TWI567569B - Natural language processing systems, natural language processing methods, and natural language processing programs - Google Patents

Description

Natural language processing system, natural language processing method, and natural language processing program product

One aspect of the present invention relates to a natural language processing system, a natural language processing method, and a natural language processing program.

As one of the basic techniques of natural language processing, it is known to divide a sentence into a morpheme string and then determine the word formation of the word of each morpheme. In connection with this, Patent Document 1 describes that the input data is decomposed into morphemes, and the morpheme dictionary is used to obtain information on the position corresponding to the morphemes to be decomposed, by using location information. The cost function determines a word formation device for the morpheme string from the candidate for the morpheme string obtained by the decomposition.

[Previous Technical Literature] [Patent Document]

[Patent Document 1] Japanese Patent Laid-Open Publication No. 2013-210856

The word formation is performed by using a segmentation model containing the scores of each prime. It can be said that the segmentation model of the knowledge required for word formation analysis is generally fixed beforehand. Therefore, if you try to parse the statement of the new domain that is not included in the segmentation model or the statement with the new property, then Of course, it will be very difficult to get the right results. On the other hand, if an attempt is made to correct the segmentation model by means of mechanical learning or the like, the time required for the correction may grow unpredictably. Therefore, the segmentation model of word formation analysis is automatically corrected within a certain period of time, which is expected by people.

The natural language processing system according to one aspect of the present invention includes: an analysis unit that performs a word formation analysis on a sentence by using a segmentation model obtained by mechanical learning using one or more pieces of training data to Each of the divided elements obtained by dividing a sentence is set to a label indicating at least a word of the word, wherein the analysis unit includes: a score indicating the output quality of the divided element and the label, and a continuous representation The fraction of the kinomism of the combination of the two labels corresponding to the two divided elements; and the correction unit, which is a label indicating the result of the analysis obtained by the analysis unit, and a correct answer for the label indicating the correct answer of one sentence, For comparison, the scores of the output primes and the kinetics scores associated with the labels of the correct answers corresponding to the incorrectly answered labels are set higher than the current values, and the output traits associated with the labels of the incorrect answers are The scores of the scores and the kinetics are set lower than the current value, and the segmentation model used by the analysis unit in the word formation analysis of the next sentence is corrected.

The natural language processing method according to one aspect of the present invention is a natural language processing method executed by a natural language processing system including a processor, which includes an analysis step and uses one or more of them. a segmentation model obtained by mechanical learning of training materials, performing parsing of a sentence, and setting a label indicating at least a word of the word for each divided element obtained by dividing the one sentence, wherein the parsing step is The model system includes: a score indicating the output prime of the divided element and the label, and a fraction of the kinomism indicating a combination of the two labels corresponding to the two consecutive divided elements; and a correction step, which is obtained in the analysis step The label of the analysis result and the correct answer data of the label indicating the correct answer of a sentence are compared, and the score of the output prime and the kinetic property associated with the label of the correct answer corresponding to the incorrectly answered label are compared. The score is set higher than the current value, and the score of the output prime associated with the label of the incorrect answer and The score of the mobilizing property is set lower than the current value, and the segmentation model used in the parsing of the next sentence in the parsing step is corrected.

The natural language processing program according to one aspect of the present invention is an analysis unit that performs word formation on a sentence by using a segmentation model obtained by mechanical learning using one or more training materials. In the analysis, a label indicating at least the word of the word is set for each of the divided elements obtained by dividing the one sentence, wherein the analysis unit includes: the input indicating that the divided element corresponds to the label The score of the elementality and the fraction of the kinomism indicating the combination of the two labels corresponding to the two consecutive divided elements; and the correction unit, which is the label of the analysis result obtained by the analysis unit, and the correctness of the statement The correct answer information of the answer label, compare and compare the score of the output prime and the score of the mobility factor associated with the label of the correct answer corresponding to the incorrectly answered label to be higher than the current value, and the incorrect answer will be The scores of the output primes associated with the labels and the scores of the mobility factors are set lower than the current values, and the segmentation model used by the analysis unit in the word formation analysis of the next sentence is corrected.

In this aspect, each time a sentence is parsed, the parsing result is compared with the correct answer data, and the segmentation model is corrected according to their differences. By correcting the segmentation model for each sentence as described above, the time required for the correction of the segmentation model when processing the complex sentence is suppressed to a substantially linear growth degree with the number of sentences, so that the segmentation of the word formation can be performed. The model is automatically corrected for a certain period of time (in other words, within a predictable time frame).

According to one aspect of the present invention, the segmentation model of the word formation can be automatically corrected within a certain period of time.

10‧‧‧Natural Language Processing System

11‧‧‧Acquisition Department

12‧‧‧ Analysis Department

13‧‧‧Amendment

20‧‧‧ segmentation model

P1‧‧‧Natural Language Processing Program

P10‧‧‧ main module

P11‧‧‧Get Module

P12‧‧‧analysis module

P13‧‧‧correction module

[Fig. 1] Processing in a natural language processing system according to an embodiment Concept map.

Fig. 2 is a diagram showing an example of word formation analysis in the embodiment.

Fig. 3 is a view showing a hardware configuration of a computer constituting the natural language processing system according to the embodiment.

Fig. 4 is a block diagram showing the functional configuration of the natural language processing system according to the embodiment.

[Fig. 5] A conceptual illustration of an example of label assignment.

[Fig. 6] (a) and (b) are schematic diagrams each showing an example of score update.

Fig. 7 is a flow chart showing the operation of the natural language processing system according to the embodiment.

Fig. 8 is a view showing the configuration of a natural language processing program according to the embodiment.

Hereinafter, embodiments of the present invention will be described in detail with reference to the accompanying drawings. In the description of the drawings, the same or equivalent elements are designated by the same reference numerals, and the repeated description is omitted.

First, the function and configuration of the natural language processing system 10 according to the embodiment will be described with reference to Figs. The natural language processing system 10 is a computer system that performs word formation analysis. The so-called word formation analysis divides the sentence into morpheme strings and then determines the processing of the words of each morpheme. The so-called statement is a unit that expresses the language expression of a finished statement and is expressed as a string. The morpheme is the smallest language list with semantics. Bit. The morpheme string is an arrangement of one or more morphemes obtained by dividing a sentence into one or more morphemes. The so-called word is a distinction between words made in accordance with the function or form of grammar.

The natural language processing system 10 uses the segmentation model 20 to perform word formation analysis for each sentence. One of the features of the natural language processing system 10 is that, for example, when the segmentation model 20 is learned, the segmentation model 20 is corrected each time a sentence is analyzed for word formation. When the correction of the segmentation model 20 is completed, the natural language processing system 10 having the determined segmentation model 20 is provided to the user. The user system can cause the natural language processing system 10 to perform word formation analysis. At this time, the word formation analysis can be performed without modifying the segmentation model 20. The "segmentation model" in this booklet is a reference (cue) at the time of dividing a sentence into one or more morphemes, and is expressed as a score of each prime. The segmentation model is obtained by mechanical learning using one or more training materials. The training material is at least a document indicating that the sentence has been divided into words and the words of each word obtained by dividing the sentence. The so-called feature is the clue needed to get the correct result in word formation. In general, there is no limit to what to use as a prime (clue). The so-called prime score is the value that represents the most reasonable appearance of the prime.

The concept of processing in the natural language processing system 10 of the present embodiment is shown in a simplified manner in FIG. Further, the gear M in Fig. 1 indicates the execution of word formation analysis. At a certain point in time, the natural language processing system 10 divides the sentence s ₁ into one or more morphemes by performing the word formation analysis using the segmentation model w ₁ . In the present embodiment, the natural language processing system 10 divides a sentence into each character and then performs processing of a character unit to divide the sentence into one or more morphemes. In other words, in the present embodiment, the divided elements to be processed are characters. The natural language processing system 10 sets a label for each character (the divided element) to indicate the result of the word formation analysis. The so-called "label" in this book is a sign indicating the attributes or functions of the characters. The labeling is described in more detail later.

Once the word formation analysis is performed, the natural language processing system 10 accepts the data indicating the correct answer of the word formation analysis of the sentence s ₁ (correctly answering the data), and corrects the segmentation model by comparing the analysis result with the correct answer data. w ₁ obtained by dividing a new model w _2. Specifically, in the natural language processing system 10, when at least a part of the word formation analysis of the sentence s ₁ is erroneous, the entire analysis result is evaluated as an error. Then, the natural language processing system 10 evaluates the primeness corresponding to each label in the correct answer data as "correct (+1)" and then sets the score of the prime to be higher than the current value, and analyzes each label in the result. The corresponding prime evaluation is "error (-1)" and then the score of the prime is set lower than the current value, thereby obtaining the segmentation model w ₂ . If a part of the label in the analysis result is the correct answer, then the two evaluations of the prime label (correct (+1)) and "error (-1)" of the label (the correct answer) will be Offset each other. Therefore, as described above, the process of setting the score of the prime to be low or high can be said to be the label of the correct answer corresponding to the incorrectly answered label (the label of the correct answer corresponding to the incorrect answer part) The score of the associated prime is set high, and the score of the prime associated with the label of the incorrect answer (the label of the incorrect part) is set low.

In addition, the natural language processing system 10 may evaluate each label in the correct answer data as "correct (+1)", and on the other hand, evaluate the label of each character in the analysis result as "error ( -1)", after the two evaluation results are offset against each label, the score of the prime corresponding to the label that is evaluated as "correct (+1)" is set to be high and will be evaluated as "error (-1 The score of the prime corresponding to the label is set to be low.

For example, suppose there are five words x _a , x _b , x _c , x _d , x _{e in} statement s ₁ . Then, suppose that the labels of the correct answers of the words x _a , x _b , x _c , x _d , x _e are t _a , t _b , t _c , t _d , t _e , respectively, and the labels of the characters are parsed by the word formation. It is t _a , t _g , t _h , t _d , t _e . In this case, the natural language processing system 10 evaluates the prime of the labels t _a , t _b , t _c , t _d , t _{e in} the correct answer as the "correct (+1)" evaluation. The score is set higher than the current value, and the prime value corresponding to the labels t _a , t _g , t _h , t _d , t _{e in} the execution result is evaluated as "error (-1)" and the score of the prime is set to be low. The value now. In this case, the scores of the primes corresponding to the labels t _a , t _d , and t _e are the same as before the update, and the scores of the primes corresponding to the correctly answered labels t _b and t _c become high. The incorrectly answered label t _g , t _h corresponds to the low score of the prime.

When the word formation analysis is performed on the next sentence s ₂ , the natural language processing system 10 uses the segmentation model w ₂ . Then, based natural language processing system 10 receives the statement s ₂ Formation of correct answer information parsed, the execution result with the correct answer data are compared with the same manner as the correction corrected segmentation model ₁ w ₂ w-division model is obtained a new The segmentation model w ₃ .

The natural language processing system 10 is such that the segmentation model is corrected each time a sentence (s ₁ , s ₂ , ..., s _t ) is processed (w ₁ → w ₂ , w ₂ → w ₃ , ..., w _t → W _t+1 ), the modified segmentation model is used in the word formation of the next sentence. This method of updating the model each time a training material is processed is also called "online learning" or "online mechanical learning".

An example of the result of the word formation analysis performed by the natural language processing system 10 is shown in FIG. In this example, the natural language processing system 10 will be equivalent to "I bought a book" in this English. buy (hon wo katte)" (buy a book) The sentence of this day is divided into five words x ₁ : "本 (hon)", x ₂ : " (wo)", x ₃ : "buy (ka)", x ₄ : " (t)", x ₅ : " (te)". Then, the natural language processing system 10 sets a label for each character by performing word formation analysis. In the present embodiment, the label is a combination of the appearance of the character in the word, the word of the word, and the sub-level of the word of the word, and is expressed using an English letter such as "SN-nc".

The appearance pattern is used to indicate whether a text is formed by a single word or combined with other characters to form a word, and when the text is part of a word formed by more than 2 characters, the text is located within the word. Which location information. In the present embodiment, the appearance pattern is expressed by any of S, B, I, and E. The appearance of the "S" system indicates that the words constitute a single word. The appearance pattern "B" indicates that the text is at the beginning of a word formed by two or more characters. The appearance "I" indicates that the text is in the middle of a word made of more than 3 characters. The appearance "E" indicates that the text is at the end of the word formed by more than 2 characters. The example of Fig. 2 illustrates that the words x ₁ , x ₂ , x ₅ are separate and one word, and the words x ₃ , x ₄ are one word formed.

In addition, the architecture regarding the appearance of the aspect is not limited. In the present embodiment, although the "SBIEO" architecture is employed, it is also possible to employ, for example, the "IOB2" architecture known to the industry.

Examples of the word can be exemplified by nouns, verbs, auxiliary words, adjectives, adjective verbs, and continuation words. In the present embodiment, the noun is represented by "N", the auxiliary word is represented by "P", and the verb is represented by "V". The example of Fig. 2 illustrates that the word x ₁ is a noun, the word x ₂ is a help word, the word x ₃ , the word formed by x ₄ is a verb, and the word x ₅ is a help word.

The sub-level of the word is the subordinate concept of the corresponding word. For example, nouns can be subdivided into general nouns and proper nouns, and auxiliary words can be subdivided into auxiliary words, subsequent auxiliary words, and auxiliary words. In the present embodiment, the general noun is represented by "nc", the proper noun is represented by "np", the auxiliary auxiliary word is represented by "k", the subsequent auxiliary auxiliary is represented by "sj", and the general verb is represented by "c". Said. The example in Fig. 2 illustrates that the character x ₁ is a general noun, the word x ₂ is a verb auxiliary word, the word x ₃ , the word formed by x ₄ is a general verb, and the word x ₅ is a continuation auxiliary word.

The fraction of the primes memorized by the segmentation model 20 is the fraction of the output feature and the kinesin (transition) The score of feature).

The output prime is a clue indicating the correspondence between the label and the text or the type of the text. In other words, the output prime is a clue as to which kind of label is relatively easy to be matched by which character or type of text. The output prime is the prime representation of the output matrix corresponding to the hidden Markov model. In the present embodiment, the output quality of a single conjunction (a string formed by only one character) and the output of a double conjunction (a string formed by two consecutive characters) are used.

Here, the type of text refers to the type of words in a certain language. Examples of Japanese language types include Chinese characters, hiragana, katakana, English letters (uppercase and lowercase), Arabic numerals, Chinese numerals, and Chinese black (‧). Further, in the present embodiment, the character type is indicated by an English letter. For example, "C" means Chinese characters, "H" means hiragana, "K" means katakana, "L" means English letters, and "A" means Arabic numerals. The example of Fig. 2 illustrates that the characters x ₁ , x ₃ are Chinese characters, the characters x ₂ , x ₄ , and x ₅ are hiragana.

The output property of a single conjunction of a character is a clue indicating the correspondence between the tag t and the character x. Further, the output property of the single conjunction of the character type is a clue indicating the correspondence between the tag t and the character type c. In the present embodiment, the score s of the most reasonable appearance corresponding to the label t and the character x is expressed by {t/x, s}. Further, the score s of the most reasonable appearance corresponding to the label t and the character type _c is expressed by {t/c, s}. The segmentation model 20 contains scores for plural tags for a text or type of text. For a text or type of text, if there is information about all kinds of labels, the segmentation model 20 will also contain a score on the combination of the label and the text or text type that is practically impossible to grammatically. However, the scores of the primes that are impossible in grammar are relatively low.

The following is an example of the score of the output quality of the Japanese word "hon". Although the text cannot be a mnemonic in Japanese grammar, as mentioned above, there is still information on the nature of "S-P-k/本(hon)" which does not exist in grammar.

{S-N-nc/本(hon), 0.0420}

{B-N-nc/本(hon),0.0310}

{S-P-k/本(hon), 0.0003}

{B-V-c/本(hon), 0.0031}

In addition, an example of the score of the output quality of the Chinese character "Chinese character" is displayed.

{S-N-nc/C,0.0255}

{E-N-np/C, 0.0488}

{S-P-k/C, 0.0000}

{B-V-c/C, 0.0299}

As for the type of text, it is also possible to prepare information indicating the grammatical non-existence. For example, although in Japanese grammar, words represented by Arabic numerals are unlikely to be auxiliary words, information on "S-P-k/A" is still available.

The output property of the double conjunction of the character is a clue indicating the correspondence between the tag t and the string x _i x _i+1 . Further, the output property of the double conjunction of the character type is a clue indicating the correspondence between the tag t and the character string c _i c _i+1 of the character type. In the present embodiment, the score s of the most reasonable appearance of the label t and the characters x _i , x _i+1 is expressed by {t/x _i /x _i+1 ,s}. Further, the score s of the most reasonable appearance of the label t and the character type c _i , c _i+1 is expressed by {t/c _i /c _i+1 , s}. For a double conjunction, if there is information about all possible labels, the segmentation model 20 will also memorize the data on the combination of labels and double conjunctions that are not grammatically impossible.

The following shows about "this (hon wo) An example of the score of the output of this double conjunction.

{SN-nc/本(hon)/ (wo), 0.0420}

{BN-nc/本(hon)/ (wo), 0.0000}

{SPk/本(hon)/ (wo), 0.0001}

{BVc/本(hon)/ (wo), 0.0009}

Also, an example of the score of the output quality of the double conjunction of the hiragana after the Chinese character is displayed.

{S-N-nc/C/H, 0.0455}

{E-N-np/C/H, 0.0412}

{S-P-k/C/H, 0.0000}

{B-V-c/C/H, 0.0054}

The so-called migration primality, based text labels representing x _i t _i of the next character and its label of x _{i + 1} t _{i + 1} of the composition (corresponding to the two characters formed by the combination of two successive labels) leads. The kinesin is a trait related to a double conjunction. The kinetics are the prime representations of the migration matrix corresponding to the hidden Markov model. In the present embodiment, the score s of the most reasonable appearance of the combination of the label t _i and the label t _i+1 is expressed by {t _i /t _i+1 , s}. If there is information on the physiology of all possible combinations, the segmentation model 20 will also memorize the data on the combination of the two tags that are practically impossible to grammatically.

Below, several examples of the scores of kinetics are shown.

{S-N-nc/S-P-k, 0.0512}

{E-N-nc/E-N-nc, 0.0000}

{S-P-k/B-V-c, 0.0425}

{B-V-c/I-V-c, 0.0387}

The natural language processing system 10 includes one or more computers. When a plurality of computers are provided, each functional element of the natural language processing system 10 to be described later is realized by distributed processing. There is no limit to the type of each computer. For example, a desktop or portable personal computer (PC), a workstation, or a mobile terminal such as a high-performance mobile phone (smartphone) or a mobile phone or a mobile information terminal (PDA) can be used. Alternatively, various types of computers can be combined to construct the natural language processing system 10. In the case of a plurality of computers, these computers can be connected via a communication network such as the Internet or an internal network.

A general hardware configuration of each computer 100 within the natural language processing system 10 is shown in FIG. The computer 100 includes a CPU (processor) 101 that executes a work system or an application, a main memory unit 102 composed of a ROM and a RAM, and an auxiliary memory unit 103 including a hard disk or a flash memory. a network of wireless cards or wireless communication modules The control unit 104 includes an input device 105 formed by a keyboard or a mouse, an output device 106 such as a display or a printer. Of course, the hardware modules that are mounted vary depending on the type of computer 100. For example, desktop PCs and workstations usually have a keyboard, a mouse, and a screen as input devices and output devices, but smart phones usually use a touch panel as an input device and an output device.

Each function element of the natural language processing system 10 to be described later reads the predetermined software into the CPU 101 or the main memory unit 102, and causes the communication control unit 104, the input device 105, the output device 106, and the like under the control of the CPU 101. The operation is performed by reading or writing data in the main memory unit 102 or the auxiliary memory unit 103. The data or database required for processing is stored in the main memory unit 102 or the auxiliary memory unit 103.

On the other hand, the segmentation model 20 is pre-stored in the memory device. The specific implementation method of the segmentation model 20 is not limited. For example, the segmentation model 20 can also be prepared in a manner related to a database or a text file. Further, the installation location of the segmentation model 20 is not limited. For example, the segmentation model 20 may exist inside the natural language processing system 10 or may exist in another computer system different from the natural language processing system 10. When the segmentation model 20 is located in another natural language processing system, the natural language processing system 10 accesses the segmentation model 20 via the communication network.

As described above, the segmentation model 20 can also be said to be a collection of scores of various primes. In the equation, the segmentation model 20 containing the scores w ₁ , w ₂ , ..., w _n of n primes can be represented by vectors w = {w ₁ , w ₂ , ..., w _n }. The scores of the primes at the time when the segmentation model 20 is newly created are all zero. That is, it is w={0,0,...,0}. This score is gradually updated gradually by the processing of the natural language processing system 10 described later. After processing a certain number of sentences, as described above, a difference gradually occurs between the scores of each prime.

As shown in FIG. 4, the natural language processing system 10 includes an acquisition unit 11, an analysis unit 12, and a correction unit 13 as functional components. The natural language processing system 10 accesses the segmentation model 20 as needed. Although each functional element will be described below, this embodiment is described on the premise that the natural language processing system 10 processes the Japanese sentence. Of course, the language of the sentence processed by the natural language processing system 10 is not limited to Japanese, and sentences of other languages such as Chinese can also be analyzed.

The acquisition unit 11 is a functional element for acquiring a sentence to be divided into a morpheme string. The method of obtaining the statement made by the acquisition unit 11 is not limited. For example, the acquisition unit 11 can also collect statements (that is, so-called crawling) from any website on the Internet. Alternatively, the acquisition unit 11 may read out the pre-stored statements in the database in the natural language processing system 10, or may pre-store the statements in the database on the computer system other than the natural language processing system 10. The statement is read via the communication network and read. Alternatively, the acquisition unit 11 may accept a sentence input by a user of the natural language processing system 10. When the instruction to analyze the first sentence is input, the acquisition unit 11 acquires one sentence and outputs it to the analysis unit 12. Then, when the completion notification is input from the correction unit 13 to be described later, the acquisition unit 11 acquires the next sentence and outputs it to the analysis unit. 12.

The analysis unit 12 is a functional element that performs word formation analysis for each sentence. The analysis unit 12 executes the following processing every time a sentence is input.

First, the analysis unit 12 divides a single sentence into individual characters, and determines the character type of each character. The analysis unit 12 preliminarily stores a comparison table of characters and character types, or a regular expression required for determining the type of characters, and determines the type of characters using the comparison table or the regular expression.

Next, the analysis unit 12 determines the label of each character using a Viterbi algorithm. For the i-th character, the analysis unit 12 determines which one of the plurality of candidate tags of the (i-1)th character is the candidate tag and the connection with respect to each of the tags (candidate tags) that are finally selected. The hourly score (this is also known as the "join score") will be the highest. Here, the connection score is a total value of various scores of the label of the calculation target (the score of the output of the single conjunction, the score of the output of the double conjunction, and the score of the kinomism). For example, when the i-th label is "SN-nc", the analysis unit 12 considers that the (i-1)th label is "SPk", and the connection score is considered to be the highest, and if the i-th label is "SVc", When the (i-1)th label is "EN-nc", the connection score is considered to be the highest...etc. Then, the analyzing unit 12 memorizes the combination having the highest connection score (for example, (S-P-k, S-N-nc), (E-N-nc, S-V-c), etc.). The analysis unit 12 performs such processing by advancing one character at a time from the first character to the end of the text.

There is only one type of label (EOS) for the end of the text, so the combination of the highest score, the last character and the end of the label determines one (for example, the combination is (E-V-c, EOS)). In this way, the label of the last character is determined (for example, the label is determined to be "E-V-c"), and as a result, the label of the second character from the last countdown is also determined. As a result, the label is determined in the order from the end of the sentence to the beginning.

The processing performed by such analysis unit 12 is schematically shown in FIG. Fig. 5 is a diagram showing an example of label assignment of a sentence made of four characters. In order to simplify the description, in this example, the label is simplified to be expressed as "A1", "B2", etc., and the number of candidate labels for each character is set to 3. The thick line in Fig. 5 indicates the combination of the label and the label whose connection score is determined to be the highest, which is obtained by processing the sentence from the front. For example, in the processing of the third character, the connection score between the tag C1 and the tag B1 is the highest, the connection score between the tag C2 and the tag B1 is the highest, and the connection score between the tag C3 and the tag B2 is the highest. In the example of Figure 5, if processed to the end of the statement (EOS), the combination (D1, EOS) is determined, then the combination (C2, D1) is determined, and then, the combination (B1, C2) , (A2, B1) is determined sequentially. Therefore, the analysis unit 12 determines that the labels of the first to fourth characters are A2, B1, C2, and D1, respectively.

The analysis unit 12 outputs a statement in which each character is given a tag as an analysis result. The analysis unit 12 outputs at least the analysis result to the correction unit 13, but the reason is that the analysis result is the repair of the segmentation model 20. It must be on the right. The analysis unit 12 can also perform more output. For example, the analysis unit 12 may display the analysis result on the screen or print the printer, write the analysis result to a text file, or the like, and store the analysis result in a memory device such as a memory or a database. Alternatively, the analysis unit 12 transmits the analysis result to any other computer system other than the natural language processing system 10 via the communication network.

The correction unit 13 corrects the functional elements of the segmentation model 20 based on the difference between the analysis result obtained from the analysis unit 12 and the correct answer analyzed by the word formation of the sentence. The "correction of the segmentation model" in this detailed book is a process of changing the score of at least one prime in the segmentation model. In addition, depending on the situation, it is possible that the result value does not change even if a certain score is to be changed. The correction unit 13 executes the following processing each time one analysis result is input.

First, the correction unit 13 acquires the correct answer data corresponding to the input analysis result, that is, the data indicating the correct answer of the word formation analysis of the sentence processed by the analysis unit 12. The so-called correct answer data in the present embodiment is a material indicating a label (a combination of appearance, word, and sub-level of appearance) of each character forming a sentence. The correct answer is made by the staff. The method of obtaining the correct answer data by the correction unit 13 is not limited. For example, the correction unit 13 may read out the correct answer data accumulated in the database in the natural language processing system 10, or may pre-populate the database on the computer system other than the natural language processing system 10. The accumulated statements are read via the communication network and read. Alternatively, the correction unit 13 can also accept natural language. The correct answer data entered by the user of the processing system 10.

When the correct answer data is obtained, the correcting unit 13 compares the input analysis result with the correct answer data to specify the difference between the two.

If the analysis result is completely different from the correct answer data, the correction unit 13 ends the processing without correcting the division model 20, and generates a completion notification and outputs it to the acquisition unit 11. This completion notification is a signal indicating that the processing of the word formation can be performed on the next sentence after the processing of the correction unit 13 is completed. The fact that the analysis result is completely consistent with the correct answer data means that at least at this point, the segmentation model 20 does not need to be modified, so the natural language processing system 10 (more specifically, the analysis unit 12) directly uses the current segmentation model 20 Parse the next statement.

For example, for the above Japanese sentence "this buy (hon wo katte)" (buy book) The correct answer is as follows. Further, it is also appropriate to express each character as x ₁ to x ₅ .

x ₁ :{SN-nc}

x ₂ :{SPk}

x ₃ :{BVc}

x ₄ :{EVc}

x ₅ :{SP-sj}

Therefore, when the analysis result shown in FIG. 2 is input, the correction unit 13 determines that the analysis result is completely identical to the correct answer data, and the non-correction analysis unit 12 outputs the completion notification to the acquisition unit 11.

On the other hand, if the analysis results and the correct answer is not If the components are completely identical (that is, when there is a difference between the analysis result and the correct answer data), the correction unit 13 updates the score of at least a part of the segmentation model 20. More specifically, the correction unit 13 sets the score of the prime relationship associated with the label of the incorrect answer corresponding to the incorrectly answered label to be higher than the current value, and sets the score of the prime relationship associated with the incorrectly answered label. It is lower than the current value.

For example, suppose the analysis unit 12 is from the Japanese sentence "this. buy (hon wo katte)" The following analysis results were obtained.

x ₁ :{SN-nc}

x ₂ :{SPk}

x ₃ :{BVc}

x ₄ :{IVc}

x ₅ :{EVc}

In this case, the analysis result is all wrong as a whole. Therefore, the correction unit 13 evaluates the primeness corresponding to each label in the correct answer data as "correct (+1)" and sets the score of the prime to be high. At the current value, the prime value corresponding to each label in the analysis result is evaluated as "error (-1)" and the score of the prime is set lower than the current value. When the part in which the result is offset is considered, it can be said that the correction unit 13 finally performs the following processing.

The correction unit 13 is an output element corresponding to the label of the correct answer to the characters x ₄ and x ₅ "EVc/ (t)""SP-sj/ The score of (te) is set to be greater than the current value, and the output factor related to the incorrectly answered label is "IVc/ (t)""EVc/ The score of (te)" is set to be smaller than the current value. Thereby, the score of the output of the single conjunction associated with the sentence being parsed (the score on the text) is updated.

Further, the correction unit 13 sets the score of the output quality "EVc/H""SP-sj/H" associated with the label of the correct answer of the characters x ₄ and x ₅ of the incorrect answer to be larger than the current value, and will be about The score of the output quality "IVc/H" and "EVc/H" associated with the incorrectly answered label is set to be smaller than the current value. Thereby, the score of the output trait of the single conjunction (the score on the type of the text) to which the sentence to be analyzed is related is updated.

Further, the correcting unit 13 is an output element "EVc/" associated with the label of the correct answer of the characters x ₄ and x ₅ of the incorrect answer. (t)/ The score of (te) is set to be greater than the current value, and the output factor related to the incorrectly answered label is "IVc/ (t)/ The score of (te)" is set to be smaller than the current value. Thereby, the score of the output of the double conjunction associated with the sentence being parsed (the score on the text) is updated.

Further, the correction unit 13 sets the score of the output prime property "EVc/H/H" associated with the label of the correct answer of the characters x ₄ and x ₅ of the incorrect answer to be larger than the current value, and labels the incorrect answer. The score of the associated output prime "IVc/H/H" is set to be smaller than the current value. Thereby, the score of the output of the double conjunction associated with the sentence being analyzed (the score on the type of the text) is updated.

Further, the correction unit 13 sets the score of the phytonomic "BVc/EVc""EVc/SP-sj" associated with the label of the correct answer of the characters x ₄ and x ₅ of the incorrect answer to be larger than the current value, and will be about The scores of the kinesin "BVc/IVc" and "IVc/EVc" related to the incorrectly answered label are set to be smaller than the current value. Thereby, the score of the kinetics associated with the sentence being parsed is updated.

Further, as described above, the correction unit 13 may evaluate each label in the correct answer data as "correct (+1)", and on the other hand, evaluate the label of each character in the analysis result as "error. (-1)", after the two evaluation results of each label are offset, the score of the prime corresponding to the label that is evaluated as "correct (+1)" is set high, and will be evaluated as "error (- The score of the prime corresponding to the label of 1) is set low.

When the score of the prime is updated, the correction unit 13 can also use SCW (Soft Confidence-Weighted Learning). The SCW system is based on a parameter with a large number of variances, and the parameter is updated to be large (not correct), and the parameter is updated to a certain degree, and the parameter is updated to a certain degree. Smaller technique. The correction unit 13 determines the amount of change in the score based on the variance of the score having the range of values. In order to execute the SCW, a Gaussian distribution is introduced to the segmentation model 20 (vector w), and the correction unit 13 updates the average of the scores and the covariance matrix in addition to the update of the scores. The average initial value of each score is 0. Regarding the initial value of the common variance matrix of each score, the diagonal element is 1 and the other elements (non-diagonal elements) are 0. Fig. 6(a) is a diagram showing a case where the score with a large variation is greatly changed (that is, the amount of change in the score is large), and Fig. 6(b) is a diagram showing that the score having a smaller variation is given. Minor changes (ie, small changes in scores) Aspect. 6(a) and 6(b) are diagrams respectively showing that the common variance number matrix Σ is also updated when the score is updated from Sa to Sb. In addition, if the update of the covariance matrix is used, the accuracy of the calculation of the score can be maintained even if the correlation between the prime and the other primes is not considered. Therefore, the non-diagonal elements of the covariance matrix are not calculated in this embodiment. Instead, only the diagonal features are calculated. In this way, the update speed of the score can be increased.

Further, the correction unit 13 may update the score of the prime degree by using a method other than the SCW. Examples of the technique other than the SCW include Perceptron, Passive Aggressive (PA), Confidence Weighted (CW), and Adaptive Regularization of Weight Vectors (AROW).

When the segmentation model 20 is corrected by updating the score of the prime factor associated with the sentence being analyzed, the correction unit 13 generates a completion notification and outputs it to the acquisition unit 11. In this case, the natural language processing system 10 (more specifically, the analysis unit 12) analyzes the next sentence using the segmentation model 20 that has been corrected.

Next, the operation of the natural language processing system 10 will be described with reference to Fig. 7, and the natural language processing method according to the embodiment will be described.

First, the acquisition unit 11 acquires one sentence (step S11). Next, the analysis unit 12 analyzes the word formation using the segmentation model 20 (step S12, analysis step). By the word formation, each character of the sentence is given a label such as "S-N-nc".

Next, the correction unit 13 obtains the difference between the result of the word formation analysis performed by the analysis unit 12 and the correct answer data of the word formation analysis. (Step S13). If there is no such difference (step S14; NO), that is, when the word formation analysis by the analysis unit 12 is completely correct, the correction unit 13 ends the processing without correcting the division model 20. On the other hand, if the analysis result differs from the correct answer data (step S14; YES), that is, when at least a part of the word formation analysis performed by the analysis unit 12 is not correct, the correction unit 13 is updated by the update unit 13 The fraction of the primes associated with the parsed statement is corrected to correct the segmentation model 20 (step S15, correction step). Specifically, the correction unit 13 sets the score of the prime relationship associated with the label of the incorrect answer corresponding to the incorrectly answered label to be higher than the current value, and sets the score of the prime relationship associated with the incorrectly answered label to Below the current value.

Once the processing of the correction unit 13 is completed, the process returns to step S11. Processing (refer to step S16). The acquisition unit 11 acquires the next sentence (step S11), and the analysis unit 12 performs word formation analysis on the sentence (step S12). At this time, if the correction of the division model 20 is performed in the processing of the previous sentence (step S15), the analysis unit 12 performs the word formation analysis using the segmentation model 20 that has been corrected. Thereafter, the correction unit 13 executes the processing in and after step S13. In this way, as long as there is a statement of the processing object, it continues (refer to step S16).

An example of an algorithm representing the operation of the natural language processing system 10 is shown below.

Initialize w ₁ For t=1,2,... Receive instance x _t Predict structure y^ _t based on w _t Receive correct structure y _t If y^ _t ≠y _t ,update w _t+1 =update(w _t ,y _t , +1) w _t+1 =update(w _t ,y^ _t ,-1)

The first line in the upper algorithm means the initialization of the segmentation model 20 (variable w ₁ ), and by this processing, for example, the score of each prime is set to zero. The For loop on line 2 indicates that the processing after line 3 is executed once for each statement. The third line means that the statement x _{t is obtained} , which corresponds to the above-described step S11. The fourth line indicates that the processing of the word formation by the word segmentation model 20(w _t ) based on the time point is equivalent to the above-described step S12. y^ _t is the analysis result. Line 5 Series means that x _t get the statement of the morphological parsing correct answer data y _t. The sixth line means that if the analysis result y^ _t differs from the correct answer data y _t , the segmentation model 20 is updated (corrected). The seventh line indicates that the correct answer data y _{t is} regarded as a positive example, and the eighth line indicates that the analysis result y^ _t containing the error is regarded as a negative example. The processing of the seventh and eighth lines corresponds to the above-described step S15.

Next, the natural language processing program P1 required to implement the natural language processing system 10 will be described using FIG.

The natural language processing program P1 includes a main module P10, an acquisition module P11, an analysis module P12, and a correction module P13.

The main module P10, the part of the system to control the word formation analysis and its related processing. By performing the functions of the acquisition module P11, the analysis module P12, and the correction module P13, respectively, the acquisition unit of the above and the above is recorded. 11. The functions of the analysis unit 12 and the correction unit 13 are the same.

The natural language processing program P1 can be provided, for example, by being fixedly recorded on a tangible recording medium such as a CD-ROM, a DVD-ROM, or a semiconductor memory. Moreover, the natural language processing program P1 can also be superimposed on the carrier's data signal mode and provided through the communication network.

As described above, the natural language processing system according to one aspect of the present invention includes an analysis unit that performs word formation analysis on a sentence using a segmentation model obtained by mechanical learning using one or more pieces of training data. A label indicating at least a word of a word is set for each of the divided elements obtained by dividing the one sentence, wherein the analysis unit includes: a score indicating an output quality of the divided element and the label And a fraction of the mobility of the combination of the two labels corresponding to the two consecutive divided elements; and the correction unit is a label indicating the analysis result obtained by the analysis unit, and a label indicating the correct solution of a sentence. Correctly answer the data, compare it, and set the score of the output prime and the score of the kinetics associated with the label of the correct answer corresponding to the incorrectly answered label to be higher than the current value, and will be associated with the label of the incorrect solution. The score of the output prime and the score of the kinetics are set lower than the current value, and the analysis part is analyzed in the word formation of the next sentence. The segmentation model used is corrected.

The natural language processing method according to one aspect of the present invention is a natural language processing method executed by a natural language processing system including a processor, which includes an analysis step and uses one or more of them. Segmentation model of mechanical learning of training materials, for a statement Performing word formation analysis, and setting a label indicating at least the word of the word for each divided element obtained by dividing the one sentence, wherein the analyzing step is: the dividing model includes: indicating the correspondence between the divided element and the label a score of the output prime, and a fraction of the kinomism indicating a combination of two labels corresponding to the two consecutive divided elements; and a correction step of labeling the result of the analysis obtained in the analysis step, and indicating a statement Correctly answer the correct answer of the label, compare it, and set the score of the output prime and the score of the kinetics associated with the label of the correct answer corresponding to the incorrectly answered label to be higher than the current value, which will be incorrect. The scores of the output primes and the scores of the kinetics associated with the label of the solution are set lower than the current value, and the segmentation model used in the parsing of the next sentence in the parsing step is corrected.

The natural language processing program according to one aspect of the present invention is an analysis unit that performs word formation on a sentence by using a segmentation model obtained by mechanical learning using one or more training materials. In the analysis, a label indicating at least a word of a word is set for each of the divided elements obtained by dividing the one sentence, wherein the analysis unit includes: an output quality indicating that the divided element and the label correspond to each other a score, and a fraction of the kinomism indicating a combination of two labels corresponding to two consecutive divided elements; and a correction unit, which is a label indicating an analysis result obtained by the analysis unit, and a label indicating a correct answer of a sentence Correctly answer the data, compare it, and set the score of the output prime and the score of the kinetics associated with the label of the correct answer corresponding to the incorrectly answered label to be higher than the current value, which will be related to the label of the incorrect solution. The scores of the related output primes and the scores of the kinetics are set lower than the current value, and the segmentation model used by the analysis unit in the word formation analysis of the next sentence is corrected.

In this aspect, each time a sentence is parsed, the parsing result is compared with the correct answer data, and the segmentation model is corrected according to their differences. By correcting the segmentation model for each sentence as described above, the time required for the correction of the segmentation model when processing the complex sentence is suppressed to a substantially linear growth degree with the number of sentences, so that the segmentation of the word formation can be performed. The model is automatically corrected for a certain period of time (in other words, within a predictable time frame).

Moreover, by setting the prime score of the label on the correct answer to a high value and lowering the score of the label of the incorrect answer, the accuracy of the word formation of the next sentence can be further improved.

In the natural language processing system described in the other aspect, the divided element system may also be a character. By using the knowledge of the word unit (output quality and mobilization) to process each character, it is possible to perform word formation analysis without using a word dictionary that is generally large. Moreover, instead of using the knowledge of the word, the knowledge of the word unit is used to correct the segmentation model for each sentence, so even if the next statement and any of the statements analyzed so far differ in domain or nature, The next statement is to perform high-precision word formation analysis. That is, the natural language processing system described in one aspect of the present invention is adaptable to statements in an unknown domain or sentences having an unknown property.

In the natural language processing system described in the other side, lose The scores of the generative score and the kinomism have a range of values, and a score is set for each score, and the correction unit determines the score when the score is set high or low based on the variance of each score. The amount of change. By adopting this method, the scores of each prime can be converged as soon as possible.

The present invention has been described in detail based on the embodiments thereof. However, the present invention is not limited to the above embodiment. The present invention can be variously modified without departing from the gist of the invention.

In general, the number of primes contained in the segmentation model 20 increases with the number of processed words, so in a language such as Japanese or Chinese, the size of the segmentation model 20 is very large. The memory capacity required to segment model 20 can also become very large. Therefore, it is also possible to introduce the technique of Feature Hashing, and each element is quantified by a hash function. In particular, the effect of digitizing characters and strings indicating one of the prime parts is high. On the other hand, the kinetics system does not contribute much to the compression of the capacity of the segmentation model 20 even if it is hashed, but may cause the processing speed to be slow. Therefore, it is also possible to hash only the output properties without mashing the kinetic properties. In addition, regarding the hash function, it is also possible to use only one class, or to use a different hash function between the output prime and the migrating element.

In this case, the segmentation model 20 is a piece of data that is expressed in terms of numerical values for each character. For example, "this (hon)" is converted to a value of 34," (wo)" This text is converted to a value of 4788. By this numericalization, a set of bounded primes can be formed. In addition, with this prime hashing, it is possible that plural characters or strings are assigned the same value, but the probability that a frequently occurring character or string is assigned the same value to each other is very low, so the collision system is Can be ignored.

That is, in the natural language processing system described in other aspects, the segmentation model may contain output factors that are quantized by a hash function. By manipulating the text as a numerical value, the memory capacity required to segment the memory of the model can be saved.

The analysis unit 12 may perform the word formation analysis without using a relatively low score (ignoring such a prime) and using a relatively high score. As a method of neglecting the relatively low-graded prime, for example, Forward-Backward Splitting (FOBOS) and Feature Quantization can be cited.

FOBOS is a technique of compressing a score to zero by normalization (for example, L1 normalization). By using FOBOS, it is possible to ignore the primes whose scores are below the specified value (for example, a prime with a score of 0, or a prime with a score close to 0).

The quantification of the prime is to multiply the value below the decimal point by 10 ⁿ (n is a natural number of 1 or more) and integerize the score of the prime. For example, if the score of "0.123456789" is multiplied by 1000 for integerization, the score is "123". By quantifying the score, the memory capacity necessary to memorize the score in this paper can be saved. Moreover, by this method, it is possible to ignore the prime of the score below the predetermined value (for example, the prime of the fraction after the integer is 0, or the prime of the score close to 0). For example, suppose the scores of a certain prime Fa and Fb are 0.0512 and 0.0003, respectively. When these scores are multiplied by 1000, the primes Fa and Fb become 51 and 0, respectively. In this case, the analysis unit 12 performs the word formation analysis without using the prime Fb.

The normalization or quantization process is performed, for example, by the correction unit 13, other functional elements in the natural language processing system 10, or a computer system other than the natural language processing system 10. When the normalization or quantization process is executed by the correction unit 13, the correction unit 13 associates one set of sentences (for example, a plurality of sentences) into the natural language processing system 10 to construct a segmentation model. After 20 corrections, the normalization or quantization process is performed once.

In other words, in the natural language processing system described in the other aspect, the analysis unit may perform the word formation analysis without using the prime value whose score becomes equal to or lower than the predetermined value by normalization or quantization. By not using a relatively low-level prime (for example, a prime with a score of 0 by normalization or quantization, or a prime of which the score is close to 0), the amount of data of the segmentation model can be suppressed, and the time for word formation can be shortened. and many more.

In the above embodiment, the analysis unit 12 divides the sentence into characters and sets a label for each character. However, the divided element may be a word instead of a word. Along with this, the analysis unit can perform the word formation analysis by using a segmentation model and a word dictionary that are not characters but a score of the prime of the word.

As described above, the natural language processing system of the present invention is applicable to word formation analysis in any language.

10‧‧‧Natural Language Processing System

11‧‧‧Acquisition Department

12‧‧‧ Analysis Department

13‧‧‧Amendment

20‧‧‧ segmentation model

Claims (7)

A natural language processing system includes an analysis unit that performs a word formation analysis on a sentence by using a segmentation model obtained by mechanical learning using one or more pieces of training data, so that the one sentence is not a word dictionary. And dividing into one or more words, and dividing each character, and setting a label indicating at least the word of the word for each of the characters obtained by the division, wherein the analysis unit includes: a score indicating the pheromone of the combination of the output prime character corresponding to the character and the preceding label, and the combination of the two preceding labels corresponding to the two consecutive characters; and the correction unit, which is represented by the analysis result obtained by the preceding analysis unit The correct answer to the label and the label indicating the correct answer to the previous statement is compared, and the score of the pre-recorded output and the score of the pre-migration element associated with the label of the correct answer corresponding to the incorrectly answered label are set. Higher than the current value, the score of the pre-recorded output that is associated with the label of the incorrect answer and the pre-record The fractional shift primality set to be lower than the present value, the morphological analytical statement used before segmentation model in mind, the correction will be referred to parse the front lower portion. The natural language processing system according to claim 1, wherein the pre-segmentation model includes a pre-recorded output property that is quantized by a hash function. The natural language processing system as recited in claim 1 or 2, wherein the scores of the pre-recorded output primes and the pre-recorded kinetics scores are respectively The range of values has a variation number for each score; the pre-correction unit determines the amount of change in the score when the score is increased or decreased based on the variance of each score. The natural language processing system according to claim 1 or 2, wherein the pre-characteristic analysis unit performs pre-construction word analysis without using a pre-recording property in which the pre-recorded score becomes equal to or less than a predetermined value by normalization or quantization. The natural language processing system according to claim 3, wherein the pre-description analysis unit performs the pre-construction word analysis without using the pre-recording property in which the pre-record score becomes equal to or less than the predetermined value by normalization or quantization. A natural language processing method is a natural language processing method executed by a natural language processing system having a processor, which includes an analysis step using mechanical learning by using one or more training materials. The segmentation model performs a word formation analysis on a sentence, so that the one sentence is divided into one or more words instead of using a word dictionary, and each text is divided, and at least each character obtained by the division is set. a label indicating a word of a word, wherein the step of analyzing is that the pre-recording model includes: a score indicating a correspondence between the character and the pre-recorded label, and a combination of two pre-recorded labels corresponding to the two consecutive characters. The fraction of kinetics; and the correction step, which compares the label shown in the analysis result obtained in the pre-analysis step with the correct answer data of the label indicating the correct answer of the previous statement, and the label is incorrectly solved. Correspondingly correct The score of the pre-recorded output prime and the score of the pre-recorded kinetics of the label associated with the solution are set higher than the current value, and the score of the pre-recorded output prime and the score of the pre-migration element associated with the label of the incorrect answer are set to be lower than the score. The current value is corrected by the pre-command segmentation model used in the parsing of the next sentence in the pre-analysis step. A natural language processing program product that enables a computer to function as an analysis unit that performs word formation analysis on a sentence by using a segmentation model obtained by mechanical learning using one or more training materials to The sentence is divided into one or more words instead of using a word dictionary, and is divided into each character, and a label indicating at least the word of the word is set for each character obtained by the division, wherein the analysis unit is The pre-segmentation model includes: a score indicating the output quality of the corresponding character of the character and the pre-recorded label, and a fraction of the kinetics indicating a combination of the two preceding-labels corresponding to the two consecutive characters; and the correction unit, which is obtained from the pre-reviewing section. The correct answer to the label shown in the analysis result and the label indicating the correct answer to the previous statement, and the score of the pre-recorded output prime and the pre-record associated with the label of the correct answer corresponding to the incorrectly answered label. The kinesin score is set higher than the current value and will be related to the label of the incorrect answer. Score Score before the primality of mind and migration of primality set to be lower than the present value, the morphological analytical statement used before segmentation model in mind, the correction will be referred to parse the front lower portion.