JPH10187185A - Device and method for language processing - Google Patents

Abstract

PROBLEM TO BE SOLVED: To eliminate the need for unnecessary computations, to precisely predict the language score of unretrieved portions and to greatly reduce the retrieval space during a language evaluation. SOLUTION: An LR table storage section 20 records the information, which makes the correspondence between an inputted character string and the result of a syntax analysis, the score of the language evaluation corresponding to applicable rewriting rules and the predicted value of the evaluation score generated after the state for each state. A syntax analysis section 12 refers to the LR table and conducts a syntax analysis. A language score computing section 14 refers to the LR table and computes the score generated by the syntax analysis. A candidate recording section 30 records the information of the candidate being processed, the evaluation score outputted by the section 14 and the predicted score read from the section 20. A syntax analysis control section 13 refers to the data of the section 30 and determines the candidates against whom analyses are conducted.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a language unit (hereinafter, sentence) composed of one or more words, such as a phrase or a sentence,
The present invention relates to a technique for performing high-speed parsing and linguistic evaluation using a certainty score.

[0002]

2. Description of the Related Art Speech recognition, character recognition, automatic translation, semantic analysis, and the like require linguistic analysis to evaluate syntax analysis and linguistic certainty. It is determined whether the result of the parsing is correct as a sentence of the language.
If necessary, if it is determined that the sentence is correct,
Analyze the language structure. Linguistic certainty is an evaluation of the likelihood of a sentence belonging to the language, and is generally evaluated by a numerical score. When a plurality of analyzes, such as voice recognition and character recognition, are simultaneously advanced, or when a language that includes ambiguity in a natural language is processed, a plurality of candidates are generated as a result of the analysis.
In particular, when a large number of recognition results occur such as voice recognition and character recognition, it is not realistic to perform linguistic evaluation on all the recognition results. In such a case, we aim to speed up the processing by taking measures such as erasing those that are unlikely as a result according to the score in the middle or proceeding with the analysis by giving priority to the one with a good score ing.

In particular, in the field of speech recognition, a large number of language candidates to be analyzed are generated in the course of recognition, so that a method for high-speed language processing is required.

[0004] In a lecture paper "A study on A * search in context-free parsing of syllable graphs" in the spring meeting of the Acoustical Society of Japan in 1996, Yoshiharu Abe and Kunio Nakajima described a directed graph of syllables output as a result of acoustic processing as context. It discloses a method of using the A * algorithm when performing free syntax analysis. The A * algorithm is an algorithm that implements the best priority search. When the search proceeds in order from the candidate with the highest evaluation, the score with the highest evaluation is determined by adding the estimated score of the score of the unsearched part to the score of the part where the search has been completed. If the score estimated at this time is always estimated in a better direction than the finally determined score, it is said that the acceptable condition is satisfied, and it is guaranteed that a solution with the best score is found. In the A * algorithm, if there is a method capable of accurately predicting the final score without requiring much time for calculating the prediction score, efficient search can be realized.

[0005] As another technique, evaluation is performed when an evaluation score for evaluating a candidate is calculated using an acoustic score obtained by acoustic verification and a linguistic score obtained by linguistic evaluation in speech recognition. Japanese Patent Application Laid-Open No. Hei 7-295587 discloses a technique that uses a part of the language score of the missing part. A word-based reference signal is constructed in a tree structure, and the language score generated when each node of the tree structure receives a word that may be determined as a result of further matching than that node is calculated. Then, it is added to the node and recorded in advance.
When the check signal is compared with the reference signal, if the score recorded in the node is used, the evaluation score for comparing the candidates becomes close to an accurate score, so that the pruning standard can be strict. Moreover, since the calculation of the language score to be recorded in the node is performed in advance, only the score is referred to at the time of the recognition processing, and the calculation load at the time of the recognition processing does not increase so much. As described above, by calculating in advance the prediction score, which is an estimated value of the score of the portion that has not been searched, the pruning criterion can be made more strict and the number of valid candidates can be reduced.

[0006]

As described above, beam search and best-priority search are often used as methods for reducing the huge search space when proceeding with language processing.
In the beam search and the best priority search, the accuracy of the pruning criterion and the score for determining the best candidate are important. When performing acoustic matching first in speech recognition and then performing linguistic evaluation on the results, the linguistic evaluation is performed together with the evaluation score of the previously evaluated part. There is a technique for further improving the accuracy of the score by using the result of the acoustic processing as a prediction score. However, since the estimated value (predicted language score) of the language score of the unsearched portion is not prepared in advance, the predicted language score cannot be used as it is. Therefore, the accuracy of the evaluation score is not sufficient, and the search space cannot be significantly reduced. If the predicted language score is calculated during the analysis, the predicted language score can be used, but extra time is required for the processing, and the effect of reducing the search space is reduced.

On the other hand, Japanese Unexamined Patent Application Publication No. 7-295587
According to the technique disclosed in Japanese Patent Laid-Open Publication No. H10-210, a method is disclosed in which a language score that would be generated if the word being collated at that time is temporarily determined is calculated before the word is determined. However, in this technique, only a score newly generated when a word to be collated at that time is determined can be preliminarily calculated. When a plurality of words are linked to form a sentence or the like, a language score cannot be predicted for words after the word being collated. Since sufficient accuracy cannot be expected with this, it is difficult to significantly reduce the search space.

As described above, in proceeding with the syntax analysis and the accompanying language processing, there is no disclosure of a method for acquiring a predicted language score generated when the analysis proceeds without taking a long time during the processing. Therefore, it has been difficult to significantly reduce the search space for language processing without lowering the processing accuracy.

The present invention has been made in view of such circumstances, and has as its object to obtain a predicted language score without taking extra time for processing.

[0010]

According to the present invention, when performing language evaluation while performing syntax rules using a chart system, a (general) LR system, a finite state automaton, or the like, all syntax rules are used. For the analysis state or a part of the analysis state particularly required, a language sequence that can be succeeded is predicted in advance. A language score associated with the predicted language sequence is calculated, and the most appropriate value as the predicted language score is stored in association with the analysis state. The analysis state is an edge in the chart system, a row or stack state of the LR table in the LR system, a state in the finite state automaton, and the like. Then, when performing language evaluation, a predicted language score can be obtained without newly calculating by referring to the corresponding predicted language score from the state of the syntax analysis at that time. By using the predicted language score for the pruning evaluation of the beam search or the heuristic score of the best priority search, the search can be efficiently advanced.

Further, the language score predicted in advance is smaller than the score obtained when the evaluation is actually advanced.
If the same or better evaluation is guaranteed, the best priority search using the predicted score is a search that satisfies the acceptable condition. At this time, there is an advantage that it is guaranteed that the best evaluation solution is found. When the absolute value of the linguistic score is smaller and the score is better, a system is often used in which the absolute value of the score increases as the number of connected vocabulary items increases. As described above, in a system in which the score worsens as the connection of the vocabulary items becomes longer, the best prediction score can be determined in advance by performing a full search when calculating the prediction score in advance. By storing the best score as a prediction score in association with the analysis state and using it as a prediction score at the time of evaluation, the evaluation becomes the same as or better than the actual score after the evaluation.

There are various language scores that can be used here. N-gra for connection of sounds such as syllables
statistical information such as m, statistical information such as n-gram relating to the connection of vocabulary items such as words and morphemes, empirical scores given to longer vocabulary items, and empirical information unique to vocabulary items and parts of speech Or a score that is statistically or empirically assigned to a syntax rule based on an expression such as a rewriting rule. Each of these knowledge sources may be used alone, or a plurality of them may be combined. When a plurality of knowledge sources are combined, a method of multiplying the score of each knowledge source by a certain constant and then taking the sum is easy to calculate, and is suitable for the best priority search. In addition, any of the score systems of the knowledge sources mentioned here can be a system in which the score becomes worse as the number of connected vocabulary items increases. That is, an acceptable best priority search can be realized.

[0013]

Embodiments of the present invention will be described below. Embodiment 1 FIG. 1 shows the configuration of a language processing apparatus according to Embodiment 1 of the present invention. In FIG. 1, the language processing apparatus includes a language processing control unit 10, an LR table storage unit 20, and a candidate storage unit 30. And so on. The language processing control unit 10 includes a syntax analysis unit 12, a syntax analysis control unit 13, a language score calculation unit 14, and the like, and performs a process of analyzing an input character string while referring to the LR table in the LR table storage unit 20. Control the flow of FIG. 2 shows the contents of the LR table in this embodiment. L
The R table includes information for associating an input character string with a result of parsing a syntax, a score of a language evaluation corresponding to a rewriting rule to be applied, and an evaluation score generated for each state after that state. And the predicted value of. In order to create the LR table, a syntax rule given in the form of a rewrite rule may be prepared in advance, and the rule may be compiled. In addition, a score is assigned to each rewriting rule according to conditions such as frequency and importance of applying the rule and naturalness as a sentence. Here, rules that are easier to apply have lower values, and rules that are less likely to be applied are assigned higher values. In general, a score that satisfies the above condition can be calculated by determining the probability that the rule is applied and taking the absolute value of the logarithm. The probability that a rule is applied can be calculated according to the number of rules used in the analysis by analyzing a sample set of sentences using those rules. The calculated score is LR
Write in the score table along with the reduce numbers corresponding to each rewrite rule in the table. The predicted value of the evaluation score in each state of the LR table is a predicted value of a score generated by a reduce operation applied from that state to the end of the sentence. The flow of the process of calculating the prediction score will be described later.

The syntax analyzer 12 performs syntax analysis with reference to the LR table. The language score calculator 14 calculates a score generated by the syntax analysis with reference to the LR table. Candidate recording unit 30
Then, information on the candidate being processed, the evaluation score output by the language score calculation unit 14, and the prediction score read from the LR table storage unit 20 are recorded. The syntax analysis control unit 13 refers to the data in the candidate recording unit 30 to determine a candidate for which analysis is to be performed.

FIG. 3 shows a flowchart of the language processing in the first embodiment. First, one candidate for the initial state of syntax analysis is prepared for the character string to be processed (step S1).
01). Next, the first input character is read as a look-ahead symbol (step S102). The LR table is referred to for the initial state candidate and the read-ahead symbol (step S).
103). If the operation on the look-ahead symbol in each candidate is possible as a result of referring to the LR table, the operation is performed to create a new candidate (step S105). If a new candidate is not created as a result, it indicates that the syntax analysis has failed (step S104). If the syntax analysis fails, the result is output and the processing is terminated (step S
109). If a new candidate remains without failure, score evaluation is performed on the candidate (step S106).
In the evaluation of the score, for the candidate who performed the reduce operation when referring to the LR table, the score corresponding to the reduce operation is added to the evaluation score of the candidate up to that time. Since the score corresponding to the reduce operation is recorded in the reduce portion in the LR table, the score is read. Then, the prediction score is read for the state in the new LR table. Since the prediction score is recorded in the action table portion in the LR table, it is read.
After the evaluation score and the prediction score are newly calculated, the candidate whose total score has a difference greater than or equal to a predetermined value from the score of the candidate with the highest score is deleted (step S107). The deleted candidate is no longer processed. The operations from step S102 to step S107 are performed until the last input character is reached or the analysis fails. When the processing for the last input character is completed, the candidate having the highest evaluation score is output as a recognition result (step S10).
9).

FIG. 4 shows a flowchart of a process for calculating a prediction score from each state recorded in the LR table. First, one state for parsing is extracted (step S201). In this embodiment, the target state is a state corresponding to a row of the LR table. From the extracted state, all states that can be changed next are listed (step S2).
02). A linguistic score generated at the time of the transition is added to all the transitionable states, and a score required to transition from the original state to the state is calculated (step S203). The operations of step S202 and step S203 are repeated until it is determined that an appropriate prediction score has been found (step S204). In the present embodiment, the smaller the absolute value of the score is, the better the score is, and the score is set to be worse as the state of the parsing is traced, so that it is determined that an appropriate prediction score is found. The following three exist. In the first case, a candidate for an analysis that has reached an acceptable state as the completion of parsing is created, and its predicted score is the score of a candidate that continues to perform another analysis and the predicted score of a candidate that has reached acceptance until then. Are smaller values. The second case is a case where the score of the candidate for which the analysis is continued increases and becomes worse than the candidate of the prediction score which has already reached the acceptable state. The third case is when there are no more candidates to be analyzed. Step S2
When it is determined in step 04 that the prediction score has been found, the found prediction score is recorded in the corresponding recording location (step S205). If it is determined that an appropriate prediction score has not been found, the process returns to step S202 to advance the search. When the operations from step S201 to step S205 have been performed for all the states, the process ends. In the above process, when calculating the prediction score for a certain state, if the prediction score of another state that has been passed before the acceptance is already calculated, the calculation score is calculated by using the prediction score. Can be reduced.

[Second Embodiment] Next, a description will be given of a second embodiment in which the present invention to which the best priority search is applied is used for speech recognition language processing. FIG. 5 is a configuration diagram of the second embodiment. In FIG. 5, a phoneme recognition unit 60 performs signal processing on an input voice and checks it against a syllable-level voice model represented by an HMM (Hidden Markov Model) or the like. Then, speech is recognized in units of syllables, mora, and phonemes. In the present embodiment, the recognized result is expressed by a network expression (hereinafter, a mora graph) using a mora as an arc. The output of the phoneme recognition unit based on the mora graph is as shown in FIG. The language processing control unit 11 receives a mora graph as input, and stores the dictionary storage unit 22, the LR table storage unit 21, the candidate recording unit 3
1. A linguistic evaluation is performed together with the score calculation unit 40 and the linguistic knowledge source 50, and a speech recognition result is output according to the evaluation result. The dictionary storage unit 22 has information on what morphemes can be formed by the arrangement of mora. LR
The table storage unit 21 stores information obtained by compiling information on what morpheme sequence can be accepted as a sentence into an LR table format. The candidate recording unit 31 holds information on the candidate under search. The linguistic knowledge source 50 holds information on the degree of linguistic likelihood. In the present embodiment, a monogram of mora (mono-
gram, bi-gram, tri-gram, and morpheme monogram, bigram, trigram, and part of speech minimum cost methods. Monograms, bigrams, and trigrams of mora and morphemes count the number of mora and morphemes contained in text data in advance,
Give it statistically. The minimum cost of the part of speech is empirically given in advance for each part of speech so that the recognition result is improved. The score calculation unit 40 refers to the linguistic knowledge source 50 and calculates an evaluation score of a candidate that is being searched or has been searched.

FIG. 7 is a flow chart showing the flow of processing of the language processing control section 11 of the second embodiment. First, the phoneme recognition unit 6
The mora graph output from 0 is received (step S30).
1). Next, a candidate in the initial state is generated, and the candidate recording unit 31
(Step S302). The candidates for the initial state in the present embodiment are candidates for which no morpheme has been received, the reaching node is the start node of the mora graph, and the evaluation score is 0. Then, the number of candidates stored in the candidate storage unit 31 is checked, and if the number is 0, the process ends (step S303). However, initially, step S30
Since the candidate is generated in step 2, the number of candidates is not zero. Next, one candidate with the best score is extracted (step S304). Here, the best score is a score having the smallest sum of the evaluation score and the prediction score up to that point. If the extracted candidate is established as a solution,
The candidate is output as a recognition result (step S30)
5, step S306). The fact that the candidate is established as a solution means that the reaching node has reached the end node of the mora graph, and the parsing is acceptable. The output recognition result may be displayed on the display each time it is output, or may be stored together and stored in the language processing control unit 11.
May be displayed collectively when the processing for one utterance is completed. When the number of output solutions reaches a predetermined number such as one or five, the process ends there (step S307). If the number of solutions has not reached the predetermined number, the process proceeds to step S303 and the process proceeds to the next candidate. If it is determined in step S305 that the candidate is not a solution, dictionary matching is performed on the mora graph from the node reached by the candidate (step S30).
8). If the plurality of morphemes are successfully matched in step S308, the morphemes that have been successfully matched for each result, their parts of speech, the reached nodes on the mora graph,
And a phoneme prediction score attached to the reaching node. Next, syntax analysis is performed on the morpheme for which the dictionary search was successful in step S308 with reference to the LR table (step S309). As a result of referring to the LR table, candidates that failed in analysis are deleted. A candidate that has been successfully analyzed reads its parsing state and the predicted language score in that state. For the candidate for which the syntax analysis was successful, the score calculating unit 40 calculates a score according to the mora and morpheme sequence of the candidate (step S310).
The score calculation unit 40 calculates an evaluation score with reference to the language knowledge source 50. The newly created candidate is recorded in the candidate recording unit 31 according to the number of candidates in the candidate recording unit 31 and the evaluation score of the newly created candidate (Step S).
311). Among the candidates already stored in the candidate recording unit 31, candidates with a score that is worse than a certain difference from the best score and candidates are arranged in the order of the score, and the candidates are ranked in a predetermined number. The bad candidate is deleted from the candidate recording unit 31. The processing from step S303 to step S311 is repeated until there are no more candidates in the candidate recording unit 31 or the number of candidates output as solutions exceeds a predetermined number.

The flow of the process of calculating the prediction score is in accordance with the first embodiment. However, it is different from the first embodiment in that the points obtained by summarizing the scores of various knowledge sources are evaluated.

[0020]

As described above, according to the present invention,
When performing language evaluation, the language score of the unsearched portion can be accurately predicted without requiring extra calculation, and the prediction space is significantly reduced by using the predicted score. Therefore, the processing time for language evaluation can be performed in a significantly short time. In particular, when there is ambiguity such as the result of voice recognition or character recognition, and the search space is wide, it is highly effective in speeding up the processing.

[Brief description of the drawings]

FIG. 1 is a block diagram illustrating a configuration of a first exemplary embodiment of the present invention.

FIG. 2 is a diagram illustrating an example of an LR table including a prediction score according to the first embodiment.

FIG. 3 is a flowchart illustrating an operation according to the first exemplary embodiment.

FIG. 4 is a flowchart illustrating an operation of calculating a prediction score according to the first embodiment.

FIG. 5 is a block diagram illustrating a configuration of a second exemplary embodiment of the present invention.

FIG. 6 is a diagram illustrating an example of a mora graph according to the second embodiment.

FIG. 7 is a flowchart illustrating the operation of the second embodiment.

[Explanation of symbols]

 10, 11 language processing control unit 12 syntax analysis unit 13 syntax analysis control unit 14, 40 language score calculation unit 20, 21 LR table storage unit 30, 31 candidate storage unit 60 voice recognition unit

Claims (6)

[Claims] 1. A language processing apparatus for performing language evaluation together with syntax analysis, comprising: a syntax analysis means for performing syntax analysis; and an analysis state that can be taken at the time of syntax analysis. A means for recording the predicted value of the language evaluation score to be generated; a means for generating a language evaluation score for each processing during the syntax analysis; and an analysis state that has been reached before the analysis state is reached. A language processing apparatus comprising: a syntax analysis control unit that controls the syntax analysis unit based on a score of a language evaluation and a predicted value of the score of the analysis state. 2. The parsing control means performs a parsing process after the parsing state when the sum of the score of the language evaluation and the predicted value of the score is poorly evaluated for the reached parsing state. 2. The language processing apparatus according to claim 1, wherein said language processing apparatus is not provided. 3. The language according to claim 1, wherein the syntax analysis control means continues the syntax analysis processing after the analysis state in the order in which the sum of the score of the language evaluation and the predicted value of the score is evaluated well. Processing equipment. 4. The language processing apparatus according to claim 3, wherein a best predicted value that can be predicted from a predetermined analysis state of the syntax analysis is a predicted value recorded in correspondence with the analysis state. 5. The language processing apparatus according to claim 1, further comprising a speech recognition unit, wherein the syntax analysis and the language evaluation are performed on a speech recognition result by the speech recognition unit. 6. A language processing method for performing a language evaluation together with a syntax analysis, comprising the steps of: performing a syntax analysis; Recording a predicted value of the evaluation score; generating a language evaluation score for each process during the syntax analysis; and determining the analytic state that has been reached before reaching the analytic state. A language processing method, comprising: controlling a step of performing the syntax analysis based on a score and a predicted value of the score of the analysis state.