KR101788303B1 - Apparatus and meyhod for large scale parallel temporal reasoning - Google Patents

Abstract

The massively parallel qualitative time inferring apparatus according to the present invention includes an HDFS unit for storing a plurality of knowledge bases in the form of a Hadoop Distributed File System (HDFS), a task management unit for assigning a task to each of the two or more processing node units, And a processing node unit for performing an inference operation on the received knowledge base to calculate an expanded knowledge base and storing the calculated expanded knowledge base in the HDFS unit.

Description

[0001] APPARATUS AND MEYHOD FOR LARGE SCALE PARALLEL TEMPORAL REASONING [0002]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to time inference, and more particularly, to massively parallel qualitative time inference.

As the number of users who use the Web and the amount of information increase exponentially, the demands of the web users are becoming more diverse and complex. Accordingly, researches on semantic web technology that efficiently extracts only information that meets the user's purpose from large-scale unstructured web data and efficiently responds to user's needs is being actively conducted. Semantic Web technology utilizes ontological knowledge that systematically defines basic concepts for applications, relationships among them, common vocabularies, and inference rules, The program itself allows the user to understand the semantics of the web content and perform the work. RDF (Resource Description Framework) and OWL (Web Ontology Language) are mainly used as representative ontology presentation languages.

Recently, a lot of new facts and discoveries have been published on the web in a variety of applications, including engineering, medicine and drug technology, security technology, and so on. Therefore, in order to cope with such a rapid change of web information, the semantic web technology is moved away from expressing only partial facts and knowledge, and the time and place where the fact and knowledge are actually valid are expressed There is a need for technological extension to be able to infer the relationship between them.

In many existing systems that process time knowledge and queries, computation of temporal data, which represents a time point and a time interval, is used to obtain an answer to a query. , Which is mainly based on quantitative temporal reasoning.

Quantitative temporal representation and reasoning can be used more effectively in applications such as production automation and control systems that require real-time and precision, while natural language understanding, question-answering systems, In certain applications, such as decision support systems, qualitative temporal reasoning that qualitatively expresses and inferences of temporal order relationships between events can be used more effectively.

Quantitative temporal representation and reasoning can be used more effectively in applications such as production automation and control systems that require real-time and precision, while natural language understanding, question-answering systems, In certain applications, such as decision support systems, qualitative time inference, which expresses and inferences the qualitative order relation between events, can be used more effectively.

 GQR, SOWL, and CHRONOS, almost all of the qualitative time inference techniques developed so far are assumed to have a single computer execution environment, and they have a limited performance to infer a large scale knowledge base of Web scale in common.

"Time Ontology Language for Representation and Inference of Time-based Time Information," Sang Kyun Kim, Kyu Chul Lee, and Mi Mi Young, Journal of KISS: Software and Applications, Vol. 36, No. 7, pp. 509-522, 2009.

SUMMARY OF THE INVENTION It is an object of the present invention to provide a parallel qualitative time inference device and a reasoning method that automatically infer large-scale time order relations between various daily events.

The massively parallel qualitative time inferring apparatus according to the present invention includes an HDFS unit for storing a plurality of knowledge bases in the form of a Hadoop Distributed File System (HDFS), a task management unit for assigning a task to each of the two or more processing node units, And a processing node unit for performing an inference operation on the received knowledge base to calculate an expanded knowledge base and storing the calculated expanded knowledge base in the HDFS unit.

The processing node unit of the present invention performs a pre-encoding operation on the knowledge base, performs an inverse & equality reasoning on the encoded knowledge base, and performs a relational inference < RTI ID = 0.0 > It is possible to infer the parallel qualification time by repeating Transitive Reasoning and Refining, and then storing the inferred knowledge base in the HDFS unit.

At this time, the processing node unit converts the knowledge base into a text form through a decoding operation, and stores the knowledge base converted into a text form in the HDFS unit. The processing node unit converts each triple statement in the form of a string included in the knowledge base into a combination of a subject ID, a predicate ID, and an object identifier (ID) .

The processing node applies the migration relationship inferencing rules defined through the migration relationship inference to the knowledge base and can reduce the set of disjunctive relations through the migration relationship inference. Then, the processing node unit can check the contradiction between facts inferred through the relationship refinement operation and refine the neighbor relation facts.

The processing node unit disassembles each constituent element including at least one of the predicate and the object by giving a triple sentence constituting the knowledge base, assigns a unique identifier to the unified resource identifier (URI) of each constituent element, registers it in the dictionary At the same time, the integrated resource identifier of each component is converted into a previously registered unique identifier, and the concatenated triple sentence is reconstructed by combining the identifiers of the converted components. In addition, the processing node part simultaneously infer the inverse relationship and the same relation to the encoded knowledge base.

According to the present invention, the parallel qualitative time inference method using the parallel qualitative time inferencing apparatus first encodes the knowledge base according to the assigned task, and performs an inverse & equality reasoning on the encoded knowledge base . Then we infer the knowledge base by repeating Transitive Reasoning and Refining until new facts are not induced. Next, the inferred knowledge base is stored in HDFS.

The massively parallel qualitative time inferencing apparatus and method according to the present invention can be applied to a parallel time inference method using a Hadoop cluster system and a MapReduce parallel framework to overcome the limitations of conventional small time inference, Enables massive qualitative time inference.

In addition, the massively parallel qualitative time inferencing apparatus and method according to the present invention can be applied to an information search based on semantic web, which is an industry field in which space and time reasoning ability is indispensable, a time-space database, an intelligent service robot, And can be widely used in fields such as response systems.

1 is a block diagram of a massively parallel qualitative time inference apparatus according to an embodiment of the present invention.
2 is a diagram illustrating an example of a time ontology of a massively parallel qualitative time inference apparatus according to an embodiment of the present invention.
3 is a diagram showing an example of a set of disjunctive descriptors of a massively parallel qualitative time inferencing apparatus according to an embodiment of the present invention.
FIG. 4 is a flowchart showing the operation of the processing node unit 130 of the massively parallel qualitative time inferencing apparatus according to an embodiment of the present invention.
FIG. 5 is a diagram illustrating a process of a pre-encoding operation S401 of the processing node unit 130 of the massively parallel qualitative time inferencing apparatus according to an embodiment of the present invention.
6A and 6B are diagrams illustrating a process of inverse relation inference inference operation S402 of the processing node unit 130 of the massively parallel parallel inequality reasoning apparatus according to an embodiment of the present invention.
FIGS. 7A and 7B are diagrams showing a processing procedure of the relational reasoning operation S403 of the processing node unit 130 of the massively parallel qualitative time inferencing apparatus according to an embodiment of the present invention.
8A and 8B are diagrams illustrating a process of a relation refining operation (S404) of the processing node unit 130 of the massively parallel parallel qualitative inferring apparatus according to an embodiment of the present invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. The terms and words used in the present specification are selected in consideration of the functions in the embodiments, and the meaning of the terms may vary depending on the intention or custom of the invention. Therefore, the terms used in the following embodiments are defined according to their definitions when they are specifically defined in this specification, and unless otherwise specified, they should be construed in a sense generally recognized by those skilled in the art.

1 is a block diagram of a massively parallel qualitative time inference apparatus according to an embodiment of the present invention.

Referring to FIG. 1, a massively parallel qualitative time inferencing apparatus according to an embodiment of the present invention deduces a parallel qualification time based on a Hadoop MapReduce parallel programming framework.

The massively parallel qualitative time inference apparatus according to an embodiment of the present invention includes a task management unit 110, a name node unit 120, two or more processing node units 130, an HDFS unit 140, and a user processing unit 150 do.

The processing node unit 130 reads a large-capacity knowledge base from the HDFS (Hadoop Distributed File System) unit 140, which is a distributed file system of the Hadoop cluster system. The processing node unit 130 allocates tasks to one or more processing node units 130 constituting the cluster through the scheduler unit 131 and the node manager unit 132 and transmits the tasks.

The two or more processing node units 130 may perform inverse and equal inference tasks, transitive inference tasks, and relationship refinement tasks for a task assigned by the task management unit 110. [ And an inference work pipeline (Pipeline) including the inference operation pipeline. The two or more processing node units 130 store the extended knowledge base obtained as a result of the inference in the distributed file system unit 140 again. The details of the procedure performed by each component will be described later with reference to the drawings.

2 is a diagram illustrating an example of a time ontology of a massively parallel qualitative time inference apparatus according to an embodiment of the present invention.

Referring to FIG. 2, the time ontology shown in FIG. 2 shows a part of a temporal ontology on which a temporal knowledge representation is based. The massively parallel qualitative time inferencing apparatus according to an embodiment of the present invention is based on Semantic Web standard ontology languages RDF and OWL and has a function of subject knowledge, predicate, object, As shown in FIG. The massive parallel qualitative time inference apparatus according to an embodiment of the present invention extends the interval algebra theory of Allen and the point algebra theory of Kautz using the time ontology and the time knowledge representation system as shown in FIG.

In the ontology of FIG. 2, the rdfs (Statement) class represents a set of facts, and each sentence representing a fact is simply a subject, a predicate, an object objects are represented in a triple form.

However, in this ontology, each sentence can be associated with a time event (kgu: Time Event) through the occursOn property to give the facts the time information that is valid. On the other hand, a time event (kgu: Time Event) can be a time point (kgu: Time Point) or a time interval (kgu: Time Interval) as in the point logarithmic theory. One time interval (kgu: Time Interval) consists of two time points (kgu: Time Points) corresponding to the starting and ending points through the kgu: tm_begin property and the kgu: tm_end property, Lt; / RTI > And the most important part of this ontology is thirteen temporal properties that represent the order relation between two time events (kgu: Time Events).

3 is a diagram showing an example of a set of disjunctive descriptors of a massively parallel qualitative time inferencing apparatus according to an embodiment of the present invention.

Referring to FIG. 3, FIG. 3 shows a set of 28 disjunctive relations that can be derived through spatial transition relationship inferences from 13 temporal relationship properties, i.e., temporal properties. There are a total of 2 ¹³ theoretically possible all possible relations from the 13 time order relation descriptors.

However, when constructing a composition table of all possible relations for temporal relationship inferences, this combination table has a size of 2 ¹³ * 2 ¹³ , so it is very difficult to store this combination table and refer to it It requires a large storage space and a lot of calculation time. In order to overcome such a problem, the present invention calculates a set of real relations that can be appeared during the relational reasoning process through a simulation inference process.

As a result of calculating the set of disjunctive relations according to the embodiment of the present invention, only 28 disjunctive relations listed in FIG. 3 among the 2 ¹³ disjunctive relations occur in the inference process. On the other hand, in the massively parallel qualitative time inference apparatus according to the embodiment of the present invention, many other disconnection relations are not actually generated.

Therefore, in the present invention, instead of the wasteful combination table of 2 ¹³ * 2 ¹³ size, the transfer relation reasoning rule can be reduced and defined as a 28 * 28 size combination table representing only 28 combinations of the disjunctive relations listed in FIG. have. The discovery of such an optimized disjunctive set and the reduction of the combinatorial table can improve performance by reducing the amount of computation over time reasoning.

FIG. 4 is a flowchart showing the operation of the processing node unit 130 of the massively parallel qualitative time inferencing apparatus according to an embodiment of the present invention.

Referring to FIG. 1 and FIG. 4, a task is transferred to two or more processing node units 130 by a task management unit 110, respectively, in a large-scale parallel qualitative time inferencing apparatus according to the present invention.

The processing node unit 130 first performs a pre-encoding operation (S401). The mapping task performed first in the processing node unit 130 of the parallel qualitative time inferencing apparatus according to the present invention is a pre-encoding task. The dictionary encoding task converts each triple statement in the form of a string contained in the knowledge base into a combination of a numeric subject ID, a predicate ID, and an object identifier (object ID). This effectively reduces the size of the knowledge base before inferencing. This pre-encoding operation can increase the efficiency of the whole reasoning work.

Next, the processing node unit 130 performs an inverse & equality reasoning operation on the encoded knowledge base (S402). In step S402, the inverse relation and the same relation are simultaneously inferred for the encoded knowledge base.

The processing node unit 130 then performs a Transitive Reasoning operation S403 and a Refining operation S404 until no further new facts are derived. In the relational reasoning work, we apply the previously defined relational reasoning rules to induce new facts. In the relationship refining work, we check the contradictions between the inferred facts and refine the adjacency facts concisely.

The processing node unit 130 determines whether or not the triple sentence has been acquired (S405), and repeats the migration relationship reasoning operation (S403) and the relationship refining operation (S404) in order.

When the triple sentence is obtained, the processing node unit 130 performs a pre-decoding operation (S406). The processing node unit 130 converts the final result knowledge base, whose inference is completed through the decoding operation, back to the text form of the original appearance. Then, the processing node unit 130 stores the knowledge base converted into the text form again in the HDFS unit 140.

FIG. 5 is a diagram illustrating a process of a pre-encoding operation S401 of the processing node unit 130 of the massively parallel qualitative time inferencing apparatus according to an embodiment of the present invention.

Referring to FIGS. 4 and 5, the processing node unit 130 of the massively parallel qualitative time inferencing apparatus according to an embodiment of the present invention performs a pre-encoding operation with the mapping task performed first. The dictionary encoding task converts each triple statement in the form of a string contained in the knowledge base into a combination of a numeric subject ID, a predicate ID, and an object identifier (object ID). This effectively reduces the size of the knowledge base before inferencing.

For example, triple sentences in the form of long strings containing URIs such as < http: // kgu // a> <http: // kgu // after> ) Into a combination of numeric short identifiers (IDs) such as "100 41 101 &quot;. The dictionary stores the URIs of the original subject, the predicate, the object and their corresponding identifiers together.

Such an encoding operation consists of two sequential detailed MapReduce operations as shown in FIG. In the first operation 510, each triple sentence constituting the knowledge base is given, and decomposed into constituent elements such as predicate and object, and an identifier (ID) unique to the URI of each constituent element is given and registered in advance. At the same time, the URI of each component is converted into a unique identifier registered in advance.

The second task 520 combines the identifiers of the transformed components to reconstruct a concise triple sentence consisting only of the identifiers. After all the time inference tasks are completed, a decoding operation is performed to retranslate the final result knowledge base back to the original character string form. In the decoding task, the IDs included in each sentence are converted back to the original URL by using a dictionary created during the encoding operation.

6A and 6B are diagrams illustrating a process of inverse relation inference inference operation S402 of the processing node unit 130 of the massively parallel parallel inequality reasoning apparatus according to an embodiment of the present invention.

Referring to FIGS. 6A and 6B, the processing node unit 130 performs an inverse & equality Reasoning operation on the encoded knowledge base. At this time, the processing node unit 130 simultaneously infer the inverse relation and the same relation to the encoded knowledge base.

FIG. 6A shows an example of an inverse relation and an identical relation inference to be implemented by one mapping task. In this work, for example, a fact indicating the order relation of two events a and b, "a after b", represents an equal relation to "b before a" indicating an inverse relation "a equal a" and "b equal b", respectively. To do this, the map step generates a (key, value) pair having "b before a" representing the inverse relationship as a key and null as a value. In the next step, the same relations "a equal a" and "b equal b" are additionally generated for one key "b before a", and the results are output. 6b shows an inverse relation and an inference relation operation algorithm implemented by the pair of map functions and the reduction function shown in FIG. 6A.

FIGS. 7A and 7B are diagrams showing a processing procedure of the relational reasoning operation S403 of the processing node unit 130 of the massively parallel qualitative time inferencing apparatus according to an embodiment of the present invention.

Referring to FIGS. 7A and 7B, the processing node unit 130 performs a Transitive Reasoning operation (S403) and a Refining operation (S404) until new facts are no longer derived do. In the migration relation reasoning work, new rules are derived by applying the migration relation reasoning rules defined above.

Figure 7A shows an example of Transitive Reasoning implemented with one MapReduce operation. The work of relational reasoning is to derive new relational facts from two facts represented by triple sentences using the relational combination table. For example, "a overlappedby (oi) | contains (di) | startedby (si)" from two existing facts "a contains (di) b" and "b metby c ", a new disjunctive relation.

To do this, in the map step, the subject and object are extracted from each sentence, and the subject and the object are used as a key, and a flag indicating the role of the key in the original sentence is combined with the original sentence to be a value (Key, value) pairs. If the role of the key is the subject in the original sentence, the flag is denoted by s, and if it is an object, the flag is denoted by o.

That is, two (key, value) pairs such as (a, <s, a contains b>) and (b, <o, a contains b>) are created from "a contains b". Then, in the redes step, if the role flag indicates object (o) for the same key (b, <o, a contains b>) and (b, <s, b metby c> (Mi) column of the combination table when the flag indicates the subject (s), and a new inter-relationship (A) overlappedby (oi) | contains (di) | startedby (si) | c ".

FIG. 7B shows an example of a migration relation reasoning algorithm implemented by the above-described pair of map functions and a reduction function.

8A and 8B are diagrams illustrating a process of a relation refining operation (S404) of the processing node unit 130 of the massively parallel parallel qualitative inferring apparatus according to an embodiment of the present invention.

Referring to FIGS. 8A and 8B, the processing node unit 130 then performs the relational reasoning operation S403 and the relational refinement operation S404 until no more new facts are derived. In relation refinement work, we check contradictions between inferred facts and perform refinement of mutual relations facts.

Figure 8A shows an example of a relationship refining operation implemented with one MapReduce operation. Relational refinement works to check for discrepancies between inferred facts and to refine interdependent facts. For example, suppose that during the inference, two facts are derived: "a overlappedby (oi) b" and "a contains (di) | In this case, the two facts for the same two events a and b can be judged to show discrepancies or contradictions because they do not contain any satisfactory order relation at the same time.

On the other hand, for two facts, "a overlappedby (oi) b" and "a overlappedby (oi) And that two different facts are replaced by a fact "a overlappedby b" which contains only common order relations, so that knowledge can be refined more concisely.

To do this, in a relational refinement task implemented by a single MapReduce task, "a overlappedby (oi) b", "a during (d) b" (Ab, <a oi b>), which combines the subject and object of each sentence into a key and the original sentence as a vlaue, (key, value) pairs such as (ac, <adc>), (ab, <a oi | di | si b>).

Through this process, sentences with the same subject and object are grouped together. In the redusing phase, we obtain the intersection of disjunctive relations for sentences with the same key, the same subject and object, and then a more refined sentence "a overlappedby (oi) b" As a result. 8B shows a detailed process of a relation refinement algorithm implemented with a pair of map functions and a reduction function.

The present invention including the above-described contents can be written in a computer program. And the code and code segment constituting the program can be easily deduced by a computer programmer of the field. In addition, the created program can be stored in a computer-readable recording medium or an information storage medium, and can be read and executed by a computer to implement the method of the present invention. And the recording medium includes all types of recording media readable by a computer.

While the present invention has been particularly shown and described with reference to exemplary embodiments thereof, it is to be understood that the invention is not limited to the disclosed exemplary embodiments, but, on the contrary, It is possible.

110:
120: Name node unit
130: Processing node unit
140: HDFS section
150:

Claims (12)

An HDFS unit for storing a plurality of knowledge bases in the form of a Hadoop Distributed File System (HDFS);
A task management unit for assigning a task to each of the two or more processing node units; And
At least two processing node units for performing an inference operation on the knowledge base received from the HDFS unit based on the assigned task to calculate an expanded knowledge base and storing the calculated expanded knowledge base in the HDFS unit;
Lt; / RTI &gt;
Each triple statement in the form of a string contained in the received knowledge base includes a time property indicating an order relation between two time events consisting of one time point such as a point algebra theory or one time interval Massive parallel qualitative time deduction device. The method according to claim 1,
The processing node unit
Performing a pre-encoding operation on the knowledge base, performing an inverse & equality Reasoning on the encoded knowledge base, and performing Transitive Reasoning &lt; RTI ID = 0.0 &gt; Wherein the refining is performed repeatedly, and then the inferred knowledge base is stored in the HDFS unit. 3. The method of claim 2,
Wherein the processing node unit converts the knowledge base into a text form through a decoding operation, and stores the knowledge base converted into a text form into the HDFS unit. 3. The method of claim 2,
The processing node unit,
It is necessary to convert each triple statement in the form of a string included in the knowledge base into a combination of a numerical subject ID, a predicate ID, and an object identifier (object ID) Massively parallel qualitative time deduction device. 3. The method of claim 2,
Wherein the processing node unit applies the migration relationship reasoning rules defined through the migration relationship inference to the knowledge base and reduces the set of disjunctive relations through the migration relationship inference. 3. The method of claim 2,
Wherein the processing node unit checks the contradiction between the facts inferred through the relationship refining operation and refines the adjacency relationship fact. 3. The method of claim 2,
The processing node unit disassembles each constituent element including at least one of predicates and objects by giving a triple sentence constituting a knowledge base, assigns a unique identifier to the unified resource identifier (URI) of each constituent element, And concurrently reconstructing a concise triple sentence by converting an integrated resource identifier of each constituent element into a previously registered unique identifier and combining the identifiers of the converted constituent elements. The method according to claim 2, wherein
The processing node unit,
And simultaneously infer the inverse relation and the same relation to the encoded knowledge base. A parallel qualitative time inference method using the massively parallel qualitative time inference apparatus according to claim 1,
Encoding a knowledge base according to an assigned task;
Performing inverse & Equal Reasoning on the encoded knowledge base;
Repeating Transitive Reasoning and Refining until new facts are not induced; And
Storing the inferred knowledge base in the HDFS;
/ RTI &gt;
Each triple statement in the form of a character string included in the knowledge base includes a time property indicating an order relation between two time events composed of one time point such as the point logarithmic theory or one time interval A large parallel concurrency time inference method. 10. The method of claim 9,
Wherein the step of storing the inferred knowledge base in the HDFS comprises converting the knowledge base into a text form through a decoding operation and storing the knowledge base converted into a text form in the HDFS. 10. The method of claim 9,
Wherein the encoding step comprises:
It is necessary to convert each triple statement in the form of a string included in the knowledge base into a combination of a numerical subject ID, a predicate ID, and an object identifier (object ID) A method for massively parallel qualitative time inference. 10. The method of claim 9,
Wherein the task of the relationship refinement is to check contradictions between inferred facts and to refine adjacency facts.

Images