JP2008242607A - Device, method and program for selecting proper candidate from language processing result - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a language processor for selecting proper processing results from candidates of processing results. <P>SOLUTION: This language processor is provided with a first storage part 121 for storing constitutional units and a generation probability of a sentence; a second storage part 122 for storing a dependency relation expressed by the constitutional units of a dependency destination and a dependency origin and a conditional probability of the appearance of the constitutional units of the dependency origin with respect to the dependency destination; an input accepting part 101 for accepting the candidate of the processing result; an analyzing part 103 for analyzing the dependency structure of the candidate of the processing result; a calculation part 103a for acquiring the generation probability corresponding to the constitutional units of a sentence end from the first storage part 121 about each of the candidates of the dependency structure, and for acquiring the conditional probability corresponding to the dependency relation from the second storage part 122, and for calculating the generation probability of the candidate of the dependency structure as a product of all the conditional probabilities and a product of the generation probabilities; and a selection part 104 for searching the candidates of the dependency structure maximizing the calculated generation probability, and for selecting the candidates of the processing result corresponding to the searched candidate. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to an apparatus, method, and program for selecting an appropriate sequence by dependency analysis using a plurality of recognition candidate sequences obtained from speech recognition, character recognition, and the like as input.

  2. Description of the Related Art Conventionally, natural language recognition processing techniques such as speech recognition and character recognition for converting speech uttered by humans and character images written on paper into character strings and word strings are widely known. In such a recognition process, it is difficult to realize a process that correctly recognizes a character string or word string intended by the input person 100%. For example, in speech recognition, recognition errors often occur due to the presence of phonemes with similar features and background noise.

  In general speech recognition processing, first, feature amount data is extracted from a user's speech signal captured by a microphone or the like by applying FFT (Fast Fourier Transform) analysis or the like. Subsequently, using a phoneme dictionary storing a standard pattern of feature value data of each phoneme stored in advance or a word dictionary storing correspondence information between phoneme symbol strings and word headings constituting each word, Convert to string.

  In the HMM (Hidden Markov Model) method, which is a typical speech recognition method, the transition relationship between phonemes is expressed by a word network, and data with probability values assigned to links between nodes (corresponding to phonemes) in the network is represented. Store in word dictionary. Then, by comparing the input feature data of the voice with the standard pattern stored in the phoneme dictionary, the similarity with each phoneme is obtained, and the word network stored in the word dictionary is obtained based on the similarity. The word candidate having the highest similarity to the input speech is used.

  When a sentence composed of a plurality of words is used as a recognition target, it is necessary to perform word recognition as described above and obtain a correct word string for the input voice signal. For example, in order to obtain the most probable word sequence from a plurality of word sequences obtained as candidates, a language model such as n-gram expressing the ease of connection of a plurality of words is used.

  According to the HMM method, the candidate narrowing process using the n-gram can be performed in a unified manner with the word recognition process. The transition probability between phonemes in the word dictionary and the transition probability between words in the n-gram can be learned in advance from speech data or a text corpus. Moreover, highly reliable speech recognition can be performed by using a large amount of speech data and a text corpus.

In the above recognition processing, when the input sequence is described as I and the word sequence obtained as an output is described as W, the conditional probability P (W | I) when the input sequence I is given is maximized. This is equivalent to obtaining a word sequence W such as argmax _WP (W | I). Here, P (W | I) can be transformed to P (W | I) = P (I | W) · P (W) / P (I) by Bayes' theorem. Further, since the input sequence I is a given value, the denominator P (I) can be considered to be fixed. Therefore, the above recognition process can be formulated as a problem of obtaining a word sequence W that is argmax _W P (I | W) · P (W).

  On the other hand, for example, when the recognition process is performed by inputting Japanese, it is necessary to obtain a dependency structure between phrases obtained from the recognized word series by a dependency analysis process. The dependency analysis is a process for obtaining a dependency structure by analyzing a modification / modification relationship between words. Dependency analysis methods include syntactic analysis to find the syntactic structure of sentences using part-of-speech as a clue for analysis, then narrow down the candidates for dependency destinations to obtain dependency structure, and direct dependency analysis from word sequences. There are methods to go and find the dependency structure. Non-Patent Document 1 proposes a technique that models the strength of dependency between words as a dependency probability and performs dependency analysis stochastically.

Masahiro Ohno, Shigeki Matsubara, Nobuo Kawaguchi, Yasuyoshi Inagaki “Statistical Dependency Analysis of Japanese Spoken Dialogue and Its Evaluation”, Information Processing Society of Japan 65th Annual Conference, Vol. 2, pp. 1-2, 2003.

  However, since the method of Non-Patent Document 1 selects an optimal dependency structure from one input phrase string, it cannot be applied to the process of selecting an optimal phrase string from a plurality of phrase strings. There was a problem.

  For example, if a candidate having an optimum dependency structure can be selected from a plurality of recognition result candidates obtained by speech recognition, it is considered that the accuracy of speech recognition can be improved. However, in the first place, the method of Non-Patent Document 1 seeks a maximum likelihood dependency structure for a given phrase string. Further, since the obtained dependency structure is simply the maximum likelihood structure for the input word sequence, it cannot be compared with a structure obtained from another word sequence. Therefore, by the method of Non-Patent Document 1, it is not possible to realize a process of selecting an optimum phrase string from a plurality of phrase strings in consideration of the dependency structure.

  The present invention has been made in view of the above, and an apparatus, method, and program capable of selecting an appropriate processing result from candidates for processing results such as recognition processing in consideration of an analysis result of dependency analysis The purpose is to provide.

  In order to solve the above-described problems and achieve the object, the present invention provides a language processing apparatus that selects a processing result from processing result candidates for a sentence constituent unit, the constituent unit and the constituent unit. A first storage unit that stores the occurrence probability of the relationship, a dependency relationship represented by the structural unit that is a dependency destination and the structural unit that is a dependency source, and the structural unit that is the dependency destination A second storage unit that stores the conditional probability of occurrence of the constituent unit that is the source of the association, an input receiving unit that receives an input of the candidate processing result, and the received candidate processing result For each of the analysis unit for analyzing the dependency structure representing the combination of the dependency relationships between the structural units, and for each of the analyzed dependency structure candidates, the analysis unit corresponds to the structural unit at the end of the sentence. The occurrence probability is acquired from the first storage unit, the conditional probabilities corresponding to the dependency relationships included in the dependency structure are acquired from the second storage unit, and all the acquired conditions are acquired. A calculation unit that calculates the occurrence probability of the dependency structure candidate that is a product of the product of the attached probability and the acquired occurrence probability, and obtains the determination of the dependency structure candidate that maximizes the calculated occurrence probability. And a selection unit that selects the processing result candidate corresponding to the dependency structure candidate as the processing result.

  Further, the present invention is a method and program capable of executing the above-described apparatus.

  According to the present invention, it is possible to select an appropriate processing result from candidates for processing results such as recognition processing in consideration of the analysis result of dependency analysis.

  Exemplary embodiments of an apparatus, a method, and a program according to the present invention will be described below in detail with reference to the accompanying drawings.

(First embodiment)
The language processing apparatus according to the first embodiment inputs a plurality of processing result candidates by language processing such as speech recognition processing, and selects an optimum processing result by referring to the result of dependency analysis. . In the following, language processing for Japanese is described as an example, but the target language is not limited to Japanese.

  FIG. 1 is a block diagram illustrating a configuration of a language processing apparatus 100 according to the first embodiment. As shown in FIG. 1, the language processing apparatus 100 includes a first storage unit 121, a second storage unit 122, a third storage unit 123, an input receiving unit 101, a control unit 102, and an output unit 105. I have.

  The 1st memory | storage part 121 memorize | stores the occurrence probability table 121a which stores the occurrence probability of a phrase. FIG. 2 is an explanatory diagram illustrating an example of a data structure of the occurrence probability table 121a according to the first embodiment. As shown in FIG. 2, the occurrence probability table 121a stores a phrase and the occurrence probability of the phrase in association with each other. As the phrase occurrence probability, a value calculated in advance using a large amount of speech data or a text corpus is stored in the occurrence probability table 121a.

  The second storage unit 122 stores a conditional probability table 122a that stores conditional probabilities of dependency relationships. The conditional probability of the dependency relationship represents the probability that the phrase that is the dependency source of the dependency relationship will appear with respect to the clause that is the dependency relationship of the dependency relationship. As for the conditional probability of the dependency relationship, a value calculated in advance using a large amount of speech data or a text corpus is stored in the conditional probability table 122a.

  FIG. 3 is an explanatory diagram illustrating an example of a data structure of the conditional probability table 122a according to the first embodiment. As shown in FIG. 3, the conditional probability table 122 a stores a relation source clause, a relation destination clause, and a conditional probability in association with each other.

  The occurrence probability table 121a and the conditional probability table 122a are referred to when the calculation unit 103a described later calculates the occurrence probability of the dependency structure.

  The third storage unit 123 stores a dictionary table 123a that stores dictionary information such as part of speech information of each word. FIG. 4 is an explanatory diagram showing an example of the data structure of the dictionary table 123a. As shown in FIG. 4, the dictionary table 123a stores word headings, parts of speech, and categories representing independent words or attached words in association with each other.

  As will be described later, in this embodiment, in principle, an input of a phrase sequence is received and dependency analysis or the like is performed, but the input of a word sequence is accepted and the same applies to a phrase sequence generated from the word sequence. It can be configured to perform processing. The dictionary table 123a is a table that is referred to when a phrase series is generated from a word series. Details of the phrase series generation processing will be described later.

  Note that the first storage unit 121, the second storage unit 122, and the third storage unit 123 are all commonly used such as a hard disk drive (HDD), an optical disk, a memory card, and a random access memory (RAM). It can be configured by a storage medium.

  The input receiving unit 101 receives a phrase series input as a constituent unit of a sentence to be subjected to dependency analysis. For example, the input receiving unit 101 receives an input of a phrase series obtained as a result of language processing such as speech recognition, character recognition, and morphological analysis.

  FIG. 5 is an explanatory diagram showing an example of an input phrase series. In the figure, five similar speech recognition result candidates are generated by speech recognition processing for Japanese speech meaning “Taro ate bad food”, and five corresponding phrase sequences are input. An example is shown. According to the method of the present embodiment, the dependency relationship between clauses is analyzed for each of these five candidates, and the most suitable candidate is selected as the speech recognition result with reference to the analysis result of the dependency relationship. It becomes possible.

  Note that the composition unit of the sentence to be accepted is not limited to the phrase series, and the input reception unit 101 is configured to accept the composition unit of other sentences such as words as long as it is a unit subject to dependency relation analysis. May be. For example, when a language that does not have a particle, such as Chinese, is to be processed, a dependency analysis can be performed by receiving an input of a word sequence.

  Further, even when dependency analysis is performed on a phrase sequence such as Japanese, first, an input of a word sequence is received, and a phrase sequence is generated from the word sequence with reference to the dictionary table 123a. May be. In this case, the input receiving unit 101 acquires the category of the word in the accepted word sequence from the dictionary table 123a, and adds the word category to the immediately preceding word that is an independent word when the category of the word is an attached word. The phrase sequence is generated by

  The control unit 102 controls processing for selecting an optimum phrase sequence from the phrase sequences received by the input receiving unit 101 as a processing result, and includes an analysis unit 103 and a selection unit 104.

  The analysis unit 103 performs dependency analysis on the phrase series received by the input receiving unit 101, and generates a dependency structure candidate represented by a combination of dependency relationships between clauses. The analysis unit 103 also includes a calculation unit 103a that calculates the occurrence probability of a generated dependency structure candidate during the dependency analysis process.

  The selection unit 104 selects an optimum phrase series as a processing result by referring to the result of dependency analysis from the inputted phrase series. Specifically, the selection unit 104 first refers to the dependency structure candidates analyzed by the analysis unit 103 and the occurrence probabilities calculated by the calculation unit 103a for each candidate, so that the occurrence probability is maximized. Find the receiving structure candidate. Then, the selection unit 104 selects a phrase series corresponding to the obtained dependency structure candidate as a processing result.

  The output unit 105 outputs the processing result selected by the selection unit 104.

  Before describing the details of candidate selection processing for selecting an optimal candidate by performing dependency analysis, the expression format of the dependency structure, the occurrence probability, and the sequential generation process of the dependency structure will be described.

  FIG. 6 is an explanatory diagram illustrating an example of a representation format of the dependency structure. In FIG. 6, an example of a phrase series including four phrases (phrase series 201 to 205), dependency structures 206 to 210 illustrating dependency relations between phrases for each phrase series, and dependency structures expressed in a list structure. (List structures 211 to 215) are shown in association with each other.

  This figure can be analyzed for a given phrase sequence, assuming two conditions: (1) the dependency relationship between clauses does not intersect; (2) the previous clause is related to the subsequent clause. An example of a dependency structure is shown.

  When a phrase series is given, there is a semantically valid dependency structure among dependency structures in which the dependency relationship from the preceding phrase to the subsequent phrase does not intersect. For example, in the case of a phrase sequence consisting of four phrases, there are five appropriate dependency structures as shown in FIG. The numerical values in the phrase series and the dependency structure are values for identifying the phrases, and are assigned serial numbers starting from 1 from the beginning to the end of the sentence.

  For example, the phrase series 201 corresponds to the dependency structure 206, and the dependency structure 206 is related to the first phrase (Taro's) related to the second phrase (elder sister's), and the second phrase (elder sister's). Is related to the third phrase (cooking), and the third phrase (cooking) is related to the fourth phrase (eating). Hereinafter, the i-th clause may be referred to as the i-th clause.

  The list structure 211 represents the dependency structure 206 in a list format. Note that d (i, j) represents a dependency relationship in which the i-th clause relates to the j-th clause.

  Next, the occurrence probability of the dependency structure will be described with reference to FIG. FIG. 7 is an explanatory diagram showing an example of the occurrence probability of the dependency structure. In FIG. 7, the occurrence probabilities 306 to 310 for the dependency structures 206 to 210 of FIG. 6 are shown.

In the figure, w _i is the i-th clause, P (w _i ) is the probability of occurrence of w _i , P (w _i , w _j ) is the probability of occurrence of w _i and w _j simultaneously, and P (w _i | w _j ) represents the conditional probability of w _i with respect to w _j . Thus, for example, the occurrence probability of the dependency structure 206 is {P (w ₁ , w ₂ ) / P (w ₂ )} · {P (w ₂ , w ₃ ) / P (w ₃ )} · P ( w ₃ , w ₄ ) = P (w ₁ | w ₂ ) · P (w ₂ | w ₃ ) · P (w ₃ | w ₄ ) · P (w ₄ ).

By generalizing this, the occurrence probability of the dependency structure S obtained from the phrase sequence consisting of N clauses becomes the conditional probability P (w) corresponding to all the dependency relationships d (i, j) included in S. _i | w _j ) and the product of the occurrence probability P (W _N ) of the sentence at the end of the sentence can be expressed by the following equation (1).

  Next, a process for obtaining a dependency structure candidate will be described with reference to FIGS. FIG. 8 is an explanatory diagram showing a process of obtaining a possible dependency structure from a four-phrase phrase sequence.

  First, consider the dependency of the last two clauses, that is, the third and fourth clauses. There is only one possible dependency relationship between two phrases, and a dependency structure 401 comprising this one dependency relationship is obtained. Next, considering the possible dependency relationship by adding the second clause, there are only two possibilities from the second clause, whether it relates to the third clause or the fourth clause. Therefore, the dependency structure 402 and the dependency structure 403 are obtained corresponding to these possibilities.

  Finally, consider the possible dependency relationships by adding the first sentence. There are three types of dependency structures derived from the dependency structure 402: dependency structures 206, 207, and 208. This is because there are three types of possibilities: when the first phrase is a dependency source, when it is dependent on the second phrase, when it is dependent on the third phrase, and when it is dependent on the fourth phrase. It corresponds.

  Further, from the dependency structure 403, two types of structures, the dependency structures 209 and 210, are derived. When the dependency structure 403 is assumed, it is possible to configure a dependency relationship that is non-intersecting from the first clause to the second and fourth clauses. On the other hand, the dependency relationship from the first clause to the third clause intersects with the dependency relationship from the second clause to the fourth clause, and does not satisfy the above-described dependency structure condition. For this reason, two types of structures are derived from the dependency structure 403.

  In FIG. 8, the process of generating dependency structure candidates in the case of four clauses has been described. However, when the number of clauses increases, a structure can be sequentially generated from the clauses at the end of the sentence.

  Next, the process for obtaining the occurrence probability shown in the equation (1) will be described with reference to FIGS. FIGS. 9-11 is a figure which shows an example of the process of calculating | requiring the occurrence probability of a dependency structure corresponding to the generation process of the dependency structure shown in FIG.

FIG. 9 is an explanatory diagram showing a process of calculating the occurrence probability of a dependency structure in which one sentence is added to two sentences at the end of the sentence. For the dependency structure 401 in FIG. 9, the occurrence probability is calculated by the product of the occurrence probability of the sentence at the end of the sentence and the conditional probability of the immediately preceding phrase with respect to the sentence at the end of the sentence. In this example, it can be represented by P (w ₃ | w ₄ ) · P (w ₄ ).

When the dependency structure 402 and the dependency structure 403 are generated based on the dependency structure 401, the dependency relationship from the second phrase to the third phrase and the dependency relation from the second phrase to the fourth phrase are newly added. Each relationship is added to the dependency structure 401. Accordingly, the occurrence probabilities of the respective dependency structures 402 and 403 are P (w ₂ | w ₃ ) and P (w ₂ | w ₄ ), which are conditional probabilities corresponding to the added dependency relationships, and the dependency structures 401. Can be calculated by multiplying the occurrence probability P (w ₃ | w ₄ ) · P (w ₄ ). Thus, the occurrence probabilities of the dependency structure 402 and the dependency structure 403 are P (w ₂ | w ₃ ) · P (w ₃ | w ₄ ) · P (w ₄ ) and P (w ₂ | w ₄ ), respectively. P (w ₃ | w ₄ ) · P (w ₄ )

  FIG. 10 is an explanatory diagram showing a process of calculating the occurrence probability of the dependency structure with one more sentence added. FIG. 10 shows a process of calculating the occurrence probabilities of the dependency structures 206 to 210 from the occurrence probabilities of the dependency structures 402 and 403.

  As for the occurrence probability of the dependency structure in which the first clause is further added to the dependency structures 402 and 403, the conditional probability corresponding to the newly added dependency relationship is also set as described above. Alternatively, it can be calculated by multiplying the occurrence probability of the dependency structure 403.

  FIG. 11 is a schematic diagram for explaining a case where the process of calculating the occurrence probability as described above is generalized. FIG. 11 shows that when the occurrence probability P of the dependency structure S (the dependency structure 1101) is obtained for a partial clause sequence from the I + 1 clause to the N clause, the clause I (the clause 1102) immediately before that is obtained. ) Is added, the calculation method of the occurrence probability is shown.

In this case, with respect to the clauses from the I + 1 clause to the N clause, if the dependency clause for the clause I whose dependency is non-intersecting is the clause i, the dependency relationship added is d (I, i). . Therefore, the newly generated dependency structure S ′ is [d (I, i) | S]. Here, [d (I, i) | S] represents a dependency structure in which the dependency relationship d (I, i) is added to the dependency structure S. Further, the occurrence probability P ′ of the dependency structure S ′ is a value obtained by multiplying P by the conditional probability P (w _I | w _i ).

  As described above, in this embodiment, a dependency structure is sequentially generated from the end of a sentence, and the occurrence probability of the generated dependency structure is calculated.

  Next, candidate selection processing by the language processing apparatus 100 according to the first embodiment configured as described above will be described with reference to FIG. FIG. 12 is a flowchart showing an overall flow of candidate selection processing in the first embodiment.

  First, the input receiving unit 101 receives an input of a plurality of phrase series (step S1201). As described above, the input receiving unit 101 may receive an input of a word series and generate a phrase series by referring to the dictionary table 123a.

  Next, the control unit 102 initializes the maximum value Pmax of the occurrence probability to 0 (step S1202). Next, the analysis unit 103 obtains one phrase series from the accepted plurality of phrase series, and executes dependency analysis processing for analyzing the dependency relation between phrases for the acquired phrase series (step S1203). In the dependency analysis process, the analyzed dependency structure candidates and the occurrence probabilities of the candidates are output. Details of the dependency analysis processing will be described later.

  Next, the selection unit 104 refers to the dependency structure candidate, which is the processing result of the dependency analysis process, and the occurrence probability of each candidate, and selects the maximum value P of the occurrence probability (step S1204). Subsequently, the selection unit 104 determines whether or not the selected maximum value P is larger than Pmax (step S1205).

  When P is larger than Pmax (step S1205: YES), the selection unit 104 sets P to Pmax, and selects a dependency structure candidate corresponding to P as an output processing result candidate (output candidate). (Step S1206).

  After selecting an output candidate or when it is determined in step S1205 that P is not greater than Pmax (step S1205: NO), the control unit 102 determines whether or not all phrase sequences have been processed (step S1205). S1207).

  If all the phrase series have not been processed (step S1207: NO), the analysis unit 103 selects the next phrase series and repeats the process (step S1203). When all the phrase sequences have been processed (step S1207: YES), the output unit 105 outputs the selected output candidate (step S1208), and ends the candidate selection process.

  As described above, in the conventional method, only the optimum dependency relationship is selected for one phrase sequence, but according to the present embodiment, each dependency relationship of a plurality of phrase sequences is selected. It is possible to select a phrase sequence from which the optimal dependency relationship is obtained as the optimal phrase sequence.

  Next, details of the dependency analysis processing in step S1203 will be described with reference to FIGS. FIG. 13 is a flowchart showing an outline of the entire flow of dependency analysis processing according to the first embodiment. FIG. 14 is a flowchart showing details of the entire flow of dependency analysis processing according to the first embodiment.

  That is, FIG. 14 corresponds to a detailed flowchart of the flowchart representing the outline of the dependency analysis process shown in FIG. The correspondence between each step in FIG. 13 and each step in FIG. 14 is shown in FIG.

  In FIG. 13, the analysis unit 103 first determines whether or not the number of phrases in the specified phrase series is 2 (step S1301). As will be described later, the dependency analysis process is recursively executed on a phrase series obtained by deleting the first phrase from the phrase series. For this reason, it is necessary to repeatedly delete the clauses and finally determine whether or not the last two clauses have been reached, and in the dependency analysis process executed in this case, it is necessary to return a specially calculated occurrence probability or the like . The determination process for this corresponds to the process of step S1301.

  When the number of clauses is 2 (step S1301: YES), the calculation unit 103a generates and outputs a dependency structure for the two clauses at the end of the sentence, and calculates and outputs the occurrence probability of the generated dependency structure. (Step S1302). In this case, the calculation unit 103a generates a dependency structure and calculates the occurrence probability by the method shown in FIGS.

  If the number of phrases is not 2 (step S1301: NO), the dependency analysis process is recursively performed on the phrase series excluding the first phrase of the specified phrase series (step S1303). In the dependency analysis process, as described above, the analyzed dependency structure candidates and the occurrence probabilities of the candidates are output.

  Note that when a phrase series having the number of phrases of 1 is input, the dependency analysis is not necessary, and thus the processing is terminated, but is omitted in FIG.

  Next, the analysis unit 103 generates a dependency structure in which the immediately preceding clause is added to each dependency structure candidate that is an analysis result of the dependency analysis processing (step S1304). At this time, the analysis unit 103 generates a dependency structure in which the newly added dependency relationship does not intersect with each dependency relationship of the existing dependency structure.

  Next, the calculation unit 103a calculates the occurrence probability of the generated dependency structure using the conditional probability corresponding to the dependency relationship with the added clause (step S1305). At this time, the calculation unit 103a acquires the conditional probability corresponding to the newly added dependency relationship from the conditional probability table 122a and uses it for calculating the occurrence probability.

  Since there may be a plurality of dependency structure candidates newly generated in step S1304, the calculation unit 103a calculates the occurrence probability for each candidate in step S1305.

  Next, the analysis unit 103 outputs the generated dependency structure candidates and the occurrence probabilities calculated for the candidates, and ends the dependency analysis processing (step S1306).

  As described above, in the present embodiment, the dependency analysis process is recursively called, whereby the dependency structure candidates are sequentially generated from the sentence end side, and the occurrence probability of the generated candidates can be sequentially calculated. In this embodiment, the occurrence probability of the dependency structure is sequentially calculated using the occurrence probability of the sentence at the end of the sentence stored in the occurrence probability table 121a prepared in advance.

  Thus, since the occurrence probability of the dependency structure is calculated using the occurrence probability of the phrase stored in the occurrence probability table 121a, the occurrence probability of the dependency structure calculated for each of the plurality of phrase sequences is calculated. It becomes possible to compare with each other. For this reason, it is possible to obtain a dependency structure that maximizes the occurrence probability by comparing the occurrence probabilities, and to select a phrase series corresponding to the obtained dependency structure as an optimum processing result.

  Next, details of the dependency analysis process will be described with reference to FIG. First, the analysis unit 103 acquires the designated start phrase I, number N of phrases, and phrase series W (step S1401).

Next, the analysis unit 103 determines whether or not the number of clauses N is 2 (step S1402). If the number of clauses is 2 (step S1402: YES), [d [1,2]] is output as a set PL of occurrence probabilities [P (w ₁ | w ₂ ) · P (w ₂ )] (step S1403), and the dependency analysis process is terminated. The output contents in step S1403 represent the output contents when the dependency analysis process is recursively executed and the dependency analysis process for the last two clauses is finally executed.

  When the number of phrases is not 2 (step S1402: NO), the analysis unit 103 further determines whether the number of phrases is 2 or more (step S1404). If it is not 2 or more (step S1404: NO), the dependency analysis process is terminated because the dependency analysis cannot be performed.

  When the number of clauses is 2 or more (step S1404: YES), the analysis unit 103 recursively designates the start clause as I + 1, the number of clauses as N-1, and the clause series as W except for the first clause. A dependency analysis process is executed (step S1405).

  Next, the analysis unit 103 acquires a set L2 of dependency structure candidates and a set P2 of occurrence probabilities of each candidate as analysis results (step S1406). Subsequently, the analysis unit 103 executes the following steps S1407 to S1417 in order to generate a dependency structure candidate when the immediately preceding phrase is added and to calculate the occurrence probability of each candidate.

  First, the analysis unit 103 initializes a set L3 for storing the candidates for the dependency structure to be generated and a set P3 for storing the occurrence probabilities of each candidate to an empty list (step S1407).

  Next, the analysis unit 103 acquires the dependency structure candidate S that is the first element of L2 and the occurrence probability P of S that is the first element of P2 (step S1408). Next, the analysis unit 103 sets I + 1 representing the position of the start phrase to the phrase position i (step S1409).

  In the following processing (steps S1410 to S1415), the immediately preceding phrase (phrase position I) that is the source phrase and the related phrase (sentence position) are moved while moving the related phrase position i toward the end of the sentence. It is determined whether the dependency relationship of i) intersects with each dependency relationship in S, and if not, a new dependency structure candidate including the dependency relationship is generated and the occurrence probability is calculated.

  First, the analysis unit 103 designates the phrase position i and the dependency structure candidate S, and executes an intersection determination process for determining the intersection of the dependency relationship (step S1410). Details of the intersection determination process will be described later.

  Next, the analysis unit 103 determines whether or not the dependency relationship to be newly added intersects with the dependency relationship in S based on the result of the intersection determination process (step S1411). If they do not intersect (step S1411: NO), the analysis unit 103 generates a dependency structure with the dependency relationship d (I, i) added to the head of the dependency structure candidate S and adds it to the set L3. (Step S1412).

Subsequently, the calculation unit 103a calculates P (w _I | w _i ) · P as the occurrence probability of the generated dependency structure and adds it to the set P3 (step S1413). P (w _I | w _i ) is a conditional probability corresponding to the dependency relationship between the phrase w _I and the phrase w _i, and can be acquired from the conditional probability table 122a.

  After adding the occurrence probability to the set P3 or when the dependency relationship intersects (step S1411: YES), the analysis unit 103 shifts the phrase position i to the end of the sentence by setting i = i + 1 ( Step S1414).

  Next, the analysis unit 103 determines whether i is larger than the number of phrases N (step S1415). If not larger (step S1415: NO), the intersection determination process is repeated for a new phrase position (step S1410). ).

  If i is larger than the number of clauses (step S1415: YES), since the intersection determination process with all the dependency relationships has been completed for the dependency structure candidate S, the analysis unit 103 determines that the set L2 and the set S and P are deleted from P2, respectively (step S1416).

  Next, the analysis unit 103 determines whether or not L2 is an empty list (step S1417). If the L2 is not an empty list (step S1417: NO), further obtains a candidate for the next dependency structure and processes it. Is repeated (step S1408).

  If L2 is an empty list, that is, if the processing is completed for all the dependency structure candidates (step S1417: YES), the analysis unit 103 adds the new dependency structure candidates added up to that point and The sets L3 and P3, which are sets of occurrence probabilities, are output (step S1418), and the dependency analysis process is terminated.

  Next, details of the intersection determination process in step S1410 will be described with reference to FIG. FIG. 15 is a flowchart showing details of the overall flow of the intersection determination process in the first embodiment.

  First, the analysis unit 103 acquires a specified phrase position i and a dependency structure candidate S (step S1501). Next, the analysis unit 103 acquires a dependency relationship d (a, b) included in the dependency structure S (step S1502).

  Subsequently, the analysis unit 103 determines whether the dependency destination is i and the dependency relationship d (a, b) based on the relationship between the phrase position i, the source phrase position a, and the destination phrase position b. Whether or not crosses. Specifically, the analysis unit 103 determines whether i is larger than a and i is smaller than b (step S1503). Unless i is larger than a and i is smaller than b (step S1503: NO), it is determined that the dependency relationship with the dependency destination i does not intersect with the dependency relationship d (a, b). it can.

  In this case, the analysis unit 103 determines whether or not all dependency relationships in S have been processed (step S1504). If not, the analysis unit 103 determines the next dependency relationship (step S1504: NO). Obtain and repeat the process (step S1502).

  If it is determined that all the dependency relationships have been processed (step S1504: YES), the analysis unit 103 determines that the dependency relationship with the dependency destination i does not intersect with all the dependency relationships in S. Determination is made (step S1505), and the intersection determination process is terminated.

  On the other hand, if i is greater than a and i is smaller than b in step S1503 (step S1503: YES), the dependency relationship with the dependency destination i intersects with the dependency relationship d (a, b). It is judged. Therefore, the analysis unit 103 determines that the dependency relationship intersects (step S1506), and ends the intersection determination process.

  As described above, in the language processing apparatus according to the first embodiment, a plurality of phrase series candidates are input as the processing result of the language processing, and a plurality of phrase series results are obtained with reference to the dependency analysis results obtained for each phrase series. The optimal candidate can be selected as the processing result from the candidates.

(Second Embodiment)
In the first embodiment, the occurrence probability of the dependency structure is calculated using the occurrence probability and the conditional probability obtained without considering the number of clauses in the sentence. On the other hand, the language processing apparatus according to the second embodiment uses the phrase occurrence probability and the conditional probability of the dependency relationship obtained for each number of clauses in the sentence, so that the dependency structure is highly accurate. The occurrence probability of is calculated.

  FIG. 16 is a block diagram illustrating a configuration of a language processing device 1600 according to the second embodiment. As illustrated in FIG. 16, the language processing device 1600 includes a first storage unit 1621, a second storage unit 1622, a third storage unit 123, an input reception unit 101, a control unit 1602, and an output unit 105. I have.

  In the second embodiment, the configurations or functions of the first storage unit 1621, the second storage unit 1622, and the control unit 1602 are different from those of the first embodiment. Other configurations and functions are the same as those in FIG. 1, which is a block diagram showing the configuration of the language processing apparatus 100 according to the first embodiment, and thus the same reference numerals are given and description thereof is omitted here.

  The 1st memory | storage part 1621 memorize | stores the occurrence probability table 1621a extended so that the occurrence probability calculated | required in advance for every number of clauses in a sentence may be stored. FIG. 17 is an explanatory diagram illustrating an example of a data structure of the occurrence probability table 1621a according to the second embodiment. As shown in FIG. 17, the occurrence probability table 1621a stores a phrase and the occurrence probability of the phrase obtained for each number of phrases in the sentence in association with each other.

In the figure, Ω _n represents a set of phrase series having the number of phrases n. If the observation space Ω is all possible phrase series, Ω can be expressed as the following equation (2).

Here, since Ω n having different _n do not have a common part with each other, the following equation (3) is established for the probability P (Ω _n ) of each subspace Ω _n .

  The second storage unit 1622 stores a conditional probability table 1622a extended so as to store conditional probabilities obtained in advance for each number of clauses in the sentence. FIG. 18 is an explanatory diagram illustrating an example of a data structure of the conditional probability table 1622a according to the second embodiment. As shown in FIG. 18, the conditional probability table 1622a stores the relational clauses, the relational clauses, and the conditional probabilities of the dependency relationships obtained for each number of clauses in the sentence in association with each other. ing.

  Similar to the control unit 102 of the first embodiment, the control unit 1602 controls the process of selecting the optimum phrase series from the accepted phrase series as a processing result. Different from the control unit 102. Since the configuration and function of the selection unit 104 are the same as those in FIG. 1 of the first embodiment, the same reference numerals are given, and description thereof is omitted here.

  The analysis unit 1603 is different from the analysis unit 103 of the first embodiment in that the analysis unit 1603 includes a calculation unit 1603a that calculates the occurrence probability of the dependency structure with reference to each of the expanded tables.

The occurrence probability of the dependency structure calculated by the calculation unit 1603a can be expressed by the following equation (4). Where P (w _i | w _j , Ω _n ) is the conditional probability of w _i for w _j in Ω _n , P (w _N, Ω _n ) is the probability of occurrence of w _N in Ω _n , P ( Ω _n ) represents the probability of occurrence of Ω _n . P (Ω _n ) refers to a value calculated in advance and stored in a storage unit (not shown).

  Next, candidate selection processing by the language processing device 1600 according to the second embodiment configured as described above will be described. The overall flow of the candidate selection process of the second embodiment is the same as FIG. 12 showing the overall flow of the candidate selection process of the first embodiment. However, the details of the dependency analysis processing in step S1203 in FIG. 12 are different from those in the first embodiment.

  Below, the dependency analysis process in 2nd Embodiment is demonstrated using FIG. FIG. 19 is a flowchart showing an overall flow of dependency analysis processing according to the second embodiment.

  First, the analysis unit 1603 determines whether or not the number of phrases in the specified phrase series is 2 (step S1901). When the number of clauses is 2 (step S1901: YES), the calculation unit 1603a generates and outputs a dependency structure for the two clauses at the end of the sentence, and calculates and outputs the occurrence probability of the generated dependency structure. (Step S1902). At this time, the calculation unit 1603a acquires, from the occurrence probability table 1621a, a value corresponding to the number of phrases in the input phrase series as the occurrence probability of the sentence at the end of the sentence.

  Here, the number of phrases in the input phrase series is a value different from the number of phrases determined in step S1901, and corresponds to the number of phrases in each phrase series in a state where the input is accepted in step S1201. . For example, this value may be specified separately from the phrase series when the dependency analysis process is executed, and may be configured so that it can be referred to in the dependency analysis process.

  The dependency analysis processing (recursion processing) and dependency structure generation processing from step S1903 to step S1904 are the same as steps S1303 to S1304 in the language processing apparatus 100 according to the first embodiment. Description is omitted.

  Next, the calculation unit 1603a calculates the occurrence probability of the generated dependency structure (step S1905). The calculation unit 1603a obtains an appropriate conditional probability from the conditional probability table 1622a in consideration of not only the dependency relationship with the added clause but also the number of clauses, and calculates the occurrence probability. Specifically, the calculation unit 1603a acquires the conditional probability corresponding to the dependency relationship with the added clause and the number of clauses in the input phrase series from the conditional probability table 1622a, and calculates the occurrence probability. Use.

  Finally, the analysis unit 1603 outputs the generated dependency structure candidates and the occurrence probabilities calculated for the candidates, and ends the dependency analysis processing (step S1906).

  Thus, in the language processing device according to the second embodiment, the occurrence probability of the dependency structure is calculated using the occurrence probability and the conditional probability obtained for each number of clauses in the sentence. The occurrence probability of the dependency structure can be calculated with higher accuracy. And it becomes possible to select an optimal process result with high precision by the occurrence probability calculated with high precision in this way.

  Next, the hardware configuration of the language processing apparatus according to the first or second embodiment will be described with reference to FIG. FIG. 20 is an explanatory diagram of a hardware configuration of the language processing apparatus according to the first or second embodiment.

  The language processing device according to the first or second embodiment communicates with a control device such as a CPU (Central Processing Unit) 51 and a storage device such as a ROM (Read Only Memory) 52 and a RAM 53 by connecting to a network. The communication I / F 54, an external storage device such as an HDD (Hard Disk Drive) and a CD (Compact Disc) drive device, a display device such as a display device, and an input device such as a keyboard and a mouse. A bus 61 is provided and has a hardware configuration using a normal computer.

  The candidate selection program executed by the language processing apparatus according to the first or second embodiment is a file in an installable format or an executable format, and is a CD-ROM (Compact Disk Read Only Memory), a flexible disk (FD). ), A CD-R (Compact Disk Recordable), a DVD (Digital Versatile Disk), and the like.

  Further, the candidate selection program executed by the language processing apparatus according to the first or second embodiment is stored on a computer connected to a network such as the Internet and is provided by being downloaded via the network. It may be configured. Further, the candidate selection program executed by the language processing apparatus according to the first or second embodiment may be provided or distributed via a network such as the Internet.

  The candidate selection program according to the first or second embodiment may be provided by being incorporated in advance in a ROM or the like.

  The candidate selection program executed by the language processing apparatus according to the first or second embodiment has a module configuration including the above-described units (input reception unit, control unit, output unit), and actual hardware. As described above, the CPU 51 (processor) reads out and executes the candidate selection program from the storage medium, whereby the above-described units are loaded onto the main storage device, and the above-described units are generated on the main storage device.

  As described above, the apparatus, method, and program according to the present invention are suitable for an apparatus, method, and program that select an optimum processing result from a plurality of processing result candidates by processing such as speech recognition, character recognition, and morphological analysis. Yes.

It is a block diagram which shows the structure of the language processing apparatus concerning 1st Embodiment. It is explanatory drawing which shows an example of the data structure of the occurrence probability table of 1st Embodiment. It is explanatory drawing which shows an example of the data structure of the conditional probability table of 1st Embodiment. It is explanatory drawing which shows an example of the data structure of a dictionary table. It is explanatory drawing which shows an example of the phrase series input. It is explanatory drawing which shows an example of the expression format of a dependency structure. It is explanatory drawing which shows an example of the occurrence probability of a dependency structure. It is explanatory drawing which showed the process of calculating | requiring the possible dependency structure from the clause series of 4 clauses. It is a figure which shows an example of the process of calculating | requiring the occurrence probability of a dependency structure. It is explanatory drawing which shows the process of calculating the occurrence probability of a dependency structure. It is a schematic diagram for demonstrating the case where the calculation process of occurrence probability is generalized. It is a flowchart which shows the whole flow of the candidate selection process in 1st Embodiment. It is a flowchart which shows the outline | summary of the whole flow of the dependency analysis process in 1st Embodiment. It is a flowchart which shows the detail of the whole flow of the dependency analysis process in 1st Embodiment. It is a flowchart which shows the detail of the whole flow of the intersection determination process in 1st Embodiment. It is a block diagram which shows the structure of the language processing apparatus concerning 2nd Embodiment. It is explanatory drawing which shows an example of the data structure of the occurrence probability table of 2nd Embodiment. It is explanatory drawing which shows an example of the data structure of the conditional probability table of 2nd Embodiment. It is a flowchart which shows the whole flow of the dependency analysis process in 2nd Embodiment. It is explanatory drawing which shows the hardware constitutions of the language processing apparatus concerning 1st or 2nd embodiment.

Explanation of symbols

51 CPU
52 ROM
53 RAM
54 Communication I / F
61 Bus 100 Language processing apparatus 101 Input reception unit 102 Control unit 103 Analysis unit 103a Calculation unit 104 Selection unit 105 Output unit 121 First storage unit 121a Occurrence probability table 122 Second storage unit 122a Conditional probability table 123 Third storage unit 123a Dictionary table 201, 202, 203, 204, 205 Phrase series 206, 207, 208, 209, 210 Dependency structure 211, 212, 213, 214, 215 List structure 306, 307, 308, 309, 310 Occurrence probability 401, 402 , 403 Dependency structure 1101 Dependency structure 1102 Clause 1600 Language processing device 1602 Control unit 1603 Analysis unit 1603a Calculation unit 1621 First storage unit 1621a Occurrence probability table 1622 Second storage unit 1622a Conditional probability table Le

Claims (10)

A language processing apparatus that selects the processing result from processing result candidates for a sentence unit,
A first storage unit that stores the structural unit and the occurrence probability of the structural unit in association with each other;
A dependency relationship represented by the constituent unit that is the destination and the constituent unit that is the source of the relationship, and a conditional probability that the constituent unit that is the source of the relation appears for the constituent unit that is the destination A second storage unit for storing
An input receiving unit that receives input of candidates for the processing result;
For each of the accepted processing result candidates, an analysis unit that analyzes a dependency structure that represents a combination of the dependency relationships between the structural units;
For each of the analyzed dependency structure candidates, the occurrence probability corresponding to the constituent unit at the end of a sentence is acquired from the first storage unit, and also corresponds to each of the dependency relationships included in the dependency structure. A calculation unit that acquires the conditional probability from the second storage unit, and calculates a probability of occurrence of the dependency structure candidate that is a product of the acquired product of all the conditional probabilities and the acquired probability of occurrence; ,
A selection unit that obtains the candidate of the dependency structure that has the maximum occurrence probability calculated, and selects the candidate of the processing result corresponding to the obtained candidate of the dependency structure as the processing result;
A language processing apparatus comprising: The calculation unit, for each non-end-end unit that is the constituent unit other than the constituent unit at the end of the sentence, in the order from the non-end-end unit at the end of the sentence toward the non-end-end unit at the beginning of the sentence, is behind the non-end-end unit. Acquiring the conditional probability corresponding to the dependency relationship with respect to the structural unit from the second storage unit, and sequentially multiplying the acquired conditional probability with respect to the occurrence probability acquired from the first storage unit. To calculate the occurrence probability of the dependency structure candidate,
The language processing apparatus according to claim 1. The calculating unit, the dependency among the candidate structures parsed from the k-th of the process result candidate, occurrence probability PL (k, l _k) of the candidate l _k-th of the dependency structure of (1) Calculated by the formula
The selection unit obtains an integer k and an integer l _k that maximize the occurrence probability PL (k, l _k ) in the equation (1), and sets the kth candidate for the processing result corresponding to the obtained integer k as the process. Select as a result,
The language processing apparatus according to claim 1.

The first storage unit associates the structural unit, the number of units representing the number of the structural units in a sentence, and the occurrence probability of the structural unit in a sentence in which the number of structural units is the unit number. Remember,
The second storage unit stores the dependency relationship, the number of units, and the conditional probability in a sentence in which the number of structural units is the number of units, in association with each other.
The calculation unit further determines, for each of the analyzed dependency structure candidates, the number of the structural units included in the processing result candidate corresponding to the dependency structure candidate, and the structural unit at the end of the sentence. The occurrence probability corresponding to the determined number is acquired from the first storage unit, and the conditional probability corresponding to the dependency relationship included in the dependency structure and the determined number is the second probability. Obtaining from the storage unit, calculating the occurrence probability of the dependency structure candidate that is the product of the acquired occurrence probability and the acquired conditional probability;
The language processing apparatus according to claim 1. The calculating unit, the dependency among the candidate structures parsed from the k-th of the process result candidate, occurrence probability PL (k, l _k) of the candidate l _k-th of the dependency structure of (2) Calculated by the formula
The selection unit obtains an integer k and an integer l _k maximizing the occurrence probability PL (k, l _k ) of the expression (2), and sets the kth processing result candidate corresponding to the obtained integer k as the process. Select as a result,
The language processing apparatus according to claim 4.

The input receiving unit receives an input of a recognition result candidate of a speech recognition process for recognizing a voice and dividing it into the constituent units as a candidate for the processing result;
The language processing apparatus according to claim 1. The input accepting unit accepts input of a recognition result candidate of a character recognition process for recognizing a character and dividing it into the constituent units as a candidate for the processing result;
The language processing apparatus according to claim 1. The input receiving unit receives input of analysis result candidates of a morpheme analysis process in which a sentence is morphologically analyzed and divided into morphemes as the constituent units as the processing result candidates,
The language processing apparatus according to claim 1. A candidate selection method in a language processing device for selecting a processing result from processing result candidates for a sentence unit,
The language processing device includes:
A first storage unit that stores the structural unit and the occurrence probability of the structural unit in association with each other;
A dependency relationship represented by the constituent unit that is the destination and the constituent unit that is the source of the relationship, and a conditional probability that the constituent unit that is the source of the relation appears for the constituent unit that is the destination A second storage unit that stores the information in association with each other,
An input receiving step of receiving input of the candidate processing result by an input receiving unit;
An analysis step of analyzing a dependency structure representing a combination of the dependency relationships between the structural units for each of the processing result candidates received by the analysis unit;
For each of the dependency structure candidates analyzed by the calculation unit, the occurrence probability corresponding to the constituent unit at the end of a sentence is acquired from the first storage unit, and the dependency relationship included in the dependency structure is obtained. The conditional probabilities corresponding to each are acquired from the second storage unit, and the occurrence probability of the candidate of the dependency structure, which is the product of the acquired product of all the conditional probabilities and the acquired occurrence probability, is calculated. A calculating step to
A selection step of obtaining a candidate for the dependency structure that maximizes the calculated occurrence probability by a selection unit, and selecting the processing result candidate corresponding to the obtained candidate for the dependency structure as the processing result;
A candidate selection method characterized by comprising: A candidate selection program in a language processing apparatus for selecting the processing result from processing result candidates for a sentence unit,
The language processing device includes:
A first storage unit that stores the structural unit and the occurrence probability of the structural unit in association with each other;
A dependency relationship represented by the constituent unit that is the destination and the constituent unit that is the source of the relationship, and a conditional probability that the constituent unit that is the source of the relation appears for the constituent unit that is the destination A second storage unit that stores the information in association with each other,
An input acceptance procedure for accepting input of candidate processing results;
An analysis procedure for analyzing a dependency structure representing a combination of the dependency relationships between the structural units for each of the accepted processing result candidates;
For each of the analyzed dependency structure candidates, the occurrence probability corresponding to the constituent unit at the end of a sentence is acquired from the first storage unit, and also corresponds to each of the dependency relationships included in the dependency structure. A calculation procedure for acquiring the conditional probability from the second storage unit, and calculating the occurrence probability of the candidate of the dependency structure, which is a product of the acquired product of all the conditional probabilities and the acquired occurrence probability; ,
A selection procedure for obtaining a candidate for the dependency structure that maximizes the calculated occurrence probability, and selecting the processing result candidate corresponding to the obtained candidate for the dependency structure as the processing result;
A candidate selection program that causes a computer to execute.

Images