Patent Application
20220101152
The present application claims the benefit under 35 U.S.C. § 119 of European Patent Application No. EP 20199308.6 filed on Sep. 30, 2020, which is expressly incorporated herein by reference in its entirety.
The present invention relates to conceptual clustering.
The task of conceptual clustering over relational data is described for example in R. E. Stepp and R. S. Michalski, “Conceptual clustering of structured objects: A goal-oriented approach,” Artificial Intelligence, 28(1):43-69 (1986).
The present invention provides a computer implemented method for solving the conceptual clustering problem over relational data, a corresponding device and a corresponding computer program.
In accordance with an example embodiment of the present invention, the computer implemented method comprises determining an embedding of a first entity, in particular of a knowledge graph, inserting a first vertex for the embedding in an in particular weighted in particular undirected graph, determining in the graph a first cluster of vertices comprising the first vertex, determining for the first cluster a second entity, in particular in the knowledge graph, determining a semantic similarity between the first entity and the second entity, in particular in the knowledge graph, determining a rule for the first cluster depending on a semantic similarity between the first entity and the second entity. This way, the conceptual clustering task is performed over entities of a Knowledge Graph, KG, where given a KG and a set of target entities, the goal is to cluster these entities into an, in particular unknown, number of distinct groups based on the quality of the descriptions computed over the KG for the groups. The rule provides description for the computed cluster that is understandable for a human and a machine alike. The method has the capability of computing high-quality clusters along with their descriptions without knowing their number a priori. Moreover, in contrast to other existing clustering methods, e.g., DBSCAN, that do not require the number of clusters as input, this method is especially appealing due to a small number of hyper-parameters to tune.
Inserting the first vertex in the graph may comprise labelling an edge of the graph that links the first vertex to a second vertex of the graph with a label. Relying on the target entities an in particular undirected complete weighted graph is constructed, where the edges between every pair of entities are labeled.
Labelling the edge may comprise determining a weight depending on a distance between a first vector for the first vertex and a second vector for the second vertex and mapping the weight to the label with a function. The edges between every pair of entities are labeled using, e.g., a cosine similarity between the respective entities in an embedding space.
Determining the first cluster may comprise determining a subset of edges of the graph for the first cluster including the first vertex, such that no cycle in the graph intersects this subset once. A multicut problem is formulated based on this graph, and solved effectively.
In accordance with an example embodiment of the present invention, determining the embedding may comprise mapping the first entity to a first vector in a vector space with a model.
The model may be an embedding model that translates entities and relations of the KG into vectors in a vector space in particular a low dimensional vector space.
Determining the rule may comprise determining a plurality of rules depending on semantic similarities to the second entity and selecting the rule from the plurality of rules.
Selecting the rule may comprise determining an amount of entities covered by the rule that belong to the first cluster, determining an amount of entities covered by the rule that belong to a second cluster, determining a measure for the rule depending on the amount of entities covered by the rule that belong to the first cluster and depending on the amount of entities that belong to the second cluster, and selecting the rule if the measure meets a condition or not selecting the rule otherwise.
Determining the measure may comprise determining a ratio of the amount of entities covered by the rule that belong to the first cluster and a cardinality of the first cluster.
Determining the measure may comprise determining a ratio of the amount of entities covered by the rule that belong to the second cluster and a cardinality of the second cluster.
In accordance with an example embodiment of the present invention, the method may comprise determining an output depending on the rule, detecting an input for the rule, and determining a label for the rule depending on the input. The rule describes a concept that is found in the KG. An expert may be provided with such a description and suggest a name for it. Such concept can then be added to the KG under the suggested name.
The method may comprise receiving an input, in particular a query or a message, selecting the rule depending on the input, determining at least one entity depending on the rule, outputting a response depending on the at least one entity.
Outputting the response may comprise indicating a state of a machine, a property of an object in a digital image or an answer to a question depending on the at least one entity.
In accordance with an example embodiment of the present invention, the device is adapted to execute the method. The computer program comprises computer readable instructions that, when executed on a computer, cause the computer to execute steps in the method.
Further advantageous embodiments are derivable from the following description and the figures.
Knowledge graphs, KGs, represent interlinked collections of factual information, and they are often encoded as a set of <subject predicate object> triples, e.g., <john worksAt Firm A>. Subjects or objects of such triples are referred to as entities. Predicates are referred to as relations. The set of triples of a KG can be naturally represented as a directed graph, whose vertices and edges are labeled.
Knowledge graph embedding, KGE, concerns embedding KG entities and relations into continuous vector spaces with a user-specified dimension n. More specifically, KGE models take as input a set of KG triples and aim at mapping the entities and relations into the n-dimensional vector space such that some features reflecting the KG structure are preserved. These features are captured by the objective function of the respective embedding model. This way from relational data, a set of numerical vectors is obtained.
Conceptual clustering is a task of grouping entities in a knowledge graph into a set of groups with the high inter-cluster similarity and low intra-cluster similarity as well as describing the commonalities for entities within each cluster in terms of their KG properties.
A multicut of a graph is a subset of its edges such that no cycle in the graph intersects this subset exactly once.
Rule induction is the task of learning rules from a KG, i.e., given a KG, the goal of rule learning is to construct a set of rules of the form H←B, where H is the entailed predicate of the form h(X,c), where c is a dummy constant corresponding to the target cluster, h is a dummy predicate reflecting the relation between X and the target cluster, and B is a conjunction of predicates of the form b1(X1,Y1), b2(X2,Y2), . . . , bn(Xn,Yn), where every X1, X2, . . . , Xn, Y1, Y2, . . . , Yn can either be a variable or a constant, such that the learned rules hold often enough in the data.
A measure for a rule is a function that takes as input a KG and a rule and outputs a value reflecting how well the given rule fits the KG.
Conjunctive query, CQ, is an expression of the form q(X1, X2, . . . , Xk)←B or sometimes simply written as X1, X2, . . . , Xk
←B, where B is the body of the CQ similar as defined for rules, and X1, X2, . . . , Xk are so-called answer variables, i.e., variables, whose positions are responsible for answers to the query.
Users formulate their information needs in the form of such queries. For example, for a KG storing in Firm A employees, a possible conjunctive query could ask for all people, who work for Department D of Firm A, are married and have a child. Formally, such query could be written as
Q(X)←type(X,person),worksFor(X,D),marriedTo(X,Y),hasChild(X,Z).
In the above example X is an answer variable of the given CQ. Conjunctive queries can be naturally represented as KG patterns.
The device 100 is adapted for executing a method that described below with reference to
The method for conceptual clustering described below does not require the information about the number of clusters.
The method takes as input a knowledge graph KG, a set of entities T, as well as other parameters including a maximum length of cluster descriptions m, a minimum coverage μ and a threshold value θ, which are used to determine a quality of the computed descriptions.
The method comprises a step 1 of determining an embedding for at least one entity of the KG 202. The embedding may be a vector in a vector space. The embedding may be determined with a model, in particular an embedding model.
A corresponding algorithm may start by embedding the entities and relations into a low dimensional vector space using a certain embedding model. Any embedding model can be used. An exemplary embedding model is describe in Ren, H., Hu, W., Leskovec, J., “Query2box: Reasoning over knowledge graphs in vector space using box embeddings,” ICLR (2020).
The method comprises a step 2 of determining a graph 208. In the example, the graph 208 is determined for the embeddings of the entities of the KG 202. The step 2 comprises inserting a first vertex 210 for the embedding 206 in the graph 208. When the target entities, i.e. a subset of the entities of the KG 202, are used, the graph 208 is incomplete in the sense that there are less vertices than entities. This means, the construction of the graph 208 for the entire KG 200 can be avoided in practice.
The graph 208 in the example is a weighted undirected graph.
The step 2 comprises determining labels for edges of the graph 208. In the example, inserting the first vertex 210 in the graph 208 comprises labelling an edge 212 of the graph that links the first vertex 210 to a second vertex 214 of the graph 208 with a label.
Labelling the edge 212 may comprise determining a weight depending on a distance [cosine similarity] between the first vector for the first vertex 210 and a second vector for the second vertex 214 and mapping the weight to the label with a function.
In the example, the graph 210 is a similarity graph that can be used as input to a multicut algorithm. Once the entities and relations are mapped to the embedding space, a complete undirected weighted graph G=(V,E) may be computed. In this example, E=T, and every pair of entities in E are connected. Then the method computes pairwise distances between every pair of entities using the cosine similarity of their vectors.
More specifically, let a,b be two target entities, and let va,vb be their respective numerical vectors in the embedding space.
Then the distance between a and b is computed using the cosine similarity function as follows:
Every edge (a,b)∈E in the constructed complete graph G is then labeled with the function Φ:E→R
The method comprises a step 3 of determining in the graph 208 at least one cluster. Preferably, the step 3 comprises determining at least two clusters in the graph 208. In the example a first cluster 216 of vertices comprising the first vertex 210 is determined. In the example, a second cluster 218 and a third cluster 220 are determined as well.
Determining the first cluster 216 may comprise determining a subset of edges of the graph 208 for the first cluster 216 including the first vertex 210, such that no cycle in the graph 208 intersects this subset once.
The multicut clustering algorithm may be used to detect prominent regions in the embedding space. Aspects of the multicut clustering approach are disclosed in
A multicut of the graph 210 is a subset of its edges, such that no cycle in the graph intersects this subset exactly once. Assigning 1 to the edges in the multicut, and 0 to all other edges, a set of all valid multicuts can be formalized by the following set of linear inequalities:
Considering only chordless cycles is sufficient in this example. Any valid multicut y∈YG uniquely defines a graph decomposition.
Given the above definitions, a minimum cost multicut problem is as follows:
This problem is solvable using effective heuristics methods based on local search strategies, e.g., as described in Keuper, M., Levinkov, E., Bonneel, N., Lavoue, G., Brox, T., Andres, B., “Efficient decomposition of image and mesh graphs by lifted multicuts,” ICCV (2015).
The method comprises a step 4 of determining for the first cluster 216 a second entity 222. The step 4 comprises determining a semantic similarity 224 between the first entity 204 and the second entity 222. The second entity 222 and the semantic similarity 224 are determined in the knowledge graph 202.
The method comprises a step 5 of determining a rule 226 for the first cluster 216 depending on a semantic similarity between the first entity 204 and the second entity 222.
Determining the rule 226 may comprise determining a plurality of rules depending on semantic similarities to the second entity and selecting the rule 226 from the plurality of rules.
The constructed regions, i.e., clusters, in the vector space may be mapped to conjunctive queries, i.e., descriptions, by learning Horn rules. More specifically, for every entity a that has been clustered into a cluster c in the step 3, the method may add a fact belongsTo(a,c) to the KG and learn rules of the form
belongsTo(X,c)←p1(X,Y),p2(Y,Z), . . . ,pm(W,U)
where m is the desired description length specified by the user as mentioned above, while p1, p2, . . . , pm are relations appearing in the KG. This means that the entity a is covered by the rule.
An exemplary algorithm that may be used for learning such rules is described in Galarraga, L., Teflioudi, C., Hose, K., Suchanek, F. M., “Fast rule mining in ontological knowledge bases with amie++,” The VLDB Journal 24(6), 707-730 (2015).
The method described therein is modified as described below to be capable of capturing constants in rule heads. More specifically in an exemplary implementation the rule is modified to start from belongsTo(X,c) rather than belongsTo(X,Y) and a body of the rule is constructed by applying expansion operators described in “Fast rule mining in ontological knowledge bases with amie++” apart from the “add closing atom operator” operator.
The learned rules can then be pruned on various rule measures, such as confidence, coverage or weighted relative accuracy.
Selecting the rule 226 may comprise determining an amount of entities covered by the rule 226 that belong to the first cluster 216 and determining an amount of entities covered by the rule that belong to a second cluster 218. A measure for the rule 226 is determined depending on the amount of entities covered by the rule 226 that belong to the first cluster 216 and depending on the amount of entities that belong to the second cluster 218. The rule 226 is selected if the measure meets a condition. Otherwise the rule 226 is not selected.
Determining the measure may comprise determining a ratio of the amount of entities covered by the rule 226 that belong to the first cluster 216 and a cardinality of the first cluster 216.
Determining the measure may comprise determining a ratio of the amount of entities covered by the rule that belong to the second cluster 218 and a cardinality of the second cluster 218.
The third cluster 220 may be used alike to determine the measure.
In an example an exclusive coverage measure may be used to estimate the quality of the obtained cluster descriptions. The main advantage of this measure is that it accounts for all computed clusters when evaluating a given rule, which is in contrast to the majority of other rule measures that estimate the quality of a rule considered in isolation. This ensures that a given rule is exclusive for a cluster at hand, which contributes to its quality. For a given ruler, cluster c, collection of clusters S and a KG the exclusive coverage measure is defined as follows:
Where cov(r,c,KG) is the standard coverage of a rule r for a cluster c and the knowledge graph KG defined as the ratio of entities covered by r within the cluster c over the cardinality of c.
Finally, among all computed cluster descriptions, those may be selected that have the highest average exclusive coverage.
The method may comprise determining an output depending on the rule 226 and detecting an input for the rule 226, in particular in response to the output. The method may comprise determining a label for the rule 226 depending on the input.
Rules can serve as explanations for clusters constructed over embeddings, and thus can be seen as assets that contribute to the explainability of machine learning models.
Rules represent human-interpretable labels for concepts, which are prominent sets of entities grouped based on their semantic similarity.
For example, assuming a KG is given that is extracted from scientific publications in the context of material science. Obviously, after such an extraction process many concepts, e.g., types of SOFC, might be missing, as they might not be directly mentioned in the text.
After applying the method described above on such a KG, one might obtain the cluster of entities with the rule 226 describing it as a set of “electrochemical conversion devices producing electricity directly from oxidizing a fuel, which have dimensions from needle-like shapes to lengths of about 1.5-2 m for rapid start-up times and large gross power.”
When material scientists are provided with such a description they can immediately realize that this label describes “tubular SOFC”, and such concept can be added to the KG under the suggested name.
The rule 226 and the output in this example is “electrochemical conversion devices producing electricity directly from oxidizing a fuel, which have dimensions from needle-like shapes to lengths of about 1.5-2 m for rapid start-up times and large gross power” or a representation thereof. The label for the rule 226 in this example may be “tubular SOFC”.
The method and the rules in particular contribute to the process of semi-automatic ontology construction.
This enables creating new concepts, i.e. new rules. Adding the new concepts to the KG explicitly as described above can optimize a process of question answering.
Aspects of an application are described below with reference to
A corresponding method comprises a step 302 of receiving an input, in particular a query.
The query may be the question formulated by users in the natural language. The question may be translated into a formal representation.
The method comprises a step 304 of selecting the rule 226 depending on the input.
The task in this example is to find an answer to the question. The answer to the question is found, when at least one entity in the KG is found using a rule that corresponds to the concept. Since the questions are formulated by users in the natural language, they might often contain concepts, for which there is no rule.
In this case, the method comprises automatically detecting the rule 226 for the concept as described above.
The method comprises a step 306 of determining at least one entity depending on the rule 226.
The method comprises a step 308 of outputting a response depending on the at least one entity.
The rule 226 may be presented to the user as a description of the concept that was detected when the rule 226 was determined. The user can then better understand the response.
The query may concern a state of a machine, a property of an object in a digital image or an answer to a question.
Outputting the response may comprise indicating a state of a machine, a property of an object in a digital image or an answer to a question depending on the at least one entity.
In a production line having a plurality of machines that are adapted to communicate via messages, the input may be a message. The message may contain information regarding the state of the machine. The rule may in this case relate to a concept for failure recognition. The output may in this case state a sanity of a product that is produced with the line or a state of health of the machine or of another machine of the production line.
For digital image processing, the KG may be a description of objects recognized in an object recognition for the image. Entities in the KG may represent the objects and/or properties thereof.
In a street view, an object may be a car, a person, a house or other part of an infrastructure. In the street view, the property may describe the object and/or a relation of the object to another object in particular in the digital image.
| Number | Date | Country | Kind |
|---|---|---|---|
| 20199308.6 | Sep 2020 | EP | regional |
