limits, based on the user input itself and the user’s previous visits to baggage limit-related web pages (i.e., the previous user activity provides context about the user). The WAS may also take into consideration the flight booked by the user (i.e., another previous user activity), to provide context as to the flight for which user wished to know the baggage limits of. By considering the user’s past user activities, in additionto the user input in and of itself, the embodiments herein can generate a satisfactory response to the user’s input. Additionally, the embodiments herein may also navigate the user to a web page that provides information on the baggage limit for the flight that the user is on. In other words, the web application system can direct the user to a web page, with the requested information, such that the content (i.e., the requested information) on the navigated-to web page substantiates the content of the generated response.

[0018] Referring now to the drawings, and more particularly to Figures 1 through 8, where similar reference characters denote corresponding features consistently throughout the figures, there are shown embodiments.

[0019] FIG. 1 illustrates the manner in which a user is able to access the web application system (WAS) 10, according to an embodiment of the present disclosure. The user may use their user device 20 to connect to a network 30, through which the user can access the WAS 10. Examples of the user device 20 include, but are not limited to, a smartphone, a personal computer, a laptop, a tablet etc. The user device 20 may use one or more general or special purpose processors, memory, storage, and networking components (either wired or wireless).

[0020] The network 30 can be the Internet, or can utilize dedicated or private communication networks (e.g., WAN, MAN, or LAN) that are not necessarily part of the Internet. In other embodiments, the network 30 can use standard communication technologies and / or protocols. Once the user device 20 connects to the network 30, they may be able to access the WAS 10. On the user device 20, the WAS 10 may, for example, be embodied in a web page on a web browser or in an application that is installed in the user device 20.General Overview Of The Web Application System

[0021] FIG. 2 illustrates the various components within the WAS 10, according to an embodiment of the present disclosure. The WAS 10 comprises a natural language processing unit (NLPU) 11 and a graphical user interface (GUI) navigator 12. The NLPU 11 comprises a conversational interface 110, a core language engine (CLE) 112, and an activity and context recognition engine (ACRE) 114.

[0022] The conversational interface 110 can be displayed on the user device 20 and can accept an utterance (i.e., an input) from the user. The conversational interface 110 can be a user- friendly interface. The utterance could be verbal (e.g., the words spoken by the user) or written (e.g., a text input from the user through a keyboard). In other words, the conversational interface 110 can accept an input in the form of a natural language statement via a voice or text command, wherein the user sends their input to the conversational interface 110 via the user device 20.

[0023] The core language engine (CLE) 112 can process the utterance from the user to determine what the user wishes to know (i.e., the user’s intent associated with their utterance). Determining the intent of the user can comprise analyzing the input of the user (via the conversational interface) and the previous activities of the user. The previous activities of the user can be indicative of a context associated with the user.

[0024] The activity and context recognition engine (ACRE) 114 is responsible for identifying the past activities of the user, and transmitting the same to the CLE 112. Examples of the past activities can include the following: previous web pages visited, the actions performed on the web pages etc.

[0025] Once the CLE 112 determines the user intent (based on the user input and the past activities), based on the determined intent, the GUI navigator 12 navigates the user to the appropriate web page. In other words, the GUI navigator 12 displays to the user a web page having the information that the user is seeking. If there are any named entities in the user input, then the CLE 112 may also extract these named entities and feed the same to the GUI navigator 12, such that depending on the web page to be displayed, any input fields on the displayed web page can be pre-filled based on the extracted named entities.

[0026] Those of skill in the art will appreciate that the WAS 10 may comprise other components / modules that are not described herein.Detailed Overview Of The Web Application System

[0027] The components within the WAS 10 will now be described in greater detail. FIG. 3A illustrates the various modules within the conversational interface 110 of the NLPU 11. The conversational interface 110 comprises the following modules: spell check module 1102, grammar correction module 1104, auto-completion module 1106, and voice-to-text module 1108.Detailed Overview Of The Conversational Interface

[0028] The conversational interface 110 may present to the user (on a display of the user device 20) a text area — with an option for voice input — for the user to provide their utterance (for example, as shown in FIG. 5). The spell check module 1102 can compare the text input from the user and the words in a reference (e.g., a dictionary), to check if the user’s text input is spelt correctly. In embodiments where the user provides a voice input (e.g., in the form of a natural language statement or query), the voice-to-text module 1108 may convert the user’s voice input to text, which may then be received by the spell check module 1102. The spell check module 1102 may make any corrections to the user’s input if there is no match between the user’s input and the words in the reference. The spell check module 1102 may make the corrections to the text input based on the closest matching word in the reference.

[0029] Once the spell check module 1102 has reviewed all the words present in the user’s input, the grammar correction module 1104 may review the words (in the order that the words were input) to see if they form a coherent sentence. If the sentence lacks coherence, then the grammar correction module 1104, along with the auto-completion module 1106, may make the necessary changes to achieve coherence. For example, if there is a missing word, then the auto-completion module 1106 may predict the next word based on the user’s input. This module 1106 may use artificial intelligence (Al) models to generate the word prediction. Similarly, if the user’s input is grammatically incorrect, then the grammar correction module 1104 may, for example, alter any prepositions or change the tense of certain words. io

[0030] The term “utterance” or “user input” or “user query” can be used interchangeably. Moreover, the terms “utterance”, “user input”, and “user query” can be interpreted broadly. For example, a textual input from a user can include, but is not limited to, receiving an image file with text in it, that can then undergo optical character recognition (OCR) to recognize the text. Further, in some embodiments, the user input can also be construed as a query. For example, for a user input “I want to cancel my ticket”, this can be interpreted as the following query “how do I cancel my ticket?”, wherein the user input and its interpretation have the same intent, i.e., cancellation of a ticket.Detailed Overview Of The Activity & Context Recognition Engine

[0031] FIG. 3C illustrates the various modules within the activity & context recognition engine (ACRE) 114, according to an embodiment of the present disclosure. The ACRE 114 comprises the following modules: a user action module 1142, a user action metadata module 1144, an action-to-label module 1146, and an activity identification module 1148.

[0032] The user action module 1142 can keep track of the various actions performed by the user on a web page (provided by the WAS 10). The following is a non-exhaustive list of actions that a user can perform:• Searching for a specific product on a web page by typing into a search bar of the web page.• Adding the specific product to a cart upon being displayed various products.• Clicking on a sub-menu from a drop-down menu to obtain more information on a product / serve (e.g., clicking on a sub-menu “Howto apply for a credit card?” from a drop-down menu of “Credit cards”).

[0033] The user action metadata module 1144 may keep track of various metadata pertaining to the user actions, such as the date on which the user action was performed, along with the time spent by the user on a specific web page etc.

[0034] The action-to-label module 1146 can map a user action to a specific label. In other words, each action performed by a user, can be mapped to a predetermined label. For example, the below table illustrates an action performed and its corresponding label.Table 1

[0035] The activity identification module 1148 can group a specific sequence of labels to an activity that is performed (i.e., a unit of activity). For example, the following table illustrates a mapping of a sequence of labels to an activity.Table 2

[0036] For the activity “PURCHASE PRODUCT”, the sequence of labels (i.e., sequence of actions) can be associated with a set of actions (i.e., searching for a product, viewing a product, and adding product to a cart) performed by a user for purchasing a product. For the activity “LOYALTY ENQUIRY”, the sequence of labels can be associated with a set of actions (i.e., checking the loyalty program web page, and further checking a web page relating to loyalty points) performed by a user for enquiring information about a loyalty program, and more specifically enquiring about the benefits associated with a loyalty points value. For the activity “BAGGAGE LIMITS ENQUIRY”, the sequence of labels can be associated with a set of actions (i.e., checking an airline’s web page, checking a baggage information web page about the airline, and further checking the baggage limits of the airline) performed by a user for enquiring about the baggage limits for a specific airline.

[0037] In some embodiments, even a partial sequence of actions can be mapped to at least one user activity. Stated differently, even though the sequence of actions of a user do not fully map to an activity (due to one or more missing actions / labels), if the sequence of the user’s actions are a part of an activity, i.e., map to a partial sequence of actions for said activity, then the activity identification module 1148 may still map this sequence of user actions to the identified activity. For instance, assume a case where a single web page can lead to other internal web pages, and the desired information that the user seeks is in one of the internal web pages. In case the user is unable to locate any of the internal web pages, then the user will not have executed a full sequence of actions (i.e., actions that completely map to an activity). Nonetheless, this sequence of user actions can be mapped partially to an activity that is associated with the internal web pages.

[0038] For example, assuming a web page associated with the baggage information has internal web pages for baggage limits and baggage contents, then for the activities with the labels “BAGGAGE LIMITS ENQUIRY” and “BAGGAGE CONTENTS ENQUIRY” the user may at least execute a sequence of actions, i.e., actions associated with the labels of “AIRLINE WEBPAGE” and “BAGGAGE INFORMATION”, wherein the actionscorrespond to visiting an airline’ s web page and viewing the web page for the baggage information for the airline. Subsequently, if the user cannot find the internal web pages on the web page associated with the airline’s baggage information, the activity identification module 1148 may nonetheless map this sequence of user actions to the activities having the labels of “BAGGAGE LIMITS ENQUIRY” and “BAGGAGE CONTENTS ENQUIRY” because said sequence of user actions forms a partial sequence of user actions required for the activities having the aforementioned labels.

[0039] The identified (or mapped) activity, or a combination of identified activities can provide context behind a user’s actions. For example, if the activities “BAGGAGE LIMITS ENQUIRY” and “BOOK FLIGHT” occur sequentially, , the context that can be recognized from this set of actions is that the user has booked a flight and the user may be a heavy packer, and accordingly has booked a flight that meets the user’s baggage requirements.

[0040] Essentially, the output of the activity identification module 1148 (more specifically, the output of the ACRE 114) may be an identified activity, pertaining to a past activity of the user, wherein the identified activity is indicative of a context associated with the user. In some instances, the output of the activity identification module 1148 can also be influenced by the user action metadata module 1144. For example, if a user spends very less time on a web page having a lot of information (i.e., time spent on a web page is less than a threshold), it is likely that the user accidentally visited that web page (i.e., the user did not intend to visit that web page or it is not the intent or a part of the intended activity of the user), and therefore the labels associated with the action of accidentally visiting said web page may not be considered by the activity identification module 1148 for mapping to an activity.

[0041] The activity identification module 1148 can be implemented by at least one of: a rules block and a first machine learning model. The rules block can maintain a set of rules for mapping the sequence of labels to an activity. The rules block may function akin to the table 2, and output the unit of activity based on the sequence of labels. An administratorof the WAS 10 may be responsible for implementing the logic for mapping a specific sequence of labels to a unit of activity. In some embodiments, the rules block may output a list of possible activities associated with the sequence of labels.

[0042] In an embodiment where the first machine learning model predicts the unit of activity, the first machine learning model can be trained based on clickstream data. The clickstream data can involve historical data of past interactions of the user with one or more web pages. Consequentially, this can also include the labels associated with each interaction, and an activity associated with a specific sequence of labels. This type of training of the first machine learning model can involve supervised learning (i.e., indicating that a sequence of labels corresponds to an activity). However, other embodiments may utilize other training techniques like unsupervised learning.

[0043] Prior to feeding the training data of clickstream data to the first machine learning model for its training, the clickstream data may undergo preprocessing. The preprocessing can involve transforming the clickstream data into a suitable format, and cleaning the data to (i) remove any irrelevant or erroneous entries and (ii) normalizing numerical data.

[0044] Upon training the first machine learning model, it can predict a unit of activity based on an input of the clickstream data of the user. Examples of the first machine learning model include, but are not limited to, recurrent neural networks (RNNs), long short-term memory (LSTM) networks, or transformer-based models. In some embodiments, the output of the first machine learning model can be a weighted list of possible activities.

[0045] In embodiments where the activity identification module 1148 is implemented by the rules block and the first machine learning model, the weighted list from the rules block and the first machine learning model are combined, wherein the combined list is analyzed, and the activity with the highest weightage (i.e., the activity with the highest likelihood) is input to the CLE 112.

[0046] The identified activity may be stored in a data structure, which captures the user’s past web page clicks, transactions, and other user interactions, along a timeline.Detailed Overview Of The Core Language Engine

[0047] FIG. 3B illustrates the various modules within the CLE 112 of the NLPU 11, according to an embodiment of the present disclosure. The CLE 112 comprises the following modules: language detection and translation module 1121, input preprocessing module 1122, intent discerning module 1123, entity extraction module 1124, and response generation module 1125.

[0048] The language detection and translation module 1121 may detect the language of the user’s input. If the user’s input is in a foreign language, the module 1121 may also convert the user’s input to another language (e.g., English) for further processing (e.g., intent discerning or entity extraction). For example, if the user’s utterance is in Spanish, then upon detecting that the user’s utterance is in Spanish, the module 1121 may convert the user’s utterance to English.

[0049] In one embodiment, the input preprocessing module 1122 takes the following as an input: (i) the user’s query; (ii) a list of all (possible / available) intents that could be attributed to any user query; and (iii) a description associated with each of intent that could be attributed to any user query. The list of all intents may be prepared based on the various possible / available services offered by the WAS 10. In case a previous activity of the user is available (e.g., the user is a repeated user of the WAS 10), the input preprocessing module 1122 also takes as input the previous activity of the user.

[0050] The input preprocessing module 1122, which may utilize a vector embedding model, may perform the following preprocessing of the inputs:1) Step 1: The description of all possible user intents that are available (on the WAS 10) are chunked into logical blocks. A logical block is considered to convey useful information. For example, the if list of intents include “buying a product,” “buying / cancelling an airflight ticket,” “opening / closing a bank account,” making an appointment” etc., then the description for each of the aforementioned intents is chunked into logical blocks.2) Step 2: Each logical block is vectorized into a mathematical representation, using a vector embedding model. The mathematical representation captures the semantic meaning of the chunk. Once the chunk is vectorized, the same could be leveraged for different mathematical operations (e.g., finding similar vectors, finding vectors in similar direction etc.)3) Step 3: The user query is also vectorized using the vector embedding model. In some embodiments, the identified activity of the user may also be vectorized by the vector embedding model.4) Step 4: The mathematical vectors (of the description of the intents) are scanned to find the ones that are close to the vectorized user query. This can be achieved by calculating the cosine similarity of the mathematical vectors (of the description of intents) to the vectorized user query. Based on this, a list of the top vectors (i.e., intent description vectors that are closest in distance to the user query vector) are identified, wherein the top intent description vectors are indicative of the closest intents matching the user’s query.

[0051] In some embodiments, the details of all of the possible intents and their respective description can be stored in a database / table, which can be retrieved by the input preprocessing module 1122. Similarly, the description of each intent may be previously vectorized (by the input preprocessing module 1122) and stored in the database / table, wherein the input preprocessing module 1122 retrieves the vectorized intent description upon receiving the user query. The terms “intent description vector” and “vectorized intent description” may be used interchangeably. Similarly, the terms “user query vector” and “vectorized user query” may be used interchangeably.

[0052] The intent discerning module 1123 can receive its input from the language detection and translation module 1121 (the input being the user query) and the input preprocessing module 1122 (the input being the top intents and their corresponding description). The intent discerning module 1123 may also accept as input a previous activity from the user, which may be indicative of a context surrounding the user, from the ACRE 114. The closest intents (based on the top intent description vectors) from the input preprocessingmodule 1122may be those intents that are deemed to be closest to the user query based on the distance between the vectorized intent description and the vectorized user query being smaller than a threshold. In order to guide the intent discerning module 1123 in discerning the user intent, the intent discerning module 1123 may also receive a set of instructions that define how the intent determination should occur.

[0053] Essentially, the intent discerning module 1123 may receive the following as inputs: (i) the user query; (ii) the closest intents (i.e., intents corresponding to the top intent description vectors from the input preprocessing module 1122); (iii) a set of instructions for determining the intent associated with the user query; and optionally (iv) the past activity of the user. Based on these inputs, the user intent can be determined, wherein the intent discerning module 1123 may utilize a large language model (LLM) (an example of a machine learning model) for determining the user intent associated with the user query. The LLM utilized by the intent discerning module 1123 may be a previously trained LLM that can determine the user input using its natural language processing capabilities.

[0054] By performing the intent determination in this manner, the embodiments herein solve two technical problems pertaining to usage of LLMs.

[0055] Firstly, the information on which an LLM is trained may not be accurate (due to the fact that the LLMs may be trained on outdated information). To overcome this limitation, the embodiments herein utilize Retrieval-Augmented Generation (RAG) and a specific set of instructions that dictate that the LLM should only utilize the information (passed as part of the RAG) for its decision-making. The information passed as part of the RAG includes the closest intents that correspond to the top intent description vectors that match with the vectorized user query.

[0056] Secondly, although RAG techniques make LLMs more accurate (because the information passed to the LLM is updated), there are limitations on the number of characters that could be passed onto the LLM. Depending on the LLM, for example an LLM hosted as a service, the cost is directly proportional to the number of tokens or words passed into the LLM and returned by the LLM. In order to make this more resource-efficient, accurate,and cost-effective, the inputs passed (as part of the RAG) to the LLM are preprocessed to ensure that the most relevant information is passed in the most concise manner, thereby reducing the processing required by the LLM for intent determination.

[0057] In other words, by controlling the number of intents and the information that the LLM (as part of the intent discerning module 1123) is supposed to consider for intent determination, the LLM utilizes lesser processing power and can output a more accurate result in lesser inference time.

[0058] The entity extraction module 1124 can receive as input the user query (from the language detection and translation module 1121) and use natural language processing techniques to analyze the text of a user’s utterance and identify whether it contains a named entity. A named entity can be a word or phrase that refers to a specific entity, such as a person, organization, location, date, product etc. For example, when a user purchases a ticket and subsequently wishes to cancel it, then by uttering the PNR (Passenger Name Record) number, the entity extraction module 1125 may extract the PNR number from the utterance. The entity extraction module 1124 may also utilize a LLM that is identical to or different from the LLM utilized by the intent discerning module 1123.

[0059] Based on (i) the output of the intent discerning module 1123 (ii) the corresponding determined intent), and (iii) the entity extraction module 1124, the response generation module 1125 (also utilizing an LLM) generates a suitable response to the user’s query using the logical reasoning capabilities of the LLM.

[0060] As previously stated, the inputs (i.e., the prompt) to the LLM can include the user query, the identified previous activity of the user, the closest intents (along with their respective description), and a set of instructions. The respective intent description can also include the answers that may be relevant to the user’s query. For example, if the intent corresponds to “Baggage contents,” then the respective intent description can include (i) an explanation that briefly explains that the intent corresponds to the desire of the user know what items are permissible to be carried in a baggage, and (ii) the permissible baggage contents.

[0061] The following is an example of a prompt that is input to the LLM. Table 3 is illustrative of the instructions that are fed to the LLM. Table 4 is illustrative of the closest intents (determined after the preprocessing performed by the input preprocessing module 1122), wherein each intent can have a description associated with it. Table 5 is illustrative of a previously identified activity of the user, which provides context of the user. Table 6 is illustrative of a user query.Table 3

[0062] The instructions to the LLM have the grounding information, i.e., information dictating how the LLM should behave.Table 4Table 5Table 6

[0063] The output of the LLM, based on the prompt (as per tables 3 to 6) can be illustrated by table 7.Table 7

[0064] In an embodiment where a single LLM is utilized as part of the intent discerning module1123, entity extraction module 1124, and response generation module 1125, the prompt provided to the LLM may include a combination of the content of tables 3-5.

[0065] In embodiments where separate LLMs are utilized for each of the intent discerning module 1123, entity extraction module 1124, and response generation module 1125, the prompts provided to the respective LLMs may vary. For example, for the intent discerning module 1123, the prompt provided to it may include instructions that are limited to onlyidentifying the user intent associated with the user query. In other words, the instructions do not require the LLM (as part of the intent discerning module 1123) to also identify any named entities in the user query. For the entity extraction module, the instructions provided to it may be to only extract the named entities from the user query. In other words, the LLM (as part of the entity extraction module 1124) need not discern the user intent, and therefore would not have to be fed the content of tables 4 and 5. For the response generation module 1125, the instructions provided to it may only be to generate a response based on the identified intent (by the intent discerning module 1123), the description associated with the identified intent, and the extracted entities (by the entity extraction module 1124). As such, the prompt to the LLM (as part of the response generation module 1125) can include a subset of the instructions in table 3 and the identified intent & description from table 4.

[0066] Each web page may be associated with an intent, wherein on determining the intent, the corresponding desired web page details is then input to the GUI Navigator 12 for directing the user to the desired web page having the information requested by the user. The response generated by the response generation module 1125 can be displayed to the user over the conversational interface 110.

[0067] The modules 1121 to 1125, within the CLE 112, may employ advanced natural language processing techniques, machine learning algorithms, and artificial intelligence (Al) models for performing their respective functions.Detailed Overview Of Process Flow Within The Web Application System

[0068] FIG. 4 illustrates a process flow of the various engines within the NLPU 11 and the GUI navigator 12, according to an embodiment of the present disclosure. The process flow has four phases, referenced by the numerals 400, 410, 420, and 430.

[0069] In the phase 400, which is performed by the ACRE 114, at step 402, a series of user clicks / actions are stored within or kept track of by the user action module 1142. At step404, the action-to-label module 1146 maps each user click / action to a predetermined label. When the user performs more than one click, the output of the action-to-label module 1146 is a sequence of labels, including at least two labels. At step 406, the activity identification module 1148 identifies an activity performed by the user based on the sequence of labels. The identified activity (which is a past activity of the user) can be representative of a context associated with the user.

[0070] In the phase 410, which is performed by the conversational interface 110, at step 412, a user query is received by the conversational interface 110, for which the CLE 112 is to analyze and accordingly generate a response. The various modules within the conversational interface 110 can process / transform the user query into a form that is suitable to be accepted by the CLE 112.

[0071] In the phase 420, which is performed by the CLE 112, at step 422, the user query and the identified activity (if available) are analyzed to determine the user’s intent behind the user query. The input preprocessing module 1122 of the CLE 112 receives three inputs: the user query; a list of all possible intents associated with the user query, along with their corresponding description; and if available, the identified user activity. The CLE 112 receives the user query from the conversational interface 110. The CLE 112 may receive the list of all intents (including their description) from a database storing the same The CLE 112 receives the identified user activity from the ACRE 114. As previously stated, the identified user activity is indicative of a context about the user. After converting the user query and the description of each possible intent to a vector, the input preprocessing module 1122 determines those vectors (corresponding to the description of intents) that are closest in distance to the vectorized user query. Based on this, the closest intents to the user’s query can be determined. The output of the input preprocessing module 1122, i.e., the closest intents, may be fed to the intent discerning module 1123, along with the description of the closest intents, the user query, user’s past activity, and a set of instructions for guiding the intent discerning module 1123 to determine the intent associated with the user query.

[0072] In some embodiments, the input preprocessing module 1122 and / or the intent discerning module 1123 may also receive the metadata of the previous user actions from the module 1144, in order to discern the user intent. For instance, if the metadata of the previous user action(s) is indicative of the fact that the identified user activity was performed within a short span of time from when the user input their query via the conversational interface 110, it is likely that the user’s prior identified activity and their query are related. For example, if the user was previously browsing web pages for cancelling a ticket, but was unable to find the cancellation form, the user may then request the WAS 10 to pull up the cancellation form. However, if there is a time difference of days or weeks between the time / date on which the identified activity occurred and when the user input their query, the CLE 112 may provide less weightage to the input of the identified activity because it is less likely that the identified activity is relevant to the user’s query.

[0073] In some embodiments, the input processing module 1122 and / or the intent discerning module 1123 may also consider a user’s (i.e., a logged-in user) chat / conversational history with the WAS 10 for determining additional context surrounding the user’s query.

[0074] The entity extraction module 1124 extracts any named entities present in the user’s query.

[0075] The analysis performed at step 422 by the various modules in the CLE 112 can result in three outputs. A first output, at step 424, can involve determining the intent associated with the user query by the intent discerning module 1123. A second output, at step 426, can involve extracting any named entities present in the user query by the entity extraction module 1124. A third output, at step 428, can involve generating a suitable response, in line with the determined intent, the determined intent’s description, and extracted named entities (i.e., the output at steps 424 and 426), by the response generation module 1125.

[0076] In the phase 430, the first output, i.e., the determined intent of the user query can be mapped to a corresponding web page, whose details (e.g., URL) are passed to the GUI navigator 12, so that at step 432, the GUI navigator 12 can display the appropriate webpage (i.e., the mapped-to web page). For example, for a query such as “What is the baggage limit for flight XYZ?”, the GUI navigator may direct the user to the web pagedisplaying baggage limits for the flight XYZ. The response generated by the CLE 112 may match the information displayed on the navigated-to web page. In other words, the response generated by the CLE 112 and the information displayed on the navigated-to web page may corroborate each other. For the response generated by the CLE 112, the information to be included in the response can be obtained from the description of the relevant intent. In other words, not only can each intent be mapped to a specific webpage, but the description of each intent can include the content of the webpage that the intent is mapped to. Hence, the intent description portion of the prompt can include the answers (e.g., the baggage limits) to the user’s query.

[0077] In some embodiments, the second output, i.e., the extracted named entities from the user query, can be passed to the GUI navigator 12 so that at step 434, any input fields (e.g., text fields, checkboxes, or radio buttons) present on any forms of the web page can be pre-filled based on the extracted named entities.

[0078] In the case of a logged-in user, who has previously entered their data or whose data is already collected in the course of the execution of the identified activity, the web page displayed by the GUI navigator 12 can also have some entries on it pre-filled, based on the data provided by the user, and the named entities extracted by the CLE 112 from the user input. In other words, the GUI navigator 12 may also receive as input the identified activity from the ACRE 114 along with data that was collected during the course of the identified activity or any other past activity. For example, upon determining that the user’s intent is to apply for a credit card, the GUI navigator 12 may redirect the user to a web page with a form for applying for the credit card, wherein the web page may have details of the user (e.g., name, address etc.) already pre-filled at the time of account creation on the WAS 10.

[0079] In another example, assuming in a first interaction with the WAS 10, the user has booked the flight ticket from location A to location B. Upon a user input of “I want to cancel my ticket”, the GUI navigator 12 may redirect the user to a cancellation form for the ticket from location A to location B, with various input fields (e.g., text fields, checkboxes, or radio buttons) filled by the data (e.g., passenger name, contact info etc.) provided by theuser at the time of purchasing the ticket (i.e., the identified activity being “BOOK FLIGHT”) and any named entities (if available, e.g., PNR number) identified from the user query.

[0080] By directing the user to the desired web page, the GUI navigator 12 enables the conservation of resources that would have otherwise been expended by the user device 20 while traversing through multiple web pages in order to land on the desired web page. Similarly, resources that would have been expended by the WAS 10 for displaying various web pages are now conserved and can be utilized for other purposes.

[0081] In some embodiments, the user may also be able to provide feedback (positive or negative) with regards to the output of the CLE 112 and the GUI navigator 12. For example, if the generated response from the CLE 112 was not satisfactory, or the web page displayed is incorrect, the user may provide negative feedback. On a similar note, if the response and the displayed web page are what the user sought out for, a positive feedback may be provided. As the WAS 10 continues to receive more data and feedback, it can continuously learn from the same and improve its understanding of the user intentions and preferences over time. This adaptive learning feature can enable the WAS 10 to provide more accurate and personalized results for each user, thereby enhancing the overall user experience. For instance, if a user provides negative feedback, which can be indicative of the intent discerning module 1123 having determined the incorrect intent, it can be deduced that the actual intent of the user’s query may not have been captured as a possible intent in the first place. In other words, the actual intent of the user may not have been available in the list of all intents corresponding to services provided by the WAS 10. In which case, the WAS 10 may need to be updated to add the new intent to the list of all intents, and continually update this list. The process of adding a new intent can also include creating a new web page that is specific to the new intent.

[0082] The WAS 10 also maintains a past chat / conversation history so that the user can revisit the previous web pages displayed to the user by the GUI navigator 12. This feature may only be available for those users logged into the WAS 10. In some embodiments, the pastchat / conversation history may be leveraged by the intent discerning module 1123 for determining the intent associated with the user’s query.

[0083] FIG. 5 illustrates an example layout of the web page that is displayed by the GUI navigator 12, according to an embodiment of the present disclosure. The layout of the web page can be divided into the following sections: (i) the website header; (ii) the web page content; (iii) user conversation history (over the conversational interface 110); and (iv) user utterance area.

[0084] The website header may display details about the web page that the user is visiting, including any menus (e.g., drop-down menus or layered menus) about the services that are offered. The menus could have clickable links to redirect to the user to other web pages.

[0085] The web page content can include information that the user is seeking. Additionally, the web page content can substantiate (or be substantiated by) the content of the response generated by the CLE 112 with respect to the user query.

[0086] The user conversation history section can allow the user to revisit any past conversations that the user has had with the NLPU 11, thereby obviating the need for the user to re-enter their past input / prompts to access any previously displayed content by the GUI navigator 12, as such avoiding the need for the WAS 10 to process the same query twice.

[0087] The user utterance area is the area where the user can provide their utterance / input to the NLPU 11, based on which the user is navigated to the appropriate web page. In some embodiments, there may be a microphone symbol (in the form of a standardized icon) located near the user utterance area that indicates to the user that they may be able to provide voice input.

[0088] FIG. 6 illustrates a method 600 performed by the ACRE 114, according to an embodiment of the present disclosure.

[0089] At step 602, the user actions are tracked across a plurality of pages, by the user action module 1142. In other words, each click performed by a user, that navigates the user from one web page to another, is kept track of by the user action module 1142.

[0090] At step 604, the user action metadata module 604 stores the metadata associated with the user actions across the plurality of web pages. The metadata can include, but is not limited to, the date and time on which the user action has occurred, and also the time spent on a web page.

[0091] At step 606, the actions tracked by the user action module 1142 are mapped, by the action- to-label module 1146, to a respective label. When there are multiple actions performed by the user, multiple labels may be generated, where the labels are generated sequentially, i.e., in the order in which the user click has occurred.

[0092] At step 608, an activity identification module 1148 maps a specific sequence of labels to a specific activity. The activity identification module 1148 can be implemented by a rule block and / or a first machine learning model. The mapped-to activity is indicative of a user intent associated with said activity and a context associated with said activity.

[0093] FIG. 7 illustrates a method 700 performed by the CLE 112, according to an embodiment of the present disclosure.

[0094] At step 702, the CLE 112 receives a user query via the conversational interface 110. The user query may be a processed user query. In other words, the user query fed to the CLE 112 may undergo changes such as voice-to-text transformation (in case the user query was input by the user device 20 via a voice command), spell check etc.

[0095] At step 704, the CLE 112 also receives at least one identified activity of the user that has occurred prior to the user inputting their query to the WAS 10.

[0096] At step 706, the CLE 112 analyses the user query. The analysis comprises vectorizing a list having the description of all possible intents that can be mapped to a service offered by the WAS 10 or an activity performed by the user via the WAS 10. The analysis alsocomprises vectorizing the user query. Based on the cosine similarity between the intent vectors and the user query, those intent description vectors which are closest in distance to the user query can be determined. In other words, the cosine values indicate those intent description vectors that are closest to the user query vector (based on a threshold), and the closest intent description vectors may relate to those intents that are likely to be the intent associated with the user query.

[0097] At step 708, upon receiving the user query, the identified activity, the closest intents, and a set of instructions based on which the intent should be determined, the CLE 112 outputs at least one of the following: (i) an intent associated with the user query; (ii) named entities extracted from the user query; and (iii) a response to the user query. The response to the user query may be based on the determined intent associated with the user query and the extracted named entities.

[0098] FIG. 8 illustrates a method 800 performed by the GUI navigator 12, according to an embodiment of the present disclosure.

[0099] At step 802, the GUI navigator 12 receives at least one of: (i) the determined intent associated with the user query; and (ii) extracted entities from the user query.

[0100] At step 804, the GUI navigator 12 maps the determined intent associated with the user query to a web page. In some embodiments, the GUI navigator 12 may directly receive the web page details corresponding to the determined intent of the user query. The web page details can include a uniform resource location (URL) of the web page to navigate the user to. Although not shown in the figures, there may be a separate web page mapping module that refers to a table that maps an intent to a web page, wherein the table includes various details of each web page.

[0101] At step 806, the GUI navigator 12 pre-fills one or more input fields on the mapped-to web page with the extracted named entities.

[0102] At step 808, the GUI navigator 12 displays the mapped-to web page to the user.

[0103] In some embodiments, the methods 600, 700, and 800 may comprise further steps not shown and / or may omit certain steps as shown, therefore this should not be construed as limiting the scope of the present disclosure.Technical Advantages

[0104] The following is a non-exhaustive description of the technical advantages achieved by the embodiments disclosed herein. The embodiments herein enable a user to access a desired web page simply by inputting a query, instead of having the user traverse through multiple web pages in order to land onto the desired web page. As such, resources utilized by the user device 20 for traversing through the multiple web pages and resources utilized by the WAS 10 for displaying the multiple web pages can now be conserved, because now only a single web page (i.e., the desired web page) is displayed to the user. Further, the information utilized by the LLM (as part of the intent discerning module 1123 and entity extraction module 1124) is a limited set of information that is fed to the LLM via RAG. As such, not only is this cost-effective to the LLM, but also improves the technical accuracy of the LLM as the information fed to the LLM is updated. Further, resources utilized by the LLM are also conserved, as the LLM does not have to take into consideration information it had learnt during its training, but rather only has to consider the information submitted to the LLM via RAG.

[0105] The following is a non-exhaustive list of practical advantages of the embodiments disclosed herein:• Saving of costs by businesses as they might need lesser number of customer support representatives;• Ease of accessibility to disabled users or elderly users that are either unable to use or struggle with the traditional navigation methods / standard navigational paradigms; and• Ability to seamlessly switch between the “traditional web application navigational paradigm” (i.e., site maps and search-based approach) or “dynamic GUI rendering based on natural language processing that considers user intent.”

[0106] The following is a non-exhaustive list of industries and sectors in which the embodiments disclosed herein may be applicable to: i) Travel and Booking Websites: The embodiments disclosed herein can be applied to travel and booking websites, enabling users to search for flights, hotels, and activities using natural language statements. This can help users find the best deals and make bookings more efficient. ii) Customer Support: The embodiments disclosed herein can be employed to improve customer support services on websites and applications. By understanding the user's intention, the WAS 10 can provide tailored solutions, reducing the need for users to navigate through multiple pages or interact with customer support representatives. iii) E-commerce Platforms: The embodiments disclosed herein can be used to enhance the shopping experience on e-commerce websites by allowing users to interact with the platform using natural language statements. This can help users quickly find products, access relevant information, and complete transactions more efficiently. iv) Online Banking and Financial Services: The embodiments disclosed herein can be applied to online banking and financial services platforms, enabling users to access account information, make transactions, and receive personalized financial advice through a conversational interface (powered by Al). v) Healthcare Portals: The embodiments disclosed herein can be integrated into healthcare portals, allowing patients to access medical records, schedule appointments, and receive personalized health recommendations based on their input. vi) E-learning Platforms: The embodiments disclosed herein can be used in e-learning platforms to provide students with a more intuitive and personalized learning experience. By understanding the user's intention, the WAS 10 can provide relevant study materials, quizzes, and assessments tailored to the student's needs.vii) Job Portals: The embodiments disclosed herein can be integrated into job portals, allowing job seekers to search for job opportunities, submit applications, and receive personalized job recommendations based on their input. viii) Content Management Systems (CMS): The embodiments disclosed herein can be used in CMS platforms to help users create, edit, and manage content more efficiently by providing a tailored GUI based on their specific needs.

[0107] By way of example, rather than limitation, the WAS 10 can operate in the capacity of a server, or a client computer in a client-server network. The server may be a physical or virtual server, and the server may be a web server, an application server, or a cloud server.

[0108] The WAS 10 includes at least one processor 102 and at least one memory 104.

[0109] By way of example, rather than limitation, the at least one processor 102 can include, for example, a microprocessor, a central processing unit, or the like. The at least one processor 110 can also include a special -purpose processor such as an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or the like. The at least one processor 102 is capable of executing a set of programming instructions / computer program code stored in the at least one memory 104 to result in the steps / actions performed by the various components 11 and 12 within the WAS 10, as described herein, along with their functionalities as recited in methods 600, 700, and 800.

[0110] By way of example, rather than limitation, the at least one memory 104 can include volatile memory (e.g., Random Access Memory (RAM)) and / or non-volatile memory (e.g., Read-Only Memory (ROM)). The at least one memory 104 stores the programming instructions / computer program code, that when executed by the at least one processor 102, result in the performance of the steps / actions described herein.

[0111] In the drawings and specification, there have been disclosed exemplary embodiments of the invention. Although specific terms are employed, they are used in a generic and descriptive sense only and not for purposes of limitation. It will be apparent to thosehaving ordinary skill in this art that various modifications and variations may be made to the embodiments disclosed herein, consistent with the present invention, without departing from the spirit and scope of the present invention. Other embodiments consistent with the present invention will become apparent from consideration of the specification and the practice of the description disclosed herein.

Claims

We claim:

1. A system, comprising: a conversational interface, configured to receive a query from a user that is input via a user device; a core language engine (CLE), configured to: receive the query from the conversational interface; and determine an intent associated with the user query; and a graphical user interface (GUI) navigator, configured to navigate the user to a desired web page based on the determined user query intent, wherein, the CLE determines the user query intent by: converting the user query into a mathematical vector, using a vector embedding model; converting a description of each intent, among a plurality of intents, to a mathematical vector, using the vector embedding model; comparing the user query vector with each intent description vector, based on a cosine similarity, to determine the distance between the user query vector and each intent description vector; selecting a subset of the intent description vectors, whose distance from the user query vector is below a threshold; and feeding to a first machine learning model (MLM) the following inputs: the user query; the intent and the description associated with each intent description vector in the subset of intent description vectors; and a set of instructions to be followed for determining the user query intent, wherein the first MLM is trained to determine the user query intent, based on the inputs, and wherein the determined user query intent is one of the intents associated with the subset of intent description vectors.

2. The system as claimed in claim, wherein: the CLE is configured to extract one or more named entities from the user query using the first MLM or a second MLM; and the GUI navigator is configured to pre-fill at least one input field on the desired web page based on the extracted one or more named entities.

3. The system as claimed in claim 1, further comprising: an activity and context recognition engine (ACRE), configured to identify at least one activity performed by the user prior to inputting the query, wherein the at least one activity is indicative of a context associated with the user, wherein the determination of the user query intent, by the first MLM, is further based on the identified at least one activity.

4. The system as claimed in claim 3, wherein the ACRE, utilizing a rule engine, identifies the at least one activity by: tracking a plurality of user actions across a plurality of web pages; mapping each user action, among the plurality of user actions, to a label, to generate at least one sequence of labels; and mapping the at least one sequence label to an activity, wherein the mapped-to activity is the identified at least one activity.

5. The system as claimed in claim 3, wherein the ACRE utilizes a third MLM for identifying the at least one activity, and wherein the trained machine learning model was trained by: receiving clickstream data indicative of past interactions of the user on one or more web pages, wherein each interaction is associated to a label; and receiving an association of a sequence of the labels to a predetermined activity.

6. The system as claimed in claim 2, wherein the first MLM is configured to generate a response to the user query based on the determined user query intent, the description of the determined user query intent, and the one or more named entities.

7. The system as claimed in claim 6, wherein a content of the desired web page substantiates the response generated by the first MLM.

8. The system as claimed in claim 1, wherein the user query is input through a text and / or a voice command.

9. A method, comprising: receiving, via a conversational interface, a query from a user; converting, using a vector embedding model, the user query into a mathematical vector; converting a description of each intent, among a plurality of intents, to a mathematical vector using the vector embedding model; comparing the user query vector with each intent description vector, based on a cosine similarity, to determine the distance between the user query vector and each intent description vector; selecting a subset of the intent description vectors, wherein the subset comprises those intent description vectors that are distanced below a threshold from the user query vector; feeding to a first machine learning model (MLM) the following inputs: the user query; the intent and the description associated with each intent description vector in the subset of intent description vectors; and a set of instructions to be followed for determining the user query intent, wherein the first MLM is trained to determine the user query intent based on the inputs, and wherein the determined user query intent is one of the intents associated with the subset of the intent description vectors; and navigating, by a graphical user interface (GUI) navigator, the user to a desired web page based on the determined user query intent.

10. The method as claimed in claim 9, comprising:extracting, using the first MLM or a second MLM, the one or more named entities from the user query; pre-filling, by the GUI navigator, at least one input field on the desired web page based on the extracted one or more named entities.

11. The method as claimed in claim 9, comprising: identifying, by an activity and context recognition engine (ACRE), at least one activity performed by the user prior to inputting the query, wherein the at least one is indicative of a context associated with the user, wherein the determination of the user query intent, by the first MLM, is further based on the identified at least one activity.

12. The method as claimed in 11, comprising: tracking, by a rule engine in the ACRE, a plurality of user actions across a plurality of web pages; mapping, by the rule engine in the ACRE, each user action, among the plurality of user actions, to a label, to generate at least one sequence of labels; and mapping, by the rule engine in the ACRE, the at least one sequence label to an activity, wherein the mapped-to activity is the identified at least one activity.

13. The method as claimed in claim 11, comprising: training a third MLM in the ACRE for identifying the at least one activity, wherein the training comprises: receiving clickstream data indicative of past interactions of the user on one or more web pages, wherein each interaction is associated to a label; and receiving an association of a sequence of the labels to a predetermined activity.

14. The method as claimed in claim 10, wherein the first MLM is configured to generate a response to the user query based on the determined user query intent, the description of the determined user query intent, and the one or more named entities.

15. The method as claimed in claim 14, wherein a content of the desired web page substantiates the response generated by the first MLM.

16. The method as claimed in claim 9, wherein the user query is input through a text and / or a voice command.

Images