JP4521631B2 - Storage medium recording tree structure dictionary and language score table creation program for tree structure dictionary - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a tree structure capable of suppressing a search error while efficiently reducing necessary capacity. <P>SOLUTION: It is decided whether the number of words sharing each non-terminal node of the tree structure dictionary stored on the storage medium (the number of terminal nodes which can be reached from the non-terminal node through child nodes) is larger than a specified threshold (444) and when it exceeds the threshold, a nonapproximate bigram language score table is held (446), but when not, scores which are approximated by unigram are requested of all of the child nodes (448), so that a maximum value among obtained apprximiate scores is held as an approximate score of the non-terminal node (450). <P>COPYRIGHT: (C)2005,JPO&amp;NCIPI

Description

  The present invention relates to a method for creating a tree structure dictionary used in large vocabulary speech recognition and the like, and in particular, to reduce the amount of memory consumed, a tree structure dictionary that introduces approximate values into language scores assigned to each node, and It relates to its creation method.

  In a large vocabulary speech recognition system, a language score using a probabilistic language model is calculated in a speech recognition process or a verification process after recognition. The probabilistic language model is a model of a natural language with a probability of occurrence of a word string or a character string in a certain language. The language score is obtained by calculating the likelihood of a word string or the like obtained as a result of speech recognition according to a language model.

  A model in which a natural language is modeled by the probability that a word string or character string composed of N words or characters is generated is called an N-gram model. In particular, the cases of N = 1, 2, 3 are called unigram, bigram and trigram, respectively. A bigram or trigram is often used from the viewpoint of calculation amount and accuracy for parameter estimation. The following description is for word strings.

  In order to calculate the language score of a recognition candidate, a word dictionary with a pre-calculated language score is often prepared. In order to improve the efficiency of searching for recognition candidates, it is common to represent a word dictionary with a tree structure network.

  An outline of the tree structure dictionary will be described with reference to FIG. Assume that the acoustic model is a monophone and only “red”, “bright”, and “blue” are considered as vocabularies. The phoneme strings 20, 22, and 24 of these words are shown in the upper part of FIG. As is apparent from the figure, the sound “a” 30 at the beginning of the word is shared by three words, and the phoneme string 32 of “aka” including the subsequent two sounds is shared by the two words.

  A tree structure dictionary 40 corresponding to the above word group is shown in the lower part of FIG. As shown in the figure, the tree structure dictionary 40 uses a portion where the phonemes from the beginning of a word are common as a shared node, and branches to another node when the word no longer shares a phoneme. For example, since the three words described above share the phoneme “a”, the first node 50 of the tree structure dictionary 40 is shared by the three words. However, the next phoneme branches to two nodes 52 (k) and 62 (o). Thereafter, the tree structure is similarly formed. As a result, the word “Akarai” is traced by following the node strings 50, 52, 54, 56, 58, 60, and the word “Akairu” is traced by following the node strings 50, 52, 54, 64. The word “Aoi” is searched by tracing the node strings 50, 62 and 66 respectively.

  The above is the basic structure of the tree structure dictionary.

  The end node of the tree structure dictionary (nodes 60, 64, and 66 in the case of FIG. 1) is assigned a bigram language score table for each word (in the example of FIG. 1, the language score tables 70, 72, and 74). ing. This table shows the probability that the word of the terminal node will subsequently occur for all the words that may precede. Therefore, this language score table has as many entries as the number of all words in the language model. That is, the capacity of the language score table depends on the number of vocabularies.

  In speech recognition, the development of hypotheses proceeds from the first node of the tree structure dictionary to the end node in synchronization with the speech recognition. The word hypothesis reaching the end node is registered in the word graph, and the search for the subsequent word hypothesis is started again from the first node of the dictionary.

  In searching the tree structure dictionary, if all possible paths are searched, there is a problem that the number of searches increases. Therefore, pruning based on the score is necessary. Therefore, for nodes other than the terminal node (for example, the nodes 52, 54, 56, 58, and 62 in the example shown in FIG. 1), the maximum likelihood among words (terminal nodes) sharing the node is A language score table (tables 80, 82, 84, 86, 88 in the case of FIG. 1) calculated for all preceding words is attached. In this case, for example, if the number of vocabularies is N, the number of entries in the language score table is N. This state is shown as a language score table 120 in FIG.

The language score table 120 shown in FIG. 3 is a language score table to be given to the node s when, for example, the node s is shared by a plurality of words w _i . If a set of words sharing the node s is represented by S, it can be written as w i εS. As shown in FIG. 3, the language score table 120 includes the conditional occurrence probability p (w _i ) of the word w _i (w _{i εS} ) calculated for all of the preceding words (words w _{p1 to} w _pN ). | W _pj ) (W _i εS, j = 1 to N) including N entries of the maximum value max p (w _i | w _pj ) (wiεS).

  At the time of branching, when branching at each node, the node score is updated by referring to the language score table of the branch destination node, and pruning is performed by not searching for branches whose score is smaller than a predetermined value. To do.

  If the process for obtaining the maximum language score is repeated frequently, an enormous amount of calculation is required. Therefore, all preceding words are pre-calculated and tabulated in advance. This is the language score table described above. The table size depends on the vocabulary number N as described above. By the way, when the shared word sets are different nodes, the contents of the language score table to be given to them are also different from each other. Therefore, it is necessary to prepare the language score table described above for each node having a different shared word.

  However, since the number of nodes in the tree structure dictionary is enormous and the number of vocabularies is large, a large amount of memory is required to perform the above-described table formation, which may be substantially impossible. Therefore, it is necessary to reduce the memory capacity required for tabulation. A technique for this is disclosed in Non-Patent Document 1 described later.

  The method disclosed in Non-Patent Document 1 approximates the above-described language score table in a tree structure dictionary using the depth from the first node as a criterion. Specifically, in the method described in Non-Patent Document 1, a certain threshold value is defined, and the above-mentioned language score table is assigned without approximation to a shared node whose depth from the first node is equal to or less than this threshold value. However, in a node having a depth exceeding the threshold value, a predetermined approximation is performed instead of the language score table described above. As an approximation method, N entries are approximated by the maximum value (scalar value) of a unigram of a terminal node sharing the node, and a language score table is created by a class bigram instead of a word bigram. An approximation method has been proposed.

  When a language score table is approximated by a unigram, a table having N entries is replaced with one scalar value, so that the storage capacity is greatly reduced. When approximating with a class bigram, the storage capacity cannot be reduced as much as a unigram, but there is an advantage that the accuracy does not decrease so much.

Kokubo et al., “Compacting factoring table in continuous speech recognition system and garbage collection of unnecessary word hypotheses”, IEICE Transactions D-II Vol. 86 no. 6, pp. 787-795, 2003

  The method disclosed in Non-Patent Document 1 can reduce memory consumption efficiently. However, if a search error occurs by approximating the language score, the amount of loss due to that error cannot be predicted. For this reason, there is a possibility that the recognition performance may be greatly reduced.

  Accordingly, an object of the present invention is to provide a storage medium storing a tree structure dictionary capable of suppressing the occurrence of a search error while efficiently reducing the capacity of the language score table for the tree structure dictionary, and a tree structure dictionary. To provide a language score table creation program.

  The storage medium according to the first aspect of the present invention stores a tree structure dictionary composed of a plurality of non-terminal nodes having child nodes and a plurality of terminal nodes corresponding to words without having child nodes. It is a medium. Each of the plurality of non-terminal nodes corresponds to a predetermined phoneme. In the tree structure dictionary, the child nodes are traced from the first node of the tree structure dictionary, so that the phoneme corresponding to the non-terminal node becomes a word that is part of its own phoneme sequence via each non-terminal node. All corresponding terminal nodes are configured to be reachable. Each of the plurality of terminal nodes is provided with a first language score table of a corresponding word according to a predetermined language model. A second language created by taking the maximum value of the language score table from all the language score tables of the terminal node that can be reached by tracing the child nodes from the non-terminal node to each of the plurality of non-terminal nodes A score table is attached. The second language score table of a plurality of non-terminal nodes is simpler than the second language score table when the number of terminal nodes that can be reached by tracing child nodes from the non-terminal nodes is less than a predetermined threshold value. Language language information has been replaced.

  Preferably, the predetermined language model is a bigram language model.

  More preferably, the simple language score information is a maximum value of a unigram score for all terminal nodes that can be reached from the non-terminal node. A class bigram score may be used instead of the unigram score.

  The second language score table has the same number of entries as the first language score table, and each entry of the second language score table is a terminal node reachable by following the child node from the non-terminal node. It may consist of the maximum value of the corresponding entry in the first language score table.

  A language score table creation program according to a second aspect of the present invention is a tree structure dictionary composed of a plurality of non-terminal nodes having child nodes and a plurality of terminal nodes corresponding to words without having child nodes. A language score table creation program for creating a language score table assigned to a plurality of non-terminal nodes. Each of the plurality of non-terminal nodes corresponds to a predetermined phoneme. The tree structure dictionary corresponds to a word in which the phoneme corresponding to the non-terminal node is part of its own phoneme string via each non-terminal node by tracing the child node from the first node of the tree structure. It is configured to be able to reach all of the end nodes that perform. Each of the plurality of terminal nodes is provided with a first language score table of a corresponding word according to a predetermined language model. The language score table creation program includes a program part for calculating the number of terminal nodes that can be reached by following a child node from each non-terminal node, and a plurality of non-terminal nodes. In response to receiving a language score table request from a parent node of the non-terminal node, a program part that requests the language score table of the child node for each of the child nodes of the non-terminal node, and a plurality of non-terminals By executing a program part that requests a language score table in each of the nodes, a second language score table for the non-terminal node is created based on the language score table returned from each of the child nodes of the non-terminal node. Create program part and language score table A program part that returns the second language score table created by the program part to the parent node, and a condition that the number of terminal nodes that can be reached by tracing the child node from the non-terminal node is predetermined with respect to a predetermined threshold value. In response to determining that the condition is satisfied, and a program part for determining whether or not the condition is satisfied, the second language score table created for the non-terminal node is used as the language score of the non-terminal node. A program part to be stored in the storage device as a table, and a program part that performs processing for requesting simple language score information to a child node of the non-terminal node in response to determining that the condition is not satisfied And simple language score information returned from a child node in response to a process requesting simple language score information. Accordingly, a simplified program stored in the storage device in response to receiving a request for the simple language score information from the parent node, and a program part for creating the simple language score information related to the non-terminal node and storing it in the storage device And a program part that returns linguistic score information to the parent node.

  Preferably, the predetermined language model is a bigram language model.

  More preferably. The simple language score information is the maximum value of the unigram score of each terminal node that can be reached by following the child nodes from each non-terminal node. A class bigram score may be used instead of the unigram score.

  The second language score table has the same number of entries as the first language score table, and each entry of the second language score table can be reached by following the child node from the non-terminal node. It may consist of the maximum value of the corresponding entry in the first language score table of the node. In this case, the program part for creating the second language score table of the non-terminal node is executed by executing the program part that requests the language score table in each of the plurality of non-terminal nodes. For each entry of the language score table returned from, a program portion for taking the maximum value and creating the second language score table of the non-terminal node may be included.

  The loss due to the search error by approximating the language score increases as the number of terminal nodes sharing the node increases. Therefore, in the present embodiment, the language score table is not approximated based on the depth from the head node as in the method disclosed in Non-Patent Document 1, but the number of words sharing each node is used as a reference. The language score table is approximated as follows. This concept will be described with reference to FIG.

FIG. 2 shows an example of the tree structure dictionary 100. Referring to FIG. 2, this tree structure dictionary 100 is shared by terminal nodes W _{1 to} W ₁₄ (hereinafter, these terminal nodes are referred to as “words” for the sake of simplicity). A plurality of nodes (for example, nodes S _A and S _B ). Node S _A is shared by _five words W _{1 to} W ₅ . Node S _B is shared by two words W ₁₃ and W ₁₄ . In this case, the node S _B is preferentially approximated between the node S _A and the node S _B.

  In this way, by determining whether or not the language likelihood is approximated based on the number of words that share a node, it is possible to efficiently reduce the capacity of the language score table for the tree structure dictionary while suppressing search errors. can do.

  The tree structure dictionary creation system according to this embodiment is realized by computer hardware, a program executed by the computer hardware, and data stored in the computer hardware. FIG. 4 shows the external appearance of the computer system 330, and FIG. 5 shows the internal configuration of the computer system 330.

  Referring to FIG. 4, this computer system 330 includes a computer 340 having an FD (flexible disk) drive 352 and a CD-ROM (compact disk read only memory) drive 350, a keyboard 346, a mouse 348, and a monitor 342. including.

  Referring to FIG. 5, in addition to the FD drive 352 and the CD-ROM drive 350, the computer 340 includes a CPU (central processing unit) 356 and a bus 366 connected to the CPU 356, the FD drive 352, and the CD-ROM drive 350. And a read only memory (ROM) 358 for storing a boot-up program and the like, and a random access memory (RAM) 360 connected to the bus 366 for storing a program command, a system program, work data, and the like. Computer system 330 further includes a printer 344.

  Although not shown here, the computer 340 may further include a network adapter board that provides a connection to a local area network (LAN).

  A computer program for causing the computer system 330 to implement the tree structure dictionary creation method according to the present embodiment is stored in the CD-ROM 362 or FD 364 inserted in the CD-ROM drive 350 or FD drive 352, and further the hard disk 354. Forwarded to Alternatively, the program may be transmitted to the computer 340 through a network (not shown) and stored in the hard disk 354. The program is loaded into the RAM 360 when executed. The program may be loaded directly into the RAM 360 from the CD-ROM 362, from the FD 364, or via a network.

  This program includes a plurality of instructions for causing the computer 340 to realize the tree structure dictionary creation method according to this embodiment. Some of the basic functions necessary to perform this method are provided by operating system (OS) or third party programs running on the computer 340, or modules of various toolkits installed on the computer 340. Therefore, this program does not necessarily include all functions necessary to realize the system and method of this embodiment. This program may include only instructions that can realize the above-described method for creating a tree structure dictionary by calling an appropriate function or “tool” in a controlled manner so as to obtain a desired result. That's fine. Since the operation of computer system 330 is well known, details thereof will not be repeated here.

  Hereinafter, it is assumed that the tree structure dictionary itself has already been created, and a computer program for realizing processing for creating a language score table will be described. These programs are all prepared as node class methods.

  There are three major programs. The first is a program for inquiring child nodes how many words (terminal nodes) exist under each node. This program is executed recursively. The second is processing for determining a language score table that each node should store in its own node. In order to perform this process, it is necessary to query the language score table of the child node. Therefore, this program is also executed recursively. The third is executed in connection with the second program. When the parent node requests to return the language score after approximation, the language score after approximation stored in the own node is used as the parent language score. Processing to return to the node. Since the third process can be easily realized, only the control structure of the program (method) for realizing the first and second processes will be described below.

  FIG. 6 shows a program for checking the number of end nodes (number of words sharing the own node) under each node (excluding the end node) by inquiring the child nodes. The control structure of is shown. Referring to FIG. 6, this program is obtained from all the child nodes of the self in step 400 that inquires all of the child nodes under the self node for the number of shared words under the node. Calculating the total number of shared words, and storing 402 the number of shared words as its own attribute. After step 402, the process is terminated with the total number of shared words underneath as a return value. When querying the number of shared words for each node in step 400, each child node executes this processing shown in FIG. That is, this process is executed recursively.

  In the case of a terminal node, there is no child node under self. When the terminal node receives an inquiry about the number of shared words from the parent node, it may simply return “1”. Accordingly, the processing related to the terminal node is not particularly illustrated here.

  With the above processing, the number of terminal nodes existing below the attribute value of each node in the tree structure dictionary is checked.

  FIG. 7 is a flowchart of processing executed in each node (excluding the terminal node) when a request for the language score table of the own node is received from the parent node. First, in step 440, each language score table is queried with respect to all of its immediate child nodes. Using the language score table returned from each child node in step 440, a language score table of its own node is created in step 442.

  Here, as shown in FIG. 3, for each preceding word, the maximum value of the language score of the terminal node below it is obtained. Therefore, one entry in the language score table is created for one preceding word. If the vocabulary number is N, N language scores are stored in the language score table.

  Subsequently, at step 444, it is determined whether or not the number of shared words under the self calculated by the process shown in FIG. 6 is equal to or greater than a predetermined threshold value. If it is equal to or greater than the threshold value, the process proceeds to step 446, and the language score table created in step 442 is stored as it is as its own language score table. If the number of shared words is less than the threshold, in step 448, an approximate score (scalar) is requested to all of its own child nodes. In step 450, the maximum value among the approximate scores obtained from all the child nodes in step 448 is stored as its own approximate score.

  After step 450, the process is terminated with the language score table created in step 442 as a return value.

  In this embodiment, a unigram is used as an approximate score. Each terminal node stores not only the bigram language score table described above but also a unigram score (scalar). Since the process shown in FIG. 7 finally requests an approximate score for each terminal node, each terminal node may return its own unigram score to the parent node.

  In addition, in the node (other than the terminal node) that has received the request for the score approximated in step 448, the processing in FIG. 7 has been completed before that, and the score approximated in step 450 has already been calculated and stored. Returns the stored approximate score to the parent node. If there is no approximate value, the process of steps 448 and 450 is performed to calculate and store the approximate score, and return it to the parent node.

  With the program having the control structure as described above, a language score table or an approximate language score (unigram) can be assigned to each node of the tree structure dictionary.

  The tree structure creation system according to the above-described embodiment operates as follows. First, a tree structure is created as in the prior art. A language score table calculated in advance is assigned to each terminal node.

  First, the number of shared words under the first node of each tree structure is inquired. In the head node, the process shown in FIG. 6 is executed. That is, the top node receives this inquiry and inquires each child node immediately below itself of the number of shared words below them. Each of the child nodes that have received the inquiry further inquires the number of shared words of the child nodes immediately below them. Thus, this query is recursively executed by each node, and finally this request is given to each terminal node. Each terminal node returns “1” in response to this request. Upon receiving this response, the parent node calculates and stores the sum of the values returned from the child nodes, and returns it to the parent node. In this way, the number of shared words below itself is returned in order from the lowest node to the upper layer node, and finally the total number of shared words below it is returned to the top node. In this process, each node stores the number of shared words under itself by the process of step 402 in FIG.

  Next, the language score table is queried for the first node. In the head node, the processing shown in FIG. 7 is executed. That is, the top node receives this inquiry and requests a language score table from each of its immediate child nodes (step 440 in FIG. 7). These child nodes further request a language score table for each of their immediate child nodes. These processes are repeated, and this request is finally given to the end node.

  Each of the terminal nodes that have received this request returns the language score table held by itself to the parent node. The parent node that has received the language score table from all the child nodes creates its own language score table based on the obtained language score table, as shown in step 442 of FIG. As shown in FIG. 3, this language score table is obtained by taking the maximum language score obtained from the child node for each preceding word.

  Further, the parent node determines whether or not the number of shared words under itself is equal to or greater than a threshold value (step 444). This determination is possible because the number of shared words under itself has already been calculated in step 402 of FIG. If the number of shared words is greater than or equal to the threshold value, the table created in step 442 is stored as its own language score table, and the process ends. Otherwise, step 448 requests approximate scores for all of its immediate child nodes. If the immediate child node is a terminal node, the terminal node returns a word unigram corresponding to the terminal node to the parent node. The parent node that receives the unigram from all the child nodes stores the maximum value in the received unigram as its own approximate score, and ends the process.

  If it is determined in step 444 in FIG. 7 that the number of shared words is less than the threshold value, the same determination is performed in the lower nodes. Therefore, when the processing of step 448 is executed in a certain node, the processing of steps 448 and 450 is executed in the child node at the time when the query of the language score table has already been performed for all of the child nodes in step 440 before that. Completed. Therefore, when a certain node executes the process of step 448, a score that is immediately approximated is returned from the node immediately below it, and the process of step 450 can be executed.

  In this way, the language score table of each node is recursively created and approximated starting from the top node. By the processing from step 444 onward in FIG. 7, a language score table that is not approximated to nodes whose number of shared words is equal to or greater than a threshold value is retained, and approximate scores are retained at other nodes. Therefore, the capacity of the tree structure dictionary is reduced, and the memory capacity necessary for speech recognition can be reduced. The approximation of the language score table is limited when the number of shared words of the node is less than a threshold value. Therefore, the influence of the search error due to the approximation is limited to a limited range, and there is little possibility that the accuracy of the speech recognition is unexpectedly deteriorated.

For example, in the example shown in FIG. 2, the case of both threshold = 3-5, approximation of language score table is not performed for the node S _A, Note carried out only for node S _B, and the In the embodiment, a unigram is used as an approximate score. However, the present invention is not limited to such an embodiment. For example, as proposed in Non-Patent Document 1, a class bigram may be used as an approximation.

  In the above embodiment, the language score table is based on the bigram language score. However, this embodiment is not limited to the one using bigram in principle. For example, the language score table may be a table composed of trigrams, and the language score table may be approximated by a bigram language score table or unigram.

  The embodiment disclosed herein is merely an example, and the present invention is not limited to the above-described embodiment. The scope of the present invention is indicated by each claim in the claims after taking into account the description of the detailed description of the invention, and all modifications within the meaning and scope equivalent to the wording described therein are intended. Including.

It is a schematic diagram for demonstrating the concept of a tree structure dictionary. It is a figure for demonstrating the approximation of the language score table in a tree structure dictionary. It is a figure which shows typically the structure of the language score table by a bigram. It is an external view of a computer. It is a block diagram of the computer shown in FIG. It is a flowchart which shows the control structure of the program (method) performed in each node, when the inquiry regarding the number of shared words under a node is received. It is a flowchart which shows the control structure of the program (method) performed by each node, when the inquiry regarding a language score table is received from the parent node.

Explanation of symbols

40,100 tree structure dictionary, 50, 52, 54, 56, 58, 62, S _A , S _B nodes, 60, 64, 66, W _{1 to} W ₁₄ terminal nodes, 70, 72, 74, 80, 82, 84, 86, 88, 120 Language Score Table

Claims (10)

A storage medium recording a tree structure dictionary composed of a plurality of non-terminal nodes having child nodes and a plurality of terminal nodes corresponding to words without having child nodes, wherein the plurality of non-terminal nodes are respectively It corresponds to a predetermined phoneme,
The tree structure dictionary
By following the child node from the first node of the tree structure dictionary, the terminal node corresponding to the word whose phoneme corresponding to the non-terminal node is part of its own phoneme sequence via each non-terminal node It is configured to reach all of the
Each of the plurality of terminal nodes is provided with a first language score table of a corresponding word according to a predetermined language model,
Each of the plurality of non-terminal nodes is created by taking the maximum value of the language score table from all the language score tables of the terminal nodes that can be reached by following the child nodes from the non-terminal node. With a language score table
The second language score table of the plurality of non-terminal nodes is the second language score table when the number of terminal nodes that can be reached by following the child nodes from the non-terminal nodes is less than a predetermined threshold value. A storage medium storing a computer-readable tree structure dictionary, characterized in that it is replaced with language score information having a smaller required storage capacity . The storage medium according to claim 1, wherein the predetermined language model is a bigram language model. The language score information having a smaller storage capacity than that of the second language score table is a maximum value of a unigram score for all terminal nodes that can be reached from the non-terminal node. Storage media. The language score information having a smaller storage capacity required than the second language score table is a maximum value of a class bigram score for all terminal nodes that can be reached from the non-terminal node. Storage media. The second language score table has the same number of entries as the first language score table, and each entry of the second language score table can be reached by following a child node from the non-terminal node. The storage medium according to claim 1, comprising a maximum value of a corresponding entry in the first language score table of a terminal node. Create a language score table to be assigned to a plurality of non-terminal nodes in a tree structure dictionary composed of a plurality of non-terminal nodes having child nodes and a plurality of terminal nodes corresponding to words without having child nodes. A language score table creation program for causing a computer to function , wherein each of the plurality of non-terminal nodes corresponds to a predetermined phoneme,
The tree structure dictionary
By following the child node from the first node of the tree structure dictionary, the terminal node corresponding to the word whose phoneme corresponding to the non-terminal node is part of its own phoneme sequence via each non-terminal node It is configured to reach all of the
Each of the plurality of terminal nodes is provided with a first language score table of a corresponding word according to a predetermined language model,
The language score table creation program includes:
For each of the plurality of non-terminal nodes, a program part that causes the computer to function to calculate the number of terminal nodes that can be reached by following the child nodes from the non-terminal node;
In response to receiving a request for a language score table from a parent node of the non-terminal node in each of the plurality of non-terminal nodes, a language score table of the child node is requested to each of the child nodes of the non-terminal node A program part that causes the computer to function ,
By executing a program part that requests the language score table in each of the plurality of non-terminal nodes, the second language of the non-terminal node is based on the language score table returned from each of the child nodes of the non-terminal node. A program part that causes the computer to function to create a score table;
A program part for causing a computer to return the second language score table created by the program part for creating the language score table to the parent node;
A program part that causes a computer to function as to whether or not the number of terminal nodes that can be reached by following a child node from the non-terminal node satisfies a predetermined condition with respect to a predetermined threshold value;
In response to determining that the condition is satisfied , the computer stores the second language score table created for the non-terminal node in a storage device as the language score table of the non-terminal node. A program part to function ,
In response to determining that the condition is not satisfied, a process of requesting language score information having a smaller storage capacity than the second language score table to a child node of the non-terminal node. A program part that causes the computer to function as it does , and
The language score information having a smaller storage capacity than the second language score table returned from the child node in response to the process of requesting the language score information having a smaller storage capacity than the second language score table. Based on the non-terminal node , a program part that causes a computer to function to create language score information having a smaller storage capacity than the second language score table and store the language score information in a storage device;
In response to receiving the request for Do language score information small in required storage capacity than the second language score table from the parent node, required than the second language score table stored in the storage device A language score table creating program including a program part that causes a computer to function to return language score information with a small storage capacity to the parent node. The language score table creation program according to claim 6, wherein the predetermined language model is a bigram language model. The language score information having a smaller storage capacity than that of the second language score table is a maximum value of a unigram score of each terminal node that can be reached by following a child node from each non-terminal node. Or the language score table preparation program in any one of Claim 7. The language score information having a smaller storage capacity than that of the second language score table is a maximum value of a class bigram score of each terminal node that can be reached by following a child node from each non-terminal node. Or the language score table preparation program in any one of Claim 7. The second language score table has the same number of entries as the first language score table, and each entry of the second language score table can be reached by following a child node from the non-terminal node. Consisting of the maximum value of the corresponding entry in the first language score table of the terminal node,
The program part that causes the computer to function to create the second language score table of the non-terminal node is executed by executing a program part that requests the language score table in each of the plurality of non-terminal nodes. Including a program portion that causes a computer to function to create a second language score table for the non-terminal node with a maximum value for each entry in the language score table returned from each of the child nodes of the non-terminal node. The language score table creation program according to any one of Items 6 to 9.

Images