First node, second node and methods performed thereby, for handling a first knowledge graph (KG)

Abstract

A computer-implemented method performed by a first node (111). The method is for handling a first knowledge graph. The first node (111) operates in a communications system (100). The first node (111) determines (307), using an artificial neural network (121), a first knowledge graph representing words in a conversational speech and relations of the words. The first node (111) then initiates (308) outputting a first indication of the determined a first knowledge graph.

Description

[0001] FIRST NODE, SECOND NODE AND METHODS PERFORMED THEREBY, FOR HANDLING A FIRST KNOWLEDGE GRAPH (KG)

[0002] TECHNICAL FIELD

[0003] The present disclosure relates generally to a first node and methods performed thereby for handling a first knowledge graph (KG). The present disclosure further relates generally to a second node and methods performed thereby, for handling the first KG. The present disclosure also relates generally to computer programs and computer-readable storage mediums, having stored thereon the computer programs to carry out these methods.

[0004] BACKGROUND

[0005] Computer systems in a communications network or communications system may comprise one or more nodes. A node may comprise one or more processors which, together with computer program code may perform different functions and actions, a memory, a receiving port, and a sending port. A node may be, for example, a server. Nodes may perform their functions entirely on the cloud.

[0006] Computer systems may be comprised in a telecommunications network. The telecommunications network may cover a geographical area which may be divided into cell areas, each cell area being served by a type of node, a network node in the Radio Access Network (RAN), radio network node or Transmission Point (TP), for example, an access node such as a Base Station (BS), e.g., a Radio Base Station (RBS), which sometimes may be referred to as e.g., gNB, evolved Node B (“eNB”), “eNodeB”, “NodeB”, “B node”, or Base Transceiver Station (BTS), depending on the technology and terminology used. The base stations may be of different classes such as e.g., Wide Area Base Stations, Medium Range Base Stations, Local Area Base Stations and Home Base Stations, based on transmission power and thereby also cell size. A cell may be understood to be the geographical area where radio coverage may be provided by the base station at a base station site. One base station, situated on the base station site, may serve one or several cells. Further, each base station may support one or several communication technologies. The telecommunications network may also comprise network nodes which may serve receiving nodes, such as user equipments, with serving beams.

[0007] The standardization organization Third Generation Partnership Project (3GPP) is currently in the process of specifying a New Radio Interface called Next Generation Radio or New Radio (NR), as well as a Fifth Generation (5G) Packet Core Network, which may be referred to as 5G Core Network (5GC). The advantages of 5G NR may include higher bandwidth, more resources, low latency and network slicing. 5G may provide services to various applications, such as enhanced Mobile Broad Band (eMBB), machine to Machine type communication (mMTC), Ultra Reliable Low Latency Communication (URLLC), etc.

[0008] 5G may be understood to bring in sizeable flexibility with technological advancements along with innovations of cloud and Al. This may be understood to bring a whole new set of opportunities in the enterprise segment.

[0009] For many enterprises, mobile cellular technology has already proven to bring great value to their digitalization process, which may include numerous use cases, such as autonomous robotics, enhanced video services, connected vehicles, remote operations, hazard, and maintenance sensors etc. This may be understood to not only enhance productivity in connected factories, but also make workplaces safer.

[0010] In the course of operations of the telecommunications network, data may be collected via the telecommunications network, which may enable to monitor and manage different functions.

[0011] The advent of for example, the Internet of Things (loT) has exponentially increased the amount of data to be monitored. The availability of large amounts of data, such as those collected for example, from loT devices, may be understood to enable the possibility of analysing such data to make predictions on events, with a high predictive power. To make predictions on events may be understood to refer to building mathematical models that may fit those data, which mathematical models may then be used to make predictions for such events. Within this context, machine learning models may be used to analyze the data collected, and enable an improved management of different types of operations via the telecommunications network.

[0012] Machine Learning

[0013] Machine learning (ML) may be understood as the study of computer algorithms that may improve automatically through experience. It is seen as a part of Artificial Intelligence (Al). ML algorithms may build a model based on sample data, known as "training data", in order to make predictions or decisions without being explicitly programmed to do so. ML algorithms may be used in a wide variety of applications, such as email filtering and computer vision, where it may be difficult or unfeasible to develop conventional algorithms to perform the needed tasks.

[0014] There may be basically 3 types of ML Algorithms: Supervised Learning, Unsupervised Learning, and Reinforcement Learning (RL).

[0015] Supervised Learning algorithms may comprise a target / outcome variable, or dependent variable, which may have to be predicted from a given set of predictors, that is, independent variables. Using this set of variables, a function may be generated that may map inputs to desired outputs. The training process may continue until the model may achieve a desired level of accuracy on the training data. Once an ML model may have been trained, an inference process may begin, whereby new data may be run through the ML model to calculate an output. Examples of Supervised Learning may be Regression, Decision Tree, Random Forest, KNN, Logistic Regression etc.

[0016] In Unsupervised Learning algorithms, there may be no target or outcome variable to predict / estimate. It may be used for clustering a population into different groups, which may be widely used for segmenting customers in different groups for specific intervention. Examples of Unsupervised Learning may be K-means, mean-shift clustering, Density-Based Spatial Clustering of Applications with Noise (DBSCAN), Expectation-Maximization (EM) Clustering using Gaussian Mixture Models (GMM), Agglomerative Hierarchical Clustering, etc....

[0017] Cluster analysis or clustering may be understood as an ML technique which may comprise grouping a set of objects in such a way that objects in the same group, which may be called a cluster, may be understood to be more similar, in some sense, to each other than to those in other groups, that is, other clusters. It may be understood as a main task of exploratory data mining, and a common technique for statistical data analysis, used in many fields, including pattern recognition, image analysis, information retrieval, bioinformatics, data compression, computer graphics and ML.

[0018] Using an RL algorithm, a machine may be trained to make specific decisions. It may be understood to work as follows: the machine may be exposed to an environment where it may train itself continually using trial and error. This machine may learn from past experience and may try to capture the best possible knowledge to make accurate decisions. An example of RL may be a Markov Decision Process (MDP). The training using RL may comprise generating an ML model. To train such an ML model, an agent, given a state of the environment, may take an action in this environment and receive a reward. The action may result in a new state of the environment. This process may be repeated in a loop. Over time, the agent may learn to take actions that may result in larger immediate and future rewards, meaning that it may be understood to be in the best interest of the agent not to take the action that may only lead to the highest reward in the next state, but the action that may cumulatively lead to the highest reward in the next state and in a future number of states.

[0019] The agent may comprise a neural network which may input the state and may produce an action. There may be several ML algorithms that may be used for training the network of the agent, e.g., policy-learning based, such as actor-critic approaches, or value-based learning, such as deep-q networks.

[0020] KGs

[0021] Since the primary mode of human communication is speech signal, information available / transmitted in the form of speech signal is enormous. Hence, it may be understood to be important to develop methods to obtain knowledge graph (KG) automatically from speech signal. Such methods may be understood to have a wide range of applications in semantic communication such as data mining, speech analytics, etc. The existing KG creation algorithms are primarily developed for written text content, which possesses a linguistically well-defined sentence structure. The KG creation algorithms may be understood to exploit such structures present in the text to obtain the KG with reasonable accuracy.

[0022] Traditional KG creation algorithms from text contain different sub-modules such as sentence extraction, co-reference resolution, entity extraction, relation extraction, etc.

[0023] Most of the existing approaches for KG creation, such as those listed in survey paper [2] may be understood to correspond to a first type of KG creation algorithm or KG Type 1, and may be represented through the block diagram depicted in Figure 1. As schematically depicted in the diagram of Figure 1 , such KG Type 1 algorithms may take text as an input at 11. From the text, a sentence extraction module 12 may extract sentences. A co-reference resolution module 13 may then find all mentions of the same entity in the text, even if it may not be explicitly mentioned. In such algorithms, a triplet extraction module 14, of the variant <h,r,t>, may then extract a set of triplets of format (h, r, t). An entity extraction module 15 in the triplet extraction module 14 may extract the two entities, h and t from the text. A relation extraction module 16 in the triplet extraction module 14 may then extract a relation r, wherein each triplet may be understood to link the two entities, h and t through the relation r. The triplets may then be input to the graph by another module 17. A triplet extraction module of the variant <h,r,t> may assume that the input text has grammatical coherence for the task of deriving the correct KG. There may be variants of such triplet representation. Typically, a KG may have verbs, predicates representative of edges of the graph and the subject, objects and other Named Entity Recognitions (NERs) being represented as nodes of the KG. The triplets may be represented in embedded space, also known as KG embeddings (KGE), and this may aid in bringing in semantical aspects of the nodes / relations and graphs from such triplets, that is, it may improve semantic representation.

[0024] Alternately, graph creation may also be possible using Graph Neural Networks (GNNs) [3,4] in a second type of KG creation algorithm, or KG Type 2. In this approach, input data may be considered as a graph, and the output may be understood to be to predict graph level, node level or edge level properties of the graph. The input graph may be constructed using fixed length of input tokens represented either as one-hot vectors or using embedding vectors. GNNs may be also referred to as having “graph-in, graph-out” architecture. The information that may be present in the input may be loaded into the edges and nodes, and the weights of neural networks may be learnt so as to transform this information, directly or through embeddings, without changing directivity of the graph.

[0025] To obtain a KG from text data in a traditional approach, the modules in the traditional KG algorithms, such as those depicted in Figure 1 , may be independently optimized for the corresponding tasks. Any errors produced by the individual submodules propagate to the downstream modules, and result in an error compounding. Hence, existing KG creation algorithms, may result in prediction error.

[0026] SUMMARY

[0027] As part of the development of embodiments herein, one or more problems with the existing technology will first be identified and discussed.

[0028] Existing KG creation algorithms are mostly developed for written text documents where the linguistic structure may be understood to be well defined. On the other hand, conversational speech may be understood to be spontaneous in nature, and it may contain disfluencies. Conversational speech does not possess any easily learnable linguistic structure and makes the creation of KG from conversational text a challenging task.

[0029] In real-life conversations, humans may communicate through speech without following any strict rule and / or ordering of words and / or sentences, and speech communication may contain disfluencies. Hence, text corresponding to a conversational speech may not possess a well-defined sentence structure, such as the one present in the written text documents. Even defining and / or identifying a sentence boundary in conversational text is a challenging task. Constructing an accurate KG using these types of text data is difficult, as sentence extraction, co-reference resolution, and triplet, that is nodes and edges, extraction may be understood to rely on well-defined word relations, e.g., grammar, in sentences and / or text document.

[0030] Moreover, the errors that may be introduced in the speech-to-text conversion further complicate the process. Furthermore, for conversational type of data the error produced by the individual modules of traditional KG creation algorithms, may be understood to be high, and result in error compounding, making the KG creating even more error-prone.

[0031] Embodiments herein may address the problems of the existing methods described by providing a framework to automatically obtain KG from speech signals, where an Automatic Speech Recognition system (ASR) followed by a Neural- Network (NN) module may be used to obtain the KG from speech data.

[0032] According to a first aspect of embodiments herein, the object is achieved by a computer- implemented method, performed by a first node. The method is for handling a first knowledge graph. The first node operates in a computer system. The first node determines, using an artificial neural network, a first KG. The first KG represents words in conversational speech and relations of the words. The determining uses, as input, data comprising the text derived, using speech recognition algorithm, from first data encoding conversational speech. The first node then initiates outputting a first indication of the determined first KG.

[0033] According to a second aspect of embodiments herein, the object is achieved by a computer-implemented method, performed by the second node. The method is for handling a respective first knowledge graph. The second node operates in the computer system. The second node obtains a plurality of sets of first data. Each set of first data encodes respective conversational speech. The second node then obtains, from the plurality of sets of first data, and using the speech recognition algorithm, a plurality of sets of second data. Each set of second data encodes respective text corresponding to respective conversational speech. The second node then obtains a respective second knowledge graph for each set of third data of a plurality of sets of third data. Each set of third data corresponds to a set of second data having been processed to remove errors and unnecessary words. The second node then trains the artificial neural network to determine a respective first knowledge graph for each respective set of second data. The training is performed by minimizing a loss function representing a difference between each respective first knowledge graph and the determined respective second knowledge graph. The training is performed iteratively until the loss function is below a threshold. The second node then initiates outputting a second indication of the trained artificial neural network.

[0034] According to a third aspect of embodiments herein, the object is achieved by the first node. The first node may be understood to be for handling the first knowledge graph. The first node is configured to operate in the computer system. The first node is further configured to determine, using the artificial neural network, the first knowledge graph configured to represent words in the conversational speech and relations of the words. The determining is configured to use, as input, the data configured to comprise text configured to be derived, using the speech recognition algorithm, from the first data configured to encode the conversational speech. The first node is also configured to initiate outputting the first indication of the first knowledge graph configured to be determined.

[0035] According to a fourth aspect of embodiments herein, the object is achieved by the second node. The second node may be understood to be for handling the first knowledge graph. The second node is configured to operate in the computer system. The second node is configured to obtain the plurality of sets of first data. Each set of first data is configured to encode the respective conversational speech. The second node is further configured to obtain, from the plurality of sets of first data, and using the speech recognition algorithm, the plurality of sets of second data. Each set of second data is configured to encode the respective text configured to correspond to the respective conversational speech. The second node is further configured to, obtain the respective second knowledge graph for each set of third data of the plurality of sets of third data. Each set of third data is configured to correspond to the set of second data, processed to remove errors and unnecessary words. The second node is further configured to train the artificial neural network to determine the respective first knowledge graph for each respective set of second data. The training is performed by minimizing the loss function representing the difference between each respective first knowledge graph and the determined respective second knowledge graph. The training is configured to be performed iteratively until the loss function is below the threshold. The second node is further configured to initiate outputting the second indication of the artificial neural network configured to be trained.

[0036] According to a fifth aspect of embodiments herein, the object is achieved by a computer program, comprising instructions which, when executed on at least one processing circuitry, cause the at least one processing circuitry to carry out the method performed by the first node.

[0037] According to a sixth aspect of embodiments herein, the object is achieved by a computer-readable storage medium, having stored thereon the computer program, comprising instructions which, when executed on at least one processing circuitry, cause the at least one processing circuitry to carry out the method performed by the first node.

[0038] According to a seventh aspect of embodiments herein, the object is achieved by a computer program, comprising instructions which, when executed on at least one processing circuitry, cause the at least one processing circuitry to carry out the method performed by the second node.

[0039] According to an eighth aspect of embodiments herein, the object is achieved by a computer-readable storage medium, having stored thereon the computer program, comprising instructions which, when executed on at least one processing circuitry, cause the at least one processing circuitry to carry out the method performed by the second node.

[0040] By obtaining the plurality of sets of first data encoding the respective conversational speeches, the second node is enabled to obtain, using the speech recognition algorithm, the plurality of sets of second data encoding the respective text corresponding to respective conversational speech. The plurality of sets of second data encoding the respective text may then be used as input to the artificial neural network to obtain the first KG.

[0041] By the second node obtaining the respective second knowledge graph for each set of third data, wherein each set of third data corresponds to a set of second data processed to remove errors and unnecessary words, the second node is then enabled to train the artificial neural network to determine the respective first knowledge graph. By the training being performed by minimizing the loss function representing the difference between each respective first knowledge graph and the determined respective second knowledge graph, iteratively, until the loss function is below a threshold, the second node is enabled to obtain the trained artificial neural network, which may automatically obtain an accurate KG from unstructured text data, such as conversational data.

[0042] By the second node then initiates outputting the second indication of the trained artificial neural network, the second node may enable a recipient, such as the first node, to use the trained artificial neural network to accurately predict KGs from unstructured text data, such as conversational data. By determining the first KG using the artificial neural network using as input the data comprising the text derived, using the speech recognition algorithm, from the first data encoding the conversational speech, the first node may be enabled to automatically obtain an accurate KG from unstructured text data, such as conversational data.

[0043] By the first node then initiating outputting the first indication of the determined first KG, the first node enables a receiving party to use the accurate first KG to effectively process the conversational speech in a variety of applications.

[0044] The approach used may be understood to be robust to lack of entities and relations that may be extracted from conversational inputs or their modality (text / speech). Furthermore, the approach may be understood to be agnostic to the grammatical mistakes of the user. The first node may be understood to automatically handle the issues with the speech on misspelled words, if the mis-spelled words are not part of the KG. In fact, in the speech-to- graph creation process, the first node may enable to reduce the effect of certain errors happening in the speech-to-text process. Embodiments herein may be extended to use in different domains related to the establishment of explainable and robust semantic communication. The first node may enable to easily understand domain specific global and local structures present in the given data.

[0045] BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Examples of embodiments herein are described in more detail with reference to the accompanying drawings, according to the following description.

[0047] Figure 1 is a schematic diagram illustrating a conventional knowledge graph (KG) creation process, according to existing methods.

[0048] Figure 2 is a schematic diagram illustrating two non-limiting examples, in panels a) and b), of a computer system, according to embodiments herein.

[0049] Figure 3 is a flowchart depicting a method in a second node, according to embodiments herein.

[0050] Figure 4 is a flowchart depicting a method in a first node, according to embodiments herein. Figure 5 is a schematic diagram depicting a non-limiting example of the method performed by the second node, according to embodiments herein.

[0051] Figure 6 is a schematic diagram depicting particular aspects of a non-limiting example of the method performed by any of the first node and the second node, according to embodiments herein. Figure 7 is a schematic diagram depicting particular aspects of another non-limiting example of the method performed by any of the first node and the second node, according to embodiments herein.

[0052] Figure 8 is a schematic diagram depicting particular aspects of a further non-limiting example of the method performed by any of the first node and the second node, according to embodiments herein.

[0053] Figure 9 is a schematic block diagram illustrating an embodiment of a first node, according to embodiments herein.

[0054] Figure 10 is a schematic block diagram illustrating an embodiment of a second node, according to embodiments herein.

[0055] DETAILED DESCRIPTION

[0056] Certain aspects of the present disclosure and their embodiments address the challenges identified in the Background and Summary sections with the existing methods and provide solutions to the challenges discussed.

[0057] Embodiments herein may be understood to relate to end-to-end neural network based knowledge graph (KG) for semantic communication.

[0058] Embodiments herein may be understood to enable to provide a framework to automatically generate KG from conversational speech data. This may be achieved in embodiments herein via an end-to-end neural network approach, to obtain a text-to-graph module, which may enable KG based semantic communication in a wide variety of problems where human speech may be the primary source of information sharing.

[0059] Some of the embodiments contemplated will now be described more fully hereinafter with reference to the accompanying drawings, in which examples are shown. In this section, the embodiments herein will be illustrated in more detail by a number of exemplary embodiments. Other embodiments, however, are contained within the scope of the subject matter disclosed herein. The disclosed subject matter should not be construed as limited to only the embodiments set forth herein; rather, these embodiments are provided by way of example to convey the scope of the subject matter to those skilled in the art. It should be noted that the exemplary embodiments herein are not mutually exclusive. Components from one embodiment may be tacitly assumed to be present in another embodiment and it will be obvious to a person skilled in the art how those components may be used in the other exemplary embodiments.

[0060] Several embodiments and examples are comprised herein. It should be noted that the embodiments and / or examples herein are not mutually exclusive. Components from one embodiment or example may be tacitly assumed to be present in another embodiment or example and it will be obvious to a person skilled in the art how those components may be used in the other exemplary embodiments and / or examples. Figure 2 depicts two non-limiting examples, in panels “a” and “b”, respectively, of a computer system 100, in which embodiments herein may be implemented. In some example implementations, such as that depicted in the non-limiting examples of Figure 2, the computer system 100 may be a computer network. In other example implementations, such as that depicted in panel b) of Figure 2, the computer system 100 may be implemented in a telecommunications system, sometimes also referred to as a cellular radio system, cellular network or wireless communications system. In some examples, the telecommunications system may comprise network nodes which may serve receiving nodes, such as wireless devices, with serving beams.

[0061] In some examples, the telecommunications system may be, for example, a communications network, such as 5G system, or Next Gen network. The telecommunications system may also, or alternatively, support other technologies, such as LTE, e.g., LTE Frequency Division Duplex (FDD), LTE Time Division Duplex (TDD), LTE Half-Duplex Frequency Division Duplex (HD-FDD), LTE operating in an unlicensed band, Wideband Code Division Multiple Access (WCDMA), Universal Terrestrial Radio Access (UTRA) TDD, Global System for Mobile communication (GSM) / Enhanced Data Rate for GSM Evolution (EDGE) Radio Access Network (GERAN) network, Ultra-Mobile Broadband (UMB), EDGE network, network comprising of any combination of Radio Access Technologies (RATs) such as e.g., Multi-Standard Radio (MSR) base stations, multi-RAT base stations etc., any 3rd Generation Partnership Project (3GPP) cellular network, Wireless Local Area Network / s (WLAN) or WiFi network / s, Worldwide Interoperability for Microwave Access (WiMax), Internet of Things (loT), Machine Type Communication (MTC), IEEE 802.15.4-based low-power short-range networks such as IPv6 over Low-Power Wireless Personal Area Networks (6LowPAN), Zigbee, Z-Wave, Bluetooth Low Energy (BLE), or any cellular network or system.

[0062] The computer system 100 comprises nodes, whereof a first node 111 and a second node 112 are depicted in Figure 2. In some examples, such as that depicted in Figure 2 a), the first node 111 and the second node 112 may be co-located or be the same node. The computer system 100 may comprise additional nodes. As depicted in the non-limiting example of Figure 2 b), the computer system 100 may comprise a third node 113.

[0063] Any of the first node 111 , the second node 112 and the third node 113 may be understood, respectively, as a first computer system or server, a second computer system or server, and a third computer system or server. Any of the first node 111 , the second node 112 and the third node 113 may be implemented as a standalone server in e.g., a host computer in the cloud 115, as depicted in the non-limiting example of Figure 2b). In other examples, any of the first node 111, the second node 112 and the third node 113 may be a distributed node or distributed server, such as a virtual node in the cloud 115, and may perform some of its respective functions locally, e.g., by a client manager, and some of its functions in the cloud 115, by e.g., a server manager. In other examples, any of the first node 111 , the second node 112 and the third node 113 may perform its functions entirely on the cloud 115, or partially, in collaboration or collocated with a radio network node. Yet in other examples, any of the first node 111 , the second node 112 and the third node 113 may also be implemented as processing resources in a server farm. In yet other examples, any of the first node 111 , the second node 112 and the third node 113 may be comprised in a device, such as the device 140 described below, or on the edge. Any of the first node 111 , the second node 112 and the third node 113 may be under the ownership or control of a service provider or may be operated by the service provider, or on behalf of the service provider.

[0064] Any of the first node 111 , and the second node 112 may have a capability to perform machine-implemented learning procedures, which may be also referred to as “machine learning” (ML).

[0065] Any of the first node 111, and the second node 112 may have a capability to manage an artificial neural network 121. The artificial neural network 121 may be understood as a machine learning framework, which may comprise a collection of connected nodes, where in each node or perceptron, there may be an elementary decision unit. Each such node may have one or more inputs and an output. The input to a node may be from the output of another node or from a data source. Each of the nodes and connections may have certain weights or parameters associated with it. In order to solve a decision task, the weights may be learnt or optimized over a data set which may be representative of the decision task. The most commonly used node may have each input separately weighted, and the sum may be passed through a non-linear function which may be known as an activation function. The nature of the connections and the node may determine the type of the neural network, for example a feedforward network, recurrent neural network etc. That any of the first node 111 and the second node 112 may have the capability to manage the artificial neural network 121 may be understood herein as having the capability to store the training data set and the models that may result from the machine learning, to train a new model, and once the model may have been trained, to use this model for prediction. In some embodiments, such as those depicted in Figure 2 b), the system that may be used for training the model and the one used for prediction may be different. That is, the first node 111 may be different from the second node 112. In some embodiments, the first node 111 may be understood as a node having a capability to use the model, once the model may have already been trained. However, in other embodiments, the first node 111 may also have a capability to train the artificial neural network 121.

[0066] Any of the second node 112, and in some embodiments, the first node 111, used for training the artificial neural network 121 may require more computational resources than the first node 111 that may use the trained model to make predictions. Therefore, any of the second node 112, and in some embodiments, the first node 111, used for training the artificial neural network 121 may, for example, support running python / Java with Tensorflow or Pytorch, Theano etc... Any of the first node 111 and the second node 112 may also have GPU capabilities.

[0067] In some embodiments, any of the first node 111 and the second node 112 may be a core network node, such as, e.g., a network data analytics function (NWDAF), a Service management and orchestration (SMO) node, a positioning node, a coordinating node, a Self- Optimizing / Organizing Network (SON) node, a Minimization of Drive Test (MDT) node, etc.... In 5G, for example, any of the first node 111 and the second node 112 may be located in the Operations Support Systems (OSS).

[0068] In other examples not depicted in Figure 2, any of the first node 111 and the second node 112 may be a radio network node, such as radio network node 130, depicted in Figure 2 b). The radio network node 130 may be, e.g., comprised in a Radio Access Network of the telecommunications system. That is, the radio network node 130 may be a transmission point such as a radio base station, for example a gNB, an eNB, or any other network node with similar features capable of serving a wireless device, such as a user equipment or a machine type communication device, in the computer system 100. In typical examples, the radio network node 130 may be a base station, such as a gNB or an eNB. In other examples, the radio network node 130 may be a distributed node, such as a virtual node in the cloud 115, and may perform its functions entirely on the cloud 115, or partially, in collaboration with a radio network node.

[0069] The telecommunications system may cover a geographical area, which in some embodiments may be divided into cell areas, wherein each cell area may be served by a radio network node 130, although, one radio network node 130 may serve one or several cells. In the example of Figure 2, the cells are not depicted to simplify the figure. The radio network node 130 may be of different classes, such as, e.g., macro eNodeB, home eNodeB or pico base station, based on transmission power and thereby also cell size. In some examples, the radio network node 130 may serve receiving nodes with serving beams. The radio network node 130 may be directly connected to one or more core networks.

[0070] Any of the first node 111, the second node 112 and the third node 113, and / or any of the nodes comprised in the computer system 100 may support one or several communication technologies, and its name may depend on the technology and terminology used.

[0071] A device 140 may be comprised in the telecommunication network. The device 140 comprised in the computer system 100 may be a wireless communication device such as a 5G UE, or a UE, which may also be known as e.g., mobile terminal, wireless terminal and / or mobile station, a Customer Premises Equipment (CPE) a mobile telephone, cellular telephone, or laptop with wireless capability, just to mention some further examples. The device 140 comprised in the telecommunications system may be, for example, portable, pocket-storable, hand-held, computer-comprised, or a vehicle-mounted mobile device, enabled to communicate voice and / or data, via the RAN, with another entity, such as a server, a laptop, a Personal Digital Assistant (PDA), or a tablet, Machine-to-Machine (M2M) device, device equipped with a wireless interface, such as a printer or a file storage device, modem, sensor, loT device, or any other radio network unit capable of communicating over a radio link in a communications system. In typical examples, the device 140 may be, or comprise, a microphone. The device 140 comprised in the telecommunications system may be enabled to communicate wirelessly in the telecommunications system. The communication may be performed e.g., via a RAN, and possibly the one or more core networks, which may be comprised within the telecommunications system.

[0072] It may be understood that the telecommunications network 100 may comprise additional radio network nodes 130 and / or additional devices 140.

[0073] The first node 111 may be configured to communicate within the computer system 100 with the second node 112 over a first link 141, e.g., a radio link, or a wired link. The first node 111 may be configured to communicate within the computer system 100 with the radio network node 130 over a second link 142, e.g., a radio link, or a wired link. The second node 112 may be configured to communicate within the computer system 100 with the radio network node 130 over a third link 143, e.g., a radio link, or a wired link. The radio network node 130 may be configured to communicate within the computer system 100 with the device 140 over a fourth link 144, e.g., a radio link. The second node 112 may be configured to communicate within the computer system 100 with the third node 113 over a fifth link 145, e.g., a radio link, or a wired link. The radio network node 130 may be configured to communicate within the computer system 100 with the third node 113 over a sixth link 146, e.g., a radio link, or a wired link. The device 140 may be configured to communicate within the computer system 100 with the third node 113 over a seventh link 147, e.g., a radio link, or a wired link. The first node 111 may be configured to communicate within the computer system 100 with the third node 113 over an eighth link 148, e.g., a radio link, or a wired link.

[0074] Any of the first link 141, the second link 142, the third link 143, the fourth link 144, the fifth link 145, the sixth link 146, the seventh link 147 and the eighth link 148 may be a direct link or may be comprised of a plurality of individual links, wherein it may go via one or more computer systems or one or more core networks in the computer system 100, which are not depicted in Figure 2, or it may go via an optional intermediate network. The intermediate network may be one of, or a combination of more than one of, a public, private or hosted network; the intermediate network, if any, may be a backbone network or the Internet; in particular, the intermediate network may comprise two or more sub-networks, which is not shown in Figure 2.

[0075] In general, the usage of “first”, “second”, “third”, “fourth”, “fifth”, “sixth” and / or “seventh”, herein may be understood to be an arbitrary way to denote different elements or entities, and may be understood to not confer a cumulative or chronological character to the nouns they modify.

[0076] Some of the embodiments contemplated herein will now be described more fully with reference to the accompanying drawings. Other embodiments, however, are contained within the scope of the subject matter disclosed herein, the disclosed subject matter should not be construed as limited to only the embodiments set forth herein; rather, these embodiments are provided by way of example to convey the scope of the subject matter to those skilled in the art.

[0077] Embodiments of a computer-implemented method, performed by the first node 111 , will now be described with reference to the flowchart depicted in Figure 3. The method is for handling a first KG. The first node 111 operates in the computer system 100.

[0078] Several embodiments are comprised herein. In some embodiments all the actions may be performed. In some embodiments, some actions may be optional. In Figure 3, optional actions are indicated with dashed lines. It should be noted that the examples herein are not mutually exclusive. Components from one embodiment may be tacitly assumed to be present in another embodiment and it will be obvious to a person skilled in the art how those components may be used in the other exemplary embodiments. One or more embodiments may be combined, where applicable. All possible combinations are not described to simplify the description.

[0079] Action 301

[0080] Embodiments herein may be understood to provide an approach for KG creation from conversational speech data by using the first artificial neural network 121.

[0081] In this Action 301 , the first node 111 may obtain a plurality of sets of first data. Each set of first data may encode a respective conversational speech.

[0082] The first data may encode recordings of conversations. A set of first data may correspond to the same topic, or the same speaker, or a certain time period, or may have a trait in common, so that the first data within each set may be analyzed together in order to create a respective KG for that first set of data.

[0083] Obtaining in this Action 301 may comprise receiving from another node in the computer system 100, for example, from the device 140, via the radio network node 130, or fetching or retrieving from a memory. Action 302

[0084] In this Action 302, the first node 111 may obtain, from the plurality of sets of first data, and using a speech recognition algorithm, a plurality of sets of second data. Each set of second data may encode respective text corresponding to a respective conversational speech.

[0085] The speech recognition algorithm may be understood to refer to an algorithm for Automatic Speech Recognition (ASR). In other words, in this Action 302, an ASR may be employed to convert the speech signal encoded in the sets of first data into the corresponding text. Any state-of-the-art ASR may be used in this Action 302 to to estimate text automatically from speech [5][6][7],

[0086] By, in this Action 302, obtaining the plurality of sets of second data, the first node 111 may be enabled to then use the obtained plurality of sets of second data as input to the artificial neural network 121 in order to train the artificial neural network 121 to determine a respective first KG for each respective set of second data, as will be described in the next actions.

[0087] Action 303

[0088] The text in any of the sets of second data, corresponding to the speech signal, may have grammatical errors and unnecessary inserted words. As described earlier, since the estimated text obtained in Action 302 may then be used as input to generate a KG, errors present in the ASR output may cascade into the KG creation module. Thus, in a cascaded approach such as that used in existing methods, speech recognition algorithm errors which propagate to the downstream module, that is, a text-to-graph creation module, may be understood to have a compounding effect on the overall error present in a predicted KG.

[0089] In order to mitigate the effect of speech recognition algorithm error propagation, in embodiments herein, a human annotator may use the text created using the speech recognition algorithm from the speech file to create an optimum KG, referred to herein as a “second” KG, from the text, by considering the grammatical mistakes and avoiding the unnecessary inserted words. The second KG may be optimum in the sense that the second KG may be understood to need to closely match the content in the speech signal while using only the words present in the speech recognition algorithm generated text. For this purpose, while creating the second KG, the annotator may need to refrain from adding new words or corrections to the existing words. The human annotator may then provide the second KG to the third node 113. The processed second data by the annotator, in order to correct the grammatical mistakes and avoid using the unnecessary inserted words may be referred to herein as third data. In accordance with the foregoing, in this Action 303, the first node 111 may obtain a respective second KG for each set of third data of a plurality of sets of third data. Each set of third data may correspond to a set of second data having been processed to remove errors and unnecessary words.

[0090] In some embodiments, at least one of the following may apply. In some embodiments, the respective second KG may be received from the third node 113 operating in the computer system 100, e.g., via the eighth link 148. The respective second KG may be also referred to herein as the “oracle KG”.

[0091] Action 304

[0092] In this Action 304, the first node 111 may train the artificial neural network 121 to determine a respective first KG for each respective set of second data. In other words, the speech recognition algorithm output, may be used as input to a Text-to-Graph generating module. The Text-to-Graph module in embodiments herein may be understood to comprise the artificial neural network 121.

[0093] The artificial neural network 121 has an end-to-end neural network architecture, where the input to the artificial neural network 121 may be understood to be text and the output may be understood to be a corresponding KG. An end-to-end architecture may be understood as a single artificial neural network which may directly predict a KG from input text, and it may be trained to explicitly minimize the KG prediction error. A non-limiting example of the end-to-end neural network architecture for KG prediction according to embodiments herein is depicted in Figure 6.

[0094] In this Action 304, the speech recognition algorithm generated text may be passed through a neural network-based text-to-graph module to obtain a respective first KG, where input may be understood to be the text data, and the output may be understood to be the corresponding, or respective, first KG.

[0095] More precisely, a first input into the end-to-end neural network architecture may comprise feature vectors corresponding to words in the text data. For example, input x_1 , x_(2 ), . . . , x_M may represent the features corresponding to the M words present in the text document corresponding to a respective set of second data. The features corresponding to the words may be word embeddings. To create input for the training, any standard word embedding technique on the ASR generated text, such as BERT, Word2Vec, etc., may be applied to obtain the word features.

[0096] A first output of the end-to-end neural network architecture may comprise a first adjacency matrix, A, and a second adjacency matrix, D. The first adjacency matrix and the second adjacency matrix, together, may be understood to correspond to a first KG, or a respective first KG. That is, the outputs A and D may be understood to represent graph adjacency matrices of the predicted KG. The first node 111 may uniquely obtain the corresponding first KG from the first adjacency matrix, A, and the second adjacency matrix, D. Both of the first adjacency matrix and the second adjacency matrix may have the same size M xM.

[0097] The first adjacency matrix and the second adjacency matrix may be used to encode the words and the their relations in the first KG, where the nodes corresponding to the words and the edge between two nodes may be understood to represent the word which may define the relation between the nodes, head and tail.

[0098] The first adjacency matrix may encode information of nodes of the first KG, or a respective first KG. The second adjacency matrix may encode the edges between the nodes in the first KG, or a respective first KG.

[0099] The first adjacency matrix, A, may be understood to represent an adjacency matrix of an undirected graph. A may be understood to be symmetric, where words corresponding to the relations may be understood to not be encoded in the matrix. That is, words corresponding to the edges may be understood to not be considered in the triplets. The first adjacency matrix, A, may be understood to represent the connection of node words, but may be understood to not encode any information of words which may represent the relations between these nodes.

[0100] In the first adjacency matrix, a first matrix element Ai, j may be understood to denote a presence of an undirected edge from word / to word j.

[0101] The second adjacency matrix, D, may be understood to be an adjacency matrix of the directed graph, representing the connected nodes, that is, words, in the first KG, or a respective first KG. In the second adjacency matrix, a second element DiJ may be understood to denote a presence of a directed edge from word / to word j. Value one may indicate the presence of edge, and zero may denote the absence of edge.

[0102] Example

[0103] Consider a speech utterance, “eirene has a boat she goes for fishing she catches tuna fish she lives in Israel where tuna is costlier". The corresponding adjacency matrices A and D have a size 20x20. That is, the number of words is M =20. Consider the corresponding example KG given in Figure 6, where the nodes ‘eirene’ and ‘boat’ are related by the word ‘has’, the matrix elements Aeirene bas, , Abas boat and Aboatbasis zero, that is, the relation between ‘eirene’ and ‘boat’, that is, ‘has’, is not being encoded in A. On the other hand, the matrix D, which, as stated earlier, may be understood to correspond to the adjacency matrix of the directed graph, may be used to encode the relation between nodes, where the value one for the elements Deirene hasand Dhas boatmay encode the relation {eirene, has, boat}. From the nonzero elements Aeirene iboat, Aboai eirene, Deirene ias* and Dbas,boat the first node 111 may infer the KG branch {eirene, has, boat}. To represent the complete graph, in D , the elementsDeirene, has ,Dhas, boat , Deirene, lives ,Dlives, IsraelDeirene,goes ,Dgoes, fishing ,Deirene, catches , D catches, tuna ,Dtuna, cos t Her , and °cos t Her, Israel may be understood to be one, and all other elements may be understood to be zero. In A , values of the elements

[0104] Aeirene, boat ,Aboat, eirene ,Afishing, eirene , AeireneiShing ,

[0105] Aeirene, Israel ,Aisrael, eirene ,Aeirene, tuna ,Atuna, eirene ,Atuna, Israel , and A Israel, tuna may be understood to be one and all other elements may be understood to be zero. The first node 111 may construct the corresponding first KG, as depicted in Figure 6, uniquely from A and D.

[0106] In embodiments herein, in order to mitigate the effect of speech recognition algorithm error propagation to the KG module, the first node 111 may use a supervised training strategy for the text-to-graph generation module, end-to-end NN architecture. During training, the plurality of sets of second data, that is, the speech recognition algorithm generated text, may be used as input, as well as the respective second KG for the same respective set of second data, obtained in Action 303 from the speech recognition algorithm text by considering the speech content. For example, if an insertion error occurred at the speech recognition algorithm output, so that the words in the speech recognition algorithm output may not have been present in the corresponding speech signal, while constructing the respective second KG, the insertion error may be ignored by the human annotator and the respective second KG constructed based on the remaining words present in the speech recognition algorithm output, as the output. For example, if ASR generated text has M words, then the corresponding A and D, may both have size M x M. If r th word is an insertion error, then r th column and row of both A and D may be understood to need to be zero. The respective second KG may have been obtained on similar lines as that of the example shown in Figure 6.

[0107] The training in this Action 304 may be performed by minimizing a loss function representing a difference between each respective first KG and the obtained respective second KG in Action 303, corresponding to the same set of second data. The loss function may be as follows: wherein:

[0108] N may be understood to be the number of training samples

[0109] M(n) may be understood to denote the number of words in nth training sample, bce may be understood to denote the binary cross entropy loss, >0, A may may be understood to denote the first adjacency matrix, D may may be understood to denote the second adjacency matrix, and

[0110] A and D may may be understood to correspond to the artificial neural network 121 predictions of A and D, respectively.

[0111] Any standard optimization algorithm such as Adam may be used to obtain the parameters of the artificial neural network 121.

[0112] The training in this Action 304 may be performed iteratively until the loss function may be below a threshold.

[0113] This training strategy may be understood to enable the end-to-end first artificial neural network 121 to learn the patterns present in the speech recognition algorithm errors and reduce the effect of speech recognition algorithm errors while predicting the respective first KG, from speech recognition algorithm generated text data. Since the artificial neural network 121 may be understood to learn the relation between the second data and the corresponding respective first KG through a supervised training, by minimizing the error between the predicted respective first KG and the respective second KG obtained in Action 303 for the same respective set of second data, it may be possible to use the artificial neural network 121 to better exploit the domain specific local, as well as the global patterns / contexts present in the input text data for the prediction of an accurate KG. Hence, the respective first graph generated using the proposed formulation may be understood to be accurate even, in the case of conversational data. By, in this Action 304, training the artificial neural network 121 to determine the respective first KG for each respective set of second data by minimizing the loss function representing the difference between each respective first KG and the obtained respective second KG, the first node 111 may be enabled to directly minimize the error between the obtained respective second KG and the predicted graph through a supervised learning strategy. Since the first node 111 may be understood to perform the direct optimization of the error, this may be understood to automatically enable the artificial neural network 121 to learn all the relevant information for predicting the correct graph, that is, the respective first KG. Since the conversational text is difficult data, because of grammatical errors, presence of disfluencies, ASR induced errors, etc., the artificial neural network 121 may be enabled to learn all the relevant information for the correct prediction. That is, this training strategy may enable the artificial neural network 121 to learn the pattern of the errors produced by the ASR and help to reduce its effect during the inferencing stage.

[0114] Hence, the first node 111 may be enabled to construct a respective first KG from unstructured, conversation speech data, even if it may contain many noisy words, grammatical issues, etc.

[0115] In some embodiments, the first node 111 may refrain from training 304 the artificial neural network 121 , and the trained artificial neural network 121 may be received in a second indication from the second node 121 operating in the computer system 100. Action 305

[0116] Once the first artificial network 121 may have been trained, the first node 111 may use it to make predictions on newly obtained data. Accordingly, in this Action 305, the first node 111 may obtain first data encoding conversational speech. That is, the first node 111 may obtain new conversational, unstructured data, for which the first node 111 may want to obtain a KG.

[0117] The first data obtained in this Action 305 may be understood to be equivalent to the first data obtained in Action 301, but this time it may be understood to be new data. The first data may be obtained, e.g., from the device 140, in a similar way as to how it was obtained in Action 301.

[0118] Action 306

[0119] In this Action 306, the first node 111 may obtain, from the first data, and using the speech recognition algorithm, data comprising text. The data comprising the text may be understood to be second data.

[0120] The first data obtained in this Action 306 may be understood to be equivalent to the second data obtained in Action 302, but this time it may be understood to be new data.

[0121] During the testing phase, that is, during the performance of e.g., Action 306, the first node 111 may apply word embedding over the ASR generated text to obtain input for the KG- predicting artificial neural network 121. Since the artificial neural network 121 may only take numbers as the input, word embeddings may be understood to be suitable representations of words as numerical values.

[0122] Action 307

[0123] In this Action 307, the first node 111 determines, using the artificial neural network 121, the first KG. The first KG represents words in the conversational speech and relations of the words. The determining in this Action 307 uses, as input, the data, that is, the second data, comprising the text derived, using the speech recognition algorithm, from the first data encoding the conversational speech. The first KG determined in this Action 307 may be understood to be equivalent to a respective first KG, as described, before, but for the new first data obtained in Action 305. Moreover, the first KG may have been determined by the first node 111 using the artificial neural network 121 , once the artificial neural network 121 may have been trained.

[0124] Determining may be understood as calculating, estimating, deriving, or similar, or obtaining or receiving from another node. The determining in this Action 307 of the first KG may be performed using the trained artificial neural network 121. During inferencing, the goal may be understood to be to predict A and D given the input word features derived from the second data, where the decision regarding the presence and absence of an edge may be decided by comparing the artificial neural network 121 output value against a predefined threshold value, for example 0.5. According to examples of embodiments herein, the first node 111 may incorporate domain knowledge by a multi-task learning approach through an auxiliary objective function to the KG loss function. Auxiliary tasks may be understood to be associated tasks that may be performed on the same input data. For example, the first node 111 may predict to which domain, for example sports, science, finance, etc., the conversation may belong to.

[0125] In this Action 307, the first node 111 may, in some embodiments, use the artificial neural network 121 it may have trained itself as described in the previous actions. Alternatively, the first node 111 may used the artificial neural network 121 , as trained by the second node 112. In such embodiments, the first node 111 may receive a second indication from the second node 112. The second indication may indicate the trained artificial neural network 121. That is, Actions 301-304 may be understood to be performed optionally by the first node 111.

[0126] By determining the first KG using the artificial neural network 121 in this Action 307, the first node 111 may be enabled to automatically obtain an accurate KG from unstructured text data by exploiting domain specific attributes. The approach used may be understood to be robust to lack of entities and relations that may be extracted from conversational inputs modality (text / speech). Furthermore, the approach may be understood to be agnostic to the grammatical mistakes of the user. The first node 111 may be understood to automatically handle the issues with the speech on mis-spelled words, if the mis-spelled words are not part of the KG. In fact, in the speech-to- graph creation process, the first node 111 may enable to reduce the effect of certain errors happening in the speech-to-text process. Embodiments herein may be extended to use in different domains related to the establishment of explainable and robust semantic communication. The first node 111 may enable to easily understand domain specific global and local structures present in the given data.

[0127] Action 308

[0128] In this Action 308, the first node 111 initiates outputting a first indication of the determined first KG.

[0129] Initiating may be understood as triggering, enabling, beginning, or similar.

[0130] Outputting in this Action 308 may be understood as displaying, e.g., on an interface or sending, for example, via the a link to another node. The first indication may be for example, as depicted in the non-limiting example of Figure 6. In another example, the first indication may be an interface through which the determined first KG may be searchable. For the search, standard searching techniques such as elastic search, etc. may be used.

[0131] By initiating outputting, the second indication in this Action 308, the first node 111 may be enabled to share the outcome of the determination of Action 307 efficiently with another node or among users, enabling to achieve the advantages already explained in relation to Action 307. This will now be illustrated with a few use cases.

[0132] Embodiments herein may have wide variety of use cases, and accordingly, the first indication may be outputted in way or another, based on the use case. Next, different nonlimiting examples of use cases are described to illustrate the usage of embodiments herein.

[0133] The first indication of the determined first KG may be outputted in one of the following.

[0134] Example Use case 1 : Connected Healthcare: Intelligent way to process conversations

[0135] According to a first option, the first indication of the determined first KG may be outputted in a healthcare application. In such embodiments, the conversational speech may describe a condition of a patient.

[0136] Most hospitals have a four-step process to understand the condition of a patient and dispatch an ambulance. As the request may be typically received via a telephone call to the hospital, it may usually a caller agent that may take the details of the request, such as location, seriousness of the situation, e.g., accident, and patient details, subject to the medical condition of the patient, or in case an attendee to the patient may be able to provide. Based on the location, which may be a Global Positioning System (GPS) enabled system, the nearest ambulance may be dispatched to the site. In these steps, one of the most important aspects may be understood to be to understand the condition of the patient from the patient, or from the attendee, who may be in a state of distress or panic. The initial conversation may be understood to be the most important in order to understand the condition of the patient. However, it has been found that in many cases it is very difficult to understand the symptoms explained by the patient or the attendee due to multiple reasons, such as network quality, the ability to explain the real problem, the state the caller may be in, e.g., in a panic, and may not be clear with speech. Many times lingual factors may also play a role in understanding the real picture. Additionally, and with such impediments, the caller may have to repeat the same set of questions / answers, first to the caller agent, and then to a doctor who may then identify which type of ambulance may need to be deployed in this scenario.

[0137] The approach described embodiments herein may now be used to generate KG from the speech, and thereby enable to understand the problem described better. A KG may be generated based on the discussion between patient or patient attender with different persons in the hospital, e.g., front desk employee, supervisor, doctor, etc. The information shared earlier with the emergency responder may be shared with the doctor, instead of the patient explaining, in form of the graph. A doctor or nurse may now be able to access the first indication output according to embodiments herein, and go through this knowledge graph to understand the problem affecting the patient better, which may help to provide a better diagnosis.

[0138] Optionally, to make the communication in an intelligent way, conversation analytics may be performed in sematic-aware networking and at the time of response, any biases may be avoided, and the exact knowledge may be advantageously enabled to be shared with the requested parties.

[0139] Finally, the first node 111 may, according to action 308, share the generated first KG among different parties, which may be understood to facilitate information sharing and enable to provide better treatment. For example, sharing the first indication may enable to setup the equipment in a laboratory in advance, before the arrival of the patient.

[0140] Example Use case 2: Tele-operated Driving: 5G Connected and Automated Driving

[0141] According to a second option, the first indication of the determined first KG may be outputted in a tele-operated driving application. In such embodiments, the conversational speech may describe instructions to a vehicle by a remote driver.

[0142] Tele-operated Driving (ToD) may be understood to be an application that may operate with different configurations and Quality of Service (QoS) levels. ToD may be understood to enable a remote driver to control a vehicle through proper conversations. The environmental conditions as observed by sensors in the vehicle may be transferred to the remote driver as perception data. In such a setup, it may be possible to understand and develop the conversation through sensing through different sensors in a vehicle. The remote driver may use the sensors details and operate the driving either via text inputs or speech input. To realize ToD, data may be exchanged through a cellular network. The appropriate configuration of the perception may be understood to be needed for enabling the remote control of the vehicle at a command center (CC). Different sensor outputs in a vehicle, such as speech texts, may be used to generate a KG regarding the specific issue the vehicle may have encountered, or for the solution of the issue. Embodiments herein may therefore enable to avoid the occurrence of accidents through error-free graph generation process.

[0143] Example Use case 3: Vehicular-to- Everything Communication

[0144] According to a third option, the first indication of the determined first KG may be outputted in a vehicle-to-everything application. In such embodiments, the conversational speech may describe vehicle-related problems between stakeholders.

[0145] The anticipated development of autonomous driving and connected vehicles will bring numerous vehicle-to-everything (V2X) applications, including both safety services and entertainment-related services which may be provided through conversation. Communication in vehicles may be understood to be used in many forms. This may comprise the traditional cellular mobile connectivity utilizing 3GPP technologies 2G, 3G, and 4G LTE. For the past decade, vehicles have been including direct connectivity, known as vehicle-to-everything (V2X). V2X may be understood to enable vehicles to communicate with other road participants, e.g., other vehicles, pedestrians, etc. In such a context, the problem may be understood to be how to establish the conversation intelligently to solve the problems, and also to avoid any biases in the procedure. The first node 111 of embodiments herein may be introduced as an end-to-end module to attend any vehicle related problems on the route through robust semantic communication established between different stakeholders, such as repair shops, payments, required parts, etc. Through search in the KG, an FSO may then get the details that may be necessary to sort out the issue.

[0146] Example Use case 4: FSOs Solving Network Problems in Customer Site

[0147] According to a third option, the first indication of the determined first KG may be outputted in a field service operation application. In such embodiments, the conversational speech may describe a problem to be solved at a customer site.

[0148] In the present servicing industry, a solution document may be usually provided to Field Service Operations (FSOs) to apply the knowledge, that is, the explanation and relatedness of the problem, with their experiences to shorten the life cycle of addressing the issues. These FSOs may not be proficient in English and may have different accents different.

[0149] A search may be based on the KG which may be generated from existing documents and different use cases. To give the better search results to the FSOs, it may be beneficial to construct a KG from the speech, utilizing suitable translators. Since grammatical errors may be made, the KG creation may be understood to need to be immune to the grammatical errors present in the text. Hence, embodiments herein may be used to arrive at the knowledge graph, which may be further used to generate search results.

[0150] Also, some problems may be handled with different FSOs together. In such scenarios, information sharing among FSOs may be understood to be important to solve the problem in finite time. In such a scenario, embodiments herein may also help in solving the problem in earlier time. Whenever a first FSO, or whoever may solve the problem earlier, may input speech signal as input, the speech signal may be translated to KG. This KG may then be stored in a database. A second FSO may use the KG and understand the problem. The end- to-end system of embodiments herein may be completely utilized to enhance the performance of FSOs during field visits. Embodiments herein may enable to solve the problem faster and effectively. Also, the information storage in a KG format may be a vital data to solve future problems through easy search and retrieval methodology. The sequence diagram of a nonlimiting example of the use case 4 is shown later, in Figure 8.

[0151] Example Use case 5: Solving problem at Telecommunication provider sites

[0152] In another use case according to the third option, the customer site may be a telecommunication provider site.

[0153] Typically, the customer may inform the problem to the telecommunication provider engineer, and depending on the type of problem, the engineer may raise a ticket to solve it. During the ticket creation, the telecommunication provider engineer may give a description of the problem and he / she may have more details on the problem.

[0154] In such a scenario, the ticket may be assigned to another engineer, and he / she may read only the description of the problem to understand the issue. The solution may take time, as the description of the problem may not have all the details on what the first engineer attending to the issue may have had. According to embodiments herein, details may be shared between engineers, so that the problem may be solved faster and in an efficient way. In this process, multiple sub-graphs may be connected to arrive at a suitable solution to close the ticket asap. In some examples, the KG may not be visible to a mechanic. The mechanic may only search through the KG to understand the problem and arrive at a solution.

[0155] Embodiments of a computer-implemented method, performed by the second node 112, will now be described with reference to the flowchart depicted in Figure 4. The method is for handling the respective first KG. The second node 112 operates in the computer system 100.

[0156] Several embodiments are comprised herein. In some embodiments all the actions may be performed. In some embodiments, some embodiments of the actions may be optional. It should be noted that the examples herein are not mutually exclusive. Components from one embodiment may be tacitly assumed to be present in another embodiment and it will be obvious to a person skilled in the art how those components may be used in the other exemplary embodiments. One or more embodiments may be combined, where applicable. All possible combinations are not described to simplify the description. The detailed description of some of the following corresponds to the same references provided above, in relation to the actions described for the second node 112, and will thus not be repeated here. For example, the first data may encode recordings of conversations.

[0157] Action 401 , 402, 403 and 404 may be performed by the second node 112 as they have been described for the first node 111 , respectively in Action 301 , 302, 303, and 304. Action 401

[0158] As was mentioned earlier in relation to the description of Figure 1 , the second node 112 may be understood to be a node having a capability to train the artificial neural network 121.

[0159] In this Action 401 , the second node 112 obtains the plurality of sets of first data. Each set of first data encodes the respective conversational speech.

[0160] Action 402

[0161] In this Action 402, the second node 112 obtains, from the plurality of sets of first data, and using the speech recognition algorithm, the plurality of sets of second data. Each set of second data encodes the respective text corresponding to the respective conversational speech.

[0162] Action 403

[0163] In this Action 403, the second node 112 obtains the respective second knowledge graph for each set of third data of the plurality of sets of third data. Each set of third data corresponds to a set of second data having been processed to remove errors and unnecessary words.

[0164] Action 404

[0165] In in this Action 404, the second node 112 trains the artificial neural network 121 to determine the respective first knowledge graph for each respective set of second data. The training in this Action 404 is performed by minimizing the loss function representing the difference between each respective first knowledge graph and the determined respective second knowledge graph. The training in this Action 404 is performed iteratively until the loss function is below the threshold.

[0166] In some embodiments, the artificial neural network 121 may have an end-to-end neural network architecture. The first input into the end-to-end neural network architecture may comprise the feature vectors corresponding to the words in the text data, and the first respective output of the end-to-end neural network architecture may comprise the first respective adjacency matrix and the second respective adjacency matrix. The first respective adjacency matrix and the second respective adjacency matrix, together, may correspond to the respective first knowledge graph. The first respective adjacency matrix may encode the respective information of the respective nodes of the respective first knowledge graph. The second respective adjacency matrix may encode the respective edges between the respective nodes in the respective first knowledge graph.

[0167] In some embodiments, in the first adjacency matrix, the first matrix element Ai, j may denote the presence of the undirected edge from word i to word j, and in the second adjacency matrix, the second element Di J may denote the presence of the directed edge from word i to word j.

[0168] Action 405

[0169] In this Action 405, the second node 112 initiates outputting a second indication of the trained artificial neural network 121.

[0170] In some embodiments, at least one of the following may apply. According to a first option, the respective second knowledge graph may be received from the third node 113 operating in the computer system 100. According to a second option, the initiating in this Action 405 outputting the second indication may comprise sending the second indication to the first node 111 operating in the computer system 100.

[0171] The second indication of the trained artificial neural network 121 may be outputted in one of: i) the healthcare application, and wherein the conversational speech may describe the condition of the patient, ii) the tele-operated driving application, and wherein the conversational speech may describe instructions to the vehicle by the remote driver, iii) the vehicle-to-everything application, and wherein the conversational speech may describe vehicle- related problems between stakeholders, and iv) the field service operation application, and wherein the conversational speech may describe the problem to be solved at the customer site.

[0172] Figure 5 is a schematic diagram depicting a non-limiting example of the end-to-end neural network architecture for KG prediction from text, text-to-graph module, where the first input is the feature vectors corresponding to the words in text document and the first respective output is the adjacency matrix corresponding to the KG. In Figure 6, input x_1 , x_(2 ), . . . , x_M may represent the features corresponding to the M words present in the text document corresponding to a respective set of second data, o may represent the sigmoid function, and the outputs A and D may be understood to represent graph adjacency matrices of the predicted KG, the first KG or the respective first KG. E may be understood to represent the output of a last hidden layer. T may be understood to represent the transpose operation. Z may be understood to represent the output of a last hidden layer. Bi-LSTM may be understood to represent Bi-directional Long-Short Term Memory Network. Bi-GRLI may be understood to represent Bi-directional Gated Recurrent Unit Network. E may be understood to represent the output of a last hidden layer. D may be understood to represent the prediction of D by the artificial neural network 121. A may be understood to represent the prediction of A by the artificial neural network 121. An appropriate word embedding, such as BERT, Word2Vec, etc., may be used to obtain the word features. The second adjacency matrix, D, may be understood to be an adjacency matrix of the directed graph, representing the connected nodes, that is, words, in the first KG, or a respective first KG. The first adjacency matrix, A, may be understood to represent an adjacency matrix of an undirected graph. A may be understood to be symmetric, where words corresponding to the relations may be understood to not be encoded in the matrix. That is, words corresponding to the edges may be understood to not be considered in the triplets. In the second adjacency matrix, a second element D / ,y may be understood to denote a presence of a directed edge from word / to word j. Value one may indicate the presence of edge, and zero may denote the absence of edge. Similarly, in the first adjacency matrix, a first matrix element AiJ may be understood to denote a presence of an undirected edge from word / to word j. The first node 111 may uniquely obtain the corresponding first KG from the first adjacency matrix, A, and the second adjacency matrix, D. Both of the first adjacency matrix and the second adjacency matrix may have the same size M xM. The first adjacency matrix and the second adjacency matrix may be used to encode the words and the their relations in the first KG, where the nodes corresponding to the words and the edge between two nodes may be understood to represent the word which may define the relation between the nodes, head and tail. The matrices A and D may be understood to be derived by the human annotator, and may correspond to the true label (oracle) of the matrices A and D, respectively. During the training, the parameters of the artificial neural network 121 may be optimized to minimize the error between the oracles and the predictions.

[0173] Figure 6 is a schematic diagram depicting a non-limiting example of the KG corresponding to the ASR text, “eirene has a boat she goes for fishing she catches tuna fish she lives in Israel where tuna is costlier". The KG depicted may be understood to correspond to the a speech utterance, “eirene has a boat she goes for fishing she catches tuna fish she lives in Israel where tuna is costlier The corresponding adjacency matrices A and D may be understood to have a size 20x20, as the number of words is M =20. The nodes ‘eirene’ and ‘boat’ are related by the word ‘has’, hence the matrix elements Aeireneihas, AhaSieirene, Ahas,boat and Aboat hasare zero, that is, in A, may be understood to not be encoding the relation between ‘eirene’ and ‘boat’, that is ‘has’. On the other hand, the matrix D, which may be understood to correspond to the adjacency matrix of the directed graph, may be used to encode the relation between nodes, where the value one for the elements Deirene hasand DhaSiboat may be understood to encode the relation {eirene, has, boat}. From the nonzero elements Aeirene iboat , Aboai eirene, Deirene ias* and Dbas,boat the KG branch {eirene, has, boat} may be inferred. To represent the complete graph, in D , the elementsDeirene, has , D has, boat ,Deirene, lives ,Dlives, IsraelDeirene, goes ,Dgoes, fishing , D eirene, catches ,Dcatches, tuna ,Dtuna, cos t Her , and °cos t Her, Israel may be understood to be one and all other elements may be understood to be zero. In A, values of the elementsAeirene, boat ,Aboat, eirene ,Afishing, eirene , Aeireneifishing ,

[0174] Aeirene, Israel ,Aisrael, eirene ,Aeirene, tuna ,Atuna, eirene ,Atuna, israel , and

[0175] AIsrael, tuna may be understood to all be one and all other elements may be understood to be zero. The corresponding KG depicted in Figure 6 may be able to be constructed uniquely from A and D.

[0176] Figure 7 is a schematic diagram depicting a non-limiting example of the cascaded approach to obtain the KG from speech, that may be followed according to embodiments herein, by any of the first node 111 or the second node 112 during the training phase. The cascaded approach to automatically obtain the KG from a speech utterance may involves two major steps. First, speech obtained according to Action 301 may be converted to text, according to Action 302, using an ASR. In a second step, according to Action 304, the ASR output may be understood to be the input to a Text-to-Graph generating module to create a graph using this text data. The Text-to-Graph module may be understood to be an end-to-end neural network architecture, where input to the artificial neural network 121 may be understood to be unstructured text data, conversation type, and the output, according to Action 304 or Action 307, may be understood to be the corresponding respective first KG. The text-to-graph module may be understood to be a neural network architecture which may predict the respective first KG, or the first KG, from the ASR generated text, where A and D correspond to the adjacency matrices, for undirected and directed graphs, which together may uniquely represent the respective first KG, or the first KG. In Figure 7, A and D correspond to the artificial neural network 121 predictions of A and D, respectively. Once the second KG may be ready, then the corresponding A and D , output for the artificial neural network 121 training, may be obtained on similar lines as that of the example shown in Figure 6. During the training of the text-to-graph module, that is, the end-to-end artificial neural network 121 module, the first node 111 may consider the ASR generated text as the input and the respective second KG, corresponding to the speech signal, as the output. This training strategy may help the artificial neural network 121 to learn the pattern of the errors produced by the ASR and to reduce its effect during the inferencing stage. It should be noted that the traditional approaches such as generating triples followed by graph creation may be a competent approach when text input may be a written type of structured text, which may be grammatically correct. However, en embodiments herein, the objective may be to construct a graph from conversation speech data, which may contain many noisy words, grammatical issues, etc., for which direct graph construction method may be good because it may directly minimize the error between the respective second KG and the predicted graph through a supervised learning strategy. Since the first node 111 may perform the direct optimization of the error, this may automatically enable the machine to learn all the relevant information for predicting the correct graph. Since the conversational text is a difficult data, because of grammatical errors, presence of disfluencies, ASR induced errors, etc., the first node 111 may be allowed to learn all the relevant information for the correct prediction.

[0177] Figure 8 is a signalling diagram depicting a non-limiting example of the method performed by the first node 111 during the inference phase, according to embodiments herein. A first FSO, FSO 1 , may record a speech and send it to the first node 111. In accordance with Action 305, an ASR engine 801 comprised in the first node 111 may convert the speech to text in accordance with action 306. The ASR engine 801 may then at 802, send the converted text to the artificial neural network 121, depicted as a “KG model”, already trained to construct the first KG. The artificial neural network 121 may construct the first KG in accordance with Action 307, and then, in accordance with Action 308, send the constructed first KG to a database 803. In this example, the database 803 is comprised in the first node 111. However, the database 803 may be a separate node. Saving the constructed first KG in the database 803 may enable that at 804, a second FSO 805, FSO 2, may send his or her suggestion or request as input in form of text corresponding to the speech to the database 803. The database 803 may then search on the KG at 806 and send the results to the FSO2 at 807. The same process may be repeated at 808, 809 and 810 for a third FSO 811 , FSO 3.

[0178] As a summarized overview of the foregoing, embodiments herein may be understood to provide a new framework which may automatically obtain a KG from speech signals, where an automatic speech recognition system followed by a neural-network module may be used to obtain the KG from speech data.

[0179] Certain embodiments herein may provide one or more of the following technical advantage(s). Embodiments herein may be understood to particularly provide an end-to-end neural-network-based text-to-graph creation module, which may enable to make an accurate KG from unstructured text data by exploiting domain specific attributes. Traditional KG creation algorithms from text may contain different sub-modules, such as sentence extraction, co-reference resolution, entity extraction, relation extraction, etc. To obtain the KG from the text data, in a traditional approach these modules are independently optimized for the corresponding tasks, and it is not explicitly optimizing any cost function directly related to the KG prediction error. Since there is no joint optimization present in the conventional approaches, the errors produced by the individual submodules, propagate to the downstream modules, and result in an error compounding. On the other hand, the end-to-end neural network-based approach may be understood to explicitly minimize the error between the predicted KG, the respective first KG, and the second KG. The end-to-end approach of embodiments herein may be understood to avoid the error compounding present in the traditional approaches, through the direct minimization, and the direct minimization of KG prediction error leads to a more accurate KG. The technique used by any of the first node 111 and the second node 112 may be understood to be different from the first type of KG creation algorithm or KG type 1 , as the first node 111 or the second node 112 may not assume grammatical coherence, which may be violated during conversations. The approach followed by the first node 111 or the second node 112 may be also understood to be different from the second type of KG creation algorithm or KG type 2, as the input text for both training and inferencing may be understood to correspond to raw text, and not graphs as in GNN. In [1], an end-to-end neural model may jointly learn to generate triples and resolve entity types as a multi-label classification task through deep learning neural networks. This approach only predicts the triplets but does not exploit the structural property of the KG for a specific domain, that is, how the triplets may be connected to form the graph. On the other hand, the proposed artificial neural network 121 -based text-to-graph module, which may be understood to directly minimize the prediction error between the predicted KG and the oracle KG, may learn both local and global properties of the graph, that is, how the triplets may be connected to form a graph for the specific domain of interest. Direct minimization of KG error may be understood to lead to better accuracy.

[0180] Another advantage may be understood to be that embodiments herein may be robust to lack of entities and relations that may be extracted from conversational inputs modality (text / speech).

[0181] A further advantage may be understood to be that embodiments herein may be agnostic to the grammatical mistakes of the user.

[0182] An additional advantage may be understood to be that embodiments herein may be agnostic to the domain, making it easier to use, as the artificial neural network 121 may be trained to obtain a supervised graph.

[0183] Yet another advantage may be understood to be that embodiments herein may automatically handle the issues with the speech on mis-spelled words, if the mis-spelled words are not part of the KG.

[0184] A further advantage may be understood to be that embodiments herein may enable to share the knowledge efficiently among users.

[0185] An additional advantage may be understood to be that embodiments herein may be extended to use in different domains related to the establishment of explainable and robust semantic communication. Yet another advantage may be understood to be that embodiments herein may enable to minimize the error between the predicted KG, that is, the first KG, and the oracle KG, that is, the second KG.

[0186] A further advantage may be understood to be that embodiments herein may enable to easily understand domain specific global and local structures present in the given data.

[0187] Yet another advantage may be understood to be that embodiments herein, in the speech-to- graph creation process, may enable to reduce the effect of certain errors happening in the speech-to-text process.

[0188] Figure 9 depicts an example of the arrangement that the first node 111 may comprise to perform the method described in Figure 3 and / or Figures 5-8. The first node 111 may be understood to be for handling the first knowledge graph. The first node 111 is configured to operate in the computer system 100.

[0189] Several embodiments are comprised herein. It should be noted that the examples herein are not mutually exclusive. One or more embodiments may be combined, where applicable. All possible combinations are not described to simplify the description. Components from one embodiment may be tacitly assumed to be present in another embodiment and it will be obvious to a person skilled in the art how those components may be used in the other exemplary embodiments. The detailed description of some of the following corresponds to the same references provided above, in relation to the actions described for the first node 11, and will thus not be repeated here. For example, the first data may be configured to encode recordings of conversations.

[0190] The first node 111 is configured to determine, using the artificial neural network 121, the first knowledge graph configured to represent words in the conversational speech and relations of the words. The determining is configured to use, as input, the data configured to comprise text configured to be derived, using the speech recognition algorithm, from the first data configured to encode the conversational speech.

[0191] The first node 111 is also configured to initiate outputting the first indication of the first knowledge graph configured to be determined.

[0192] In some embodiments, the artificial neural network 121 may be configured to have the end-to-end neural network architecture. The first input into the end-to-end neural network architecture may be configured to comprise the feature vectors corresponding to the words in the text data. The first output of the end-to-end neural network architecture may be configured to comprise the first adjacency matrix and the second adjacency matrix. The first adjacency matrix and the second adjacency matrix, together, may be configured to correspond to the first knowledge graph. The first adjacency matrix may be configured to encode information of nodes of the first knowledge graph. The second adjacency matrix may be configured to encode the edges between the nodes in the first knowledge graph.

[0193] In some embodiments, in the first adjacency matrix, the first matrix element AiJ may be configured to denote the presence of the undirected edge from word i to word j, and in the second adjacency matrix, the second element Di J may be configured to denote the presence of the directed edge from word i to word j.

[0194] In some embodiments, the first node 111 may be further configured to at least one of the following.

[0195] In some embodiments, the first node 111 may be further configured to obtain the first data configured to encode the conversational speech.

[0196] In some embodiments, the first node 111 may be further configured to obtain, from the first data, and using the speech recognition algorithm, the data configured to comprise the text, wherein the data configured to comprise the text may be configured to be second data.

[0197] In some embodiments, the first node 111 may be further configured to at least one of the following.

[0198] In some embodiments, the first node 111 may be further configured to obtain the plurality of sets of first data. Each set of first data may be configured to encode the respective conversational speech.

[0199] In some embodiments, the first node 111 may be further configured to obtain, from the plurality of sets of first data, and using the speech recognition algorithm, the plurality of sets of second data. Each set of second data may be configured to encode respective text configured to correspond to the respective conversational speech.

[0200] In some embodiments, the first node 111 may be obtain the respective second knowledge graph for each set of third data of the plurality of sets of third data. Each set of third data may be configured to correspond to the set of second data, processed to remove errors and unnecessary words.

[0201] In some embodiments, the first node 111 may be further configured to train the artificial neural network 121 to determine the respective first knowledge graph for each respective set of second data. The training may be configured to be performed by minimizing the loss function configured to configured to represent the difference between each respective first knowledge graph and the respective second knowledge graph configured to be obtained. The training may be configured to be performed iteratively until the loss function may be below the threshold. The determining of the first knowledge graph may be configured to be performed using the artificial neural network 121 configured to trained.

[0202] In some embodiments, at least one of the following may apply: a) the respective second knowledge graph may be configured to be received from the third node 113 configured to operate in the computer system 100, b) the first node 111 may be configured to refrain from training the artificial neural network 121 , and the trained artificial neural network 121 may be configured to be received in the second indication from the second node 121 configured to operate in the computer system 100.

[0203] In some embodiments, the first indication of the determined first knowledge graph may be configured to be outputted in one of: i) the healthcare application, and the conversational speech may be configured to describe the condition of the patient, ii) the tele-operated driving application, and the conversational speech may be configured to describe instructions to the vehicle by the remote driver, iii) the vehicle-to-everything application, and the conversational speech may be configured to describe vehicle-related problems between stakeholders, and iv) the field service operation application, and the conversational speech may be configured to describe the problem to be solved at the customer site.

[0204] The embodiments herein in the first node 111 may be implemented through one or more processors, such as a processing circuitry 901 in the first node 111 depicted in Figure 9, together with computer program code for performing the functions and actions of the embodiments herein. A processor, as used herein, may be understood to be a hardware component. The program code mentioned above may also be provided as a computer program product, for instance in the form of a data carrier carrying computer program code for performing the embodiments herein when being loaded into the first node 111. One such carrier may be in the form of a CD ROM disc. It is however feasible with other data carriers such as a memory stick. The computer program code may furthermore be provided as pure program code on a server and downloaded to the first node 111.

[0205] The first node 111 may further comprise a memory 902 comprising one or more memory units. The memory 902 is arranged to be used to store obtained information, store data, configurations, schedulings, and applications etc. to perform the methods herein when being executed in the first node 111.

[0206] In some embodiments, the first node 111 may receive information from, e.g., the second node 112, the third node 113, the radio network node 130, the device 140, and / or another structure in the computer system 100, through a receiving port 903. In some embodiments, the receiving port 903 may be, for example, connected to one or more antennas in first node 111. In other embodiments, the first node 111 may receive information from another structure in the computer system 100 through the receiving port 903. Since the receiving port 903 may be in communication with the processing circuitry 901 , the receiving port 903 may then send the received information to the processing circuitry 901. The receiving port 903 may also be configured to receive other information.

[0207] The processing circuitry 901 in the first node 111 may be further configured to transmit or send information to e.g., the second node 112, the third node 113, the radio network node 130, the device 140, and / or another structure in the computer system 100, through a sending port 904, which may be in communication with the processing circuitry 901 , and the memory 902.

[0208] Those skilled in the art will also appreciate that the units comprised within the first node 111 described above as being configured to perform different actions, may refer to a combination of analog and digital circuits, and / or one or more processors configured with software and / or firmware, e.g., stored in memory, that, when executed by the one or more processors such as the processing circuitry 901 , perform as described above. One or more of these processors, as well as the other digital hardware, may be included in a single Application-Specific Integrated Circuit (ASIC), or several processors and various digital hardware may be distributed among several separate components, whether individually packaged or assembled into a System-on-a-Chip (SoC).

[0209] Also, in some embodiments, the different units comprised within the first node 111 described above as being configured to perform different actions described above may be implemented as one or more applications running on one or more processors such as the processing circuitry 901.

[0210] Thus, the methods according to the embodiments described herein for the first node 111 may be respectively implemented by means of a computer program 905 product, comprising instructions, i.e. , software code portions, which, when executed on at least one processing circuitry 901, cause the at least one processing circuitry 901 to carry out the actions described herein, as performed by the first node 111. The computer program 905 product may be stored on a computer-readable storage medium 906. The computer-readable storage medium 906, having stored thereon the computer program 905, may comprise instructions which, when executed on at least one processing circuitry 901 , cause the at least one processing circuitry 901 to carry out the actions described herein, as performed by the first node 111. In some embodiments, the computer-readable storage medium 906 may be a non-transitory computer- readable storage medium, such as a CD ROM disc, or a memory stick. In other embodiments, the computer program 905 product may be stored on a carrier containing the computer program 905 just described, wherein the carrier is one of an electronic signal, optical signal, radio signal, or the computer-readable storage medium 906, as described above.

[0211] The first node 111 may comprise a communication interface configured to facilitate, or an interface unit to facilitate, communications between the first node 111 and other nodes or devices, e.g. the third node 113, The interface may, for example, include a transceiver configured to transmit and receive radio signals over an air interface in accordance with a suitable standard.

[0212] In other embodiments, the first node 111 may comprise a radio circuitry 907, which may comprise e.g., the receiving port 903 and the sending port 904. The radio circuitry 907 may be configured to set up and maintain at least a wireless connection with the second node 112, the third node 113, the radio network node 130, the device 140, and / or another structure in the computer system 100. Circuitry may be understood herein as a hardware component.

[0213] Hence, embodiments herein also relate to the first node 111 operative to operate in the computer system 100. The first node 111 may comprise the processing circuitry 901 and the memory 902, said memory 902 containing instructions executable by said processing circuitry 901 , whereby the first node 111 is further operative to perform the actions described herein in relation to the first node 111 , e.g., in Figure 3 and / or Figures 5-8.

[0214] Figure 10 depicts an example of the arrangement that the second node 112 may comprise to perform the method described in Figure 4 and / or Figures 5-7. The second node 112 may be understood to be for handling the respective first knowledge graph. The second node 112 is configured to operate in the computer system 100.

[0215] Several embodiments are comprised herein. It should be noted that the examples herein are not mutually exclusive. One or more embodiments may be combined, where applicable. All possible combinations are not described to simplify the description. Components from one embodiment may be tacitly assumed to be present in another embodiment and it will be obvious to a person skilled in the art how those components may be used in the other exemplary embodiments. The detailed description of some of the following corresponds to the same references provided above, in relation to the actions described for the second node 112, and will thus not be repeated here. For example, the first data may be configured to encode recordings of conversations.

[0216] The second node 112 is configured to obtain the plurality of sets of first data. Each set of first data is configured to encode the respective conversational speech.

[0217] The second node 112 is further configured to obtain, from the plurality of sets of first data, and using the speech recognition algorithm, the plurality of sets of second data. Each set of second data is configured to encode the respective text configured to correspond to the respective conversational speech.

[0218] The second node 112 is further configured to, obtain the respective second knowledge graph for each set of third data of the plurality of sets of third data. Each set of third data is configured to correspond to the set of second data, processed to remove errors and unnecessary words.

[0219] The second node 112 is further configured to train the artificial neural network 121 to determine the respective first knowledge graph for each respective set of second data. The training is performed by minimizing the loss function representing the difference between each respective first knowledge graph and the determined respective second knowledge graph. The training is configured to be performed iteratively until the loss function is below the threshold.

[0220] The second node 112 is further configured to initiate outputting the second indication of the artificial neural network 121 configured to be trained.

[0221] In some embodiments, the artificial neural network 121 may be configured to have the end-to-end neural network architecture. The first input into the end-to-end neural network architecture may be configured to comprise the feature vectors configured to correspond to the words in the text data. The first respective output of the end-to-end neural network architecture may be configured to comprise the first respective adjacency matrix and the second respective adjacency matrix. The first respective adjacency matrix and the second respective adjacency matrix, together, may be configured to correspond to the respective first knowledge graph. The first respective adjacency matrix may be configured to encode the respective information of respective nodes of the respective first knowledge graph. The second respective adjacency matrix may be configured to encode the respective edges between the respective nodes in the respective first knowledge graph.

[0222] In some embodiments, in the first adjacency matrix, the first matrix element AiJ may be configured to denote the presence of the undirected edge from word i to word j, and in the second adjacency matrix, the second element Di J may be configured to denote the presence of the directed edge from word i to word j.

[0223] In some embodiments, at least one of the following may apply: a) the respective second knowledge graph may be configured to be received from the third node 113 configured to operate in the computer system 100, b) the initiating outputting the second indication may be configured to comprise sending the second indication to the first node 111 configured to operate in the computer system 100.

[0224] In some embodiments, the second indication of the artificial neural network 121 configured to be trained may be configured to be outputted in one of: i) the healthcare application, and the conversational speech may be configured to describe the condition of the patient, ii) the tele-operated driving application, and the conversational speech may be configured to describe instructions to the vehicle by the remote driver, iii) the vehicle-to- everything application, and the conversational speech may be configured to describe vehicle- related problems between stakeholders, and iv) the field service operation application, and the conversational speech may be configured to describe the problem to be solved at the customer site.

[0225] The embodiments herein in the second node 112 may be implemented through one or more processors, such as a processing circuitry 1001 in the second node 112 depicted in Figure 10, together with computer program code for performing the functions and actions of the embodiments herein. A processor, as used herein, may be understood to be a hardware component. The program code mentioned above may also be provided as a computer program product, for instance in the form of a data carrier carrying computer program code for performing the embodiments herein when being loaded into the second node 112. One such carrier may be in the form of a CD ROM disc. It is however feasible with other data carriers such as a memory stick. The computer program code may furthermore be provided as pure program code on a server and downloaded to the second node 112.

[0226] The second node 112 may further comprise a memory 1002 comprising one or more memory units. The memory 1002 is arranged to be used to store obtained information, store data, configurations, schedulings, and applications etc. to perform the methods herein when being executed in the second node 112.

[0227] In some embodiments, the second node 112 may receive information from, e.g., the first node 111 , the third node 113, the radio network node 130, the device 140, and / or another structure in the computer system 100, through a receiving port 1003. In some embodiments, the receiving port 1003 may be, for example, connected to one or more antennas in second node 112. In other embodiments, the second node 112 may receive information from another structure in the wireless communications network 100 through the receiving port 1003. Since the receiving port 1003 may be in communication with the processing circuitry 1001, the receiving port 1003 may then send the received information to the processing circuitry 1001. The receiving port 1003 may also be configured to receive other information.

[0228] The processing circuitry 1001 in the second node 112 may be further configured to transmit or send information to e.g., the first node 111 , the third node 113, the radio network node 130, the device 140, and / or another structure in the computer system 100, through a sending port 1004, which may be in communication with the processing circuitry 1001 , and the memory 1002.

[0229] Those skilled in the art will also appreciate that the units comprised within the second node 112 described above as being configured to perform different actions, may refer to a combination of analog and digital circuits, and / or one or more processors configured with software and / or firmware, e.g., stored in memory, that, when executed by the one or more processors such as the processing circuitry 1001 , perform as described above. One or more of these processors, as well as the other digital hardware, may be included in a single Application-Specific Integrated Circuit (ASIC), or several processors and various digital hardware may be distributed among several separate components, whether individually packaged or assembled into a System-on-a-Chip (SoC).

[0230] Also, in some embodiments, the different units comprised within the second node 112 described above as being configured to perform different actions described above may be implemented as one or more applications running on one or more processors such as the processing circuitry 1001. Thus, the methods according to the embodiments described herein for the second node 112 may be respectively implemented by means of a computer program 1005 product, comprising instructions, i.e. , software code portions, which, when executed on at least one processing circuitry 1001, cause the at least one processing circuitry 1001 to carry out the actions described herein, as performed by the second node 112. The computer program 1005 product may be stored on a computer-readable storage medium 1006. The computer- readable storage medium 1006, having stored thereon the computer program 1005, may comprise instructions which, when executed on at least one processing circuitry 1001 , cause the at least one processing circuitry 1001 to carry out the actions described herein, as performed by the second node 112. In some embodiments, the computer-readable storage medium 1006 may be a non-transitory computer-readable storage medium, such as a CD ROM disc, or a memory stick. In other embodiments, the computer program 1005 product may be stored on a carrier containing the computer program 1005 just described, wherein the carrier is one of an electronic signal, optical signal, radio signal, or the computer-readable storage medium 1006, as described above.

[0231] The second node 112 may comprise a communication interface configured to facilitate, or an interface unit to facilitate, communications between the second node 112 and other nodes or devices, e.g., the first node 111 , the third node 113, the radio network node 130, the device 140, and / or another structure in the computer system 100. The interface may, for example, include a transceiver configured to transmit and receive radio signals over an air interface in accordance with a suitable standard.

[0232] In other embodiments, the second node 112 may comprise a radio circuitry 1007, which may comprise e.g., the receiving port 1003 and the sending port 1004.

[0233] The radio circuitry 1007 may be configured to set up and maintain at least a wireless connection with the first node 111 , the third node 113, the radio network node 130, the device 140, and / or another structure in the computer system 100. Circuitry may be understood herein as a hardware component.

[0234] Hence, embodiments herein also relate to the second node 112 operative to operate in the wireless communications network 100. The second node 112 may comprise the processing circuitry 1001 and the memory 1002, said memory 1002 containing instructions executable by said processing circuitry 1001 , whereby the second node 112 is further operative to perform the actions described herein in relation to the second node 112, e.g., in Figure 4 and / or Figures 5-7.

[0235] When using the word "comprise" or “comprising”, it shall be interpreted as non- limiting, i.e., meaning "consist at least of'. The embodiments herein are not limited to the above-described preferred embodiments. Various alternatives, modifications and equivalents may be used. Therefore, the above embodiments should not be taken as limiting the scope of the invention.

[0236] Generally, all terms used herein are to be interpreted according to their ordinary meaning in the relevant technical field, unless a different meaning is clearly given and / or is implied from the context in which it is used. All references to a / an / the element, apparatus, component, means, step, etc. are to be interpreted openly as referring to at least one instance of the element, apparatus, component, means, step, etc., unless explicitly stated otherwise. The steps of any methods disclosed herein do not have to be performed in the exact order disclosed, unless a step is explicitly described as following or preceding another step and / or where it is implicit that a step must follow or precede another step. Any feature of any of the embodiments disclosed herein may be applied to any other embodiment, wherever appropriate. Likewise, any advantage of any of the embodiments may apply to any other embodiments, and vice versa. Other objectives, features and advantages of the enclosed embodiments will be apparent from the following description.

[0237] As used herein, the expression “at least one of:” followed by a list of alternatives separated by commas, and wherein the last alternative is preceded by the “and” term, may be understood to mean that only one of the list of alternatives may apply, more than one of the list of alternatives may apply or all of the list of alternatives may apply. This expression may be understood to be equivalent to the expression “at least one of:” followed by a list of alternatives separated by commas, and wherein the last alternative is preceded by the “or” term.

[0238] Any of the terms processor and circuitry may be understood herein as a hardware component.

[0239] As used herein, the expression “in some embodiments” has been used to indicate that the features of the embodiment described may be combined with any other embodiment or example disclosed herein.

[0240] As used herein, the expression “in some examples” has been used to indicate that the features of the example described may be combined with any other embodiment or example disclosed herein.

[0241] REFERENCES

[0242] [1] Michael Stewart and Wei Liu, Seq2KG: An End-to-End Neural Model for Domain Agnostic KG (not Text Graph) Construction from Text, Proceedings of the 17th International Conference on Principles of Knowledge Representation and Reasoning (KR 2020).

[0243] [2] Abu-Salih, B., 2021. Domain-specific KGs: A survey. Journal of Network and Computer Applications, 185, p.103076. [3] Sanchez-Lengeling, B., Reif, E., Pearce, A. and Wiltschko, A.B., 2021. A gentle introduction to graph neural networks. Distill, 6(9), p.e33.

[0244] [4] Joshi, C., 2020. Transformers are graph neural networks. The Gradient, 12.

[0245] [5] Watanabe, S., Hori, T., Karita, S., Hayashi, T., Nishitoba, J., llnno, Y., Enrique Yalta Soplin, N., Heymann, J., Wiesner, M., Chen, N., Renduchintala, A., Ochiai, T. (2018) ESPnet: End-to-End Speech Processing Toolkit. Proc. Interspeech 2018, 2207-2211, doi: 10.21437 / 1 nterspeech.2018-1456

[0246] [6] Yu Zhang, Wei Han, James Qin, Yongqiang Wang, Ankur Bapna, Zhehuai Chen, Nanxin Chen, Bo Li, Vera Axelrod, Gary Wang, Zhong Meng, Ke Hu, Andrew Rosenberg, Rohit Prabhavalkar, Daniel S. Park, Parisa Haghani, Jason Riesa, Ginger Perng, Hagen Soltau, Trevor Strohman, Bhuvana Ramabhadran, Tara Sainath, Pedro Moreno, Chung-Cheng Chiu, Johan Schalkwyk, Frangoise Beaufays, Yonghui Wu, Google USM: Scaling Automatic Speech Recognition Beyond 100 Languages, https: / / doi.org / 10.48550 / arXiv.2303.01037, March 2023. [7] S. Watanabe, T. Hori, S. Kim, J. R. Hershey and T. Hayashi, "Hybrid CTC / attention architecture for end-to-end speech recognition", IEEE J. Sei. Topics Signal Process., vol. 11 , no. 8, pp. 1240-1253, Dec. 2017.

Claims

CLAIMS:

1. A computer-implemented method performed by a first node (111), the method being for handling a first knowledge graph, the first node (111) operating in a computer system (100), the method comprising:- determining (307), using an artificial neural network (121), a first knowledge graph representing words in a conversational speech and relations of the words, wherein the determining (307) uses, as input, data comprising text derived, using a speech recognition algorithm, from first data encoding conversational speech, and- initiating (308) outputting a first indication of the determined first knowledge graph.

2. The method according to claim 1 , wherein the artificial neural network (121) has an end-to-end neural network architecture, wherein a first input into the end-to-end neural network architecture comprises feature vectors corresponding to words in the text data, and wherein a first output of the end-to-end neural network architecture comprises a first adjacency matrix and a second adjacency matrix, wherein the first adjacency matrix and the second adjacency matrix, together, correspond to the first knowledge graph, wherein the first adjacency matrix encodes information of nodes of the first knowledge graph, and wherein the second adjacency matrix encodes the edges between the nodes in the first knowledge graph.

3. The method according to claim 2, wherein:- in the first adjacency matrix, a first matrix element AiJ denotes a presence of an undirected edge from word i to word j, and- in the second adjacency matrix, a second element Di J denotes a presence of a directed edge from word i to word j.

4. The method according to any of claims 1-3, further comprising at least one of:- obtaining (305) the first data encoding the conversational speech, and- obtaining (306), from the first data, and using the speech recognition algorithm, the data comprising the text, wherein the data comprising the text is second data.

5. The method according to claim 4, further comprising at least one of:- obtaining (301) a plurality of sets of first data, wherein each set of first data encodes a respective conversational speech,- obtaining (302), from the plurality of sets of first data, and using the speech recognition algorithm, a plurality of sets of second data, wherein each set of second data encodes respective text corresponding to a respective conversational speech,- obtaining (303) a respective second knowledge graph for each set of third data of a plurality of sets of third data, wherein each set of third data corresponds to a set of second data having been processed to remove errors and unnecessary words, and- training (304) the artificial neural network (121) to determine a respective first knowledge graph for each respective set of second data, wherein the training (304) is performed by minimizing a loss function representing a difference between each respective first knowledge graph and the obtained respective second knowledge graph, wherein the training (304) is performed iteratively until the loss function is below a threshold, and wherein the determining (307) of the first knowledge graph is performed using the trained artificial neural network (121).

6. The method according to claim 5, wherein at least one of:- the respective second knowledge graph is received from a third node (113) operating in the computer system (100), and- the first node (111) refrains from training (304) the artificial neural network (121), and wherein the trained artificial neural network (121) is received in a second indication from a second node (121) operating in the computer system (100).

7. The method according to any of claims 1-6, wherein the first indication of the determined first knowledge graph is outputted in one of: i. a healthcare application, and wherein the conversational speech describes a condition of a patient, ii. a tele-operated driving application, and wherein the conversational speech describes instructions to a vehicle by a remote driver, iii. a vehicle-to-everything application, and wherein the conversational speech describes vehicle-related problems between stakeholders, and iv. a field service operation application, and wherein the conversational speech describes a problem to be solved at a customer site.

8. A computer-implemented method performed by a second node (112), the method being for handling a respective first knowledge graph, the second node (112) operating in a computer system (100), the method comprising:- obtaining (401) a plurality of sets of first data, wherein each set of first data encodes a respective conversational speech,- obtaining (402), from the plurality of sets of first data, and using a speech recognition algorithm, a plurality of sets of second data, wherein each set of second data encodes respective text corresponding to a respective conversational speech,- obtaining (403) a respective second knowledge graph for each set of third data of a plurality of sets of third data, wherein each set of third data corresponds to a set of second data having been processed to remove errors and unnecessary words,- training (404) an artificial neural network (121) to determine a respective first knowledge graph for each respective set of second data, wherein the training (404) is performed by minimizing a loss function representing a difference between each respective first knowledge graph and the determined respective second knowledge graph, wherein the training (404) is performed iteratively until the loss function is below a threshold, and- initiating (405) outputting a second indication of the trained artificial neural network (121).

9. The method according to claim 8, wherein the artificial neural network (121) has an end-to-end neural network architecture, wherein a first input into the end-to-end neural network architecture comprises feature vectors corresponding to words in the text data, and wherein a first respective output of the end-to-end neural network architecture comprises a first respective adjacency matrix and a second respective adjacency matrix, wherein the first respective adjacency matrix and the second respective adjacency matrix, together, correspond to the respective first knowledge graph, wherein the first respective adjacency matrix encodes respective information of respective nodes of the respective first knowledge graph, and wherein the second respective adjacency matrix encodes respective edges between the respective nodes in the respective first knowledge graph.

10. The method according to claim 9, wherein:- in the first adjacency matrix, a first matrix element AiJ denotes a presence of an undirected edge from word i to word j, and- in the second adjacency matrix, a second element Di J denotes a presence of a directed edge from word i to word j.

11. The method according to any of claims 8-10, wherein at least one of:- the respective second knowledge graph is received from a third node (113) operating in the computer system (100), and- the initiating (405) outputting the second indication comprises sending the second indication to a first node (111) operating in the computer system (100).

12. The method according to any of claims 8-11, wherein the second indication of the trained artificial neural network (121) is outputted in one of: i. a healthcare application, and wherein the conversational speech describes a condition of a patient, ii. a tele-operated driving application, and wherein the conversational speech describes instructions to a vehicle by a remote driver, iii. a vehicle-to-everything application, and wherein the conversational speech describes vehicle-related problems between stakeholders, and iv. a field service operation application, and wherein the conversational speech describes a problem to be solved at a customer site.

13. A first node (111), for handling a first knowledge graph, the first node (111) being configured to operate in a computer system (100), the first node (111) being further configured to:- determine, using an artificial neural network (121), a first knowledge graph configured to represent words in a conversational speech and relations of the words, wherein the determining is configured to use, as input, data configured to comprise text configured to be derived, using a speech recognition algorithm, from first data configured to encode conversational speech, and- initiate outputting a first indication of the first knowledge graph configured to be determined.

14. The first node (111) according to claim 13, wherein the artificial neural network (121) is configured to have an end-to-end neural network architecture, wherein a first input into the end-to-end neural network architecture is configured to comprise feature vectors corresponding to words in the text data, and wherein a first output of the end-to-endneural network architecture is configured to comprise a first adjacency matrix and a second adjacency matrix, wherein the first adjacency matrix and the second adjacency matrix, together, are configured to correspond to the first knowledge graph, wherein the first adjacency matrix is configured to encode information of nodes of the first knowledge graph, and wherein the second adjacency matrix is configured to encode the edges between the nodes in the first knowledge graph.

15. The first node (111) according to claim 14, wherein:- in the first adjacency matrix, a first matrix element AiJ is configured to denote a presence of an undirected edge from word i to word j, and- in the second adjacency matrix, a second element Di J is configured to denote a presence of a directed edge from word i to word j.

16. The first node (111) according to any of claims 13-15, being further configured to at least one of:- obtain the first data configured to encode the conversational speech, and- obtain, from the first data, and using the speech recognition algorithm, the data configured to comprise the text, wherein the data configured to comprise the text is configured to be second data.

17. The first node (111) according to claim 16, being further configured to at least one of:- obtain a plurality of sets of first data, wherein each set of first data is configured to encode a respective conversational speech,- obtain, from the plurality of sets of first data, and using the speech recognition algorithm, a plurality of sets of second data, wherein each set of second data is configured to encode respective text configured to correspond to a respective conversational speech,- obtain a respective second knowledge graph for each set of third data of a plurality of sets of third data, wherein each set of third data is configured to correspond to a set of second data, processed to remove errors and unnecessary words, and- train the artificial neural network (121) to determine a respective first knowledge graph for each respective set of second data, wherein the training is configured to be performed by minimizing a loss function configured to configured to represent a difference between each respective first knowledge graph and the respective second knowledge graph configured to be obtained, wherein the training is configured to be performed iteratively until the loss function is belowa threshold, and wherein the determining of the first knowledge graph is configured to be performed using the artificial neural network (121) configured to trained.

18. The first node (111) according to claim 17, wherein at least one of:- the respective second knowledge graph is configured to be received from a third node (113) configured to operate in the computer system (100), and- the first node (111) is configured to refrain from training (304) the artificial neural network (121), and the trained artificial neural network (121) is configured to be received in a second indication from a second node (121) configured to operate in the computer system (100).

19. The first node (111) according to any of claims 13-18, wherein the first indication of the determined first knowledge graph is configured to be outputted in one of: i. a healthcare application, and wherein the conversational speech is configured to describe a condition of a patient, ii. a tele-operated driving application, and wherein the conversational speech is configured to describe instructions to a vehicle by a remote driver, iii. a vehicle-to-everything application, and wherein the conversational speech is configured to describe vehicle- related problems between stakeholders, and iv. a field service operation application, and wherein the conversational speech is configured to describe a problem to be solved at a customer site.

20. A second node (112), for handling a respective first knowledge graph, the second node (112) being configured to operate in a computer system (100), the second node (112) being further configured to:- obtain a plurality of sets of first data, wherein each set of first data is configured to encode a respective conversational speech,- obtain, from the plurality of sets of first data, and using a speech recognition algorithm, a plurality of sets of second data, wherein each set of second data is configured to encode respective text configured to correspond to a respective conversational speech,- obtain a respective second knowledge graph for each set of third data of a plurality of sets of third data, wherein each set of third data is configured tocorrespond to a set of second data, processed to remove errors and unnecessary words,- train an artificial neural network (121) to determine a respective first knowledge graph for each respective set of second data, wherein the training is performed by minimizing a loss function representing a difference between each respective first knowledge graph and the determined respective second knowledge graph, wherein the training is configured to be performed iteratively until the loss function is below a threshold, and- initiate outputting a second indication of the artificial neural network (121) configured to be trained.

21. The second node (112) according to claim 20, wherein the artificial neural network (121) is configured to have an end-to-end neural network architecture, wherein a first input into the end-to-end neural network architecture is configured to comprise feature vectors configured to correspond to words in the text data, and wherein a first respective output of the end-to-end neural network architecture is configured to comprise a first respective adjacency matrix and a second respective adjacency matrix, wherein the first respective adjacency matrix and the second respective adjacency matrix, together, are configured to correspond to the respective first knowledge graph, wherein the first respective adjacency matrix is configured to encode respective information of respective nodes of the respective first knowledge graph, and wherein the second respective adjacency matrix is configured to encode respective edges between the respective nodes in the respective first knowledge graph.

22. The second node (112) according to claim 21, wherein:- in the first adjacency matrix, a first matrix element AiJ is configured to denote a presence of an undirected edge from word i to word j, and- in the second adjacency matrix, a second element Di J is configured to denote a presence of a directed edge from word i to word j.

23. The second node (112) according to any of claims 20-22, wherein at least one of:- the respective second knowledge graph is configured to be received from a third node (113) configured to operate in the computer system (100), and- the initiating outputting the second indication is configured to comprise sending the second indication to a first node (111) configured to operate in the computer system (100).

24. The second node (112) according to any of claims 20-23, wherein the second indication of the artificial neural network (121) configured to be trained is outputted in one of: i. a healthcare application, and wherein the conversational speech is configured to describe a condition of a patient, ii. a tele-operated driving application, and wherein the conversational speech is configured to describe instructions to a vehicle by a remote driver, iii. a vehicle-to-everything application, and wherein the conversational speech is configured to describe vehicle- related problems between stakeholders, and iv. a field service operation application, and wherein the conversational speech is configured to describe a problem to be solved at a customer site.

25. A computer program (905), comprising instructions which, when executed on at least one processing circuitry (901), cause the at least one processing circuitry (901) to carry out the method according to any of claims 1-7.

26. A computer-readable storage medium (906), having stored thereon a computer program (905), comprising instructions which, when executed on at least one processing circuitry (901), cause the at least one processing circuitry (901) to carry out the method according to any of claims 1-7.

27. A computer program (1005), comprising instructions which, when executed on at least one processing circuitry (1001), cause the at least one processing circuitry (1001) to carry out the method according to any of claims 8-12.

28. A computer-readable storage medium (1006), having stored thereon a computer program (1005), comprising instructions which, when executed on at least one processing circuitry (1001), cause the at least one processing circuitry (1001) to carry out the method according to any of claims 8-12.

Images