JP2016009428A - Morphological analysis turning apparatus, voice synthesis system, and morphological analysis tuning method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a technique for reducing a burden of a user applied to correction work of a morpheme by optimizing a morphological dictionary while retaining the morpheme to a minimum unit.SOLUTION: A morphological analysis tuning apparatus performs morphological analysis of an input text, and compares an optimum morphological analysis result as a morphological analysis result in which a cost of the morphological analysis result takes a minimum value, with a correct morphological analysis result that a user inputs related to the input text as correct data. As the result of the comparison processing, when there is a different part between the correct morphological analysis result and the optimum morphological analysis result, the morphological analysis tuning apparatus creates a cost adjustment rule such that the cost of the correct morphological analysis result becomes smaller than that of the optimum morphological analysis result.

Description

  The present invention relates to a morpheme analysis tuning apparatus, a speech synthesis system, and a morpheme analysis tuning method, for example, for correcting (tuning) a morpheme misanalysis that occurs during text analysis when synthesizing input text speech. Regarding technology.

  In recent years, there are increasing opportunities to hear voices that have undergone voice synthesis processing in many scenes of daily life. With the introduction of the waveform connection method, the synthesized sound has been improved considerably in terms of sound quality. Accordingly, services that automatically provide information using voice, such as an in-vehicle navigation device, an automatic broadcasting device in a public facility, a mail reading device, and an automatic interpretation system, are widely used.

  Text-to-speech synthesis technology consists of two steps: text analysis and speech generation. In text analysis, the input natural language is decomposed into the smallest meaningful units (morphemes), and then reading, accent determination, prosodic hierarchy analysis, etc. are performed (morphological analysis). This morphological analysis is the process that most affects the sound quality because it affects the operation of all subsequent text analysis modules. About morphological analysis, it is indicated by patent documents 1, for example.

  Currently, there are business tools developed by Yahoo! (registered trademark), google (registered trademark), etc. as morpheme analysis tools often used in Japanese morpheme analysis, including JUMAN, ChaSen and MeCab. Open source tools to do this. The latter is more widely used in linguistic research because the former is charged or has severe restrictions on its use.

Morphological analysis methods are roughly classified into rule-based morpheme analysis methods and statistical-based morpheme analysis methods. The statistical-based morphological analysis method is a method that statistically performs parameter learning based on data, and is frequently used recently. On the other hand, the rule-based morphological analysis method performs analysis based on manual rules (heuristics), and was prevalent before the 1980s. Specifically, the rule-based morphological analysis method enumerates the possibility that words or parts of speech are concatenated and appear in a sentence as rules (jointability rule table), and obtains only possible part of speech as a solution. Is the method. The connectability rule table is represented as a two-dimensional matrix with parts of speech (or words). By using this connection possibility table, it is possible to limit the connection of parts of speech that are grammatically unnatural, so that some ambiguity can be resolved, but there is still the possibility that multiple solutions exist. There are challenges. In order to solve this problem, in the rule-based morphological analysis method, the strength of the possibility that words or parts of speech are connected and appear in a sentence is quantified as a cost, and the part of speech sequence with the lowest cost is obtained as a solution. I have to. As shown in Table 1, the connection cost table is represented by a two-dimensional matrix having parts of speech (or words) in rows and columns and cost as an element. Therefore, superiority or inferiority among a plurality of solutions generated during morphological analysis can be distinguished by cost. Such costs are generally determined manually. In addition to the cost, there is also the expression of the score, but the difference between the two is the concept that the strength of the connection increases as the cost decreases, whereas the connection increases as the score increases. It is a concept that increases the strength of. Both are concepts that can be used complementarily. In morphological analysis, it is common to minimize the cost rather than maximizing the score so that the number of morphemes does not increase.
However, when the cost is determined manually, there is a problem that the burden on the user is large and it is difficult to maintain consistency in the cost determination method.

  In recent years, it has become possible to learn cost (probability value) from large-scale text data (tagged corpus) against the backdrop of improvements in the processing power of personal computers. For this reason, a technique (corpus-based morphological analysis technique) has been proposed in which the ease of appearance of words and the ease of connection of parts of speech (words) are obtained as probability values from an analyzed corpus assigned parts of speech. This is a technique for modeling a sentence in a language on the assumption that the appearance probability of a word constituting the sentence depends only on the immediately preceding n-1 word (tri-gram model when n = 3, n = 2). Is called a bi-gram model). Typical modeling methods include “ChaSen” that uses a mixture of bi-gram and tri-gram, “JUMAN” that realizes an automatic learning function using HMM, and Conditional Random Fields for parameter estimation. Examples include “MeCab” using (CRF). As a result, the burden on the user can be reduced.

JP-A-56-17467

  However, such a corpus-based morphological analysis method generally requires a large-scale text corpus, and it is very difficult to maintain this large-scale text corpus. If there is a mistake in text analysis, the text corpus will be tuned, but the statistical model cannot be corrected manually. For this reason, the correct data is added to the text corpus and re-learning is performed. However, even if this method is used, it cannot be surely guaranteed that an analysis error can be corrected.

  By the way, in recent years, the sound quality of synthesized speech has been remarkably improved and applied in various scenes. For example, it is a scene of car navigation route guidance or railway premises broadcasting. In these situations, there is a customer's desire to reduce reading errors as much as possible. In order to meet this demand and improve the reading accuracy, tuning work such as daily dictionary maintenance and correction of reading mistakes is indispensable.

  However, since the statistical-based morphological analysis method uses a statistical probability model from a large-scale database (learning data), manual tuning cannot be performed. For documents with morphological analysis errors, correct labels are assigned, and then added to the training data, and the model is re-learned. I cannot guarantee it.

  On the other hand, with the rule-based morphological analysis method, it is very difficult to tune the cost table using only part-of-speech information, and it is not possible to deal with individual special cases. For this reason, tuning for morphological analysis has been performed by registering longer morphemes.

  However, in such a tuning method, the number of morphemes (individual morphemes having no versatility) becomes enormous, causing system enlargement and performance degradation. When word registration over a plurality of parts of speech is necessary or there are a plurality of accents, the registration becomes difficult, the burden on the user becomes excessive, and the user's subjectivity easily enters. In addition, it is desirable that morphemes are composed of the smallest units that are meaningful as much as possible. However, if special words are registered as morphemes each time they appear, the matching between the morphological analysis methods used so far and the processing operations deteriorates. there is a possibility.

  The present invention has been made in view of such a situation, and provides a technique for optimizing a morpheme dictionary and reducing a burden on a user's morpheme correction work while keeping the morpheme in a minimum unit. .

  In order to solve the above problems, the present invention provides a morphological analysis tuning apparatus that generates a cost adjustment rule used for morphological analysis based on a morphological analysis result of a text sentence. The morpheme analysis tuning apparatus includes at least a correct morpheme analysis result database that stores correct morpheme analysis results that have been determined by the user to be correct, a correct morpheme analysis result, and a morpheme read from the morpheme dictionary. And a processor that creates a cost adjustment rule while optimizing it. Further, the processor compares the optimal morpheme analysis result, which is the morpheme analysis result having the lowest cost among the morpheme analysis results of the text sentence, with the correct morpheme analysis result input by the user as the correct data regarding the text sentence. If the comparison process and the result of the comparison process show that the correct morpheme analysis result and the optimal morpheme analysis result have different parts, a cost adjustment rule is generated so that the correct morpheme analysis result is less expensive than the optimal morpheme analysis result. The adjustment process is executed.

  Further features related to the present invention will become apparent from the description of the present specification and the accompanying drawings. The embodiments of the present invention can be achieved and realized by elements and combinations of various elements and the following detailed description and appended claims.

  It should be understood that the description herein is merely exemplary and is not intended to limit the scope of the claims or the application of the invention in any way.

  According to the present invention, the user automatically gives the correct morpheme analysis result of the entire sentence to the system, and the system automatically optimizes the correction target morpheme sequence, the optimal correction range (the degree of versatility), and the optimal score. Information adjustment values can be determined. For this reason, the burden on the user is small, and consistency of correction can be maintained for different users.

It is a figure which shows schematic structure of the speech synthesis system by embodiment of this invention. It is a figure for demonstrating the structure and procedure outline | summary of a general text-to-speech synthesis process. It is a figure which shows the content of the natural language process. It is a figure which shows the hardware constitutions at the time of comprising a text analysis device and a morphological analysis tuning device by one computer system. It is a figure which shows the system configuration example which provided the user interface in the device of the terminal side, and provided the hardware which implement | achieves all or one part of a text analysis function and a morphological analysis tuning function in the server. It is a figure which shows the structural example at the time of comprising a text analysis apparatus and a morphological analysis tuning apparatus with a separate computer. It is a block diagram about the functional structure of text analysis and morphological analysis tuning. It is a figure which shows the detailed structure (FIG. 8A) and the example of presentation (FIG. 8B) of the analysis result presentation part 703. FIG. It is a figure which shows the example of a presentation screen of a correct morphological analysis result. It is a figure for demonstrating the process performed by the target morpheme automatic extraction part. It is a figure which shows the example of the part from which the correct morphological analysis result and the optimal morphological analysis result extracted by the morphological analysis result comparison module. It is a figure for demonstrating the process performed in the correction method determination part. It is a figure for demonstrating the process performed in a word registration part. It is a figure for demonstrating the process performed by the cost adjustment rule preparation part.

  Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the accompanying drawings, functionally identical elements may be denoted by the same numbers. The attached drawings show specific embodiments and implementation examples based on the principle of the present invention, but these are for understanding the present invention and are not intended to limit the present invention. Not used.

  This embodiment has been described in sufficient detail for those skilled in the art to practice the present invention, but other implementations and configurations are possible without departing from the scope and spirit of the technical idea of the present invention. It is necessary to understand that the configuration and structure can be changed and various elements can be replaced. Therefore, the following description should not be interpreted as being limited to this.

  Furthermore, as will be described later, the embodiment of the present invention may be implemented by software running on a general-purpose computer, or may be implemented by dedicated hardware or a combination of software and hardware.

  Hereinafter, each process in the embodiment of the present invention will be described with each process “unit” (for example, the cost adjustment rule creating unit 709) as the subject (operation subject), but the process by each process unit is executed by the processor. Therefore, the description may be made with the processor as the subject.

<Speech synthesis system>
FIG. 1 is a diagram showing a schematic configuration of a speech synthesis system 100 according to an embodiment of the present invention. The speech synthesis system 100 includes a text analysis device 101, a morphological analysis tuning device 102, and a speech synthesis device 103.

  The text analysis apparatus 101 receives text data (including keyboard input, scanner input, etc.) input by the user, and decomposes and analyzes the input text by referring to the morpheme dictionary and executing morpheme analysis.

  The morphological analysis tuning apparatus 102 acquires a text analysis result and accepts a determination input by the user as to whether or not the text analysis result is correct. Then, the morphological analysis tuning apparatus 102 determines whether or not to correct the cost (score) value of the morphological analysis result based on the determination input by the user, and determines a rule for the correction. When a new rule is determined, the text analysis apparatus 101 performs morphological analysis again according to the new rule, and outputs a correct analysis result to the speech synthesizer 103.

  The speech synthesizer 103 acquires a text analysis result from the text analysis device 101, predicts a prosody using a phonetic symbol string included in the text analysis result, and generates speech.

<Structure and procedure of text-to-speech synthesis>
FIG. 2 is a diagram for explaining an outline of a configuration and procedure of a general text-to-speech synthesis process. Text-to-speech synthesis (also referred to as TTS) technology is a series of processes for synthesizing speech from text, and is roughly composed of “natural language processing” and “speech generation”. FIG. 3 is a diagram showing details of natural language processing.

In natural language processing, as shown in FIG. 3, the input natural language is decomposed into the smallest meaningful units (morphemes), readings are added, accents are determined, and prosodic hierarchy analysis is performed. Output phonetic symbol string.
In the speech generation process, the prosody is predicted using the input phonetic symbol string, and speech is generated.

  An object of the embodiment of the present invention is to improve reading errors in synthesized speech processing by tuning a morphological analysis part in text speech synthesis processing. Therefore, as described above, the morphological analysis tuning apparatus 102 is included in the speech synthesis system 100 in addition to the text analysis apparatus 101 and the speech synthesis apparatus 103.

  First, the morphological analysis tuning apparatus 102 presents a text analysis result (or synthesized speech) to a tuning operator (user). The user verifies the presented text analysis result and determines whether there is an analysis error (in the case of speech, reading error). And when there is a mistake, the said morphological analysis tuning apparatus receives the correct morphological analysis result (what a user thinks is a correct answer) which a user inputs. Further, the morpheme analysis tuning apparatus extracts an optimal target morpheme series in comparison with a morpheme analysis result (optimal morpheme analysis result having a minimum cost) analyzed by the text analysis apparatus. For the target morpheme sequence, the apparatus automatically separates the case where word registration is necessary and the case where cost adjustment is necessary, and performs word registration and cost adjustment rule creation, respectively. When creating a cost adjustment rule, the device performs cost adjustment only on the extracted morpheme sequence so that the cost of the correct morphological analysis result specified by the user is minimized, and the cost adjustment rule is converted into morphological analysis data. Add to In creating the cost adjustment rule, it is desirable to set an optimal cost adjustment value and rule application range using a correct morphological analysis database prepared in advance.

<Hardware configuration example>
Hereinafter, the hardware configuration of the text analysis device 101 and the morphological analysis tuning device 102 included in the speech synthesis system 100 according to the embodiment of the present invention will be described.

(1) Configuration Example when Constructed with One Computer FIG. 4 is a diagram showing a hardware configuration when the text analysis device 101 and the morphological analysis tuning device 102 are configured with one computer system. The text analysis / morpheme analysis tuning apparatus 400 includes a CPU 401, a memory 402 as a main storage device, a storage device 403, a text input I / F (interface) 404, a voice input I / F 405, and a voice connected to a speaker. An output I / F 406 and a text presentation I / F 407 are included, and these components are connected to each other by a bus 408. The text analysis / morpheme analysis tuning apparatus 400 is incorporated as a text analysis / morpheme analysis tuning unit in a device such as a car navigation apparatus, a mobile phone, or a personal computer. Therefore, each hardware shown in FIG. 4 may be realized by using a device configuration in which the text analysis / morpheme analysis tuning apparatus is incorporated, or a device in which the text analysis / morpheme analysis tuning apparatus is incorporated. It may be provided separately.

(2) Configuration example when configured with terminals and servers Although all functions may be realized by only one computer device as shown in FIG. 4, a user interface (text input I / F as shown in FIG. 5). , Voice output I / F, text presentation I / F) are provided in the device on the terminal side, and the server is provided with hardware for realizing all or part of the text analysis function and the morphological analysis tuning function. It is also conceivable that they are connected to each other by / F.

(3) Configuration example when the text analysis device and the morphological analysis tuning device are configured separately FIG. 6 is a diagram illustrating a configuration example when the text analysis device 101 and the morphological analysis tuning device 102 are configured by separate computers. . In order to perform text analysis using morpheme analysis data (morpheme dictionary, cost adjustment rule, etc.) created by the morpheme analysis tuning device 102, text from the hardware (morpheme analysis tuning device 102) that implements the function of morpheme analysis It must be accessible to the storage device of the analysis device 101 that stores morphological analysis data. Therefore, the text analysis device 101 and the morphological analysis tuning device 102 are connected via the network 600. It should be noted that the connection may be made directly by a bus or the like without going through a network, or data may be exchanged between both devices via a recording medium (DVD, CD, USB memory, etc.). .

  As shown in FIG. 6, the text analysis apparatus 101 includes a CPU 1011, a memory 1012, a storage device 1013, a voice input I / F 1014, a voice output I / F 1015, and a communication I / F 1016. These are connected to each other via a bus 1017. The morphological analysis tuning apparatus 102 includes a CPU 1021, a memory 1022, a storage device 1023, a voice input I / F 1024, a voice output I / F 1025, a text presentation I / F 1027, and a communication I / F 1026. These are connected to each other via a bus 1028.

  As described above, since the text analysis device 101 and the morphological analysis tuning device 102 can exchange data with each other, the text analysis device 101 directly stores the morphological analysis tuning stored in the storage device 1013. Cost adjustment rules created by the device 102 can be accessed.

  The CPUs 1011 and 1021 govern overall control of the user voice DB creation device. The memories 1012 and 1022 are used as a work area for the CPU. The storage devices 1013 and 1023 are nonvolatile storage media. Specifically, for example, an HDD (hard disk), an FD (flexible disk), a flash memory, or the like can be used. In the storage devices 1013 and 1023, for example, various programs such as a sound quality evaluation program and a segment replacement program described later, and various data such as an existing speech database for speech synthesis are recorded. The voice input I / Fs 1014 and 1024 are interfaces for connecting a voice input device (not shown) such as a macrophone, and receive voice input from the voice input device. Audio output I / Fs 1015 and 1025 are interfaces for connecting an audio output device (not shown) such as a speaker. The communication I / Fs 1016 and 1026 are hardware for exchanging data (communication) between apparatuses. For example, a wired LAN card, a wireless LAN card, a modem, etc. can be considered. The text presentation I / F 1027 is a device that can display text, such as a personal computer monitor or a mobile screen. When the text is presented to the user by voice, the text presentation I / F 1027 can be substituted (combined) with the voice output I / F.

<Functional configuration of text analysis and morphological analysis tuning>
FIG. 7 is a block diagram showing the functional configuration of text analysis and morphological analysis tuning. As shown in FIG. 7, the functions for realizing the text analysis device include a text input unit 701, a morpheme analysis unit 702, an analysis result presentation unit 703, and a morpheme dictionary 704. The morphological analysis tuning function includes an analysis result determination unit 705, a correct morpheme analysis result input unit 706, a target morpheme automatic extraction unit 707, a correction method determination unit 708, a cost adjustment rule creation unit 709, and a correct morpheme analysis result. A database 710 and a cost adjustment rule group 711 are included. Details of each will be described below.

(I) Text Input Unit The text input unit 701 has a function of accepting text input by the user. Here, manual input with a keyboard, automatic input from a text database prepared in advance, automatic input of text from the web, and the like can be considered. Alternatively, text data may be obtained by taking a sentence with a scanner.

(Ii) Morphological Analysis Unit The morphological analysis unit 702 performs morphological analysis on the text received by the text input unit 701. Here, the morpheme is a morpheme that takes into account not only the notation but also the part of speech, reading, and accent. For example, for the same notation “Tokyo”, there are cases where the part of speech = name of place and the case of part of speech = person name, which are defined as different morphemes. The morpheme dictionary 704 needs to include at least three pieces of information of “morpheme notation”, “morpheme part of speech”, and “morpheme reading (with accent)” as constituent items. Since the embodiment of the present invention specializes in speech synthesis, information on “morpheme reading (with accent)” is required. If it is only text analysis, this information is not necessary.

(Iii) Analysis Result Presentation Unit The analysis result presentation unit 703 has a function of presenting the morpheme analysis result analyzed by the morpheme analysis unit 702 to the user. As the presentation method, it is possible to present only text, present only voice, or present simultaneously with text and voice. However, if the purpose is to improve reading errors in speech synthesis, even if the morphological analysis result is incorrect, if the synthesized speech has no reading errors (including accent errors and prosodic boundary errors), tune it. Since there is no need, simultaneous presentation of voice and text is desirable.

  FIG. 8 is a diagram illustrating a detailed configuration (FIG. 8A) and a presentation example (FIG. 8B) of the analysis result presentation unit 703. FIG. 8 shows an example in which the input text is “University of Tokyo”. The text generation module 7031 analyzes the input morpheme analysis result. Then, the text display device 7032 presents only the generated text (“morpheme 1: notation = Tokyo, part of speech = place name, reading = Tokyo; morpheme 2: notation = university, part of speech = general noun, reading = dialog”) to the user. To do. On the other hand, the speech synthesis non-joule 7033 receives the analyzed morpheme sequence and synthesizes speech. The speech synthesis / playback device 7034 outputs the synthesized speech generated by the speech synthesis module 7033.

  FIG. 8B is a diagram illustrating an example in which both text display and synthesized speech are presented to the user. In this way, both text and synthesized speech may be presented, or only one of them may be presented. However, when the synthesized speech is presented, it is considered that the user does not need special skills or expertise, and the correctness of the morphological analysis result can be determined. In addition, when it is assumed that morphological analysis is performed for speech synthesis, it is desirable to present synthesized speech as a presentation method that is closer to the actual application situation.

(Iv) Analysis Result Determination Unit The analysis result determination unit 705 has a function of receiving a determination as to whether or not the result of the presented morpheme analysis result is correct by the user. The analysis result determination unit 705 includes an interface such as a button (for example, “OK” and “NG” buttons). The analysis result determination unit 705 inputs a determination as to whether the analysis result is acceptable using the interface, and accepts the input. Here, when the user determines that the “analysis result is correct” (when the “OK” button is pressed), the entire process of text analysis tuning ends. Then, the text input unit 701 enters a standby state until the next text is input. At the same time, it is desirable to accumulate this analysis result (correct morpheme analysis result) in the correct morpheme analysis result database 710 in order to determine a cost adjustment value by the cost adjustment rule creation unit 709 described later. On the other hand, when the user determines that “the analysis result is incorrect” (when the “NG” button is pressed), the process proceeds to the next process (correct morphological analysis result input unit 706).

  For example, the analysis result of the text of “Tokyo Graduate School Director” “morpheme 1: notation = Tokyo, part of speech = place name, reading = Tokyo; morpheme 2: notation = university, part of speech = general noun, reading = diak; morpheme 3: notation = director If the part of speech = general noun, reading = incho ”is presented, the user determines that it is correct, clicks“ OK ”, ends the tuning process, and stores the analysis result in the correct morphological analysis corpus. (Here, the “correct morphological analysis result” is what the user thinks is correct and is not necessarily the correct result. For example, in the case of “Tokyo Graduate School Director”, “morpheme 1: notation” = Tokyo, part of speech = place name, reading = Tokyo; morpheme 2: notation = graduate school, part of speech = general noun, reading = digaquin; morpheme 3: notation = long, part of speech = suffix, reading = butterfly Maybe)

  On the other hand, the analysis result of the text of "Tokyo Graduate School Director" "morpheme 1: notation = east, part of speech = general noun, reading = hibashi; morpheme 2: notation = Kyoto University, part of speech = proper noun, reading = Kyoto; morpheme 3: notation = Gakuin, part of speech = general noun, reading = cuckin; morpheme 4: notation = long, part of speech = suffix, reading = butterfly ”is presented, the user determines that the morphological analysis error and clicks“ NG ” Then proceed to the next process.

(V) Correct morpheme analysis result input unit The correct morpheme analysis result input unit 706 operates only when the user determines that the morpheme analysis result determined to be optimal by the morpheme analysis unit 702 is “incorrect”. The input of the correct morphological analysis result is obtained.

  The correct morpheme result input unit 706 has an interface. For example, the user may directly input the correct morphological analysis result using a keyboard. However, considering the burden on the user, it is better to present the user with a list of all morpheme combinations that can compose the input document, starting from the lowest cost (in descending score) desirable. In this case, it is also necessary to prepare an option “There is no correct answer in the list”. Here, processing in the target morpheme automatic extraction unit 707 depends on whether or not all morphemes constituting the correct morpheme analysis result exist in the morpheme dictionary (or whether or not “No correct answer exists in the list” is selected). Will be different. When one or more of the morphemes constituting the correct morphological analysis result does not exist in the morpheme dictionary 704 (or “No correct answer exists in the list” is selected), the target morpheme automatic extraction unit 707 inputs the user Get the correct morpheme. On the other hand, when all the morphemes constituting the correct morpheme analysis result exist in the morpheme dictionary (or an option other than “No correct answer exists in the list” is selected), the target morpheme automatic extraction unit 707 The correct path selected by the user from among the above is acquired.

  FIG. 9 is a diagram illustrating an example of a screen for presenting a correct morphological analysis result. In the example shown in FIGS. 9A and B, when “He is a Tokyo Graduate School Director” is input as the text, the morpheme analysis unit 702 calls all paths (here, morpheme combinations that can constitute the text as paths). ), And the cost value of each path (the cost value is the sum of the part-of-speech cost of all morphemes and the part-of-speech connection cost of all neighboring morphemes from the pre-defined part-of-speech cost list and part-of-speech connection cost table) Is the total sum). Then, the analysis result presentation unit 703 presents each path to the user. When presenting only text, it is as shown in FIG. 9A, and when presenting simultaneously with text and voice, it is as shown in FIG. 9B. The user selects the correct answer path from the presented 1 to N paths using the prepared user interface. However, if the correct path does not exist, “No correct path exists in the list” is selected, and the correct morphological analysis result is input using the keyboard or the like included in the correct morphological analysis result input unit 706.

(Vi) Target Morphological Automatic Extraction Unit FIG. 10 is a diagram for explaining processing executed by the target morpheme automatic extraction unit. The target morpheme automatic extraction unit 707 has a morpheme analysis result comparison module 7071, and the optimal morpheme analysis result obtained by the morpheme analysis unit 702 (the lowest cost among all morpheme combinations that can constitute text ( The one with the highest score)) is compared with the correct morphological analysis result obtained by the correct morphological analysis result input unit 706 (what the user thinks is the correct answer). Then, the morpheme analysis result comparison module 7071 extracts both different parts (target morpheme string) and outputs them to the correction method determination unit 708. When there are a plurality of different portions, a plurality of target morpheme strings are extracted. On the other hand, if there are no different parts (Yes in the comparison step), the entire process is terminated.

  For example, as shown in FIG. 11, the optimal morpheme analysis result is “He / Higashi / Kyoto University / Gakuin / Director /” (hereinafter, in order to simplify the description, even if it is simply expressed as morpheme, It refers to a morpheme that contains information such as notation, part of speech, reading, etc. In other words, even if “Tokyo / University” is simply described, “morpheme 1: notation = Tokyo, part of speech = place name, reading = Tokyo; Notation = university, part of speech = general noun, reading = diagram ”, and the correct morphological analysis result is“ He / Ha / Tokyo / University / Director / ”is the morphological analysis result comparison module 7071 A different portion 1101 is extracted, and the correct morphological analysis result (“Tokyo / University / Director”) is set as the correct target morpheme. On the other hand, the optimum morpheme analysis results (“East / Kyoto University / Gakuin / Head”) of different parts are set as the optimum target morpheme. That is, the target morpheme is composed of a correct target morpheme and an optimal target morpheme.

(Vii) Correction Method Determination Unit FIG. 12 is a diagram for explaining processing executed by the correction method determination unit. The correction method determination unit 708 has a morpheme search module 7081, and checks whether or not all morphemes constituting the target morpheme obtained from the target morpheme automatic extraction unit 707 exist in the morpheme dictionary 704. If there is at least one morpheme (unregistered morpheme) that is not listed in the morpheme dictionary (No in the search determination step), the unregistered morpheme is written as a word registration unit (in this specification, “word”) , Which is synonymous with morpheme) and is registered in the morpheme dictionary 704. On the other hand, when all the morphemes are listed in the morpheme dictionary (Yes in the search determination step), the correction method determination unit 708 outputs the correct target morpheme and the optimal target morpheme to the cost adjustment rule creation unit 709.

  For example, when the correct morpheme analysis result is “Tokyo / University / Director”, the morpheme search module 7081 searches whether each of the three morphemes “Tokyo”, “University”, and “Director” exists in the morpheme dictionary 704. . If everything exists in the morpheme dictionary 704, the process proceeds to the cost adjustment rule creation unit 709. For example, when “Tokyo” and “University” are listed in the dictionary but “Director” is not listed in the dictionary, the morpheme “Director” is output to the word registration unit 712.

(Viii) Word Registration Unit FIG. 13 is a diagram for explaining processing executed by the word registration unit. The word registration unit 712 has a function of registering correct morphemes that are not registered in the morpheme dictionary 704. The word registration unit 712 has a morpheme registration module 7121, registers unregistered words in the morpheme dictionary 704 using this, and updates the morpheme dictionary 704. Also, the morpheme registration module 7121 outputs the morpheme registered in the morpheme dictionary 704 to the target morpheme automatic extraction unit 707.

  For example, when the morpheme “Director” is input, the word registering unit 712 registers the “Director” in the morpheme dictionary 704 and then outputs it to the target morpheme automatic extraction unit 707.

  By updating the dictionary, the optimum morpheme analysis result performed by the target morpheme automatic extraction unit 707 may be different from the optimum morpheme analysis result using the dictionary before being updated. For example, before updating the dictionary, the optimal morphological analysis result was "He / Higashi / Kyoto University / Gakuin / Chief /", but after updating the dictionary, the optimal morphological analysis result was "Hi / ha / Tokyo / University / Director / ". In this case, since it is the same as the correct morphological analysis result input by the user, the entire process is completed. On the other hand, even after updating the dictionary, if the optimal morphological analysis result is “His / Higashi / Kyoto University / Gakuin / Chief /”, the process further proceeds to the correction method determination unit 708, but the unregistered morpheme Does not exist (the unregistered morpheme has already been registered), and the process further proceeds to the cost adjustment rule creation unit 709.

(Ix) Cost Adjustment Rule Creation Unit The cost adjustment rule creation unit 709 has a function of automatically creating a cost adjustment rule for the target morpheme sequence input from the correction method determination unit 708. The said cost adjustment rule is a rule described using language feature-values (Table 2), such as description of a morpheme which comprises a target morpheme string, a part of speech, a reading, and an accent position.

  FIG. 14 is a diagram for explaining processing executed by the cost adjustment rule creation unit. Here, the versatility and accuracy of the rule differ depending on the difference in the feature values used for the description. Here, versatility (size) is defined for all language features. In particular, when it is intended to improve reading errors in speech synthesis, it is desirable to use features that affect the synthesized speech, such as accent type and accent position, as the characteristics of the dismantling elements.

  For example, when “Tokyo / University / Director” is described using notation, “Rule: Morphological notation is“ Tokyo ”,“ University ”, and“ Director ”are arranged in order. The cost is -30. " In this case, it applies to “He is a Tokyo Graduate School Director” but cannot apply to “He is a Nanjing Graduate School President”. (Here, "Nanjing" is a place name in China.) In other words, this rule has high accuracy but low versatility.

  For example, if the same “Tokyo / University / Director” is described using parts of speech, then “rule: the part of speech of the morpheme is“ genoun noun ”,“ general noun ”, and“ general noun ”. If they are lined up, the cost is -30. " In this case, it applies to “He is the Tokyo Graduate School Director”. On the other hand, when applying the morphological combination of “Tokyo: place name noun”, “king: general noun (meaning of king)” and “department store: general noun” to “here is Tokyo Oo department store”, “here / ha / tokyo / Oh / Department Store / ”is analyzed as the optimal path. However, since “Here / Higashi / Keio / Department Store /” is the correct answer, the registration of this rule would have resulted in misanalysis. Thus, the highly versatile rule lacks accuracy. In general, versatility and accuracy are in a trait-off relationship, and it is very difficult to achieve both.

  In addition, the cost value greatly affects the accuracy. For example, “rule: the part of speech of a morpheme is“ place name noun ”,“ general noun ”,“ general noun ”, and the cost is set to −30 when three morphemes are arranged in order. ”Is the lowest cost of“ here / is / East / Keio / department store / ”compared to“ here is Tokyo Oo Department Store ”, but it can be analyzed correctly, but“ rule: morpheme part of speech ” If the three morphemes “place name noun”, “general noun”, and “general noun” are arranged in order, the cost is −130. ”Is registered, the cost of“ here / is / Tokyo / king / department store / is ”is the lowest with respect to“ here is Tokyo King department store ”, and a morphological analysis error occurs.

  In order to solve these problems, the cost adjustment rule creation unit 709 includes a rule creation module 7091, a cost optimization module 7092, and a rule registration module 7093 as shown in FIG. In the rule creation module 7091, at least the adjustment cost value is added to the cost value of the correct target morpheme (if there are a plurality of cost values, it is effectively subtracted), and then it may become smaller than the cost value of the optimal target morpheme. This is a rule creation condition. In other words, for "He is a graduate student of Tokyo", for example, the cost value of "He / has / Higashi / Kyoto University / Gakuin / Head /" of the optimal morphological analysis result is 500, and the correct morphological analysis result of "He" If the cost value of “/ is / Tokyo / University / Director /” is 525, the adjustment cost value must be set to a value smaller than at least “−26”.

  However, such a cost determination (cost adjustment rule creation) method can guarantee that the correct morpheme sequence has the minimum cost (that is, the optimal morpheme analysis result) for the input text currently being processed. Cannot evaluate the effect on the sentence. In particular, if a more versatile rule is created, using the created cost adjustment rule as in the example of “Here is Tokyo Oo Department Store” may degrade the analysis accuracy of the morphological analysis system. There is sex.

  Therefore, in order to achieve the best balance between versatility and accuracy, the cost optimization module 7092 uses the correct morphological analysis result database 710 prepared in advance to convert the morphological analysis results that have conventionally been made correct by applying the rules. The presence or absence of an effect (being a result different from the correct morpheme) is confirmed (impact determination step). In the case where there is an influence (in the case where the influence determination step is Yes), the rule creation module 7091 recreates a rule with slightly reduced versatility, and the process of the cost optimization module 7092 is executed again. . On the other hand, when there is no influence (No in the influence determination step), the rule registration module 7093 registers the cost adjustment rule in the cost dormitory rule group 711. In this way, the cost optimization module 7092 adjusts the cost value so as not to affect all the correct morphological analysis results (or to have the least effect). Note that the rule creation module 7091 is preferably created from rules with higher versatility.

  For example, in order to adjust the cost of “He is a graduate student at Tokyo,” “rule: morpheme part of speech is“ place name noun ”,“ general noun ”, and“ general noun ”. The cost is assumed to be “−N”. And the value of −N is optimized using the correct morphological analysis result database. As for the optimization method, various mathematical algorithms such as the EM algorithm for obtaining an optimum solution have been proposed. As the simplest method, the value of −N is varied from 0 to a sufficiently small value, A method may be considered in which the range of the one that does not affect the correct answer database (or has the least influence) is specified, and the largest value is set from the range. Here, “adopting the rule with the least influence” means that the rule having the highest versatility is applied to the plurality of correct morphological analysis results registered in the correct morphological analysis result database 710 sequentially. This means that a rule that minimizes the number of correct morphemes (a number that results in a difference from the correct morpheme) is adopted. For example, as a result of changing the value of −N from 0 to 1000 in increments of 1 in the range of −30 to −1000, it can be confirmed that the correct morphological analysis result database is not affected. Set. On the other hand, when the value of -N is changed from 0 to 1000 in increments of 1 and it is confirmed that the correct morphological analysis result is affected in all ranges, for example, a rule with small versatility is “ Rule: The part of speech of a morpheme has a cost of −30 when three morphemes of “place name noun”, “general noun”, and “general noun” are arranged in order. ”And optimize the value of −N in the same manner as described above. In this way, rules that have high versatility are applied in order, and rules that are confirmed to have no effect on the correct morphological analysis result database 710 are adopted as cost adjustment rules. However, even if the rule is the least versatile (when all features are used), if it affects the database, you can choose between adopting the rule with the least impact or not creating a rule. . This may be determined by the user each time or may be defined in advance.

  Here, even if there is no correct morphological analysis result database 710, morphological analysis mistakes can be improved in the input sentence (input text). However, it is considered that the larger the correct morphological analysis result database 710 is, the higher quality the rule can be created (which improves the analysis accuracy of the morphological analysis system most). In the correct morphological analysis result database 710, the original text and the correct morphological analysis result (path N) are registered as a set.

<Summary>
(I) According to the embodiment of the present invention, the morphological analysis tuning apparatus performs morphological analysis on the input text, and among the morphological analysis results, the optimal morphological analysis result that is the morphological analysis result that takes the minimum value, and The correct morphological analysis result input as correct data by the user with respect to the input text is compared. Then, if the correct morpheme analysis result and the optimal morpheme analysis result have different parts as a result of the comparison process, the apparatus sets the cost adjustment rule so that the correct morpheme analysis result is less expensive than the optimal morpheme analysis result. Generate. For example, when the cost of the optimal morpheme analysis result is 20 and the cost of the correct morpheme analysis result is 30, the cost adjustment rule is “the morphological part of morpheme of the correct morpheme analysis result (“ place name noun ”+“ general noun ”+“ In the case of “general nouns” (eg, “Tokyo”, “university”, “director”), the cost can be “-11”. By doing so, the cost of the correct morphological analysis result can be set to 19 (in this case, the cost of the optimal morphological analysis result is not changed), so the optimality of the correct morphological analysis result ( Costs can be minimized (scores can be maximized).

  If there is a difference between the correct morpheme analysis result and the optimum morpheme analysis result as a result of the comparison, the morpheme of the different part is registered in the morpheme dictionary. Further, the morpheme analysis tuning apparatus uses this updated morpheme dictionary to perform morpheme analysis on the same input text and obtain an updated optimum morpheme analysis result. Then, the apparatus again compares the updated optimal morpheme analysis result and the correct answer morpheme analysis result, and executes a process of generating a cost adjustment rule when there is a different part. By doing so, it is possible to immediately correct a state where the correct morpheme analysis result is different from the optimal morpheme analysis result because it is simply not in the morpheme dictionary, and it is not necessary to generate a cost adjustment rule unnecessarily. In addition, the morpheme dictionary can be maintained in an optimum state.

  More specifically, the morphological analysis tuning device generates the cost adjustment rule so that the correct answer of the correct morphological analysis result registered in the correct morphological analysis result database is maintained when the cost adjustment rule is applied. Alternatively, a cost adjustment rule that minimizes the cost of the correct morphological analysis result registered in the correct morphological analysis result database is adopted. More specifically, the device applies the cost adjustment rule with the highest versatility to the correct morphological analysis result registered in the correct morpheme analysis result database in order to maintain the correct answer. The cost adjustment rule that minimizes the cost of the correct morphological analysis result is adopted. By doing in this way, the correct morphological analysis result already registered in the database is not affected (the correct answer is not changed to an incorrect answer) or the influence is minimized. become able to.

(Ii) The present invention can also be realized by software program codes that implement the functions of the embodiments. In this case, a storage medium in which the program code is recorded is provided to the system or apparatus, and the computer (or CPU or MPU) of the system or apparatus reads the program code stored in the storage medium. In this case, the program code itself read from the storage medium realizes the functions of the above-described embodiments, and the program code itself and the storage medium storing the program code constitute the present invention. As a storage medium for supplying such program code, for example, a flexible disk, CD-ROM, DVD-ROM, hard disk, optical disk, magneto-optical disk, CD-R, magnetic tape, nonvolatile memory card, ROM Etc. are used.

  Also, based on the instruction of the program code, an OS (operating system) running on the computer performs part or all of the actual processing, and the functions of the above-described embodiments are realized by the processing. May be. Further, after the program code read from the storage medium is written in the memory on the computer, the computer CPU or the like performs part or all of the actual processing based on the instruction of the program code. Thus, the functions of the above-described embodiments may be realized.

  Further, by distributing the program code of the software that realizes the functions of the embodiment via a network, it is stored in a storage means such as a hard disk or memory of a system or apparatus, or a storage medium such as a CD-RW or CD-R And the computer (or CPU or MPU) of the system or apparatus may read and execute the program code stored in the storage means or the storage medium when used.

  Finally, it should be understood that the processes and techniques described herein are not inherently related to any particular apparatus, and can be implemented by any suitable combination of components. In addition, various types of devices for general purpose can be used in accordance with the teachings described herein. It may prove useful to build a dedicated device to perform the method steps described herein. Various inventions can be formed by appropriately combining a plurality of constituent elements disclosed in the embodiments. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, constituent elements over different embodiments may be appropriately combined. Although the present invention has been described with reference to specific examples, these are in all respects illustrative rather than restrictive. Those skilled in the art will appreciate that there are numerous combinations of hardware, software, and firmware that are suitable for implementing the present invention. For example, the described software can be implemented in a wide range of programs or script languages such as assembler, C / C ++, perl, shell, PHP, Java (registered trademark).

  Furthermore, in the above-described embodiment, control lines and information lines are those that are considered necessary for explanation, and not all control lines and information lines on the product are necessarily shown. All the components may be connected to each other.

  In addition, other implementations of the invention will be apparent to those skilled in the art from consideration of the specification and embodiments of the invention disclosed herein. Various aspects and / or components of the described embodiments can be used singly or in any combination. The specification and specific examples are merely exemplary, and the scope and spirit of the invention are indicated in the following claims.

DESCRIPTION OF SYMBOLS 100 ... Speech synthesis system 101 ... Text analysis device 102 ... Morphological analysis tuning device 103 ... Speech synthesis device 701 ... Text input unit 702 ... Morphological analysis unit 703 ... Presentation of analysis result Unit 704 ... morpheme dictionary 705 ... analysis result determination unit 706 ... correct morpheme analysis result input unit 707 ... target morpheme automatic extraction unit 708 ... correction method determination unit 709 ... cost adjustment rule creation Unit 710 ... correct morphological analysis result database 711 ... cost adjustment rule group 712 ... word registration unit

Claims (15)

A morphological analysis tuning device for generating a cost adjustment rule used for morphological analysis based on a morphological analysis result of a text sentence,
A correct morphological analysis result database for storing correct morphological analysis results that are input by the user as determined to be correct;
A processor that creates the cost adjustment rule while optimizing using the correct morphological analysis result and the morpheme read from the morpheme dictionary,
The processor is
A comparison process for comparing the optimal morpheme analysis result, which is the morpheme analysis result having the lowest cost among the morpheme analysis results of the text sentence, and the correct morpheme analysis result input by the user as correct data regarding the text sentence When,
As a result of the comparison process, when the correct morpheme analysis result and the optimal morpheme analysis result have different portions, the cost adjustment rule is set so that the correct morpheme analysis result has a lower cost than the optimal morpheme analysis result. Adjustment process to generate,
A morphological analysis tuning apparatus characterized by executing In claim 1,
The processor registers the correct morpheme analysis result corresponding to the different part in the morpheme dictionary and updates the morpheme dictionary, and then obtains the updated optimal morpheme analysis result obtained using the updated morpheme dictionary. The correct morpheme analysis result and the updated optimal morpheme analysis result are compared again, and the adjustment process is performed when the correct morpheme analysis result and the updated optimal morpheme analysis result have different parts. A morphological analysis tuning device characterized by In claim 1,
In the adjustment process, the processor generates the cost adjustment rule so that the correct answer of the correct morphological analysis result registered in the correct morphological analysis result database is maintained when the cost adjustment rule is applied. A morphological analysis tuning device characterized by that. In claim 3,
In the adjustment process, the processor sequentially applies the most versatile cost adjustment rule to the low versatility cost adjustment rule to the correct morphological analysis result registered in the correct morphological analysis result database to maintain the correct answer. A morphological analysis tuning device characterized by adopting a cost adjustment rule. In claim 1,
In the adjustment process, the processor employs a cost adjustment rule that minimizes the cost of the correct morphological analysis result registered in the correct morphological analysis result database when the cost adjustment rule is applied. A morphological analysis tuning device. In claim 5,
In the adjustment process, the processor sequentially applies the most versatile cost adjustment rule to the correct morpheme analysis result registered in the correct morphological analysis result database, from the most versatile cost adjustment rule to the correct morpheme analysis result. A morphological analysis tuning apparatus that employs a cost adjustment rule that minimizes the cost of an analysis result. In claim 1,
The processor presents a plurality of morphological analysis results of the text sentence to a user in ascending order of cost, and allows the user to select the correct morphological analysis result from the plurality of morphological analysis results. A morphological analysis tuning device. A speech synthesis system that generates and outputs synthesized speech of an input text sentence,
A text analysis device that refers to the morpheme dictionary, executes morpheme analysis, gives a reading of the text sentence based on the result of the morpheme analysis, and determines an accent;
A morphological analysis tuning device according to claim 1;
A voice synthesizer that generates a synthesized voice of the text sentence using the reading of the text sentence from the text analysis device and the accent;
The speech analysis system, wherein the text analysis device executes the morpheme analysis using the cost adjustment rule generated by the morpheme analysis tuning device. A morphological analysis tuning method for generating a cost adjustment rule used for morphological analysis based on a morphological analysis result of a text sentence,
A processor accessing a correct morphological analysis result database that stores a correct morphological analysis result inputted by determining that the user is correct, and obtaining the correct morphological analysis result;
Using the correct morpheme analysis result and the morpheme read from the morpheme dictionary to create the processor while optimizing the cost adjustment rule,
Creating the cost adjustment rule while optimizing,
The processor includes an optimal morpheme analysis result that is a morpheme analysis result that takes a minimum value among morpheme analysis results of the text sentence, and a correct morpheme analysis result that the user has input as correct data regarding the text sentence. Executing a comparison process to be compared;
When the processor has a portion where the correct morpheme analysis result and the optimal morpheme analysis result are different as a result of the comparison process, the correct morpheme analysis result has a lower cost than the optimal morpheme analysis result. And a step of executing an adjustment process for generating a cost adjustment rule. In claim 9,
The processor registers the correct morpheme analysis result corresponding to the different part in the morpheme dictionary and updates the morpheme dictionary, and then obtains the updated optimal morpheme analysis result obtained using the updated morpheme dictionary. The correct morpheme analysis result and the updated optimal morpheme analysis result are compared again, and the adjustment process is performed when the correct morpheme analysis result and the updated optimal morpheme analysis result have different parts. A morphological analysis tuning method characterized by In claim 9,
In the adjustment process, the processor generates the cost adjustment rule so that the correct answer of the correct morphological analysis result registered in the correct morphological analysis result database is maintained when the cost adjustment rule is applied. A morphological analysis tuning method characterized by the above. In claim 11,
In the adjustment process, the processor sequentially applies the most versatile cost adjustment rule to the low versatility cost adjustment rule to the correct morphological analysis result registered in the correct morphological analysis result database to maintain the correct answer. A morphological analysis tuning method characterized by adopting a cost adjustment rule. In claim 9,
In the adjustment process, the processor employs a cost adjustment rule that minimizes the cost of the correct morphological analysis result registered in the correct morphological analysis result database when the cost adjustment rule is applied. To tune morphological analysis. In claim 13,
In the adjustment process, the processor sequentially applies the most versatile cost adjustment rule to the correct morpheme analysis result registered in the correct morphological analysis result database, from the most versatile cost adjustment rule to the correct morpheme analysis result. A morphological analysis tuning method that employs a cost adjustment rule that minimizes the cost of an analysis result. In claim 9,
The processor presents a plurality of morphological analysis results of the text sentence to a user in ascending order of cost, and allows the user to select the correct morphological analysis result from the plurality of morphological analysis results. To tune morphological analysis.

Images