Neural routing apparatus

Abstract

A content-based publish and subscribe system using a neural routing apparatus routes services, events and similar provided by publishers to subscribers. The neural router uses one or more large language models to determine recipient subscribers for services and routes them from publishers to recipient subscribers. The neural router apparatus comprises a virtual memory arrangement for managing large scale subscription sets.

Description

[0001] NEURAL ROUTING APPARATUS

[0002] DESCRIPTION OF BACKGROUND

[0003] The following disclosure relates to data communication technology . More particularly the disclosure relates to machine learning assisted content-based publish / subscribe systems and content routing in complex routing systems .

[0004] Internet-of-Things ( loT ) device is a battery- constrained node with limited computational and storing capabilities . Nevertheless , loT devices are used in a wide range of applications , such as , smart city scenarios , autonomous transportation, industrial control and automati zation systems , and other 4G / 5G / 6G and similar network applications . loT devices enable sensing and monitoring the environment and, due to their limited functions , further transmit data to a processing node using wireless technology, such as Bluetooth Low Energy (BLE ) , LoRaWAN, Wi-Fi , or cellular technologies 4G / 5G / 6G . Devices establish a connection between each other using an appropriate network protocol . loT data protocols comprise several categories . One example of a category is so called publish / subscribe category .

[0005] The publish / subscribe stands out for an asynchronous communication between different components or systems , improved scalability, low latency and simplicity compared to traditional request / response ( sender / receiver ) model . The components of publish / subscribe category includes a publisher, that provide information about the event ; a subscriber, that express the interest about the event ; and a message broker, that is responsible for acquiring, filtering and routing messages between the publisher and the subscriber . The publish / subscribe category supports several types of publishing strategies , such as a content-based strategy .

[0006] Content-based publish / subscribe systems typically filter messages based on subscriber' s interest on the specific attributes in the published data . They disseminate information to a large number of subscribers based on their interests using keywordbased filters . Usually, the subscriber expresses interest in a form of a special query, which needs to have the predetermined format and technical details for accurate match with the service from publisher .

[0007] This beforementioned way of subscribing to the particular content is inefficient and imprecise . There is a need for expressive , i . e . natural , multimodal semantic content routing from publishers of content to subscribers of content . In addition, there is a need for scalable and resource aware system .

[0008] SUMMARY

[0009] A neural routing apparatus is disclosed . In the following disclosure a content-based publish and subscribe system using a neural routing apparatus routes services , events and similar provided by publishers to subscribers . The neural router uses one or more large language models to determine recipient subscribers for services and routes them from publishers to recipient subscribers . The neural router apparatus comprises a virtual memory arrangement for managing large scale subscription sets .

[0010] In a first aspect a router is disclosed . The router comprises of at least one processor configured to execute computer programs , at least one communication interface to receive and transmit data, and at least one memory configured to store data and a computer program code comprising instructions . The router is configured to receive at least one communication indicating the interest from a subscriber, receive at least one communication indicating a service from a publisher, determine a destination for the at least one communication indicating a service based on the received at least one communication indicating an interest , and transmit the at least one communication indicating a service to the determined destination . Here and further, any communication indicating a service may comprise any events , information on the events , advertisements , or other relevant data, that was sent by publisher . Communications indicating the interest is data from a subscriber, where subscribers express their interest on event . When determining the destination, the router is configured to input the received communications to a large language model stored in the at least one memory . A large language model is a type of arti ficial intelligence that process and understand the natural human language . The disclosed apparatus benefits in improved information dissemination efficiency and accuracy by utili zing large language models . It is advantageous in management and processing of large- scale communications in content-based publish / subscribe systems .

[0011] In an embodiment of the first aspect the router is configured to generate a shared vocabulary for incoming communications . Shared vocabulary is a common set of words and phrases for natural language processing .

[0012] Alternatively, in an embodiment of the first aspect the router is configured to use pre-trained word embeddings for incoming communications . Word embeddings are a representation of words being applied in natural language processing .

[0013] In an embodiment of the first aspect the router is configured to convert content of each communication into a numerical vector using the generated vocabulary or pre-trained embeddings , normali ze the vectors to have simi lar magnitudes and store the resulting vectors in the at least one memory . These result in captured semantic meaning of communications and reduced dimensionality, making it easier to process communications further .

[0014] In an embodiment of the first aspect the router is configured to cluster communications received from subscribers in accordance with the generated vocabulary or pre-trained embeddings in order to get a first group of clusters , cluster communications received from publishers in accordance with the generated vocabulary or pre-trained embeddings in order to get a second group of clusters , merge similar clusters from the first and the second group of clusters , and output the merged clusters . This method results in grouped communications received from subscribers and publishers that can be assigned to different large language models .

[0015] In an embodiment of the first aspect the router is configured to monitor the load and latency of at least one large language model used in routing and select at least one large language model for each of the clusters based on at least one of the following : monitored load balancing metrics , monitored latency, domain of the large language model , model memory consumption or model accuracy . By doing so, several goals are achieved, such as efficient utili zation of large language models and minimi zation of false positive and false negative errors . In an embodiment of the first aspect the router is configured to determine a similarity distance threshold and a batch si ze , group services into batches based on the similarity distance to each other and at least one communication indicating an interest from a subscriber, and pack the relevant communications from a subscriber and communications from a publisher in each batch for large language model filtering . In other words , this step performs grouping similar communications from publishers with relevant communications from subscribers for further large language model processing .

[0016] In an embodiment of the first aspect the router is configured to compute service embeddings for each cluster, calculate similarity distance to communications from a subscriber for each cluster, batch similar services based on the similarity distance , and combine batched services and relevant communications from a subscriber into a packed batch . In result , the communications from publishers are matched to clusters of communications from subscribers based on their embedding distance .

[0017] In an embodiment of the first aspect the router is configured to address false positives and negatives in the packed batches . This step outlines a mechanism for handling errors .

[0018] In an embodiment of the first aspect the router is configured to create a dataset of labelled examples and train the system using supervised learning techniques . The labelled examples include communications from a subscriber, communications from a publisher and labels indicating whether they match or not . During training the system learns the mapping between communications and optimi zes its performance .

[0019] In an embodiment of the first aspect the router is configured to determine an appropriate cluster for a new communication from subscriber based on similarity distance , add the new communication from subscriber to the appropriate cluster, and run an optimi zation method for the appropriate cluster . The benefit of described embodiment includes using delta updates to identify and process only a new communication from subscriber instead of re-processing the entire set of communications after each communication addition .

[0020] In an embodiment of the first aspect the router is configured to determine a target communication from a subscriber to be removed from a cluster based on similarity distance , remove the target communication from subscriber from the cluster, and run an optimi zation method for the cluster . The benefit of described embodiment includes using delta updates to identify and process only a target communication from subscriber instead of re-processing the entire set of communications after each communication removal .

[0021] In an embodiment of the first aspect the router is configured router to receive an advertisement communication comprising availability information on at least one service available from a publisher .

[0022] In a second aspect a method for routing a communication is disclosed . It comprises receiving at least one communication indicating an interest from a subscriber, receiving at least communication indicating a service from a publisher, determining a destination for the at least one communication indicating a service based on the received at least one communication indicating an interest , and routing the at least one communication indicating a service to the determined destination . When determining the destination, the router is configured to input the received communications to a large language model stored in at least one memory . The benefit of the disclosed method includes improved information dissemination efficiency and accuracy by utilizing large language models .

[0023] In a third aspect a computer program is disclosed . A computer program comprising instructions which, when executed by a computing device , are further configured to cause the computer device to perform a method according to the second aspect .

[0024] The aspects discussed above provide an efficient arrangement for routing communications from publishers to subscribers . The arrangement is particularly beneficial when the system is a large- scale system and comprises a large number of subscriptions and services . The arrangement discussed in the aspects provides a possibility to adj ust the arrangement so that increased accuracy is used where it is needed and workload is distributed so that the resources are used more efficiently as in the conventional solutions .

[0025] BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The accompanying drawings , which are included to provide a further understanding of the neural routing apparatus and constitute a part of this specification, illustrate embodiments and together with the description help to explain the principles of the neural routing apparatus . In the drawings :

[0027] Fig . 1 is a block diagram of a system comprising a neural routing apparatus ;

[0028] Fig . 2 presents an example of a method for routing communications ;

[0029] Fig . 3 is an example of a flow chart of a method for processing a communication ;

[0030] Fig . 4 is an example of a flow chart of a method for a communication' s embeddings ;

[0031] Fig . 5 is an example of a flow chart of a method for an embedding clustering;

[0032] Fig . 6 is an example of a flow chart of a method for batching and packing communications ;

[0033] Fig . 7 is an example of a flow chart of a method for a communication matching; and

[0034] Fig . 8 illustrates an example of a distributed publish / subscribe system . DETAILED DESCRIPTION

[0035] Reference will now be made in detail to the embodiments , examples of which are illustrated in the accompanying drawings .

[0036] In figure 1 an example of a of a system 100 comprising a neural routing apparatus 102 is disclosed . The apparatus 102 comprises at least one processor 104 and at least one memory 106 . The at least one processor 104 and at least one memory 104 are configured to work together and execute computer program code 110 that is stored to the at least one memory . The at least one memory is also configured to store data relevant to the computer programs , operating systems and similar . The apparatus further comprises a communication interface 112 , which may comprise several network technologies . The communication interface may comprise one or more transceivers configured to transmit and receive data over wired or wireless network technologies .

[0037] In the example of Figure 1 one the neural routing apparatus is configured to receive communications from a subscriber 114 and a publi sher 116 . In the example of figure 1 only one of each is shown, however, in typical operation the neural routing apparatus accepts and receives communications from a plurality of subscribers and publishers .

[0038] The neural routing apparatus 102 comprises an acces s to at least one large language model . The large language model may be stored in the memory of the apparatus or it may be used as a distributed resource so that the at least one memory 106 stores a pointer or access information to at least one large language model resource .

[0039] Publishers aim to publish information on the events , advertisements on the event , or other relevant data, i . e . , any service , that would be delivered to subscriber . Subscribers aim to express their interest on the event and further receive the relevant information from publisher based on their interest . In addition , there is a middleware entity in the publish- subscribe system, called router . Router receives communications indicating a service from publishers and communications indicating an interest from subscribers , process and route them .

[0040] Figure 2 comprises an example of flow chart of a method for routing communications received from publishers and subscribers . In the method the router, such as a neural router or a neural routing apparatus , receives at least one communication from a publisher 202 and a subscriber 204 . The communications from a publisher comprise events , messages , advertisements or similar . The communications from a subscriber comprise interests . Optionally the arrangement may be extended so that advertisements that indicate the availability of a publisher of content are transmitted . The advertisements are used setup state for processing events published under the advertisements . This makes the system more deterministic and helps in the management and scaling of the operations .

[0041] When the communications have been received from both subscribers and publishers , then the method determines a destination to at least one publisher communication . The communications may have been received directly or indirectly . Thus , a communication received from a subscriber or publisher maybe received through another router . The determination of the destination may be done us ing a neural router , such as the neural router of the example of figure 1 . The neural router may use embedding-based clustering and semantic batching to efficiently route incoming messages to their corresponding routing table entries . This involves mapping both the routing table entries and incoming messages to shards based on their embedding distances , which can be computed using techniques such as word2vec or BERT . The large language model then matches the messages to their corresponding routing table entries within each shard . By grouping messages together as semantic batches within a time window the efficiency can be further improved .

[0042] Figure 3 discloses a f low chart of an example method for processing a communication . The method is initiated by creating embeddings 302 for communications indicating a service from publishers and communications indicating an interest from subscribers . Typically all received communications are processed, however, it may be sufficient that only part of the communications are processed before initiating routing according to the principles discussed in this disclosure . This can be achieved by us ing a shared vocabulary or a set of pre-trained word embeddings . Shared vocabulary i s a common set of words and phrases used to facilitate effective communication . Word embeddings is a representation of words in natural language process ing, where words or phrases from the vocabulary are mapped to vectors or real numbers . A more detailed example of an algorithm for creating embeddings 302 is described in figure 4 .

[0043] With the created embeddings the method then clusters the communications received from subscribers and communications received from publishers into groups that can be assigned to one or more different large language models 304 . Each of the clusters may be assigned one or more large language models and they may be shared with other clusters . This clustering is typically done in a way that minimi zes the distance between similar communications .

[0044] After clustering the method involves calibration 306 , to ensure that the clustering is accurate and large language models are utili zed efficiently . In this step , router may monitor the load and latency of at least one large language model used in routing and selects at least one large language model based on load balancing, latency metrics , language model domain and other possible parameters . The large language model may be chosen according to efficiency parameters or focusing on accuracy of the large language model . For example , the large language model may be application specific containing vocabulary relevant to a specific domain, for example material science , 6G telecommunications or similar . Different large language models may have different memory requirements and the geographical location of the large language model used may have an significant effect on latency and data communications efficiency .

[0045] The large language model selection capability enables subscription and service specific large language models to be utilized . The large language model selection enables to optimi ze different Quality of Experience parameters , such as performance , memory, energy consumption, accuracy and so on .

[0046] Additionally, a router may use a large language model for cluster optimization . Optimi zation of communication clusters are based on the following techniques : subsumption and merging . Large language model aggregates communications from publishers and subscribers into a fewer number of communications within a cluster, which results in improved memory utili zation . The algorithm contains identifying similar communications or similar content within communications from publishers or subscribers and combining them into a single combined communication with loss or without loss . Subsumption and merging techniques may result in false positives , but it will not result in false negatives . The optimi zation outcomes in optimi zed content filters and transformations for multi-modal content .

[0047] Once the clustering 304 and calibration 306 are done , the next step comprises batching and packing 308 . It involves determining a similarity distance threshold and a batch si ze . Furthermore , the bathing and packing comprises grouping services into batches based on the similarity distance to each other and at least one communication indicating an interest from a subscriber . Finally, packing the relevant communications from a subscriber and communications from a publisher in the each batch for large language model filtering . Batching can be done using a sliding window approach, where the system processes events in fixed-si ze batches and matches them with one or more communications indicating an interest from a subscriber that have similar embeddings within a certain distance threshold .

[0048] This step can involve an additional optimi zation step, in which communication indicating an interest are aggregated by a large language model based on semantic or logical subsumption and merging operations . The large language model optimi zation step may be calibrated as described in the previous step 306 and a specific large language model may be chosen for each communication indicating an interest type for cover / merge optimi zations . These optimi zations involve also batching and packing in itself .

[0049] This step can also involve paralleli zation of a large language model assignment to clusters based on load balancing and latency metrics . Each large language model can handle multiple clusters , but the number of subscriptions assigned to a large language model should not exceed its capacity . Likewise , each cluster may be assigned a plurality of large language models and not necessarily only one . Then the communications from subscribers and publishers are matched, step 310 . It involves computing service embeddings for each cluster and calculating similarity distance to communications from a subscriber for each cluster . Similar services are batched based on the similarity distance . Batched services and relevant communications from a subscriber are batched into a packed batch .

[0050] The matches are not necessarily always correct . Thus , handling 312 fal se positive and false negatives is important in improving accuracy of the arrangement . It comprises a mechanism for addressing false positive and false negatives . This can be done using human-in-the-loop or active learning techniques , where the system seeks feedback from a human expert to improve its accuracy over time . In some applications it is possible not to perform this step, however, typically the improved accuracy is desired and the step is performed .

[0051] In the arrangement , as in many applications involving large amounts data, load balancing is important when distributing the work to different computing entities . Like the previous step, load balancing 314 is optional , however, often desired as it provides a possibility to distribute the computing tasks more evenly and reduces the time needed for the computations . To ensure that no single large language model is overwhelmed with too many requests , a load balancing algorithm can be used to distribute the workload evenly across all available large language models . This can be done using a round-robin approach or other load balancing techniques .

[0052] Finally, the arrangement discussed in the above may include a step of system training 316 . The system may be trained using real-world data to optimi ze its performance . Thi s can be done using supervised learning techniques , where the system is trained on a dataset of label led examples to learn the mapping between subscriptions and events . This step involves creating a dataset of label led examples and training the system using supervised learning techniques . The labelled examples include communications from a subscriber, communications from a published and a label indicating whether they match or not . The training may be partially supervised or self-supervi sed and it may involve training by a human or another machine learning entity that has been assigned to train the arrangement discussed above .

[0053] In Figures 4 - 7 more detai led examples of some of the steps of the method shown in figure 3 are discussed . The following examples are to be understood j ust an examples and similar functionality may be achieved through other methods that are suitable for the purpose .

[0054] Figure 4 discloses an example of an algorithm for creating embeddings . The method comprises converting content of each communication into a numerical vector using the generated vocabulary or pre-trained embeddings 402 , normali zing the vectors to have similar magnitudes 404 , and storing the resulting vectors in the at least one memory 406 . These result in captured semantic meaning of communications and reduced dimensionality, making it easier to process communications further .

[0055] Figure 5 discloses an example of an algorithm for clustering embeddings . The communications received from subscribers and communications received from publishers are clustered into groups that can be assigned to different large language models . This clustering should be done in a way that minimi zes the distance between similar communications .

[0056] The example of an algorithm for embedding clustering comprises the following steps : clustering communications received from subscribers in accordance with the generated vocabulary or pre-trained embeddings in order to get a first group of clusters 502 , clustering communications received from publishers in accordance with the generated vocabulary or pre-trained embeddings in order to get a second group of clusters 504 , merging simi lar clusters from the first and the second group of clusters 506 , and outputting the merged clusters 508 . The clustering can be performed, for example , by using k-means clustering method . K-means clustering is a popular unsupervised machine learning algorithm used for data clustering . It aims to partition a set of observations into k clusters , where each observation belongs to the cluster with the nearest mean, also known as the cluster center or centroid .

[0057] Figure 6 discloses an example of an algorithm for batching and packing . The method comprises determining a similarity distance threshold and a batch si ze 602 , grouping services into batches based on the similarity distance to each other and at least one communication indicating an interest from a subscriber 604 , and packing the relevant communications from a subscriber and communications from a publisher in the each batch for large language model f iltering 606 . The method may be us ing a sl iding window approach, where the neural router processes services , such as events , in fixed-si ze batches and matches them with subscriptions that have similar embeddings within a certain distance threshold . The neural router may handle instructions that contain a set of subscriptions and associated events or other services to process batches . This structure enables more efficient processing and management of large- scale subscription sets .

[0058] The neural route instruments a large language model for a targeted subset of subscriptions , allowing better customi zation and tailoring of instruction sets to each individual large language model . Instructions are designed and optimi zed based on the capabilities and characteristics of the large language model . This ensures improved performance and accuracy when processing events .

[0059] Figure 7 discloses an example of a matching step . This step involves mapping communication indicating an interest from a subscriber based on their embedding distance to clusters of communications from a publisher . An example of an algorithm for matching is illustrated in Figure 7 . The algorithm comprises computing service embeddings for each cluster 702 , calculating similarity distance to communications from a subscriber for each cluster 704 , batching similar services based on the similarity distance 706 , and combining batched services and relevant communications from a subscriber into a packed batch 708 . This can be performed for multiple clusters and large language models simultaneously using parallel processing techniques .

[0060] Figure 8 discloses an example of a distributed publish / subscribe system . The system comprise two routers 802 and 804 . In the example of figure 8 subscriber devices 806 are connected to a first router 802 and publisher devices 808 are connected to a second router 804 . This , however, is only an example of such a system and typically the system comprises a higher number of routers that are communicating with each other and one router can be connected to subscriber devices and publisher devices .

[0061] The neural routing apparatus discussed in the above handles incremental updates through delta updates , lazy processing, and batch processing techniques , improving efficiency in subscription additions and removals . Instead of reprocessing the entire set after each addition or removal , delta updates are used to identify and process only the affected communications . In case of addition delta update , the router is configured to determine an appropriate cluster for a new communication from subscriber based on similarity distance , add the new communication from subscriber to the appropriate cluster, and run an optimi zation method for the appropriate cluster . In case of removal delta update , the neural routing apparatus is configured to determine a target communication from a subscriber to be removed from a cluster based on similarity distance , remove the target communication from subscriber from the cluster, and run an optimi zation method for the cluster . These methods can significantly reduce processing time and resource usage .

[0062] To minimi ze overhead and improve system' s performance , lazy processing is used . The strategy is to delay cache updates or memory management adj ustments until they are actually needed, rather than updating them immediately after each addition or removal . In other words , when adding or removing a subscription, mark the affected events as "to be processed" instead of processing them immediately . Periodically process all marked subscription clusters in batches .

[0063] In batch processing, handle multiple additions or removal s at once instead of handling them individually to reduce overhead and improve overall system throughput . This could be achieved by accumulating a list of subscriptions to be added or removed in the background, and when the li st reaches a predefined threshold, processing all additions and removals together using delta updates , lazy processing, and batch processing techniques . The above-mentioned method may be implemented as computer software which is executed in a computing device able to communicate with a mobile device . When the software is executed in a computing device it is configured to perform the above described inventive method . The software is embodied on a computer readable medium so that it can be provided to the computing device , such as the neural routing apparatus 100 of figure 1 .

[0064] As stated above , the components of the exemplary embodiments can include computer readable medium or memories for holding instructions programmed according to the teachings of the present inventions and for holding data structures , tables , records , and / or other data described herein . Computer readable medium can include any suitable medium that participates in providing instructions to a processor for execution . Common forms of computer-readable media can include , for example , a floppy disk, a flexible disk, hard disk, magnetic tape , any other suitable magnetic medium, a CD-ROM, CD±R, CD±RW, DVD, DVD-RAM, DVDiRW, DVD±R, HD DVD, HD DVD-R, HD DVD-RW, HD DVD- RAM, Blu-ray Disc, any other suitable optical medium, a RAM, a PROM, an EPROM, a FLASH-EPROM, any other suitable memory chip or cartridge , a carrier wave or any other suitable medium from which a computer can read .

[0065] It is obvious to a person skil led in the art that with the advancement of technology, the basic idea of the neural routing apparatus may be implemented in various ways . The neural routing apparatus and its embodiments are thus not limited to the examples described above ; instead they may vary within the scope of the claims .

Claims

CLAIMS1 . A router comprising : at least one processor configured to execute computer programs ; at least one communication interface to receive and transmit data ; and at least one memory configured to store data and a computer program code comprising instructions which, when executed by the at least one processor, are configured to cause the router to : receive at least one communication indicating an interest from a subscriber ; receive at least one communication indicating a service from a publisher ; determine a destination for the at least one communication indicating a service based on the received at least one communication indicating an interest , wherein when determining the destination, the router is further configured to input the received communications to a large language model stored in the at least one memory; transmit the at least one communication indicating a service to the determined destination .2 . The router according to claim 1 , wherein the instructions which, when executed by the at least one processor, are further configured to cause the router to generate a shared vocabulary for incoming communications .3 . The router according to claim 1 , wherein the instructions which, when executed by the at least one processor, are further configured to cause the router to : use pre-trained word embeddings for incoming communications .4 . The router according to any of preceding claims 1 - 3 , wherein the instructions which, when executed by the at least one processor, are further configured to cause the router to : convert content of each communication into a numerical vector using the generated vocabulary or pre-trained embeddings ; normali ze the vectors to have similar magnitudes ; store the resulting vectors in the at least one memory .5 . The router according to any of preceding claims 1 - 4 , wherein the instructions which, when executed by the at least one processor, are further configured to cause the router to : cluster communications received from subscribers in accordance with the generated vocabulary or pre-trained embeddings in order to get a first group of clusters ; cluster communications received from publishers in accordance with the generated vocabulary or pre-trained embeddings in order to get a second group of clusters ; merge similar clusters from the first and the second group of clusters ; output the merged clusters .6 . The router according to any of preceding claims 1 - 5 , wherein the instructions which, when executed by the at least one processor, are further configured to cause the router to : monitor the load and latency of at least one large language model used in routing; select at least one large language model for each of the clusters based on at least one of the following : monitored load balancing metrics , monitoredlatency, domain of the large language model , model memory consumption or model accuracy .7 . The router according to any of preceding claims 1 - 6 , wherein the instructions which, when executed by the at least one processor, are further configured to cause the router to : determine a similarity distance threshold and a batch si ze ; group services into batches based on the similarity distance to each other and at least one communication indicating an interest from a subscriber ; and pack the relevant communications from a subscriber and communications from a publisher in the each batch for large language model filtering .8 . The router according to any of preceding claims 5 - 7 , wherein the instructions which, when executed by the at least one processor, are further configured to cause the router to : compute service embeddings for each cluster ; calculate similarity distance to communications from a subscriber for each cluster ; batch similar services based on the similarity distance ; and combine batched services and relevant communications from a subscriber into a packed batch .9 . The router according to claim 8 , wherein the instructions which, when executed by the at least one processor, are further configured to cause the router to : address false positives and negatives in the packed batches .10 . The router according to any of preceding claims 1 - 9 , wherein the instructions which, when executed by the at least one processor, are further configured to cause the router to :create a dataset of labelled examples , wherein each example comprises communications from a subscriber, communications from a publisher and a label indicating whether they match or not ; train the system using supervised learning techniques to learn the mapping between communications .11 . The router according to any of preceding claims 5 - 10 , wherein the instructions which, when executed by the at least one processor, are further configured to cause the router to : determine an appropriate cluster for a new communication from subscriber based on similarity distance ; add the new communication from subscriber to the appropriate cluster ; run an optimi zation method for the appropriate cluster .12 . The router according to any of preceding claims 5 - 11 , wherein the instructions which, when executed by the at least one processor, are further configured to cause the router to : determine a target communication from a subscriber to be removed from a cluster based on similarity distance ; remove the target communication from subscriber from the cluster ; run an optimi zation method for the cluster .13 . The router according to any of preceding claims 1 - 12 , wherein the instructions which, when executed by the at least one processor, are further configured to cause the router to receive an advertisement communication comprising availability information on at least one service available from a publisher .14 . A method for routing a communication comprising : receiving at least one communication indicating an interest from a subscriber ; receiving at least communication indicating a service from a publisher ; determining a destination for the at least one communication indicating a service based on the received at least one communication indicating an interest , wherein when determining the destination, the router is further configured to input the received communications to a large language model stored in at least one memory; routing the at least one communication indicating a service to the determined destination .15 . A computer program comprising instructions which, when executed by a computing device , are further configured to cause the computer device to perform a method according to claim 14 .

Images