Learning device, program, and learning method

The learning device uses multi-agent reinforcement learning to automatically generate fiber rules in Assembly Calculus, addressing the challenge of determining dependency relationships in languages like Japanese, achieving accurate parsing of sentences with varying parts of speech.

JP7586862B2Active Publication Date: 2024-11-19SOFTBANK CORPORATION
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
JP2022113724
Authority / Receiving Office
JP · JP
Patent Type
Patents
Current Assignee / Owner
Filing Date
2022-07-15
Publication Date
2024-11-19
Estimated Expiration
2042-07-15

AI Technical Summary

Technical Problem

Existing parsers struggle to accurately determine dependency relationships in languages like Japanese, where the order of subjects and objects is not fixed, and manually setting fiber rules is time-consuming and difficult.

Method used

A learning device that uses multi-agent reinforcement learning to automatically generate fiber rules in Assembly Calculus, where agents learn which fibers to use for projecting dependencies based on part-of-speech transitions in sentences.

Benefits of technology

Enables accurate parsing of sentences with varying parts of speech and longer sentences, including Japanese, by learning optimal fiber rules through reinforcement learning, and can potentially be applied to other languages.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure 0007586862000002
    Figure 0007586862000002
  • Figure 0007586862000003
    Figure 0007586862000003
  • Figure 0007586862000004
    Figure 0007586862000004
Patent Text Reader

Abstract

To provide a learning device, a learning method, and a program for learning rules in AC (Assembly Calculus) using multiple pieces of learning sentence data.SOLUTION: In a system 10 in which an analysis device 100, a learning device 200, and a communication terminal 30 communicate via a network 20, the learning device 200 includes a storage unit that stores learning sentence data including learning sentences and dependencies between morphemes in the learning sentences, and a learning execution unit that learns rules in AC (Assembly Calculus) using multiple pieces of learning sentence data stored in the storage unit.SELECTED DRAWING: Figure 1
Need to check novelty before this filing date? Find Prior Art

Description

[Technical field]

[0001] The present invention relates to a learning device, a program, and a learning method. [Background technology]

[0002] Non-Patent Document 1 describes AC (Assembly Calculus), which is a computational model of the brain based on Assembly. Non-Patent Document 2 describes a technology for outputting labels of dependency relationships such as subject and object from a sentence. [Prior art document] [Non-patent literature] [Non-patent document 1] Christos H. Papadimitriou, Santosh S. Vempala, Daniel Mitropolsky, Michael Collins, and Wolfgang Maass. "Brain computation by assemblies of neurons". Proceedings of the National Academy of Sciences. 2020, 117(25):14464-14472. [Non-patent document 2] Daniel Mitropolsky, Michael J. Collins and Christos H. Papadimitriou. "A Biologically Plausible Parser". Transactions of the Association for Computational Linguistics. 2021, 9: 1374-1388 Summary of the Invention [Means for solving the problem]

[0003] According to one embodiment of the present invention, there is provided a learning device. The learning device may include a storage unit that stores learning sentence data including a learning sentence and dependencies between morphemes in the learning sentence. The learning device may include a learning execution unit that uses a plurality of the learning sentence data stored in the storage unit to learn rules in an Assembly Calculus (AC).

[0004] In the learning device, the learning execution unit may learn a Fiber rule in the AC.

[0005] In any of the learning devices, the learning execution unit may learn the rule by reinforcement learning. The learning execution unit may learn the rule by executing reinforcement learning for a plurality of part-of-speech agents, in which a transition of sentence dependencies represented by a plurality of element areas each representing an element of a sentence is set as an environment, the position of the part-of-speech agent in the environment and an element area in which an Assembly is present in the environment are set as states, and the transition of the plurality of element areas by the part-of-speech agent is set as an action. The learning execution unit may execute the reinforcement learning by giving a reward determined based on dependencies included in the training sentence data for the action of the part-of-speech agent.

[0006] According to one embodiment of the present invention, there is provided a program for causing a computer to function as the learning device.

[0007] According to one embodiment of the present invention, there is provided a learning method executed by a computer, which may include a learning execution step of learning rules in an Assembly Calculus (AC) by using training sentence data including training sentences and dependencies between morphemes in the training sentences.

[0008] The above summary of the invention does not list all of the necessary features of the present invention. Also, subcombinations of these features may also be inventions. [Brief description of the drawings]

[0009] [Figure 1] An example of a system 10 is shown diagrammatically. [Diagram 2] FIG. 1 is an explanatory diagram for explaining an Assembly Calculus. [Diagram 3] FIG. 1 is an explanatory diagram for explaining an Assembly Calculus. [Figure 4] FIG. 1 is an explanatory diagram for explaining an Assembly Calculus. [Diagram 5] FIG. 1 is an explanatory diagram for explaining an Assembly Calculus. [Figure 6] FIG. 1 is an explanatory diagram for explaining an Assembly Calculus. [Figure 7] FIG. 1 is an explanatory diagram for explaining an Assembly Calculus. [Figure 8] FIG. 2 is an explanatory diagram for explaining learning by a learning device 200. [Figure 9] FIG. 2 is an explanatory diagram for explaining a reward in reinforcement learning of the learning device 200. [Figure 10] FIG. 2 is an explanatory diagram for explaining a reward in reinforcement learning of the learning device 200. [Figure 11] An example of a Q-table 700 for each action for each morpheme (part of speech) is shown in schematic form. [Figure 12] 2 shows an example of a functional configuration of a learning device 200. [Figure 13] An example of a projection algorithm by the learning execution unit 206 is shown below. [Figure 14] 13 shows an example of a readout algorithm by the learning execution unit 206. [Figure 15] An example of an algorithm for multi-agent reinforcement learning by the learning execution unit 206 is shown below. [Figure 16] An example of the hardware configuration of a computer 1200 functioning as the learning device 200 is shown in schematic form. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0010] In Assembly Calculus, Assembly can be generated automatically. There is an operation called Projection that forms a new Assembly from one Area to another Area, and a parser that can obtain the dependency relationship between words in a sentence using this operation by Projection of Assembly has already been proposed.

[0011] This parser can output labels for dependencies such as the subject and object of a sentence. This parser includes areas that correspond to parts of speech of words, and is composed of a lexicon area that is thought to exist in the Left Medial Temporal Lobe (MTL), and areas that correspond to dependencies such as subjects and objects that are thought to be included in the Superior Temporal Gyrus (STG), and these areas are interconnected by multiple synapses called fibers. Projection is performed between areas via fibers, and a new assembly is formed from an assembly in a certain area in the projection destination area. Projection is performed through a fiber area that is inhibited from the lexicon area, and the dependencies are recorded.

[0012] It is very difficult to correctly output dependency relationships in Japanese. For example, the order in which subjects and objects appear in a sentence is not fixed in Japanese, so dependency relationships cannot always be correctly obtained from the order of parts of speech alone. In addition, even if the same particle is added after a noun, the noun may be either the subject or the object, and it may be difficult to determine the subject or object from the particle alone. The above-mentioned PRSer is based on try and error and linguistics, and manually sets fiber rules that indicate which fibers to disable or inhibit at each point in time. However, it is very time-consuming and difficult to manually set general-purpose fiber rules that can correctly output dependency labels for various types of sentences.

[0013] Therefore, the inventor first set the Brain Area required for parsing a simple Japanese sentence, constructed a Japanese parser using Projection, and then devised an algorithm to automatically learn Fiber rules. In the learning, an Agent is prepared for each type of part of speech in the Lexicon Brain Area, and using multi-agent reinforcement learning, each Agent learns which Fiber to use for Projection during parsing in order to correctly output dependencies. It is suggested that this learning algorithm may be used to realize a parser that can be applied to longer sentences and sentences with a large number of words and different types of parts of speech, and it is believed that it can also be used to learn parsers for other languages. In addition, it is believed that this learning algorithm and the learned model will lead to an elucidation of language acquisition in the brain.

[0014] The present invention will be described below through embodiments of the invention, but the following embodiments do not limit the scope of the invention according to the claims. Furthermore, not all of the combinations of features described in the embodiments are necessarily essential to the solution of the invention.

[0015] 1 illustrates an example of a system 10. The system 10 includes an analysis device 100 and a learning device 200. The analysis device 100 is a device that uses Assembly Calculus to analyze dependencies between words in an input sentence. The learning device 200 is a device that learns rules used in Assembly Calculus.

[0016] Learning device 200 provides rules generated by learning to analysis device 100. Analysis device 100 may execute Assembly Calculus using rules acquired from learning device 200. Analysis device 100, for example, analyzes dependencies of sentences received from communication terminal 30 using Assembly Calculus, and provides the analysis result to communication terminal 30.

[0017] Analysis device 100, learning device 200, and communication terminal 30 may communicate via network 20. Network 20 may include a mobile communication network. The mobile communication network may conform to any of the following communication methods: LTE (Long Term Evolution) communication method, 5G (5th Generation) communication method, 3G (3rd Generation) communication method, and 6G (6th Generation) communication method or later. Network 20 may include the Internet. Network 20 may include a LAN (Local Area Network). Analysis device 100, learning device 200, and communication terminal 30 may each be connected to network 20 by wire or wirelessly.

[0018] Analysis device 100 and learning device 200 may be integrated together. That is, learning device 200 may further include the functions of analysis device 100.

[0019] 2 to 7 are explanatory diagrams for explaining the Assembly Calculus by the analysis device 100. Here, the Assembly Calculus for analyzing English dependency will be explained.

[0020] In this example, a case where the dependency of the sentence "The man saw a woman" is analyzed will be taken as an example. In Assembly Calculus, in addition to the LEX (Lexicon) area 402, multiple sentence element areas are defined. There are many types of sentence elements, such as subjects, predicates, objects, determiners, adjectives, adverbs, and prepositions, but only the DET (Determiners) area 404, SUBJ (Subject) area 406, OBJ (Object) area 408, and VERB area 410, which are sentence elements related to the analysis of the sentence "The man saw a woman", are shown here.

[0021] When a word is input, the analysis device 100 performs projection via fibers between a plurality of areas according to the fiber rule 420 corresponding to the part of speech of the word, thereby analyzing the dependency relationship between the words. In this example, a case where the fiber rule 420 includes a pre-command and a post-command will be described as an example, but the fiber rule 420 does not have to include a post-command.

[0022] Fig. 2 shows the initial state. In this example, in the initial state, the LEX area 402, the DET area 404, the SUBJ area 406, and the VERB area 410 are turned on, and all the fibers are turned off. In Figs. 2 to 7, the names of the areas that are turned on are underlined to distinguish between the areas that are turned on and the areas that are turned off. Also, the fibers that are turned on are shown with solid lines, and the fibers that are turned off are shown with dashed lines to distinguish between them.

[0023] A case where "the" is input will be described with reference to Fig. 3. The input word is shown in the upper right corner of the LEX area 402 for convenience.

[0024] For "the", the pre-command turns on the DET area 404 and turns on the fiber between the LEX area 402 and the DET area 404. The assembly in the LEX area 402 and the assembly in the DET area 404 are connected. The post-command turns off the fiber between the LEX area 402 and the DET area 404.

[0025] 4 will be used to explain the case where "man" is input. For "man", the pre-command turns on the fiber between the LEX area 402 and SUBJ area 406, the fiber between the LEX area 402 and OBJ area 408, the fiber between the DET area 404 and SUBJ area 406, the fiber between the LEX area 402 and OBJ area 408, and the fiber between the OBJ area 408 and VERB area 410. Since the OBJ area 408 is off, the assembly in the LEX area 402 is connected to the assembly in the SUBJ area 406, and the assembly in the SUBJ area 406 is connected to the assembly in the DET area 404. The post command turns off the DET area 404, the fibers between the LEX area 402 and the SUBJ area 406, the fibers between the LEX area 402 and the OBJ area 408, the fibers between the DET area 404 and the SUBJ area 406, the fibers between the LEX area 402 and the OBJ area 408, and the fibers between the OBJ area 408 and the VERB area 410.

[0026] 5 will be used to explain the case where "saw" is input. For "saw", the pre-command turns on the fiber between the LEX area 402 and the VERB area 410, and the fiber between the VERB area 410 and the SUBJ area 406. The assembly in the LEX area 402 is connected to the assembly in the VERB area 410, and the assembly in the VERB area 410 is connected to the assembly in the SUBJ area 406. The post-command turns off the SUBJ area 406, turns on the OBJ area 408, and turns off the fiber between the LEX area 402 and the VERB area 410.

[0027] 6 explains the case where "a" is input. For "a", the pre-command turns on the LEX area 402 and the fiber between the LEX area 402 and the DET area 404. The assembly in the LEX area 402 and the assembly in the DET area 404 are connected. The post-command turns off the fiber between the LEX area 402 and the DET area 404.

[0028] 7 explains the case where "woman" is input. For "woman", the pre-command turns on the fiber between the LEX area 402 and the SUBJ area 406, the fiber between the LEX area 402 and the OBJ area 408, the fiber between the DET area 404 and the SUBJ area 406, the fiber between the LEX area 402 and the OBJ area 408, and the fiber between the OBJ area 408 and the VERB area 410. The assembly in the LEX area 402 and the assembly in the OBJ area 408, the assembly in the DET area 404 and the assembly in the OBJ area 408, and the assembly in the OBJ area 408 and the assembly in the VERB area 410 are connected. The post command turns off the DET area 404, the fibers between the LEX area 402 and the SUBJ area 406, the fibers between the LEX area 402 and the OBJ area 408, the fibers between the DET area 404 and the SUBJ area 406, the fibers between the LEX area 402 and the OBJ area 408, and the fibers between the OBJ area 408 and the VERB area 410.

[0029] Through the above process, the dependency relationship of "The man saw a woman" is analyzed. In English, the rules for the dependency relationships of sentence elements are clearly defined, so it is possible to manually generate Fiber rules 420. However, in Japanese, the rules for the dependency relationships of sentence elements are unclear because the order in which subjects and objects appear is not fixed, and it is difficult to manually generate Fiber rules 420. In the system 10 according to this embodiment, the learning device 200 generates rules through learning.

[0030] 8 is an explanatory diagram for explaining learning by the learning device 200. The learning device 200 may learn the Fiber rule in the Assembly Calculus by reinforcement learning. Here, a case will be described in which a pre-command is used but a post-command is not used.

[0031] In this example, the learning device 200 uses Q-learning as a learning algorithm. Note that the learning device 200 may use DQN (Deep Q-Network) as a learning algorithm, or may use other learning algorithms.

[0032] The learning device 200 may learn the Fiber rule by performing reinforcement learning for a multi-agent 500 including multiple part-of-speech agents, in which the transition of sentence dependencies represented by multiple element areas each representing an element of a sentence is defined as environment 510, the position of the part-of-speech agent in environment 510 and the element area in which an Assembly exists in environment 510 are defined as states 520, and the transition of the multiple element areas by the part-of-speech agent is defined as behavior 530.

[0033] The learning device 200 performs learning using a plurality of learning sentence data including Japanese learning sentences and dependencies between morphemes in the learning sentences. Here, an example will be described in which a learning sentence "I want to drink coffee" and learning sentence data including a learning sentence in which "drink" and "coffee" are in an object relationship, "drink" and "want to" are in an auxiliary verb relationship, and "coffee" and "ga" are in a case particle relationship, and a learning sentence data including a learning sentence "I want to drink" and learning sentence data including a learning sentence in which "drink" and "watashi" are in a subject relationship, "drink" and "want to" are in an auxiliary verb relationship, and "watashi" and "ga" are in a case particle relationship are used.

[0034] The multi-agent 500 includes a plurality of part-of-speech agents. FIG. 8 illustrates a noun agent 501, a verb agent 502, a case particle agent 503, and an auxiliary agent 504. The multi-agent 500 may include other parts-of-speech agents. The multi-agent 500 may also include more specific agents that classify parts of speech from various viewpoints. For example, the multi-agent 500 includes a noun (person) agent, a noun (object) agent, and a noun (place) agent that classify nouns into people, objects, places, and the like. For example, the multi-agent 500 may include an intransitive agent and a transitive agent for verbs, and may further include a potential verb agent, an auxiliary verb agent, and the like.

[0035] In the example shown in FIG. 8, the environment 510 includes a LEX area 511, and an OBJ area 512, a VERB area 513, an AUX area 514, and a CASE area 515 related to the analysis of "I want to drink coffee" and "I want to drink coffee". In this way, the learning device 200 may perform learning by including in the environment 510 only the areas of elements related to the analysis of the learning sentence used for learning, among the elements of a plurality of sentences. Then, learning may be performed for each of the different learning sentences, and the results may be merged to generate a final rule. This makes it possible to appropriately reduce the calculation load during learning. The learning device 200 may perform learning by including in the environment 510 the areas of all sentence elements. The area of ​​the elements related to the analysis of the learning sentence may be the area of ​​the elements corresponding to the dependency included in the learning sentence. The area of ​​the elements related to the analysis of the learning sentence may include the area of ​​the elements corresponding to the dependency included in the learning sentence, in addition to the area of ​​the elements corresponding to the dependency included in the learning sentence, which may be related. The likely related dependencies may be set, for example, by a user of learning device 200. Note that learning device 200 may perform learning by including the areas of all sentence elements in environment 510.

[0036] The learning device 200 gives a reward 540 to the multi-agent 500 each time the multi-agent 500 completes an action 530 for the training text. The learning device 200 gives a reward 540 to each of the multi-agents 500 using the output of Readout, which confirms which sentence element area the multi-agent 500 has transitioned to after the multi-agent 500 has completed its action 530, and the dependency relationships included in the training text data, i.e., the correct answers for the dependencies.

[0037] The learning device 200 calculates the reward R to be given to each part-of-speech agent using the following formula 1. POS(Part Of Speech) You may calculate:

[0038]

number

[0039] N is the expected number of labels (the number of correct dependency labels included in the learning sentence data), C is the number of correct labels including each morpheme among the output of Readout, and E is the number of labels in the output of Readout that exceed the expected number of labels. By configuring it to subtract E from C, it is possible to reduce the reward when an incorrect dependency is output, and more appropriate learning can be achieved.

[0040] 9 shows an example of the Readout output 610 and its reward. Here, the correct dependency labels included in the learning sentence data are "'drink', 'coffee', OBJ" indicating that "drink" and "coffee" are in an object relationship, "'drink', 'tai', AUX" indicating that "drink" and "tai" are in an auxiliary verb relationship, and "'coffee', 'ga', CASE" indicating that "coffee" and "ga" are in a case particle relationship, and the Readout output 610 is "'drink', 'coffee', OBJ" and "'drink', 'tai', AUX" are shown as examples.

[0041] Since the number of correct dependency labels included in the learning sentence data is three, N = 3. Since the number of labels in Readout output 610 is two, which is three or less, E = 0. And, since both of the two labels in Readout output 610 match the correct dependency labels included in the learning sentence data, both are correct.

[0042] For the noun agent 501, the number of correct labels containing the noun "coffee" is one, so the reward is 1 / 3. For the verb agent 502, the number of correct labels containing the verb "drink" is two, so the reward is 2 / 3. For the adp agent 503, the reward is 0, because there is no correct label containing the case particle "ga." For the aux agent 504, the number of correct labels containing the auxiliary verb "tai" is one, so the reward is 1 / 3.

[0043] 10 shows an example of the Readout output 620 and its reward. Here, the correct dependency labels included in the learning sentence data are ""drink", "coffee", OBJ", ""drink", "want", AUX", and ""coffee", "ga", CASE", and the Readout output 620 shows three examples: ""want", "coffee", OBJ", which indicates that "want" and "coffee" are in an object relationship, ""drink", "coffee", AUX", and ""coffee", "ga", CASE", which indicate that "drink" and "coffee" are in an auxiliary verb relationship.

[0044] Since the number of correct dependency labels included in the learning sentence data is three, N=3. Since the number of labels in Readout output 620 is three, which is three or less, E=0. And, of the three labels in Readout output 610, "'tai', 'coffee', OBJ" and "'nomi', 'coffee', AUX" are incorrect, and "'coffee', 'ga', CASE" is correct.

[0045] For the noun agent 501, the number of correct labels including the noun "coffee" is one, so the reward is 1 / 3. For the verb agent 502, the reward is 0, because there is no correct label including the verb "drink." For the adp agent 503, the number of correct labels including the case particle "ga" is one, so the reward is 1 / 3. For the aux agent 504, the reward is 0, because there is no correct label including the auxiliary verb "tai."

[0046] 11 shows an example of a Q-table 700 for each action for each morpheme (part of speech). Here, the Q-table 700 corresponding to the ACTION: OBJ of the noun "coffee" is shown. The horizontal axis indicates the area where the Assembly is firing, and the vertical axis indicates the current area of ​​the agent.

[0047] As the learning device 200 advances Q-learning, it updates the Q-values ​​in the Q-table 700. As learning advances with the learning sentence data of "I want to drink coffee", for example, the Q-values ​​corresponding to the Present Area "OBJ" and the Assembly Area "OBJ, CASE", the Q-values ​​corresponding to the Present Area "OBJ" and the Assembly Area "OBJ, CASE, VERB, AUX", and the Q-values ​​corresponding to the Present Area "VERB" and the Assembly Area "OBJ, CASE, VERB" increase.

[0048] The learning device 200 can update the rules of the Assembly Calculus for Japanese by updating the Q table for each action for each morpheme (part of speech).

[0049] 12 illustrates an example of the functional configuration of the learning device 200. The learning device 200 includes a storage unit 202, an input receiving unit 204, a learning execution unit 206, and a result providing unit 208.

[0050] The storage unit 202 stores various information. The storage unit 202 stores training sentence data including training sentences and dependencies between morphemes in the training sentences. The dependencies included in the training sentence data have been confirmed to be correct as dependencies for the training sentences. The storage unit 202 stores a plurality of training sentence data.

[0051] Input accepting unit 204 accepts various inputs. For example, input accepting unit 204 accepts input of learning text data. For example, input accepting unit 204 accepts learning text data input to learning device 200 by a user of learning device 200. Input accepting unit 204 accepts learning text data transmitted from a device external to learning device 200. Input accepting unit 204 stores the accepted learning text data in memory unit 202.

[0052] The training sentence data accepted by the input accepting unit 204 may include, in addition to the training sentence and dependency relations, a plurality of morphemes contained in the training sentence and each part of speech of the plurality of morphemes obtained by performing morphological analysis on the training sentence.

[0053] The learning execution unit 206 learns rules in the Assembly Calculus using multiple training sentence data stored in the storage unit 202. When the training sentence data does not include the parts of speech of each of the multiple morphemes included in the training sentence, the learning execution unit 206 may perform a morpheme analysis to identify the parts of speech of the multiple morphemes.

[0054] The learning execution unit 206 may learn the Fiber rules in the Assembly Calculus.

[0055] The learning execution unit 206 may learn rules in the Assembly Calculus by reinforcement learning. The learning execution unit 206 may use Q-learning as a learning algorithm. The learning execution unit 206 may use DQN as a learning algorithm, or may use other learning algorithms.

[0056] The learning execution unit 206 may learn rules by executing reinforcement learning for a plurality of part-of-speech agents, in which the environment is a transition of sentence dependencies represented by a plurality of element areas each representing an element of a sentence, the state is the position of the part-of-speech agent in the environment and the element area in which an Assembly exists in the environment, and the action is the transition of the plurality of element areas by the part-of-speech agent. The learning execution unit 206 may execute reinforcement learning by giving a reward determined based on the dependencies included in the learning sentence data for the action by the part-of-speech agent.

[0057] Examples of part-of-speech agents include a noun agent, a verb agent, an adjective agent, an adjectival verb agent, a particle agent, an auxiliary verb agent, an adverb agent, a conjunction agent, a conjunction agent, and an interjection agent.

[0058] The noun agents may be classified into ordinary noun agents, proper noun agents, numeral agents, formal noun agents, pronoun agents, etc. The noun agents may also be classified into noun (person) agents, noun (thing) agents, noun (place) agents, etc.

[0059] The verb agents may be classified into intransitive agents, transitive agents, potential verb agents, auxiliary verb agents, etc. The particle agents may be classified into case particle agents, conjunctive particle agents, adverbial particle agents, final particle agents, etc. The adverb agents may be classified into state adverb agents, degree adverb agents, and vocative adverb agents, etc.

[0060] The learning execution unit 206 executes Q-learning for a plurality of part-of-speech agents, in which the transition of sentence dependencies represented by a plurality of element areas is set as an environment, the position of the part-of-speech agent in the environment and the element area in which the Assembly is present in the environment are set as states, the transition of the plurality of element areas by the part-of-speech agent is set as an action, and a reward determined based on the dependencies included in the learning sentence data is given to the action by the part-of-speech agent, thereby updating the Q table for each action for each morpheme (part-of-speech). The learning execution unit 206 may generate Fiber rules in the Assembly Calculus from the updated plurality of Q tables. The learning execution unit 206 may set the updated plurality of Q tables as rules in the Assembly Calculus.

[0061] For example, the learning execution unit 206 performs learning using learning sentence data including the learning sentence "I want to drink coffee" and learning sentence data including the learning sentence "I want to drink," thereby enabling dependency analysis to be realized that distinguishes between the former, in which "coffee" is the object, and the latter, in which "I" is the subject, even though the sentence structures are the same.

[0062] The learning execution unit 206 executes learning using a large amount of learning sentence data, so that rules capable of appropriately analyzing dependency relationships can be generated even in Japanese, which has a complex dependency structure.

[0063] In addition, the learning execution unit 206 may use DQN when there is a large amount of learning sentence data or when there are many types of parts of speech to be considered. This allows learning to converge without exploding the state-action space.

[0064] Also, the learning execution unit 206 may perform learning for each subset of sentence elements and merge the results, instead of learning all of the elements of a plurality of sentences. For example, the learning execution unit 206 may use classified learning sentence data to perform Q-learning for each subset of sentence elements, thereby partially updating the Q-table, and proceed with learning so that the entire Q-table is eventually updated. This allows for effective learning of frequently occurring sentences, and allows for practical dependency learning.

[0065] Furthermore, when learning using sentences that are related to the sentences before and after, such as sentences of human conversation or sentences of a story, the learning execution unit 206 may input and learn the sentence before the sentence when learning a sentence. This makes it possible to reflect the relationship between the sentences before and after in the learning result, and to infer an implicit subject, thereby improving the accuracy of dependency analysis.

[0066] 13 shows an example of a projection algorithm by the learning execution unit 206. The learning execution unit 206 may generate a new Assembly from a certain element area to another element area according to the algorithm shown in FIG.

[0067] 14 shows an example of an algorithm for the readout by the learning execution unit 206. The learning execution unit 206 may realize the above-mentioned readout according to the algorithm illustrated in FIG.

[0068] Fig. 15 shows an example of an algorithm of multi-agent reinforcement learning by the learning execution unit 206. "project" and "readout" shown in Fig. 15 may be processes according to the algorithms shown in Fig. 13 and Fig. 14. The learning execution unit 206 may execute learning according to the algorithm exemplified in Fig. 15.

[0069] 16 is a schematic diagram showing an example of a hardware configuration of a computer 1200 functioning as the learning device 200. A program installed on the computer 1200 can cause the computer 1200 to function as one or more "parts" of the device according to the present embodiment, or cause the computer 1200 to execute operations or one or more "parts" associated with the device according to the present embodiment, and / or cause the computer 1200 to execute a process or steps of the process according to the present embodiment. Such a program can be executed by the CPU 1212 to cause the computer 1200 to execute specific operations associated with some or all of the blocks of the flowcharts and block diagrams described herein.

[0070] The computer 1200 according to this embodiment includes a CPU 1212, a RAM 1214, and a graphics controller 1216, which are connected to each other by a host controller 1210. The computer 1200 also includes input / output units such as a communication interface 1222, a storage device 1224, a DVD drive, and an IC card drive, which are connected to the host controller 1210 via an input / output controller 1220. The DVD drive may be a DVD-ROM drive, a DVD-RAM drive, etc. The storage device 1224 may be a hard disk drive, a solid state drive, etc. The computer 1200 also includes a legacy input / output unit such as a ROM 1230 and a keyboard, which are connected to the input / output controller 1220 via an input / output chip 1240.

[0071] The CPU 1212 operates according to a program stored in the ROM 1230 and the RAM 1214, thereby controlling each unit. The graphic controller 1216 acquires image data generated by the CPU 1212 into a frame buffer or the like provided in the RAM 1214 or into itself, and causes the image data to be displayed on the display device 1218.

[0072] The communication interface 1222 communicates with other electronic devices via a network. The storage device 1224 stores programs and data used by the CPU 1212 in the computer 1200. The DVD drive reads programs or data from a DVD-ROM or the like and provides them to the storage device 1224. The IC card drive reads programs and data from an IC card and / or writes programs and data to an IC card.

[0073] The ROM 1230 stores therein a boot program or the like that is executed by the computer 1200 upon activation, and / or a program that depends on the hardware of the computer 1200. The input / output chip 1240 may also connect various input / output units to the input / output controller 1220 via a USB port, a parallel port, a serial port, a keyboard port, a mouse port, and the like.

[0074] The programs are provided by a computer-readable storage medium such as a DVD-ROM or an IC card. The programs are read from the computer-readable storage medium, installed in the storage device 1224, the RAM 1214, or the ROM 1230, which are also examples of computer-readable storage media, and executed by the CPU 1212. Information processing described in these programs is read by the computer 1200, and brings about cooperation between the programs and the various types of hardware resources described above. An apparatus or method may be constructed by implementing operations or processing of information according to the use of the computer 1200.

[0075] For example, when communication is performed between the computer 1200 and an external device, the CPU 1212 may execute a communication program loaded in the RAM 1214 and instruct the communication interface 1222 to perform communication processing based on the processing described in the communication program. Under the control of the CPU 1212, the communication interface 1222 reads transmission data stored in a transmission buffer area provided in the RAM 1214, the storage device 1224, a DVD-ROM, or a recording medium such as an IC card, and transmits the read transmission data to the network, or writes reception data received from the network to a reception buffer area or the like provided on the recording medium.

[0076] Furthermore, the CPU 1212 may cause all or a necessary portion of a file or database stored in an external recording medium such as the storage device 1224, a DVD drive (DVD-ROM), an IC card, etc. to be read into the RAM 1214, and perform various types of processing on the data on the RAM 1214. The CPU 1212 may then write back the processed data to the external recording medium.

[0077] Various types of information, such as various types of programs, data, tables, and databases, may be stored in the recording medium and undergo information processing. The CPU 1212 may perform various types of processing on the data read from the RAM 1214, including various types of operations, information processing, conditional judgment, conditional branching, unconditional branching, information search / replacement, etc., as described throughout this disclosure and specified by the instruction sequence of the program, and write back the results to the RAM 1214. The CPU 1212 may also search for information in a file, database, etc. in the recording medium. For example, when a plurality of entries each having an attribute value of a first attribute associated with an attribute value of a second attribute are stored in the recording medium, the CPU 1212 may search for an entry whose attribute value of the first attribute matches a specified condition from among the plurality of entries, read the attribute value of the second attribute stored in the entry, and thereby obtain the attribute value of the second attribute associated with the first attribute that satisfies a predetermined condition.

[0078] The above-described programs or software modules may be stored in a computer-readable storage medium on the computer 1200 or in the vicinity of the computer 1200. In addition, a recording medium such as a hard disk or a RAM provided in a server system connected to a dedicated communication network or the Internet can be used as a computer-readable storage medium, thereby providing the programs to the computer 1200 via the network.

[0079] The blocks in the flowcharts and block diagrams in the present embodiment may represent stages of a process in which an operation is performed or "parts" of an apparatus responsible for performing the operation. Particular stages and "parts" may be implemented by dedicated circuitry, programmable circuitry provided with computer-readable instructions stored on a computer-readable storage medium, and / or a processor provided with computer-readable instructions stored on a computer-readable storage medium. The dedicated circuitry may include digital and / or analog hardware circuits, and may include integrated circuits (ICs) and / or discrete circuits. The programmable circuitry may include reconfigurable hardware circuits, such as, for example, field programmable gate arrays (FPGAs), programmable logic arrays (PLAs), and the like, including AND, OR, XOR, NAND, NOR, and other logical operations, flip-flops, registers, and memory elements.

[0080] A computer-readable storage medium may include any tangible device capable of storing instructions that are executed by a suitable device, such that a computer-readable storage medium having instructions stored thereon comprises an article of manufacture that includes instructions that can be executed to create means for performing the operations specified in the flowcharts or block diagrams. Examples of computer-readable storage media may include electronic storage media, magnetic storage media, optical storage media, electromagnetic storage media, semiconductor storage media, and the like. More specific examples of computer-readable storage media may include floppy disks, diskettes, hard disks, random access memories (RAMs), read-only memories (ROMs), erasable programmable read-only memories (EPROMs or flash memories), electrically erasable programmable read-only memories (EEPROMs), static random access memories (SRAMs), compact disk read-only memories (CD-ROMs), digital versatile disks (DVDs), Blu-ray disks, memory sticks, integrated circuit cards, and the like.

[0081] The computer readable instructions may include either assembler instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, state setting data, or source or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk®, JAVA®, C++, etc., and conventional procedural programming languages ​​such as the “C” programming language or similar programming languages.

[0082] Computer readable instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus, or to a programmable circuit, either locally or over a local area network (LAN), a wide area network (WAN), such as the Internet, etc., to cause the processor of the general purpose computer, special purpose computer, or other programmable data processing apparatus, or to a programmable circuit, to execute the computer readable instructions to generate means for performing the operations specified in the flowcharts or block diagrams. Examples of processors include computer processors, processing units, microprocessors, digital signal processors, controllers, microcontrollers, etc.

[0083] Although the present invention has been described above using the embodiments, the technical scope of the present invention is not limited to the scope described in the above embodiments. It is clear to those skilled in the art that various modifications and improvements can be made to the above embodiments. It is clear from the description of the claims that such modifications and improvements can also be included in the technical scope of the present invention.

[0084] It should be noted that the order of execution of each process, such as operations, procedures, steps, and stages, in the devices, systems, programs, and methods shown in the claims, specifications, and drawings is not specifically stated as "before" or "prior to," and that the process may be performed in any order unless the output of a previous process is used in a later process. Even if the operational flow in the claims, specifications, and drawings is described using "first," "next," etc. for convenience, it does not mean that the process must be performed in this order. [Explanation of symbols]

[0085] 10 system, 20 network, 30 communication terminal, 100 analysis device, 200 learning device, 202 memory unit, 204 input reception unit, 206 learning execution unit, 208 result provision unit, 402 LEX area, 404 DET area, 406 SUBJ area, 408 OBJ area, 410 VERB area, 420 Fiber rule, 500 multi-agent, 501 noun agent, 502 verb agent, 503 adp agent, 504 aux agent, 510 environment, 511 LEX area, 512 OBJ area, 513 VERB area, 514 AUX area, 515 CASE area, 520 state, 530 action, 540 reward, 610 Readout output, 620 Readout output, 700 Q table, 1200 computer, 1210 host controller, 1212 CPU, 1214 RAM, 1216 graphic controller, 1218 display device, 1220 input / output controller, 1222 communication interface, 1224 storage device, 1230 ROM, 1240 input / output chip

Claims

1. a storage unit for storing training sentence data including training sentences and dependencies between morphemes in the training sentences; a learning execution unit that learns rules in an AC (Assembly Calculus) using the plurality of learning sentence data stored in the storage unit; Equipped with The learning execution unit is a learning device that learns the rules by performing reinforcement learning for a plurality of part-of-speech agents, in which the environment is the transition of sentence dependencies represented by a plurality of element areas each representing an element of a sentence, the state is the position of the part-of-speech agent in the environment and the element area in which an Assembly is located in the environment, and the transition of the plurality of element areas by the part-of-speech agent is the action.

2. The learning device according to claim 1 , wherein the learning execution unit learns a Fiber rule in the AC.

3. 2. The learning device according to claim 1, wherein the learning execution unit executes the reinforcement learning by giving a reward for an action by the part-of-speech agent, the reward being determined based on a dependency included in the training sentence data.

4. A program for causing a computer to function as the learning device according to any one of claims 1 to 3.

5. 1. A computer-implemented method of learning, comprising: A learning execution stage in which rules in an AC (Assembly Calculus) are learned using learning sentence data including learning sentences and dependencies between morphemes in the learning sentences. Equipped with The learning execution stage is a learning method in which the rules are learned by executing reinforcement learning for a plurality of part-of-speech agents, with the environment being the transition of sentence dependencies represented by a plurality of element areas each representing an element of a sentence, the state being the position of the part-of-speech agent in the environment and the element area in which an Assembly exists in the environment, and the behavior being the transition of the plurality of element areas by the part-of-speech agent.

Citation Information

Patent Citations

  • Dependence structure information processor, dependence structure information processing method, its program and recording medium

    JP2007080117A