JP2014149786A - Natural language analysis processing device, method, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To accurately perform natural language analysis processing for each of successively input character strings by restricting an increase in an amount of calculation.SOLUTION: A problem for language analysis processing for one input unit is decomposed into minimal portion problems by a minimal portion problem creating section 22. Sets of minimal portion problems are formed. S number of groups composed by sets of portions of the all the sets of the minimal portion problems are formed by a group forming section 26, and are assigned to S number of calculation nodes 30. In each calculation node 30, the processes of a parameter updating section 32, a portion decomposition update section 34, a first flag determination section 36, a synchronization section 38, a limiting condition update section 40, and a second flag determination section 42 are repeated until the value of a limiting condition parameter is determined to have converged to the optimum value.

Description

  The present invention relates to a natural language analysis processing device, method, and program, and more particularly, to a natural language analysis processing device, method, and program for performing natural language analysis processing on an input document.

  Natural language is a language normally used by humans, such as Japanese and English. A technique for grammatically and semantically analyzing a document described in a natural language has an academic significance from a linguistic viewpoint, for example, to understand the origin and structure of the language. In recent years, various services that automatically analyze grammatically and semantically documents generated by humans and use the results have been developed mainly on the web. For example, services such as a translation site, a reputation analysis site for people and products, and a summary site for a specific event correspond to this. In these services, digitized documents generated by humans are grammatically and semantically analyzed in the system, and the actual services are provided by using them. In this sense, grammatical and semantic analysis of natural language is the core technology of these services, and has come to occupy a very important position in the information processing field.

  Generally speaking, natural language analysis includes everything from surface analysis such as word breaks and part-of-speech estimation to more advanced analysis such as estimation of dependency relationships between words and clauses. For example, there are “sentence breaks” that estimate sentence breaks from documents, “word breaks” that estimate word breaks, part-of-speech assignment that estimates word part-of-speech, and dependency analysis that estimates the relationship between words and clauses. . These examples are shown in FIG.

  Language analysis problems such as part-of-speech assignment and dependency analysis are classified as problems with “interdependence”. This is because, for example, in the problem of determining the part of speech of a word, when determining the part of speech of a word, there is a relationship in which the part of speech to be given to the target word is determined by the surrounding parts of speech and the surrounding parts of speech. And surrounding parts of speech are also problems that depend on the part of speech before and after that. If there is a rule that part of speech B should not appear after a part of speech A, there is no guarantee that if the part of speech of each word is estimated independently, a result satisfying this rule can be obtained. . That is, if the part of speech of each word is estimated independently, the consistency as a whole cannot always be obtained, which causes a problem of searching for the best part of speech string for the entire document. In particular, considering solving problems such as sentence breaks, word breaks, part-of-speech estimation, and dependency analysis, the problem of searching for the best output is a very difficult problem because it is a very complex problem. It is.

  In order to solve these problems, methods of estimating grammatical and semantic analysis results using algorithms such as history-based sequential analysis (so-called stack decoding), dynamic programming, and integer programming are widely used. (Non-patent documents 1 to 3).

Joakim Nivre. Algorithms for deterministic incremental dependency parsing.Computational Linguistics, 2008. F. Sha and F. Pereira. Shallow Parsing with Conditional Random Fields. Proc. Of HLT / NAACL, 2003. Dan Roth, Wen-tau Yih Global Inference for Entity and Relation Identi_cation via a Linear Programming Formulation Introduction to Statistical Relational Learning, 2007.

  The history-based sequential analysis method is a method for performing the next estimation using the estimation results obtained immediately before (see FIG. 10). Compared with the method of solving the problem independently, the output consistency can be maintained. However, since this is a method of deciding the next analysis result by believing in the analysis result up to the previous one, if there is an error in the previous analysis result, the analysis will continue according to the error, so the propagation of the error There is a problem that happens.

  As a method for solving the problem, a method based on dynamic programming (see FIG. 11) and a method based on integer programming have come to be used. Since this is a method for selecting the best output as a whole, there is no problem of an analysis error in the middle compared to the history-based sequential analysis method. However, there is a problem that the amount of calculation for selecting the best output as a whole becomes very large as compared with the sequential analysis method.

  In particular, a system is assumed in which text is sequentially input, such as streaming input, in which processing is sequentially performed in the order in which the texts are input and results are sequentially output. For example, applications using natural language analysis include machine translation, document proofreading, and online illegal harmful information monitoring. Assuming a situation where such an application is processed online and in real time, there is a simultaneous interpretation system in particular for translation, and there is an application such as online document proofing when creating a non-native language document in document proofreading. In order to realize such a service, it is an important requirement that text can be processed sequentially in real time.

In a situation where sequential processing corresponding to sequential input is required, the history-based method is most suitable. Conversely, when using the integer programming problem method to perform sequential processing with an analysis method, at present, optimization must be performed again every time an input is entered, which is extremely difficult from a computational viewpoint. There is a problem.

  In general, methodologies that perform overall optimization such as integer programming have a higher analysis system than methodologies that perform history-based analysis, but as described above, overall optimization such as integer programming However, the sequential processing has a problem in terms of computational complexity.

  The present invention has been made in order to solve the above-described problems, and it is possible to perform natural language analysis processing with high accuracy on an input character string that is sequentially input while suppressing an increase in the amount of calculation. An object is to provide a language analysis processing apparatus, method, and program.

  In order to achieve the above object, a natural language analysis processing device according to a first aspect of the present invention provides a language analysis process for an input character string obtained by concatenating each of character strings consisting of at least one character that is sequentially input. A natural language analysis processing apparatus that performs natural language analysis processing including: a part that performs the language analysis processing on a predetermined character unit or character string basis on a problem of performing the language analysis processing on the input character string A partial problem generating unit that sequentially generates the partial problem when decomposed into problems for each input in character units or character string units, and a portion that stores a set of partial problems generated sequentially by the partial problem generating unit Each time the partial problem is successively generated by the problem storage means, the partial problem generation means, S calculation nodes (S is a natural number of 2 or more) for performing the natural language analysis processing, Each time the subproblem is generated by the subproblem generating means, the subproblem is stored so that each subproblem belongs to at least two groups. Creating S groups composed of arbitrary subsets of all sets and assigning them to the S computation nodes, each of the S computation nodes being sent by the group creation means Update a solution of each subproblem of the subset based on a constraint condition that the solutions of the minimal subproblems belonging to the two or more groups match for a subset of the assigned subset of the subproblem A partial decomposition updating means for performing the partial decomposition update of the subset updated by the partial decomposition updating means. Group inactive state setting means for setting a group inactive state indicating that update processing by the partial update means of the computation node is not performed when all the solutions of the partial problems of When the solution of the subproblem updated by the partial decomposition updating means does not match the solution of the subproblem updated last time, the solution for matching the solution of the subproblem according to the constraint condition The constraint condition active state setting means for canceling the constraint condition inactive state, which is set for the constraint condition parameter and indicates that the update process of the constraint parameter is not performed, and the updated by the partial decomposition update means Notifying the other computation nodes of the solution of each subproblem of the subset, and each subproblem of the subset notified from the other computation node A synchronization means for obtaining a solution; a solution for each partial problem of the subset of the other computation node obtained by the synchronization means; and a solution for each partial problem of the subset updated by the partial decomposition update means; The constraint condition updating means for updating the constraint condition parameter of the partial problem for each partial problem, and the constraint condition parameter updated by the constraint condition updating means for each partial problem was updated last time A group active state setting means for canceling the group inactive state set for the computing node to which the group to which the partial problem belongs is assigned when the value does not match the value of the constraint parameter; The constraint condition parameter updated by the constraint condition update unit is the same as the constraint condition parameter updated last time. The constraint condition inactive state setting means for setting the constraint condition inactive state for the constraint parameter of the partial problem, and whether the value of the constraint parameter has converged or not Until the value of the constraint parameter is determined to have converged, update by the partial decomposition update unit, setting by the group inactive state setting unit, setting by the constraint active state setting unit, synchronization unit A convergence determination unit that repeats notification and acquisition by, update by the constraint condition update unit, setting by the group active state setting unit, and setting by the constraint condition inactive state setting unit, The computation node in which the group inactive state is set does not perform the update by the partial decomposition update unit, and does not update the group. The synchronization means of the calculation node set to the inactive state notifies the solution of each partial problem of the subset last updated by the partial decomposition update means to other calculation nodes, and the constraint The condition update means is configured as a natural language analysis processing device that does not update the constraint parameter for which the constraint condition inactive state is set.

  The natural language analysis method according to the second invention includes a subproblem generation means, a subproblem storage means, S calculation nodes (S is a natural number of 2 or more), and a group creation means, and is input at least sequentially. A natural language analysis processing method in a natural language analysis processing apparatus for performing natural language analysis processing including language analysis processing on an input character string obtained by concatenating each of character strings composed of one character, the partial problem The partial problem when the problem of performing the language analysis processing on the input character string is decomposed into a partial problem of performing the language analysis processing in a character unit or character string unit defined in advance by the generation unit For each input in units or character strings, it is generated sequentially, and the partial problem storage means stores a set of partial problems generated sequentially by the partial problem generation means. Each time the subproblem is sequentially generated by the subproblem generating means, the subproblem is set so that each subproblem belongs to at least two groups. And creating S groups composed of arbitrary subsets for the entire set of the subproblems and assigning them to the S computation nodes, and performing the natural language analysis processing by the S computation nodes, Performing the natural language analysis processing by each of the S computation nodes may be performed by the partial decomposition updating unit with respect to a subset of the subset of the partial problem assigned by the group creating unit with respect to the two or more subsets. The solutions of each subproblem of the subset are updated based on the constraint condition that the solutions of the minimal subproblems belonging to the group of And when the solutions of the partial problems of the subset updated by the partial decomposition updating means by the group inactive state setting means all coincide with the solutions of the partial problems of the subset updated last time A group inactive state indicating that the update process by the partial update unit of the calculation node is not performed, and the partial condition update unit updates the partial problem by the constraint condition active state setting unit. When the solution of the subproblem does not match the solution of the subproblem updated last time, the constraint condition parameter is set as a solution when matching the solution of the subproblem according to the constraint, The constraint condition inactive state indicating that the update process of the constraint parameter is not performed is released, and updated by the partial decomposition update unit by the synchronization unit Notifying the other computation nodes of the solution of each partial problem of the subset that has been obtained, obtaining the solution of each partial problem of the subset notified from the other computation node, by the constraint condition update means, Based on the solution of each subproblem of the subset of the other computation node obtained by the synchronization means and the solution of each subproblem of the subset updated by the partial decomposition update means, The constraint condition parameter of the partial problem is updated, and the constraint condition parameter updated by the constraint condition update unit for each partial problem is updated by the group active state setting unit with the value of the constraint condition parameter updated last time. The group inactivity set for the computing node to which the group to which the partial problem belongs is assigned. When the constraint condition parameter updated by the constraint condition update unit matches the value of the constraint parameter updated last time for each partial problem by the constraint condition inactive state setting unit In addition, the constraint condition inactive state is set for the constraint parameter of the partial problem, and it is determined whether or not the value of the constraint parameter has converged by a convergence determination unit, and the value of the constraint parameter is converged Update by the partial decomposition update unit, setting by the group inactive state setting unit, setting by the constraint condition active state setting unit, notification and acquisition by the synchronization unit, update by the constraint condition update unit, The setting by the group active state setting unit and the setting by the constraint inactive state setting unit are repeated. The computation node in which the group inactive state is set is not updated by the partial decomposition updating means, and the group inactive state is set in the natural language analysis processing method including The synchronization means of the calculation node notifies the solution of each partial problem of the subset last updated by the partial decomposition update means to other calculation nodes, and the constraint condition update means The constraint parameter for which the state is set is not updated.

  According to the first and second inventions, the problem of performing a natural language analysis process on an input character string that is sequentially input is divided into a set of subproblems to create a group, and belongs to two or more groups Based on the constraints that match the solution of the subproblem, each compute node updates the solution for a subset of the assigned subproblem depending on whether or not a group of computation nodes is inactive. Using the solution of the subproblem obtained from other calculation nodes, it is calculated by repeating updating the constraint parameter until convergence, depending on whether or not the constraint parameter is inactive. The natural language analysis process can be performed with high accuracy on the input character string that is sequentially input while suppressing an increase in the amount.

  Further, according to the first and second inventions, the partial decomposition updating means is a Lagrange undetermined multiplier updated last time for a subset of the partial problems of the group assigned by the group creating means, The Lagrange undetermined multiplier is updated so as to optimize the value of a predetermined objective function using the solution of each subproblem of the subset and the constraint parameter, and the updated Lagrange undetermined multiplier is And updating the solution of each subproblem of the subset so as to optimize the value of the objective function, and the synchronization means updates the Lagrange undetermined multiplier and the subset updated by the partial decomposition update means. The solution of each subproblem is notified to another computation node, and the Lagrange undetermined multiplier and the portion notified from the other computation node. Each constraint sub-problem is obtained, and the constraint condition update means is based on the Lagrange undetermined multiplier of the other computation node obtained by the synchronization means and the solution of each sub-problem of the subset. The constraint parameter may be updated for each subproblem so as to optimize the value of the objective function.

  According to a third aspect of the present invention, there is provided a natural language analysis processing apparatus for performing a natural language analysis process including a language analysis process on an input character string obtained by concatenating each of character strings composed of at least one character that is sequentially input. The language analysis processing device, wherein the problem of performing the language analysis processing on the input character string is decomposed into a partial problem of performing the language analysis processing in a predefined character unit or character string unit For each input of the character unit or the character string unit, a partial problem generation unit that sequentially generates, a partial problem storage unit that stores a set of partial problems generated sequentially by the partial problem generation unit, and the partial problem generation Each time the subproblem is sequentially generated by the means, S calculation nodes (S is a natural number of 2 or more) for performing the natural language analysis processing, and the subproblem generating means Each time a partial problem is generated sequentially, an arbitrary portion of the partial problem set is stored in the partial problem storage means so that each partial problem belongs to at least two groups. A group creating means for creating S groups composed of sets and assigning to the S computing nodes, a constraint condition updating means, a group active state setting means, a constraint condition inactive state setting means, and a convergence determination Each of the S computing nodes is a subset of the subset of the sub-problems assigned by the group creating means, and each of the minimal sub-problems belonging to the two or more groups. Based on the constraint condition that the solutions match, the partial decomposition updating means for updating the solution of each partial problem of the subset, and the partial decomposition updating means When the solution of each subproblem of the newly updated subset is consistent with the solution of each subproblem of the subset updated last time, the update processing by the subupdate means of the calculation node is not performed A group inactive state setting means for setting a group inactive state indicating that, and for each partial problem, the solution of the partial problem updated by the partial decomposition update means is the same as the solution of the partial problem updated last time. If not, the constraint condition inactive state is set for the constraint parameter that is a solution when matching the solution of the subproblem according to the constraint condition, and indicates that the constraint parameter update process is not performed. The constraint condition update means includes a constraint condition active state setting means for canceling the solution, and the constraint condition update means includes a solution of each partial problem of the subset of each computation node and the partial decomposition update method. And updating the constraint parameter of the partial problem for each partial problem based on the solution of each partial problem of the subset updated by the group active state flag assigning means, If the constraint parameter updated by the constraint condition update means does not match the value of the constraint parameter updated last time, the calculation node assigned to the group to which the partial problem belongs is set. The group inactive state is canceled, and the constraint condition inactive state setting means sets the constraint condition parameter updated by the constraint condition update means for each partial problem to be equal to the previously updated constraint parameter value. The constraint condition inactive for the constraint parameter of the subproblem The convergence determination means determines whether or not the value of the constraint parameter has converged, and updates by the partial decomposition update means until determining that the value of the constraint parameter has converged, the group Setting by the inactive state setting unit, setting by the constraint condition active state setting unit, notification and acquisition by the synchronization unit, update by the constraint condition update unit, setting by the group active state setting unit, and the constraint condition inactive state A natural language analysis processing apparatus that repeats setting by a setting unit, wherein the calculation node in which the group inactive state is set does not perform update by the partial decomposition update unit, and the constraint condition update unit performs the constraint The condition inactive state is set The constraint parameter is not updated, and the group inactive state is set The solution of each subproblem of the subset in the computation node is configured as a natural language analysis processing device that uses the solution of each subproblem of the subset last updated by the partial decomposition update means.

  A natural language analysis method according to a fourth aspect of the present invention is a partial problem generation means, a partial problem storage means, S (S is a natural number of 2 or more) calculation nodes, a group creation means, a constraint condition update means, a group active state Natural including a language analysis process for an input character string obtained by concatenating each of character strings composed of at least one character sequentially input, including a setting unit, a constraint condition inactive state setting unit, and a convergence determination unit A natural language analysis processing method in a natural language analysis processing apparatus for performing language analysis processing, wherein the problem of performing the language analysis processing on the input character string by the partial problem generation means is defined in character units or characters defined in advance. The partial problem when decomposed into partial problems for performing the language analysis processing in units of columns is sequentially generated for each input of the character unit or character string unit, and the partial problem storage A set of subproblems sequentially generated by the subproblem generating means is stored in each stage, and each time the partial problems are sequentially generated by the partial problem generating means by the group generating means, With respect to the stored set of subproblems, S groups are formed by forming arbitrary S subsets with respect to the entire set of subproblems so that each subproblem belongs to at least two groups. The S calculation nodes perform the natural language analysis processing, update by the constraint condition update means, set by the group active state setting means, and by the constraint condition inactive state setting means Setting and determining by the convergence determining means, each of the S computation nodes by the natural language The analysis process is performed by the partial decomposition updating unit, and the partial sub-problem belonging to the two or more groups is solved for a subset of the partial problem of the group assigned by the group creating unit. Based on the matching constraints, update the solution of each partial problem of the subset, and the solution of each partial problem of the subset updated by the partial decomposition update means by the group inactive state setting means, When all the solutions of the partial problems of the subset updated last time are consistent with each other, set a group inactive state indicating that update processing by the partial update unit of the calculation node is not performed, and the constraint condition For each subproblem by the active state setting means, the solution of the subproblem updated by the partial decomposition update means is updated last time. If the solution of the subproblem is not consistent with the solution of the subproblem, the update of the constraint parameter set for the constraint parameter that is a solution when matching the solution of the subproblem according to the constraint is not performed. The constraint condition updating means includes: releasing the constraint inactive state for expressing, for each partial problem based on a solution of each partial problem of the subset updated by the partial decomposition update means of each computation node Updating the constraint parameter of the subproblem, and the group active state setting means, for each subproblem, the constraint parameter updated by the constraint condition update means is the same as the constraint parameter updated last time. If it does not match the value, it is set for the computation node to which the group to which the subproblem belongs is assigned. The group inactive state is canceled, and the constraint inactive state setting unit is configured such that, for each partial problem, the constraint parameter updated by the constraint condition update unit matches the value of the constraint parameter updated last time. If it does, the constraint condition inactive state is set for the constraint parameter of the partial problem, the convergence determination means determines whether the value of the constraint parameter has converged, and the constraint condition Until it is determined that the value of the parameter has converged, update by the partial decomposition update unit of each calculation node, setting by the group inactive state setting unit, setting by the constraint active state setting unit, notification and acquisition by the synchronization unit , Update by the constraint condition update means, setting by the group active state setting means, and the constraint condition inactive state A natural language analysis processing method including repeating setting by a setting unit, wherein the calculation node in which the group inactive state is set does not perform update by the partial decomposition update unit, and the constraint condition update unit includes: The solution of each partial problem of the subset in the computation node in which the group inactive state is set without updating the constraint parameter in which the constraint inactive state is set is the partial decomposition. Let it be the solution of each subproblem of the subset last updated by the updating means.

  According to the third and fourth inventions, the problem of performing a natural language analysis process on an input character string that is sequentially input is divided into a set of subproblems to create a group, and belongs to two or more groups Based on the constraints that match the solution of the subproblem, each calculation node updates the solution based on the active state or inactive state set for the group for each assigned subproblem subset, and each calculation Using the solution of the subproblem updated at the node, the update of the constraint parameter is repeated until convergence based on the active state or inactive state set in the constraint parameter, thereby increasing the amount of calculation. The natural language analysis processing can be performed with high accuracy for each input character string that is sequentially input.

  According to a third aspect of the present invention, the partial decomposition updating unit is configured to update a Lagrange undetermined multiplier updated last time and each part of the subset with respect to the subset of the partial problem of the group assigned by the group generation unit. Updating the Lagrange undetermined multiplier so as to optimize the value of the objective function determined in advance using the solution of the problem and the constraint parameter, and using the updated Lagrange undetermined multiplier, the objective function The solution of each subproblem of the subset is updated so as to optimize the value of the subset, and the constraint condition update means updates the Lagrange undetermined multiplier and the subset updated by the partial decomposition update means of each computation node Updating the constraint parameter for each subproblem to optimize the value of the objective function based on the solution of each subproblem It can also be so.

  Moreover, the program of this invention is a program for functioning a computer as each means which comprises the natural language analysis processing apparatus of Claims 1-4.

  As described above, according to the natural language analysis processing device, method, and program of the present invention, an increase in the amount of calculation is suppressed, and natural language analysis processing is performed with high accuracy on input character strings that are sequentially input. It can be carried out.

It is a figure which shows the example of the difference between a "sequential processing system" and a "batch processing system". It is a figure for demonstrating the method to produce the group of a minimum subproblem. It is the schematic which shows the structure of the natural language analysis processing apparatus which concerns on the 1st Embodiment of this invention. It is a flowchart which shows the content of the natural language analysis processing routine in the calculation node of the natural language analysis processing apparatus which concerns on the 1st Embodiment of this invention. It is the schematic which shows the other example of the natural language analysis processing apparatus which concerns on the 1st Embodiment of this invention. It is the schematic which shows the structure of the natural language analysis processing apparatus which concerns on the 2nd Embodiment of this invention. It is the schematic which shows the structure of the language analysis control apparatus which concerns on the 2nd Embodiment of this invention. It is the schematic which shows the structure of the language analyzer which concerns on the 2nd Embodiment of this invention. It is a figure which shows the example of word break, part-of-speech tagging, specific expression extraction, phrase break, and dependency between phrases. It is a figure which shows the history-based sequential analysis method in a prior art. It is a figure which shows the example of the analysis based on the dynamic programming in a prior art.

  Hereinafter, the outline of the invention will be described in detail.

<Outline of the invention>
In the present invention, it is proposed in Non-Patent Document 4 (Andre FT Martins, Mario AT Figueiredo, Pedro MQ Aguiar, Noah A. Smith, Eric P. Xing An Augmented Lagrangian Approach to Constrained MAP Inference Proc. Of ICML, 2011.). Sequential analysis is realized while maintaining the global optimality using dual decomposition + extended Lagrangian relaxation (DD-ADMM) method, which is one methodology for solving the integer programming problem. A major feature of the DD-ADMM method is a method of solving a problem to be actually solved by dividing it into small groups. Further, as an important property obtained by this feature, each divided group can be solved independently of other groups, and there is an advantage that calculation is relatively easy.

  In addition, because of the nature of being able to solve independently, when solving individual groups, the analysis results of other groups are not directly affected, so the groups generated by future input are virtually unknown. However, it is possible to obtain an analysis result partially.

  In the present invention, the calculation cost is reduced by reusing the previous processing result for the sequential input. Further, a group active / inactive state control unit is introduced to control whether each group should be calculated using the calculation cost or not when searching for the optimum solution.

  As an embodiment of the present invention, it is assumed that a Japanese morphological analyzer used in an online real-time Japanese document proofreading system outputs a sequential analysis result as a human inputs a document.

  What is input at one time by sequential input is a character string after one kana-kanji conversion is confirmed. Therefore, the minimum is one character, and the longest is one document. In general, a character string input at a time by sequential input is about 1 to 3 words or about one phrase. Since the character string input at one time has a variable length, the character string input at one time is referred to as “one input unit”. Processing is performed as follows.

-Start state: No input (initial state)
Process 1: The user inputs one input unit to the system. Process 2: The system accepts one input unit and analyzes it together with the text input so far.
-Process 3: Output the obtained analysis result-Process 4: Wait for the next input-(Repeat process 1-4)
-End status: User input is complete (no explicit end declaration required)

  The present invention corresponds to a processing unit that presents an optimal natural language analysis result sequentially for one input unit that is sequentially input in the above processing 2. Hereinafter, a processing method corresponding to the above processing 2 will be described.

  First, a natural language analysis processing device by “sequential processing” targeted by the present invention will be defined. Here, as a comparison with the “sequential processing method”, a method of starting processing after all input text has been input to the end is referred to as a “batch processing method”. In the “sequential processing method” and the “batch processing method”, it is assumed that finally inputted documents are completely the same. The difference in processing method varies depending on the input method.

  In the sequential processing method, a situation is assumed in which one input unit is divided into several times and partially inputted at each time. The difference between the “sequential processing method” and the “batch processing method” is shown in FIG.

  Assume that one input unit is finally input until time T (0 ≦ t ≦ T). Let t be the current time, t-1 one hour before that, and t = 0 be a state in which no input is made (initial state).

  Here, as a point of caution, if T = 1, it means that the entire input has been input at one time, and this means that the processing is the same as the batch processing method. That is, the sequential processing method is a processing framework that indirectly includes the batch processing method. If the sequential processing method is possible, there is an advantage as a method in which the sequential processing method can cope with the batch processing method.

  On the contrary, it is difficult as a methodology to cope with sequential processing by the batch processing method. However, in practice, as a naïve application method, it means that the batch processing method is repeatedly performed for each input with respect to one input unit that is sequentially input, so that generally a computational difficulty occurs.

  As can be seen from the comparison between the sequential processing method and the batch processing method so far, the requirement as the sequential processing method is a method that can efficiently process the processing result of the previous time effectively. . For example, the conventional history-based method assumes that the analysis result up to the previous time is correct, and uses the analysis result to perform processing for one input unit newly entered at the current time. Yes.

  The present invention is based on a batch processing method as a processing method, but even when the whole is not being input, an input unit newly input at the current time while obtaining an optimal solution at that time Corresponds to a processing method for efficiently obtaining a new solution by using the processing result of the previous time. Next, the collective processing method as the basis of the present invention will be described.

  In the present invention, the minimum unit of processing is determined in advance as a basic idea. The minimum unit of processing corresponds to what can be defined as the smallest problem in defining a problem in morphological analysis processing. Since the problem of the minimum unit corresponds to the part of the problem to be solved, the problem of the minimum unit is referred to as “minimal subproblem”. The index of the minimal subproblem is represented by r. Also, let R denote the set of indexes for all the minimal subproblems. Therefore, the problem to be solved is composed of | R | An example of a subproblem is a minimal subproblem.

Next, considering one input unit as one input, morphological analysis is defined by a set of minimal subproblems. Therefore, the solution space of these problems is represented by a set Y of vectors. One output obtained when an input is given is represented by a vector Y∋z. Here, the solution of the minimal sub-problem corresponds to one element z _i of the vector.

  Next, in the present invention, a group is composed of several minimal subproblems, and the problem is expressed by a set of groups. At this time, each group can include the minimum subproblem in an overlapping manner. The method of creating a group is arbitrary, but the following two points should be satisfied. The first point is to form a group so as to cover the entire problem to be solved, and the second point is that each local subproblem belongs to a plurality of groups. Here, the created group is considered as a partial problem for the entire problem. Therefore, one group solves as one optimization problem.

Here, it is assumed that S groups are created. the subspace of the solution space included in s-th group and Y _s. That is, Y _s ⊆Y. At this time, let z _s (r) be the solution of the rth minimal subproblem included in the sth group. An example of creating a group is shown in FIG.

Let R _s be the set of indices for the minimal subproblem included in the sth group. Next, let R _{s be a set} that satisfies R ⊆ U _{s = 1} ^S R ′ _s and R _s = R ′ _s ∩R. Although this is not a minimal subproblem, there is an auxiliary subproblem that is effective in solving the whole problem, so this is an extension to cope with the case where it is desired to form a group including these subproblems. That is, here, the index included in the set corresponding to R ′ _s \ R is not a minimal subproblem but corresponds to an index of an auxiliary subproblem effective in solving the problem.

For the output obtained in the s-th group, let f _s be a function for calculating the goodness of output. Here, the value calculated under the assumption that the larger the value of f _s (z _s ) is, the better the output is. At this time, the whole problem for performing the natural language analysis processing can be formulated as an optimization problem expressed by the following equation (1).

Here, z _s is a set of solutions of the minimal subproblem included in the group s. Z is a complete set of z _s (r) representing a solution of a minimal subproblem having a value of 0 or 1. This optimization problem into S groups independently a problem of selecting the best a partial solution that seems at output z _{s (r)} for the minimum subproblem each group is responsible, partial solution of each group u This corresponds to solving as a constraint condition that it matches the vector of (r). The point of this formulation lies in the part of the constraint condition z _s (r) = u (r), and in each group s, when the rth minimal subproblem is included, all values of z _s (r) are u ( This means that the condition for the solution is to match r). That is, in each group, the solution z _s (r) is determined independently, but can be limited to a solution that can be consistent as a whole by the constraint condition z _s (r) = u (r). That is. An objective function for optimization is obtained using the constraint equation and Lagrangian relaxation. The above equation (1) means obtaining a set of solutions z _s (r) of the minimal subproblem that maximizes the score defined by an arbitrary score function f _s . However, it is a condition that the problem to be solved can be divided into arbitrary S groups. That is, the solution having the maximum score obtained by the linear sum of the divided S groups is the solution.

In an actual morphological analysis problem, for example, regarding word division, if there is a division candidate point between characters, it is divided at the division candidate point (z _s (r) = 1) or not (z _s). It can be regarded as an integer programming problem for determining (r) = 0). In the case of POS tagging, a problem of selecting one from part of speech defined in advance each word, it assigns a solution of one minimum subproblems z _{s (r)} in each part of speech, one from among them This is equivalent to solving an integer programming problem with the constraint that 1 and all others are 0. Similarly to part-of-speech assignment, reading and original shape estimation are problems of selecting one of readings and original candidates of each word, and assigning a solution z _s (r) of one minimal subproblem to each candidate This is equivalent to solving an integer programming problem with the constraint that one of them is 1 and all others are 0. A solution that is most suitable for the whole of such a division or assignment problem is selected as a result of the morphological analysis.

  When actually solving an integer programming problem, an optimization objective function of the following equation (2) is obtained by using (extended) Lagrangian relaxation.

Next, it is described in Non-Patent Document 5 (Stephen Boyd, Neal Parikh, Eric Chu, Borja Peleato, and Jonathan Eckstein. Distributed optimization and statistical learning via the alternating direction method of multipliers. Foundations and Trends in Machine Learning, 2011.) As shown, the augmented Lagrangian term ρ / 2 (z _s (r) -u (r)) ² is added and the problem is transformed into a quadratic form as shown in the following equation (3). Make the problem easier to solve.

This is an optimal value and is consistent with the original problem. In order to simplify the subsequent expression expansion, λ _s (r) = − ρα _s (r) and the following expression (4) is obtained.

The actual solution is obtained by the following iterative process.
Initialization: Variable k = 0 for managing repetition
Processing 1: Update Lagrange undetermined multiplier α ^k _s (r) Processing 2: Estimate the solution z ^k _s (r) of each local subproblem independently for each local subproblem Processing 3: Restriction parameter (each update and processing ^u k (r) is the average of the solution of the minimum partial problem) 4: When I was convergence determination convergence u ^ (r) ⁼ and ^u k (r). If not, k = k + 1 and the process returns to step 1.
-Output: u ^ is output.

  Hereinafter, the processes 1 to 4 will be described in detail.

First, processing 1 will be described. As processing 1, Lagrange undetermined multiplier α ^s (r) of each computation node is updated. When z and u are fixed, the direction of the optimum value of each α _s (r) is a partial differential direction with respect to α _s (r) of the objective function L (z, u, α).

  The following update formula is obtained from the relationship of the formula (5).

  The update formula of the above equation (6) can be calculated independently for each minimal subproblem.

Next, the process 2 will be described. The process 2 is a process for estimating the solution z ^k _s (r) of each minimal subproblem. When α and u are fixed at the time of the iterative calculation k, each z _s ( The optimal solution of r) is the problem of finding z _s (r) that maximizes the objective function L (z, u, α).

Here, it is assumed that f _s is the following equation (8).

Here, θ _s (r) in the above equation (8) represents a score regarding r in the group s, and this value is given from the outside. In other words, it is treated as a constant in this optimization. According to the definition, when only the term related to z is extracted from the above equation (4), the following equation (9) is obtained.

  The calculation of equation (7) can also be calculated for each group s. That is, the maximization problem of the following equation (11) may be solved for each group s.

  Next, processing 3 will be described. As processing 3, the constraint parameter u (r) is updated. The optimal solution for u when z and α are fixed is that the partial differential value for u (r) becomes 0 with respect to the objective function L (z, u, α). From the relationship, the following equation (12) is obtained.

That is, u means the sum of the average of z _s (r) and the average of α _s (r). For updating the constraint parameter u (r), all α _s (r) and z _s ( r) is required.

Next, the process 4 will be described. As the process 4, it is determined whether u obtained in the process 3 is an optimal value (has converged). Two small positive real numbers ε ₁ and ε ₂ are given, and it is determined that convergence has occurred when the following equations (13) and (14) are satisfied (see Non-Patent Document 5).

  If the convergence is not determined in the convergence determination, the process returns to process 1 with k = k + 1. If it is determined that it has converged to the optimum value, the iterative process is terminated and the solution u is output. The above is the description of the batch processing.

  Next, a description will be given of a method for realizing the sequential processing by formulating the above-mentioned language analysis problem by an integer programming problem and solving the problem by dual decomposition + extended Lagrangian relaxation.

  In the batch processing described above, all optimization variables are initialized to 0 or the like and the optimization is started. However, in the present invention, in order to efficiently cope with sequential input, the processing result at the previous time is initialized. Value. As a summary, with regard to the processing results up to the previous time, for the part that is not affected by the analysis of one input unit newly input at the current time, the analysis result is the same as that obtained at the previous time, The analysis result of the part having a dependency relationship with one input unit newly input at the current time is processed again. By such a method, the result of optimization at the previous time can be efficiently reused, so that the actual processing time can be greatly reduced compared to the method of performing optimization every input.

  Such an efficient processing can be effectively introduced by the property of “the individual subproblem can be solved independently” of the dual decomposition + Lagrange relaxation method. On the other hand, integer programming solvers and the like do not have such a property, and are not suitable for processing a solution that is partially, efficiently, and sequentially corrected. When the above equation (1) is expanded for sequential processing, it can be expressed as the following equation (15).

The total number of groups that are generated by the text input until the time t is written as S _t. Further, let S _T be the total number of groups that are finally generated when the input is made up to the end at the final time T. However, this is the (1) and S value equal S = S _T in formula, and that it matches the total number of groups obtained when given an input in a batch when no sequential processing.

At this time, the total score of the group newly generated at time t

Is the same as the conventional history-based method. That is, the difference from the history-based method is that the maximum value of the scores of all the groups up to the current time t is obtained while simultaneously taking into account the maximum value of the scores of the groups generated up to the time t−1. . However, unless otherwise devised, this is equivalent to simply executing the batch processing method repeatedly at each time. For this reason, in the present invention, the processing results up to time t-1 are used well, so that it is possible to perform sequential processing with a much smaller amount of calculation than simply executing repeatedly.

  First, as the simplest method, a method of using the state obtained at the optimum solution at time t-1 as the initial value at time t can be considered. If the optimal solutions at time t-1 and time t are close, this method is considered to work intuitively and well. However, since variables and constraints are increased or decreased due to time changes, it is not always obvious whether the final state at the previous time is simply the initial state. Also, depending on the optimization algorithm, the initial value may not work very effectively.

However, since the optimization algorithm based on dual decomposition used in the present invention is independent of other constraints and variables handled in each constraint and its variables as compared with other optimization algorithms, the time from time t−1 to time The optimization variable z _s (r) of the group that increases or decreases when it changes to t has an advantage that it can be easily added / deleted without considering consistency with other variables. In addition, since the algorithm for obtaining the optimal solution is an iterative calculation with fixed variables, it can be expected that the effect of using the state at time t−1 is also large because the algorithm gradually approaches the optimal solution.

Therefore, as a procedure, when changing from time t-1 to time t and shifting from integer programming problem P _t-1 to integer programming problem P _t , first of the constraints (partial problems) with P _t-1 Get the difference in perspective. At this time, regarding the addition, z _s (r) = 0, α _s (r) = 0, and if r is newly added, u (r) = 0 is set and a new optimization variable is generated. . Conversely, at the time of deletion, all z _s (r) and α _s (r) related to _s are deleted, and if the deleted z _s (r) is an element related to the last u (r), u (r ) Is also deleted. It is assumed that the groups other than these increased and decreased groups and the equality constraint variables take over the final state of P _t−1 .

  Non-Patent Document 6 (Terry Koo, Alexander M. Rush, Michael Collins, Tommi Jaakkola, and David Sontag. Dual decomposition for parsing with non-projective head automata. In Proceedings of the 2010 Conference on Empirical Methods in Natural Language Processing, pages 1288 -1298, Cambridge, MA, October 2010. Association for Computational Linguistics. And Non-Patent Document 7 (Andre Martins, Noah Smith, Mario Figueiredo, and Pedro Aguiar. Dual decom-position with many overlapping components. In Proceedings of the 2011 Con -ference on Empirical Methods in Natural Language Processing, pages 238-249. Association for Computational Linguistics, July 2011.), focusing on the fact that many optimization variables continue to be fixed at constant values during optimization. A method of improving the optimization speed by removing such variables from the optimization is adopted. In the present invention, such a method is further developed, and a more precise optimization variable state control method is proposed. As a result, the proposed state control method is effective in a situation where optimization variables and constraints change dynamically, such as the dynamic change integer programming problem sequence taken up by the present invention.

First, as a mechanism for reducing unnecessary calculation during optimization, a flag F _s having a binary state of “active” and “inactive” for each group s and each variable u (r) of the equality constraint, to grant the F _r. “Active” state (F _s = 1 or F _r = 1) means that s or r performs normal optimization processing, and “inactive” state (F _s = 0 or F _r = 0) ) Means that s or r is excluded from the calculation of the optimization process.

First, the algorithm expressed by the above equations (6) to (14) can be expressed by two types of calculation blocks, the above equations (6) to (11) and the above equations (12) to (14). , Each can be rewritten in the form of a single sentence. In the processing of the above formulas (6) to (11), u (r) is calculated as a constant, so that only independent variables z _s (r) and α _s (r) are variables for each group s. It can be calculated in units of group s. On the other hand, in the processing of the above formulas (12) to (14), z _s (r) is calculated as a constant, and therefore can be calculated in units of variable u (r) based on equality constraints. The flags of F _s and F _r are switched under the following conditions.

・ F _s = 1 → F _s = 0
It is assumed that when all variables z _s (r) included in the corresponding group s become z ^k _s (r) = z ^k−1 _s (r), the group s of its own has converged. Thereafter, as long as F _s = 0 continues, the processing represented by the above expressions (6) to (11) for the group s is not performed.

・ F _r = 1 → F _r = 0
When u ^k (r) = u ^k−1 (r) for the corresponding r, it is assumed that its variable u (r) has converged. Thereafter, as long as F _r = 0 continues, the processing of the above expressions (12) to (14) is not performed on u (r). However, the convergence determination value is cached, and even if the inactive state continues, the cached value is used to appropriately calculate the convergence determination value.

  The conditions after the above two states are set relatively loosely. Therefore, there is a possibility of shifting to an inactive state even when the signal is not converged aggressively. In order to avoid this problem, a process for returning from the inactive state to the active state is shown below.

・ F _s = 0 → F _s = 1
Let R _s ∋r. When the constraint parameter u ^k−1 (r) ≠ u ^k (r), that is, the constraint parameter u (r) that is shared with the solution z _s (r) of the minimal subproblem included in the group s is updated. Means that. That is, since the solution of the group does not match the constraint parameter u (r), the group s is returned to the active state in order to estimate the optimum value again.

・ F _r = 0 → F _r = 1
When z _s (r) in a certain group s is z ^k−1 _s (r) ≠ z ^k _s (r), that is, the value of the constraint parameter u (r) may change. In order to evaluate the optimal solution again, the active state is set.

  The point of active state and inactive state is that when you deactivate a flag, you can only change your own flag, and when you activate it, you change only a flag in a different category from your own. It is a point that cannot be done. This is designed so that the state control does not get stuck.

  In this way, by introducing the process of changing from the inactive state to the active state, it becomes possible to cope with restrictions due to time changes of dynamic change integer programming problem sequences and increase / decrease of optimization variables. In other words, since it has a framework that can easily transition from the inactive state to the active state, even if the surrounding conditions change, such as increase / decrease of the optimization variable, optimization is performed with simple flag management and minimum calculation be able to.

<First Embodiment>
<System configuration>
The natural language analysis processing apparatus 100 according to the first embodiment of the present invention receives a character string of one input unit to be sequentially applied as a natural language analysis processing target as a sequential input, and connects the sequentially input character strings. The result of the Japanese morphological analysis, which is the result of the natural language analysis processing for the input character string, is sequentially output. The natural language analysis processing apparatus 100 is configured by a computer including a CPU, a RAM, and a ROM that stores a program for executing a natural language analysis processing routine to be described later, and functionally as described below. It is configured. As shown in FIG. 3, the natural language analysis processing apparatus 100 includes an input unit 10, a calculation unit 20, and an output unit 50.

  The input unit 10 sequentially receives a character string of one input unit that is sequentially input. Further, the input unit 10 sequentially receives the calculation node number S and the parameter ρ input manually.

The computing unit 20 includes a minimal subproblem generator 22, a minimal subproblem storage unit 24, a group creation unit 26, S calculation nodes 30 ₁ to 30 _s, and a parameter storage unit 46. In addition, when showing arbitrary calculation nodes among calculation nodes 30 ₁ to 30 _s , they are referred to as calculation nodes 30.

  The minimal subproblem generation unit 22 sequentially inputs a predefined minimal subproblem obtained by dividing a problem of performing language analysis processing on an input character string obtained by concatenating character strings sequentially input by the input unit 10. A set of minimal subproblems is generated for each input input character string.

  The minimal subproblem storage unit 24 stores a set of minimal subproblems sequentially generated by the minimal subproblem generation unit 22.

Each time the minimal subproblem is sequentially generated by the minimal subproblem generating unit 22, the group creating unit 26 has at least two minimum subproblems for each set of minimal subproblems stored in the minimal subproblem storage unit 24. S groups are created so as to belong to these groups, and assigned to the S computation nodes 30 ₁ to 30 _s . In addition, the group creation unit 26 notifies the input parameter ρ to each of the _S computation nodes 30 ₁ to 30 _S. It should be noted that a group is created so that each minimum subproblem stored in the local subproblem storage unit 24 belongs to the same group as the previously assigned group.

Each of the S calculation nodes 30 ₁ to 30 _s includes a parameter update unit 32, a partial decomposition update unit 34, a first flag determination unit 36, a synchronization unit 38, a constraint condition update unit 40, and a second flag. A determination unit 42 and a convergence determination unit 44 are provided. Each S parameter update unit 32, partial decomposition update unit 34, first flag determination unit 36, synchronization unit 38, constraint condition update unit 40, second flag determination unit 42, convergence determination unit 44, However, processing units having similar functions are denoted by the same numbers. In addition, each of the S calculation nodes 30 ₁ to 30 _s includes a parameter update unit 32, a partial decomposition update unit 34, and a first flag determination unit each time a minimal partial problem is sequentially generated by the minimal partial problem generation unit 22. 36, the synchronization unit 38, the constraint condition update unit 40, the second flag determination unit 42, and the convergence determination unit 44 perform each process. The first flag determining unit 36 is an example of a group inactive state setting unit and a constraint condition active state setting unit, and the second flag determining unit 42 is a group active state setting unit and a constraint condition inactive state setting. It is an example of a means.

The parameter update unit 32 updates the Lagrange undetermined multiplier α _s (r) for each minimal subproblem r included in the own group s assigned by the group creation unit 26. Specifically, the Lagrange undetermined multiplier α _s (r) is updated according to the above equation (6) for each minimal subproblem r. Since the update formula (6) can be calculated independently at each calculation node 30, it is not necessary to communicate with the other calculation nodes 30. The initial value of the Lagrange multipliers α _{s (r)} is the value of the Lagrangian obtained as a result of the execution of the natural language analysis processing routine of the previous stored in the parameter storage unit 46 undetermined multipliers α _{s (r)} To do. When the natural language analysis processing routine has never been executed, the initial value of the Lagrange undetermined multiplier α _s (r) is set to zero. When the group flag F _s = 0 of the own group s, the parameter update unit 32 does not update the Lagrange undetermined multiplier α _s (r) for each local subproblem r, and the initial value or the previously updated value As Lagrange undetermined multiplier α _s (r).

The partial decomposition updating unit 34 updates the solution z _s (r) of each minimal subproblem in the calculation node 30 _{s with} respect to the subset of the minimal subproblem r included in the own group s assigned by the group creation unit 26. . Specifically, the solution of each local subproblem is updated by solving the maximization problem shown in the above equation (11). Note that the initial value of the solution z _s (r) of each minimal subproblem is the solution z _{s of} each minimal subproblem obtained as a result of the previous execution of the natural language analysis processing routine stored in the parameter storage unit 46. The value of (r) is assumed. If the natural language analysis processing routine has never been executed, the initial value of the solution z _s (r) of each minimal subproblem is set to zero. Further, when the group flag F _s = 0 of the own group s, the partial decomposition update unit 34 does not update the solution z _s (r) of each local subproblem, and the initial value or the value updated last time As the value of the solution z _s (r) of each minimal subproblem.

The first flag determination unit 36 calculates the solution z ^k _s (r) of each local subproblem included in the own group s obtained by the partial decomposition update unit 34 as the solution z ^k−1 _s (r) obtained last time. It is determined whether or not it matches. As the case for all minimum partial problems included in the self group _{^{s z k s (r) =}} z k-1 s (r), a group s assigned to compute node 30 _s has converged, group flag of its own group s a value of 0 representing non-activation imparts to F _s. Further, for each minimal subproblem r included in the own group s, when there is a minimal subproblem r satisfying z ^k _s (r) ≠ z ^k−1 _s (r), the constraint condition parameter of the minimal subproblem r Assuming that the value of u (r) may change, a value of 1 _indicating activation is _assigned to the constraint flag F _r for u (r).

The synchronization unit 38 transmits the Lagrange undetermined multiplier α _s (r) and the solution z _s (r) of each local subproblem that are updated or held this time at the calculation node 30 to all the calculation nodes 30 _i other than itself. Notice. In addition, the synchronization unit 38 receives the Lagrange undetermined multiplier α _s (r) and the solution z _s (r) of each local subproblem that have been notified from all the other computation nodes 30 _{i and} are updated or held this time. By this processing, each computation node 30 can obtain the value of Lagrange undetermined multiplier α _s (r) and the solution z _s (r) of each local subproblem that all computation nodes 30 _s have.

The constraint condition update unit 40 uses the Lagrange undetermined multiplier α _s (r) received from all the other computation nodes 30 _s and the solution z _s (r) of each local subproblem, according to the above equation (12). The constraint condition parameter u (r) indicating the solution when the solutions of the minimal subproblems are made to coincide with each other is updated for each of the minima partial problems r. The initial value of u (r) is the value of u (r) obtained as a result of the previous execution of the natural language analysis processing routine stored in the parameter storage unit 46. If the natural language analysis processing routine has never been executed, the initial value of the constraint parameter u (r) is set to zero. When the constraint flag F _r = 0, the constraint condition update unit 40 does not update u (r), and holds the initial value or the value updated last time as the value of u (r).

  In addition, although u (r) is obtained independently at each computation node 30, the obtained u (r) matches at all computation nodes 30. As a processing method, u (r) may be calculated by one arbitrary calculation node 30 and then notified to each calculation node 30. In this case, however, the other calculation nodes need to wait until the calculation of the selected calculation node 30 is completed and the result is notified. In the present embodiment, a case will be described as an example where a method of performing the same calculation in each calculation node 30 is taken.

The second flag determination unit 42 determines whether or not u (r) obtained by the constraint condition update unit 40 matches u ^k−1 (r) for each of the minimal partial problems r. In the case of u ^k (r) = u ^k−1 (r), the variable u (r) of its own has converged, and a value of 0 representing deactivation is given to the constraint flag F _r . When u ^k (r) ≠ u ^k−1 (r), the value of the constraint equation u (r) is updated, and the solution of the group including the minimal subproblem r is the constraint parameter u (r). Therefore, a value of 1 indicating activation is _assigned to the group flag F _s of the group s including the minimal subproblem r.

The convergence determination unit 44 determines whether the parameter u obtained by the constraint condition update unit 40 has converged to an optimum value using the above equations (13) and (14). Specifically, given two small real numbers ε ₁ and ε ₂ and satisfying the above equations (13) and (14), it is determined that u has converged to an optimum value.

  In the convergence determination, if it has not converged to the optimum value, k = k + 1 is set, and the process returns to the parameter updating unit 32. If it is determined that the value has converged to the optimum value, the iterative process is terminated.

  This convergence determination process may also be performed by any one computation node 30 in the same manner as u (r) and the result may be notified to the whole. However, since synchronization processing is necessary, in this embodiment, an example will be described in which all the convergence determinations are individually performed by the calculation node 30 and the processing is terminated if the convergence is determined. The result of this convergence determination is always the same at all the computation nodes, so even if the determination is made here, the result is the same.

When it is determined that u (r) has converged to the optimum value, the convergence determination unit 44 combines the solutions u (r) of the respective minimal subproblems r obtained at that time to generate a morphological analysis result. The output unit 50 outputs the result as a result of the natural language analysis process. Further, the solution u (r) of each minimal subproblem r obtained at that time, the solution z _s (r) of each minimal subproblem r included in the group s updated in each computation node 30, and each computation node 30 Each value of the Lagrange undetermined multiplier α _s (r) of the group s updated in step S is stored in the parameter storage unit 46. In the present embodiment, the case where the result of the natural language analysis process is output from any one calculation node 30 has been described as an example. However, the result of the natural language analysis process is output from all the calculation nodes 30. May be.

Parameter storage unit 46, the solution of the minimum subproblem r obtained in the calculation node 30 u (r), the solution of the minimum subproblem r contained in the updated group s z _{s (r),} the updated group The value of the Lagrange undetermined multiplier α _s (r) of _s is stored.

<Operation of natural language analysis processing device>
Next, the operation of the natural language analysis processing apparatus 100 according to the first embodiment of the present invention will be described. First, every time a character string of one input unit to be subjected to natural language analysis processing is input to the natural language analysis processing device 100, one character string input by the minimal subproblem generation unit 22 in the natural language analysis processing device 100. A minimal subproblem for performing a natural language analysis process on a character string in an input unit is generated and added to a set of minimal subproblems stored in the minimal subproblem storage unit 24. Each time a minimal subproblem is generated, the group creation unit 26 creates S groups for the set of minimal subproblems stored in the minimal subproblem storage unit 24, and S computing nodes 30. assigned to the _{1 ~30 s.}

Each time a minimal problem is sequentially generated, the result of the previous execution of the natural language analysis processing routine stored in the parameter storage unit 46 as an initial value for each of the constraint parameter u ⁰ (r). Each obtained value is set. When the natural language analysis processing routine has never been executed, 0 is set as an initial value for each of the constraint parameter u ⁰ (r). Further, 0 is set as an initial value in the constraint parameter u ⁰ (r) for the newly generated minimal subproblem r. Further, the value of the group flag F- _s of each group s is set to 1 representing the activation state, and the value of the constraint flag F _r of the constraint condition parameter u (r) for each minimal subproblem r is activated. Set to a value of 1 representing the conversion state. Then, the natural language analysis processing routine shown in FIG. 4 is executed by each computation node 30 of the natural language analysis processing apparatus 100. Hereinafter, a case where the calculation is executed by the calculation node 30 _s will be described.

First, in step S100, an initial value is set for each Lagrange undetermined multiplier α ⁰ _s (r) and a solution z ⁰ _s (r) of each minimal subproblem for a subset of the assigned minimal subproblem r of the own group s. As a result, the values obtained as a result of the previous execution of the natural language analysis processing routine stored in the parameter storage unit 46 are set. If the natural language analysis processing routine has never been executed, 0 is set as the initial value for each Lagrange undetermined multiplier α ⁰ _s (r) and the solution z ⁰ _s (r) of each local subproblem. Set. Further, 0 is set as an initial value for each of the Lagrange undetermined multiplier α ⁰ _s (r) and the solution z ⁰ _s (r) of the minimal subproblem for the newly generated minimal subproblem r.

  Next, in step S102, an initial value 1 is set to a variable k indicating the number of repetitions.

Next, in step S104, it is determined whether or not the value of the group flag F- _s of the own group s is zero. If the value of the group flag _{F s} of the own group s is 0, the process proceeds to step S118, (the case of 1) when the value of the group flag _{F s} is not 0, the process proceeds to step S106.

Next, in step S106, the parameter update unit 32 causes the Lagrange undetermined multiplier α ⁰ _s (r) set in step S100, the solution z ⁰ _s (r) of each local subproblem, and the constraint parameter u ⁰ ( r), or Lagrange undetermined multiplier α ^k−1 _s (r) updated last time, based on the solution z ^k−1 _s (r) of each minimal subproblem and the constraint parameter u ^k−1 (r) The Lagrange undetermined multiplier α ^k _s (r) for each of the minimal subproblems r in the own group s is updated according to the equation (6).

Next, in step S108, the Lagrangian undetermined multiplier α ^k _s (r) updated in step S106 by the partial decomposition updating unit 34 and the solution z ⁰ _s (r) of each local subproblem set in step S100 are processed. ), or on the basis of the previous respective minimum subproblems updated solutions ^{z _k-1} s (r), according to the above (11), the solution to each of the minimum partial problems r in the own group s ^z _k s (r ).

Next, in step S110, the first flag determination unit 36 determines all the solutions included in the own group s based on the solution z ^k _s (r) of each local subproblem of the own group s updated in step S108. It is determined whether or not the solution z ^k _s (r) of the minimal subproblem r is the same value as the value of z ^k−1 _s (r) updated last time. If they are the same value, the process proceeds to step S112. If they are not the same value, the process proceeds to step S114.

Next, in step S112, to impart a value of 0 representing the inactive state to the group flag F _s for its own group s.

Next, in step S114, each minimal subproblem of the own group s is determined by the first flag determination unit 36 based on the solution z ^k _s (r) of each minimal subproblem of the own group s updated in step S108. r of the solution ^z _k s (r) determines whether there is a solution ^z _k s minimum subproblems r not the same value as the value of ^{z _k-1} s was last updated (r) (r). If there is a solution z ^k _s (r) of the minimal sub-problem r that is not the same value, the process proceeds to step S116, and otherwise, the process proceeds to step S118.

Next, in step S116, with respect to the constraint parameter u (r) for the minimal subproblem r of the solution z ^k _s (r) of the minimal subproblem r of the own group s that is not the same value as the previously updated solution in step S114. A value of 1 representing the activation state is _assigned to the constraint flag _Fr.

Next, in step S118, the synchronization unit 38 determines the Lagrange undetermined multiplier α ^k _s (r) of the own group s updated in step S106 and the local sub-problems of the own group s updated in step S108. solution ^z _k s a (r) notifies the other computing nodes 30, from all the other computing nodes 30 _i, the updated Lagrange multipliers alpha ^k _i (r), and the solution ^z _{k i} of each minimum subproblem (R) is acquired (i = 1,..., S−1, s + 1,... S). If the processing of step S106 and step S108 is omitted because the own group s is in an inactive state, the Lagrange undetermined multiplier α ^k of the own group s held at that time is stored. ⁻¹ _s (r), and the solution z ^k−1 _s (r) of each minimal subproblem, the ^kth Lagrange undetermined multiplier α ^k _s (r), and the solution z ^k _s (r) of each minimal subproblem ) As a value of).

Next, in step S120, the Lagrange undetermined multiplier α ^k _s (r) updated in step S106 of the own group s by the constraint condition updating unit 40 or the Lagrange undetermined multiplier notified to the other calculation node 30 in step S118. α ^k _s (r) and the solution z ^k _s (r) of each local subproblem updated in step S108 of the own group s or the solution of each local subproblem notified to the other calculation nodes 30 in step S118 Based on z ^k _s (r), the Lagrange undetermined multiplier α ^k _i (r) obtained from all the other computation nodes 30 _i in step S118, and the solution z ^k _i (r) of each local subproblem, Only the constraint parameter u ^k (r) having a constraint flag F _r value of 1 is updated according to the above equation (12). New. For the constraint parameter u (r) in which the value of the constraint flag F _r is 0, the value of u ^k−1 (r) held at that time is set to the kth constraint parameter u ^k. (R).

Next, in step S122, the constraint condition parameter u ^k (r) is the same as the previously updated constraint parameter u ^k−1 (r) for each of the minimal subproblems r by the second flag determination unit 42. It is determined whether it is a value. For u ^k (r) for the minimal sub-problem r having the same value, the processing in step S124 described later is executed, and for u ^k (r) for the minimal sub-problem r not having the same value, it will be described later. The process in step S126 is executed.

Next, in step S124, each constraint flag F _r for the constraint parameter u ^k (r) having the same value as the previously updated constraint parameter u ^k−1 (r) is set to a value of 0 indicating an inactive state. Give.

Next, in step S126, the second flag determination unit 42 determines at least one minimal subproblem r regarding the constraint parameter u ^k (r) that is not the same value as the constraint parameter u ^k−1 (r) updated last time. to each group flag F _s for the group s containing, it imparts a value representing the active state.

Next, in step S128, based on the solutions z ^k _s (r) of all the minimal sub-problems r updated or notified and all the constraint parameters u ^k (r) updated or held. In accordance with the above equations (13) and (14), it is determined whether or not all of the constraint parameter u ^k (r) have converged to the optimum value. If the above expressions (13) and (14) are not satisfied, it is determined that they have not converged, the process proceeds to step S132, the variable k is incremented by 1, and the process returns to step S104. On the other hand, if the above expressions (13) and (14) are satisfied, it is determined that convergence has occurred, and the process proceeds to step S132.

Next, in step S132, the Lagrange undetermined multiplier α ^k _s (r) notified and acquired in step S118 of the own group s, the solution z ^k _s (r) of the minimal subproblem r, and the final in step S120 All the constraint parameter u ^k (r) updated or held in the parameter storage unit 46 is stored in the parameter storage unit 46. Also, the result of the natural language analysis process is generated using the parameter u ^k (r) of the constraint condition that is finally updated or held in step S120, and the result is output from the output unit 50. End the routine.

  As described above, according to the natural language analysis processing apparatus according to the first embodiment, the problem of performing the natural language analysis processing on the sequentially input character string is divided into a set of partial problems. A group is created and assigned to each computation node according to whether or not the inactive state is set for the group of computation nodes based on the constraint condition that the solutions of the subproblems belonging to two or more groups match. Update the constraint parameter according to whether the inactive state is set for the constraint parameter using the solution of the partial problem obtained from another computation node. By repeating this process until convergence, an increase in the amount of calculation can be suppressed, and natural language analysis processing can be performed with high accuracy on input character strings that are sequentially input.

  Natural language analysis is generally formulated as a discrete optimization problem, and it is assumed that obtaining an optimal solution is equivalent to obtaining an analysis result of natural language. By solving as an integer programming problem using the extended Lagrangian relaxation method, the natural language analysis problem can be decomposed into subproblems and solved independently, and the natural language analysis problem can be decomposed into subproblems and solved independently. Active / inactive control can be realized by utilizing the property of being able to do so.

  Further, by performing the process of changing from the active state to the inactive state, it is possible to perform the optimization by skipping the process of the group whose value does not change, and the calculation amount can be greatly reduced. Conversely, by introducing a process for changing from the inactive state to the active state, for example, when the optimal solution changes from the optimal solution of the previous time due to a change in time and a new input, the non-active state has been A value that has not been updated in the active state can be updated by making it active. In other words, this active / inactive state control makes it possible to find the optimal solution with the minimum necessary calculation, and to cope with the change of the optimal solution due to time changes with the minimum calculation cost. Is possible.

  Also, even in an environment where input is passed sequentially, such as streaming input, it is possible to continue processing while presenting the optimal solution for the input that has entered so far, so real-time information such as simultaneous interpretation is possible. It becomes possible to provide the basic technology for realizing the processing system. In addition, since the analysis accuracy is greatly improved as compared with the history-based method, the performance of an application using natural language processing can be raised. Further, by reusing the previous processing result and active / inactive control, redundant calculation costs in the sequential processing are reduced, and efficient sequential processing becomes possible. Further, since the obtained solution is the optimum solution of the input text, the best analysis result at that time is sequentially obtained.

  Further, by performing sequential analysis, optimization can be performed even in a situation where only a part of the optimization variables of the problem to be solved is known, and an optimal solution can be obtained in the situation of only the partial variables. In addition, the optimization solution of the integer programming problem to be solved can be used gradually as time changes, and the optimal solution of the next time can be efficiently reused by reusing the optimal solution of the previous time. Can be obtained. As a result, the optimal solution that is finally obtained as a whole problem basically matches the situation in which the number of optimization variables increases sequentially and the situation in which the optimization problem is solved as usual in a lump. It is equivalent to processing.

In the above embodiment, the case where each calculation node 30 includes the constraint condition update unit 40, the second flag determination unit 42, and the convergence determination unit 44 has been described as an example. However, the present invention is not limited to this. is not. For example, as illustrated in FIG. 5, each calculation node 30 includes a parameter update unit 32, a partial decomposition update unit 34, and a first flag determination unit 36, and the calculation unit 20 includes a constraint condition update unit 40, a second flag You may comprise so that the determination part 42 and the convergence determination part 44 may be provided one each. In this case, the constraint condition updating unit 40 uses the Lagrange undetermined multiplier α ^k _s (r) obtained at all the computation nodes 30 _s and the solution z ^k _s (r) of each local subproblem to update the parameter ^u k (r), on whether the second flag determination unit 42 is the same as the value of ^u values obtained ^u k (r) was last updated or maintained ^k-1 (r) Based on this, a value representing an active state or an inactive state is assigned to the target F _r and F _s , and the convergence determination unit determines whether the obtained constraint parameter u ^k (r) has converged to an optimal value. To do. If the convergence is not determined in the convergence determination, the constraint condition parameter u ^k (r) obtained for each computation node 30 is notified and the process returns to the process by the parameter updating unit 32.

<Second Embodiment>
Next, a second embodiment will be described. In addition, about the part which becomes the structure similar to 1st Embodiment, the same code | symbol is attached | subjected and description is abbreviate | omitted.

  In the second embodiment, in the distributed parallel computing environment provided with a plurality of language analysis devices connected by a network, the parameter is updated by the distributed parallel computation by the plurality of language analysis devices. This is different from the embodiment.

As shown in FIG. 6, the natural language analysis processing system 200 according to the second embodiment includes a language analysis control device 201 and S language analysis devices 202 _{1 to} 202 _S. The language analysis control device 201 and the S language analysis devices 202 _{1 to} 202 _S are connected via a network 203. In addition, when referring to any language analysis device among the language analysis devices 202 _{1 to} 202 _S , the language analysis device 202 is referred to.

  As illustrated in FIG. 7, the language analysis control device 201 includes an input unit 10, a calculation unit 220, and an output unit 230.

  The calculation unit 220 includes a minimum partial problem generation unit 22, a minimum partial problem storage unit 24, and a group creation unit 26.

Each time the minimal subproblem is sequentially generated by the minimal subproblem generating unit 22, the group creating unit 26 has at least two minimum subproblems for each set of minimal subproblems stored in the minimal subproblem storage unit 24. to belong to a group, create a S-number of groups, to the S number of the language analysis unit ₂₀₂ 1 to 202 _S via the network 203. In addition, the group creation unit 26 transmits the input parameter ρ to each of the _S language analysis apparatuses 202 _{1 to} 202 S via the network 203. It should be noted that a group is created so that each of the minimal subproblems stored in the minimal subproblem storage unit 24 has the same group configuration as the previously assigned group.

Each of the S language analysis devices 202 _{1 to} 202 _S includes an input unit 240, a calculation unit 250, and an output unit 260 as shown in FIG.

  The input unit 240 receives a subset of the minimal subproblems included in the own group s transmitted from the language analysis control device 201. The input unit 240 receives information transmitted from another language analysis apparatus 202 via the network 203.

  The calculation unit 250 includes a parameter update unit 32, a partial decomposition update unit 34, a first flag determination unit 36, a synchronization unit 38, a constraint condition update unit 40, a second flag determination unit 42, a convergence determination unit 44, and a parameter storage unit 46. It has.

The parameter updating unit 32 updates the Lagrange undetermined multiplier α _s (r) for each minimal subproblem r included in the own group s transmitted to the language analysis device 202. When the group flag F _s = 0 of the own group s, the parameter update unit 32 does not update the Lagrange undetermined multiplier α _s (r) for each minimal subproblem r, and the initial value or the value updated last time As Lagrange undetermined multiplier α _s (r).

The partial decomposition updating unit 34 updates the solution z _s (r) of each minimal subproblem for a subset of the minimal subproblem r included in the own group s transmitted to the language analysis device 202. When the group flag F _s = 0 of the own group s, the partial decomposition update unit 34 does not update the solution z _s (r) of each local subproblem, but the initial value or the value updated last time As the value of the solution z _s (r) of each minimal subproblem.

The synchronization unit 38 transmits α _s (r) and z _s (r) updated or held this time by the language analysis device 202 _s to all the language analysis devices 202 other than itself via the network 203. To do. In addition, the synchronization unit 38 transmits the Lagrange undetermined multiplier α _i (r) and the solution z _i (r) of each local subproblem, which are transmitted from all the other language analysis devices 202 _{i and} are updated or held this time. receive. By this processing, each language analysis device 202 can acquire the values of α _s (r) and z _s (r) of all the language analysis devices 202 _s .

The constraint condition update unit 40 uses the Lagrange undetermined multiplier α _i (r) and the solution z _i (r) of the minimal subproblem received from all the other language analyzers 202 _i according to the above equation (12). Update the constraint parameter u (r) for the subproblem r.

The convergence determination unit 44 determines whether or not the obtained constraint parameter u (r) has converged to an optimum value, and combines the constraint parameter u (r) obtained when it is determined that it has converged. The result of the natural language analysis processing is generated and transmitted to the language analysis control device 201 by the output unit 260. Further, the solutions u (r) of all the minimal sub-problems r obtained at that time, the solutions z _s (r) of the respective minimal sub-problems r of the own group s, and the Lagrange undetermined multiplier α _s (r) of the own group s Are stored in the parameter storage unit 46.

<Operation of natural language analysis processing system>
Next, the operation of the natural language analysis processing system 200 according to the second embodiment will be described. First, when a character string of one input unit to be subjected to natural language analysis processing is input to the language analysis control device 201, the language analysis control device 201 decomposes the problem of performing natural language analysis processing into a minimal partial problem. To generate a set of minimal subproblems. Then, the group creation unit 26 creates S groups for the set obtained by combining the set of the generated minimal subproblems and the set of the minimal subproblems stored in the minimal subproblem storage unit 24. The data is transmitted to S language analyzers 202 via 203 and assigned to S language analyzers 202.

  Then, the natural language analysis processing routine shown in FIG. 4 is executed by each language analysis device 202.

  The result of the natural language analysis process generated by combining the finally updated constraint parameter u (r) by at least one language analysis device 202 is transmitted to the language analysis control device 201 via the network 203. . The language analysis control device 201 outputs the result of the natural language analysis processing received by the language analysis device 202 by the output unit 230.

  As described above, according to the natural language analysis processing system according to the second embodiment, a plurality of language analysis devices connected via a network perform natural language analysis processing by distributed parallel processing. Can be speeded up.

  Note that the present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

  In the above embodiment, each time a character string in one input unit is input, a new problem is sequentially generated and added to an existing minimal subproblem. However, the present invention is limited to this. is not. A new minimal subproblem may be created by changing the minimal subproblem created immediately before using a newly input character string of one input unit.

  In the above embodiment, the group flag activation and constraint flag deactivation settings have been described for the case where all the groups and the constraint parameter are set for each computation node. It is not limited to. For each computation node, the activation setting for only the group flag of the own group and the deactivation setting for only the constraint parameter related to the minimal subproblem belonging to the own group may be performed.

  In the above-described embodiment, the case where the constraint parameter is updated for all the constraint parameters for each computation node has been described. However, the present invention is not limited to this. For each computation node, only the constraint parameter related to the minimal subproblem belonging to the own group may be updated.

  Moreover, although the above-mentioned natural language analysis processing apparatuses 100 and 200 have computer systems inside, if the “computer system” uses a WWW system, a homepage providing environment (or display environment) Shall be included.

  Further, in the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium or provided via a network. It is also possible to do. Moreover, you may comprise each part of the natural language analysis processing apparatus 100 of this Embodiment with hardware.

DESCRIPTION OF SYMBOLS 10 Input part 20 Calculation part 22 Minimal subproblem generation part 24 Minimal subproblem storage part 26 Group preparation part 30 Calculation node 32 Parameter update part 34 Partial decomposition update part 36 1st flag determination part 38 Synchronization part 40 Restriction condition update part 42 1st 2 flag determination unit 44 convergence determination unit 46 parameter storage unit 50 output unit 100 natural language analysis processing device 200 natural language analysis processing system 201 language analysis control device 202 language analysis device 203 network 220 arithmetic unit 230 output unit 240 input unit 250 arithmetic unit 260 Output section

Claims (8)

A natural language analysis processing device that performs natural language analysis processing including language analysis processing on an input character string obtained by concatenating each of character strings composed of at least one character that is sequentially input,
The partial problem when the problem of performing the language analysis processing on the input character string is decomposed into a partial problem of performing the language analysis processing in a predefined character unit or character string unit, the character unit or character string A sub-problem generating means for sequentially generating each unit input;
Sub-problem storage means for storing a set of sub-problems sequentially generated by the sub-problem generation means;
Each time the subproblem is generated sequentially by the subproblem generation means, S computation nodes (S is a natural number of 2 or more) for performing the natural language analysis processing;
Each time the partial problem is sequentially generated by the partial problem generating means, the partial problem is set such that each partial problem belongs to at least two groups with respect to the set of partial problems stored in the partial problem storage means. Creating S groups composed of arbitrary subsets of the total set of and assigning them to the S computation nodes,
Each of the S compute nodes is
For each subset of the subsets of the group assigned by the group creation means, each part of the subset is based on a constraint condition that the solutions of the respective minimal subproblems belonging to the two or more groups match. A partial update method for updating the solution of the problem;
The partial update means of the computation node when the solutions of the partial problems of the subset updated by the partial decomposition update means all match the solutions of the partial problems of the subset updated last time A group inactive state setting means for setting a group inactive state indicating that the update processing by is not performed,
For each subproblem, when the solution of the subproblem updated by the partial decomposition update means does not match the solution of the subproblem updated last time, the solution of the subproblem is made to match according to the constraint condition. A constraint condition active state setting means for canceling the constraint condition inactive state, which is set for the constraint parameter that is the solution at the time, and represents that the constraint parameter update process is not performed,
Notifying the other calculation nodes of the solutions of the respective partial problems updated by the partial decomposition updating means, and obtaining the solutions of the respective partial problems of the subset notified from the other calculation nodes Synchronization means;
Based on the solution of each subproblem of the subset of the other computation node obtained by the synchronization means and the solution of each subproblem of the subset updated by the partial decomposition update means, , Constraint update means for updating the constraint parameters of the partial problem;
For each subproblem, if the constraint condition parameter updated by the constraint condition update means does not match the previously updated value of the constraint condition parameter, the computation node to which the group to which the subproblem belongs is assigned Group active state setting means for canceling the group inactive state set for:
For each subproblem, when the constraint parameter updated by the constraint condition update unit matches the value of the constraint parameter updated last time, the constraint condition for the constraint parameter of the subproblem Restriction condition inactive state setting means for setting the inactive state;
It is determined whether or not the value of the constraint parameter has converged, and until it is determined that the value of the constraint parameter has converged, the update by the partial decomposition update unit, the setting by the group inactive state setting unit, the constraint Convergence determining means for repeating setting by the condition active state setting means, notification and acquisition by the synchronizing means, update by the constraint condition update means, setting by the group active state setting means, and setting by the constraint condition inactive state setting means; A natural language analysis processing apparatus including
The calculation node in which the group inactive state is set does not perform update by the partial decomposition update unit,
The synchronization means of the computation node in which the group inactive state is set notifies other computation nodes of the solution of each partial problem of the subset last updated by the partial decomposition update means,
The natural language analysis processing device, wherein the constraint condition update means does not update the constraint parameter in which the constraint condition inactive state is set. The partial decomposition updating means, for the subset of the partial problem of the group assigned by the group creating means, the Lagrange undetermined multiplier updated last time, the solution of each partial problem of the subset, and the constraint parameter To update the Lagrangian undetermined multiplier so as to optimize a predetermined objective function value, and to use the updated Lagrange undetermined multiplier to optimize the value of the objective function, Update the solution of each subproblem of the subset,
The synchronizing means notifies the Lagrange undetermined multiplier updated by the partial decomposition updating means and the solution of each subproblem of the subset to other calculation nodes, and the Lagrange notified from the other calculation nodes. Obtain the solution to the undetermined multiplier and each minimal subproblem of the subset,
Each of the constraint condition updating means optimizes the value of the objective function based on the Lagrange undetermined multiplier of the other computation node acquired by the synchronization means and the solution of each partial problem of the subset. The natural language analysis processing apparatus according to claim 1, wherein the constraint parameter is updated for a partial problem. A natural language analysis processing device that performs natural language analysis processing including language analysis processing on an input character string obtained by concatenating each of character strings composed of at least one character that is sequentially input,
The partial problem when the problem of performing the language analysis processing on the input character string is decomposed into a partial problem of performing the language analysis processing in a predefined character unit or character string unit, the character unit or character string A sub-problem generating means for sequentially generating each unit input;
Sub-problem storage means for storing a set of sub-problems sequentially generated by the sub-problem generation means;
Each time the subproblem is generated sequentially by the subproblem generation means, S computation nodes (S is a natural number of 2 or more) for performing the natural language analysis processing;
Each time the partial problem is sequentially generated by the partial problem generating means, the partial problem is set such that each partial problem belongs to at least two groups with respect to the set of partial problems stored in the partial problem storage means. A group creating means for creating S groups composed of arbitrary subsets of the entire set and assigning them to the S computing nodes;
A constraint condition update means, a group active state setting means, a constraint condition inactive state setting means, and a convergence determination means,
Each of the S compute nodes is
For each subset of the subsets of the group assigned by the group creation means, each part of the subset is based on a constraint condition that the solutions of the respective minimal subproblems belonging to the two or more groups match. A partial update method for updating the solution of the problem;
The partial update means of the computation node when the solutions of the partial problems of the subset updated by the partial decomposition update means all match the solutions of the partial problems of the subset updated last time A group inactive state setting means for setting a group inactive state indicating that the update processing by is not performed,
For each subproblem, when the solution of the subproblem updated by the partial decomposition update means does not match the solution of the subproblem updated last time, the solution of the subproblem is made to match according to the constraint condition. A constraint condition active state setting unit that cancels a constraint condition inactive state that is set for the constraint condition parameter that is a solution when the constraint condition parameter is not updated. For each subproblem based on the solution of each subproblem of the subset of each computation node and the solution of each subproblem of the subset updated by the partial decomposition update means, Update constraint parameters,
The group active state flag assigning unit, for each partial problem, if the constraint parameter updated by the constraint condition update unit does not match the value of the constraint parameter updated last time, the partial problem is Release the group inactive state set for the compute node to which the group to which it belongs is assigned,
The constraint condition inactive state setting means, for each partial problem, when the constraint condition parameter updated by the constraint condition update means matches the value of the constraint condition parameter updated last time, Set the constraint inactive state for the constraint parameter in question,
The convergence determination means determines whether or not the value of the constraint parameter has converged, and updates by the partial decomposition update means, the group inactive state setting until it is determined that the value of the constraint parameter has converged Setting by means, setting by the restriction condition active state setting means, notification and acquisition by the synchronization means, update by the restriction condition update means, setting by the group active state setting means, and setting by the restriction condition inactive state setting means Is a natural language analysis processing device,
The calculation node in which the group inactive state is set does not perform update by the partial decomposition update unit,
The constraint condition update means does not update the constraint parameter in which the constraint condition inactive state is set,
The solution of each subproblem of the subset in the computation node in which the group inactive state is set is the solution of each subproblem of the subset last updated by the partial decomposition update means. Natural language analysis Processing equipment. The partial decomposition updating means, for the subset of the partial problem of the group assigned by the group creating means, the Lagrange undetermined multiplier updated last time, the solution of each partial problem of the subset, and the constraint parameter Updating the Lagrangian undetermined multiplier so as to optimize the value of the objective function determined in advance, and using the updated Lagrangian undetermined multiplier to optimize the value of the objective function. Update the solution of each subproblem in the subset,
The constraint condition update means optimizes the value of the objective function based on the Lagrange undetermined multiplier updated by the partial decomposition update means of each computation node and the solution of each partial problem of the subset. The natural language analysis processing apparatus according to claim 3, wherein the constraint parameter is updated for each partial problem. It includes a subproblem generating means, a subproblem storing means, S (S is a natural number greater than or equal to 2) calculation nodes, and a group creating means, each of which is connected to a character string consisting of at least one character that is sequentially input. A natural language analysis processing method in a natural language analysis processing apparatus that performs natural language analysis processing including language analysis processing on an input character string obtained by:
The partial problem when the problem of performing the language analysis processing on the input character string is decomposed by the partial problem generation means into a partial problem of performing the language analysis processing in a predefined character unit or character string unit. , Each time the character unit or character string unit is input,
A set of subproblems sequentially generated by the subproblem generator by the subproblem storage means;
Each time the partial problem is sequentially generated by the group generation means by the group creation means, each partial problem belongs to at least two or more groups with respect to the set of partial problems stored in the partial problem storage means. Create S groups of arbitrary subsets for the entire set of subproblems and assign them to the S computational nodes,
The natural language analysis processing is performed by the S calculation nodes,
Performing the natural language analysis process by each of the S computation nodes is as follows:
Based on the constraint condition that the solutions of the local subproblems belonging to the two or more groups match with respect to a subset of the partial problem set of the group assigned by the group creation means by the partial decomposition update means. Update the solution of each subproblem of the subset,
When the solutions of the subproblems of the subset updated by the partial decomposition updating means by the group inactive state setting means all coincide with the solutions of the subproblems of the subset updated last time, Set a group inactive state indicating that update processing by the partial update unit of the calculation node is not performed,
When the solution of the partial problem updated by the partial decomposition update unit for each partial problem by the constraint active state setting unit does not match the solution of the partial problem updated last time, according to the constraint condition The constraint condition inactive state, which is set for the constraint parameter that is a solution when matching the solution of the subproblem and does not perform the update process of the constraint parameter, is released,
The synchronization means notifies the solution of each subproblem of the subset updated by the subdivision updating means to other calculation nodes, and also notifies each subproblem of the subset notified from the other calculation node. Get the solution,
Based on the solution of each partial problem of the subset of the other computation node acquired by the synchronization means by the constraint update means and the solution of each partial problem of the subset updated by the partial decomposition update means For each subproblem, update the constraint parameter of the subproblem,
If the constraint parameter updated by the constraint condition update unit for each partial problem by the group active state setting unit does not match the value of the constraint parameter updated last time, the group to which the partial problem belongs Release the group inactive state set for the compute node to which is assigned,
For each partial problem by the constraint condition inactive state setting means, when the constraint condition parameter updated by the constraint condition update means matches the value of the constraint condition parameter updated last time, the partial problem Setting the constraint inactive state for the constraint parameter of
It is determined whether or not the value of the constraint parameter has converged by a convergence determination unit, and the update by the partial decomposition update unit, the group inactive state setting unit until it is determined that the value of the constraint parameter has converged Setting by the constraint condition active state setting means, notification and acquisition by the synchronization means, update by the constraint condition update means, setting by the group active state setting means, and setting by the constraint condition inactive state setting means. A natural language analysis processing method including repetition,
The calculation node in which the group inactive state is set does not perform update by the partial decomposition update unit,
The synchronization means of the computation node in which the group inactive state is set notifies other computation nodes of the solution of each partial problem of the subset last updated by the partial decomposition update means,
The natural language analysis processing method, wherein the constraint condition update means does not update the constraint parameter in which the constraint condition inactive state is set. The partial decomposition updating means, for the subset of the partial problem of the group assigned by the group creating means, the Lagrange undetermined multiplier updated last time, the solution of each partial problem of the subset, and the constraint parameter To update the Lagrangian undetermined multiplier so as to optimize a predetermined objective function value, and to use the updated Lagrange undetermined multiplier to optimize the value of the objective function, Update the solution of each subproblem of the subset,
The synchronizing means notifies the Lagrange undetermined multiplier updated by the partial decomposition updating means and the solution of each subproblem of the subset to other calculation nodes, and the Lagrange notified from the other calculation nodes. Obtain the solution to the undetermined multiplier and each minimal subproblem of the subset,
The constraint condition update means optimizes the value of the objective function based on the Lagrange undetermined multiplier of the other computation node acquired by the synchronization means and the solution of each partial problem of the subset. The natural language analysis processing method according to claim 5, wherein the constraint parameter is updated for a partial problem. Sub-problem generating means, sub-problem storage means, S calculation nodes (S is a natural number of 2 or more), group creation means, constraint condition update means, group active state setting means, constraint condition inactive state setting means, and A natural language analysis apparatus that includes a convergence determination unit and performs a natural language analysis process including a language analysis process on an input character string obtained by concatenating each of character strings composed of at least one character sequentially input. A language analysis processing method,
The partial problem when the problem of performing the language analysis processing on the input character string is decomposed by the partial problem generation means into a partial problem of performing the language analysis processing in a predefined character unit or character string unit. , Each time the character unit or character string unit is input,
A set of subproblems sequentially generated by the subproblem generator by the subproblem storage means;
Each time the partial problem is sequentially generated by the group generation means by the group creation means, each partial problem belongs to at least two or more groups with respect to the set of partial problems stored in the partial problem storage means. Create S groups of arbitrary subsets for the entire set of subproblems and assign them to the S computational nodes,
The natural language analysis processing is performed by the S calculation nodes,
Updated by the constraint condition update means,
Set by the group active state setting means,
Set by the constraint inactive state setting means,
Determining by the convergence determining means,
Performing the natural language analysis process by each of the S computation nodes is as follows:
Based on the constraint condition that the solutions of the minimal sub-problems belonging to the two or more groups match with respect to a subset of the partial problem set of the group assigned by the group creation means by the partial decomposition update means. Update the solution of each subproblem of the subset,
When the solutions of the subproblems of the subset updated by the partial decomposition update means by the group inactive state setting means all match the solutions of the subproblems of the subset updated last time , Set a group inactive state indicating that update processing by the partial update means of the calculation node is not performed,
The constraint condition when the solution of the partial problem updated by the partial decomposition update unit for each partial problem does not match the solution of the partial problem updated last time by the constraint condition active state setting unit. Canceling the constraint inactive state, which is set for the constraint parameter that is a solution when matching the solution of the partial problem according to
The constraint condition update means updates the constraint parameter of the partial problem for each partial problem based on the solution of each partial problem of the subset updated by the partial decomposition update means of each computation node,
The group active state setting means, for each partial problem, if the restriction condition parameter updated by the restriction condition update means does not match the value of the restriction condition parameter updated last time, the partial problem belongs to Release the group inactive state set for the compute node to which a group is assigned,
The constraint condition inactive state setting means, for each partial problem, when the constraint condition parameter updated by the constraint condition update means matches the value of the constraint condition parameter updated last time. Setting the constraint inactive state for the constraint parameter of
The convergence determination means determines whether or not the value of the constraint parameter has converged, and updates by the partial decomposition update means for each calculation node until determining that the value of the constraint parameter has converged, the group Setting by the inactive state setting unit, setting by the constraint condition active state setting unit, notification and acquisition by the synchronization unit, update by the constraint condition update unit, setting by the group active state setting unit, and the constraint condition inactive state A natural language analysis processing method including repeating setting by a setting means,
The calculation node in which the group inactive state is set does not perform update by the partial decomposition update unit,
The constraint condition update means does not update the constraint parameter in which the constraint condition inactive state is set,
Natural language analysis in which the solution of each subproblem of the subset in the computation node in which the group inactive state is set is the solution of each subproblem of the subset last updated by the partial decomposition update means Processing method.   The program for functioning a computer as each means of the natural language analysis processing apparatus of any one of Claims 1-4.