JP2019204214A - Learning device, learning method, program and estimation device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a learning device, a learning method, a program, and an estimation device capable of identifying whether or not a word is appropriate regardless of the position in a sentence. SOLUTION: A learning device 1 includes an acquisition unit 151 that acquires a sentence, a division unit 152 that divides the acquired sentence into a plurality of elements that are characters or character strings in a predetermined unit, and appears next to each element. A first learning unit 153 that learns the elements in the order of the acquired sentences; an order conversion unit 154 that rearranges the arrangement order of the plurality of elements in reverse order from the beginning to the end of the sentence; And a second learning unit 155 that learns the element that appears next to each element in the order of the sentences. [Selection diagram] Fig. 15

Description

  The present invention relates to a learning device, a learning method, a program, and an estimation device.

  There is a technique for estimating a text following an input text, estimating an error in the text, and the like by learning a co-occurrence relationship of words appearing in a sentence. For example, in Patent Document 1, a recurrent neural network (Recurrent Neural Network, hereinafter referred to as “RNN”) in which the order in which each word appears in a sentence and the relationship such as the dependency of each word is learned is constructed. A learning device or the like is disclosed that estimates the text following the input text using the constructed RNN and outputs a sentence having an appropriate structure.

JP 2018-45656 Gazette

  However, in the invention according to Patent Document 1, since the order of words appearing in a sentence is learned in order from the beginning of the sentence, for example, a word positioned at the beginning of the sentence, a word following a punctuation mark, and the like are appropriately estimated. It is difficult.

  In one aspect, an object of the present invention is to provide a learning device or the like that can identify whether a word is an appropriate word regardless of the position in the sentence.

  In one aspect, the learning device includes an acquisition unit that acquires a sentence, a division unit that divides the acquired sentence into a plurality of elements that are characters or character strings of a predetermined unit, and the element that appears next to each element A first learning unit that learns elements in the order of the acquired sentences; an order conversion unit that rearranges the arrangement order of the plurality of elements from the beginning to the end of the sentences; and A second learning unit that learns the element that appears next to each element in order.

  In one aspect, it is possible to identify whether or not the word is appropriate regardless of the position in the sentence.

It is a block diagram which shows the structural example of a learning apparatus. It is explanatory drawing for demonstrating a language model. It is explanatory drawing for demonstrating a text learning process. It is explanatory drawing of the error estimation process based on the arrangement | sequence order of the text only in the forward direction. It is explanatory drawing for demonstrating the text learning process of reverse order. It is explanatory drawing of an error estimation process. It is a flowchart which shows an example of the process sequence of a text learning process. It is a flowchart which shows an example of the process sequence of an error estimation process. 12 is an explanatory diagram of an error estimation process according to Embodiment 2. FIG. 10 is a flowchart illustrating an example of a processing procedure of error estimation processing according to the second embodiment. 10 is a block diagram illustrating a configuration example of a learning device according to Embodiment 3. FIG. It is explanatory drawing which shows an example of the record layout of a phrase list. It is explanatory drawing for demonstrating a subword learning process. It is a flowchart which shows an example of the process sequence of a subword learning process. It is a functional block diagram which shows operation | movement of the learning apparatus of the form mentioned above.

Hereinafter, the present invention will be described in detail with reference to the drawings illustrating embodiments thereof.
(Embodiment 1)
FIG. 1 is a block diagram illustrating a configuration example of the learning device 1. In the present embodiment, a learning device 1 that generates a language model 141 that learns each word and phrase appearing in a sentence according to the arrangement order of the sentences, and estimates an error location in the sentence using the generated language model 141. Will be described.

  The learning device 1 is an information processing device capable of various information processing, and is, for example, a server device, a personal computer, a multifunction terminal, or the like. The learning device 1 divides a learning sentence into words (elements) of a predetermined unit, and learns the divided words according to the arrangement order of the sentences. Thereby, the learning apparatus 1 generates a language model 141 that predicts the appearance probability (occurrence probability) of the next phrase from the previous phrase.

  FIG. 2 is an explanatory diagram for explaining the language model 141. The language model 141 models the probability that a natural language sentence is generated. In the general language model 141, the occurrence probability of a word that appears following the one or more words is calculated from one or more words that appear in order from the top in the sentence, and the next word is estimated. In the example of FIG. 2, the occurrence probability of the next appearing phrase is calculated based on the phrases “I”, “ha”, “school”, and “ni” that appear sequentially from the top of the sentence. As shown in FIG. 2, in a general language model 141, a word such as “I” is input to calculate the occurrence probability of the next word candidate “go”, “is”, etc., and the candidate with the highest occurrence probability Predict “go” as the next phrase. Thus, in the general language model 141, the next word / phrase is predicted from the word / phrase up to immediately before.

  In the present embodiment, the learning device 1 further replaces the beginning and end of the sentence, learns the order of words when the words in the sentence are reversed, and uses one or more words that appear in order from the end of the sentence. The language model 141 that predicts the immediately preceding word is generated. That is, the learning device 1 generates two language models 141 by learning the arrangement order of words in the forward direction in the sentence and the arrangement order in the reverse direction. The learning device 1 uses the two language models 141 to estimate an error location that is estimated to be a typographical error from a sentence that is an error estimation target.

Returning to FIG. 1, the description will be continued. The learning device 1 includes a control unit 11, a main storage unit 12, a communication unit 13, and an auxiliary storage unit 14.
The control unit 11 includes an arithmetic processing unit such as one or a plurality of CPUs (Central Processing Units), MPUs (Micro-Processing Units), and GPUs (Graphics Processing Units), and stores the program P stored in the auxiliary storage unit 14. Various information processing, control processing, and the like related to the learning device 1 are performed by reading and executing. The main storage unit 12 is a temporary storage area such as a static random access memory (SRAM), a dynamic random access memory (DRAM), or a flash memory, and temporarily stores data necessary for the control unit 11 to execute arithmetic processing. Remember. The communication unit 13 includes a processing circuit for performing processing related to communication, and transmits and receives information to and from the outside.

  The auxiliary storage unit 14 is a large-capacity memory, a hard disk, or the like, and stores a program P and other data necessary for the control unit 11 to execute processing. The auxiliary storage unit 14 stores the language model 141 generated from the learning text as described above.

  Note that the auxiliary storage unit 14 may be an external storage device connected to the learning device 1. Further, the learning apparatus 1 may be a multi-computer composed of a plurality of computers, or may be a virtual machine constructed virtually by software.

  Further, in the present embodiment, the learning device 1 is not limited to the above configuration, and may include, for example, a reading unit that reads information stored in a portable storage medium, an input unit that receives operation input, a display unit that displays an image, and the like. Good.

  FIG. 3 is an explanatory diagram for explaining the text learning process. FIG. 3 illustrates a state in which the learning device 1 learns each phrase arranged in the forward direction in the learning sentence, and constructs a language model 141 based on LSTM (Long Short-Term Memory) which is a kind of RNN. Yes. Below, the outline | summary of the process which the learning apparatus 1 performs is demonstrated.

  The learning device 1 acquires a plurality of learning sentences from the outside via, for example, the communication unit 13. The learning device 1 first performs natural language processing such as morphological analysis on the acquired learning sentence, and divides the sentence into words (elements) that are characters or character strings of a predetermined unit. The division unit is, for example, a unit such as a word or a phrase, but is not particularly limited. For example, the learning device 1 may store in advance a dictionary (not shown) storing a plurality of words and divide the sentence according to the words stored in the dictionary.

  Note that the learning apparatus 1 may divide a sentence in units other than words, phrases, and the like as in Embodiment 3 described later. That is, the learning device 1 only needs to be able to divide a learning sentence into elements that are characters or character strings in a predetermined unit, and the element as a division unit is not limited to a unit such as a word or a phrase.

  The learning device 1 inputs each divided phrase to the input layer related to the RNN and performs machine learning. FIG. 3 conceptually illustrates the configuration of the RNN. As shown in FIG. 3, the RNN has an input layer, an intermediate layer (hidden layer), and an output layer. The input layer has a plurality of neurons that respectively accept inputs of words that appear in order from the beginning of the sentence. The output layer has a plurality of neurons corresponding to each neuron of the input layer and estimating and outputting words following the word inputted to each neuron. The intermediate layer has a plurality of neurons for calculating the output value (phrase) in each neuron in the output layer with respect to the input value (phrase) in each neuron in the input layer. Each neuron in the intermediate layer uses the operation result in the intermediate layer for the past input value (shown by a right-pointing arrow in FIG. 3) to perform the operation on the next input value, so that the next word to the next word Is estimated.

  The configuration of the RNN illustrated in FIG. 3 is an example, and the present embodiment is not limited to this. For example, the intermediate layer is not limited to one layer, and may be two or more layers. Further, the number of neurons in the input layer and the output layer is not limited to the same number. For example, the number of outputs is at least as good as the input.

  In the present embodiment, the learning apparatus 1 performs learning according to an RNN algorithm. For example, learning according to another algorithm such as other deep learning, N-gram model, SVM (Support Vector Machine), Bayesian network, or decision tree. And the language model 141 may be generated.

  The learning device 1 inputs each phrase of the learning sentence to each neuron in the input layer in accordance with the arrangement order in the sentence, and obtains an output value from each neuron in the output layer. In the example of FIG. 3, the learning device 1 inputs each phrase “yesterday”, “ha”, “train”, “ni”,... Of the learning sentence to each neuron of the corresponding input layer according to the order in the sentence. To do. The learning device 1 performs an operation on each neuron in the output layer through the intermediate layer, and calculates the occurrence probability of a word that appears at an arbitrary position (order) in the sentence based on the word that appears immediately before, Estimate the next word. In the example of FIG. 3, the learning device 1 performs the estimation by calculating the occurrence probability of the word that appears second based on the first word “yesterday”. The learning device 1 calculates and estimates the occurrence probability of the third phrase based on the first and second phrases “yesterday” and “ha”. Similarly, the learning device 1 estimates each word / phrase.

  The learning device 1 compares the estimated phrase with the actual phrase (correct answer value), adjusts the parameters of each neuron so that the output value from each neuron in the output layer approximates the correct answer value, and constructs the RNN. For example, the learning device 1 adjusts the weights of neurons in each network layer so that the phrase estimated as the phrase following “yesterday” becomes the actual phrase “ha”. Thereby, the learning apparatus 1 generates a language model 141 in which a word / phrase that appears next to the word / phrase is learned from a word / phrase that appears immediately before in the forward direction of the learning text.

  FIG. 4 is an explanatory diagram of an error estimation process based on the text arrangement order only in the forward direction. FIG. 4 conceptually illustrates a case in which the occurrence probability of each word / phrase is calculated based only on the forward language model 141 in order to detect an error location in the sentence.

  The learning device 1 acquires, for example, a target sentence that is a target for detection (estimation) of an error portion via the communication unit 13. The learning device 1 divides the target sentence into words of a predetermined unit, as in learning. Then, the learning device 1 refers to the generated language model 141 and calculates the occurrence probability of each word based on the words that appear until immediately before. In the example shown in FIG. 4, when the learning device 1 calculates the occurrence probability of the phrase “yesterday” that appears third, the first and second phrases “I” and “ha” are calculated using the above-described RNN. From the above, the occurrence probability of “Yesterday” is calculated. Then, the learning device 1 determines whether or not the word / phrase is an error part based on the calculated occurrence probability. For example, the learning device 1 compares the occurrence probability with a predetermined threshold, and determines that it is an error location when the occurrence probability is equal to or lower than the threshold, that is, when the occurrence probability is low.

  However, when an error part is estimated from the learning result only in the forward direction, there is a possibility that the error part cannot be appropriately estimated depending on the position of the corresponding phrase in the sentence. For example, in the case of a Japanese sentence, various words can appear at the beginning of a sentence, immediately after a punctuation mark, or immediately after a particle, so that the occurrence probability of many words is low. For example, although the sentence shown in FIG. 4 is a correct sentence, the word “I” located at the head and “School” located immediately after the reading point have low occurrence probabilities. In this case, the learning apparatus 1 may estimate the correct words “I” and “School” as erroneous parts.

  Therefore, in the present embodiment, the learning apparatus 1 further learns sentences when the words are rearranged in the reverse order to generate the language model 141, and combines the learning results in the forward direction and the reverse direction to identify the error part. The above problem is solved by estimating.

  FIG. 5 is an explanatory diagram for explaining a reverse-order text learning process. FIG. 5 illustrates a state in which the words in the learning sentence are rearranged in the reverse order and the words are learned in order from the end of the sentence.

  For example, after completing the learning process in the forward direction, the learning device 1 replaces the learning sentence from the beginning to the end, and rearranges the learning sentences in the reverse order starting with the last word / phrase. That is, as shown in FIG. 5, the learning apparatus 1 uses “yesterday”, “ha”, “train”, “ni”,. Rearrange in the order of “ni”, “Kyoto”, “ha”, “yesterday”. The learning device 1 inputs the rearranged words to the input layer of the reverse order learning RNN having the same configuration as the RNN shown in FIG. Then, the learning device 1 constructs an RNN that estimates a word that appears immediately before the one or more words / phrases based on one or more words / phrases that appear in reverse order from the end of the sentence. That is, the learning device 1 constructs an RNN that estimates a phrase at an arbitrary position in a sentence from a subsequent phrase that follows the phrase.

FIG. 6 is an explanatory diagram of the error estimation process. FIG. 6 illustrates how the occurrence probability of each word is calculated based on the forward and backward language models 141.
As described with reference to FIG. 4, the learning device 1 divides the target sentence for error estimation into words of a predetermined unit. Then, the learning device 1 inputs each word / phrase of the target sentence to the input layer of the RNN in each of the forward direction and the backward direction, and calculates the occurrence probability for each of the forward direction and the backward direction.

  When performing calculation for the forward direction, the learning device 1 inputs each word / phrase to the input layer of the RNN learned for the forward direction, and from one or more words / phrases appearing in order from the top, to the one or more words / phrases Calculate the occurrence probability of the following phrase. That is, the learning device 1 calculates the occurrence probability of the next word / phrase from the previous words / phrases.

  When performing calculation in the reverse direction, the learning device 1 first rearranges the words and phrases of the target sentence in the reverse order. Then, the learning device 1 inputs each rearranged word / phrase in the reverse order to the RNN learned about the reverse order, and the one or more words appearing in order from the end, that is, from the subsequent words / phrases, The probability of occurrence of a word that appears immediately before the word is calculated. Thus, as shown in FIG. 6, the words “I” and “School” whose occurrence probability is low in the forward direction have a high occurrence probability in the reverse direction.

  The learning device 1 determines whether or not each word is erroneous based on the occurrence probability of each word calculated for each of the forward direction and the backward direction. For example, the learning device 1 estimates an error location when the occurrence probabilities in the forward direction and the backward direction are both equal to or less than a predetermined threshold. The learning device 1 outputs the estimation result by outputting a target sentence in which the display mode of the error part is different from the display mode of the other part, for example, by color-displaying the words estimated as the error part.

  Further, when the learning device 1 detects an error location, the learning device 1 may estimate and present (output) a correct word / phrase instead of the word / phrase estimated to be an error. That is, the learning device 1 outputs a word / phrase that is a correction candidate for the word / phrase estimated to be an error part. As described with reference to FIG. 2, the learning device 1 can predict a word / phrase having a high occurrence probability from the preceding and following words / phrases by using the language model 141. For example, the learning apparatus 1 refers to the language model 141 (learning result) used for error estimation, and outputs a word / phrase having the highest occurrence probability as a correction candidate at the position (order) of the word / phrase estimated as an error location. do it. In this case, the learning apparatus 1 may estimate a correction candidate from both directions using the two language models 141 in the forward direction and the reverse direction, and the correction candidate is determined only from the language model 141 in either the forward direction or the reverse direction. It may be estimated.

  As described above, the learning device 1 can appropriately estimate the error location in the sentence by using the language model 141 in which each word / phrase is learned in the order of arrangement in the forward direction and the reverse direction.

  Note that the learning apparatus 1 may use the language model 141 as a Bi-directional RNN to simultaneously learn the order of arrangement in the forward direction and the reverse direction, and integrate the two RNNs into one. However, since a unidirectional RNN can be tuned more easily, it is preferable to generate two RNNs as described above.

FIG. 7 is a flowchart illustrating an example of a processing procedure of the text learning process. Based on FIG. 7, the processing content of the text learning process which the learning apparatus 1 performs is demonstrated.
For example, the control unit 11 of the learning device 1 acquires a plurality of learning sentences via the communication unit 13 (step S11). The control unit 11 divides the acquired sentence into a plurality of words (elements) that are characters or character strings of a predetermined unit (step S12). For example, the control unit 11 divides a sentence in units such as words and phrases, but the unit of the phrase (element) to be divided is not particularly limited.

  The control unit 11 learns a phrase that appears next to each phrase in the order from the beginning to the end of the sentence, and generates a forward language model 141 (step S13). For example, as described above, the control unit 11 inputs one or a plurality of words that appear in order from the head of the sentence, and constructs an RNN that outputs the occurrence probability of the word that follows the one or more words.

  The control unit 11 rearranges the order of the words in the reverse order in which the sentence is replaced from the beginning to the end (step S14). The control unit 11 learns a phrase that appears next to each phrase in the order of the rearranged sentences, and generates a language model 141 in the reverse direction (step S15). For example, the control unit 11 receives, as an input, one or more words that appear in order from the end of the sentence, and constructs an RNN that outputs the occurrence probability of the word that is located immediately before the one or more words. The control unit 11 ends the series of processes.

FIG. 8 is a flowchart illustrating an example of a processing procedure of error estimation processing. Based on FIG. 8, the processing content of the error estimation processing which the learning apparatus 1 performs is demonstrated.
The control unit 11 of the learning device 1 acquires a target sentence that is an error estimation target via, for example, the communication unit 13 (step S31). The control unit 11 divides the acquired target sentence into a plurality of words (elements) that are characters or character strings of a predetermined unit (step S32).

  The control unit 11 refers to the language model 141 generated for the forward direction, and calculates the occurrence probability of each word that appears in the target sentence (step S33). That is, the control unit 11 inputs one or a plurality of words that appear in order from the top of the target sentence to the input layer of the RNN constructed as the language model 141, and the occurrence probability of the words following the one or more words Is calculated.

  The control unit 11 rearranges the order of the words in the target sentence in the reverse order (step S34). Then, the control unit 11 refers to the language model 141 generated in the reverse direction and calculates the occurrence probability of each word that appears in the target sentence (step S35). That is, the control unit 11 inputs one or a plurality of words that appear in order from the end of the target sentence to the input layer of the RNN constructed in the reverse direction, and generates a word that appears immediately before the one or more words. Probability is calculated.

  The control unit 11 estimates an error location in the target sentence based on the occurrence probability of each phrase calculated for each of the forward direction and the backward direction (step S36). For example, the control unit 11 estimates a word / phrase whose occurrence probabilities in the forward direction and the backward direction are both equal to or less than a threshold value as an error location. The control unit 11 outputs an estimation result (step S37). For example, the control unit 11 outputs a target sentence in which the display mode of the word / phrase estimated to be an erroneous part is different from the display mode of other words / phrases by color-coded display or the like. In addition to the step S37, the control unit 11 may estimate and output a correction candidate for the word / phrase estimated to be an error part. For example, the control unit 11 refers to the forward and / or backward language model 141 and outputs, as a correction candidate, a phrase having the highest occurrence probability at the position (order) of the phrase estimated as an error location. The control unit 11 ends the series of processes.

  In the above description, it is assumed that the same learning apparatus 1 performs text learning and error estimation. However, the present embodiment is not limited to this. The learning device 1 may perform only text learning, install a learning result (language model 141) in another device, and configure as an estimation device, and the estimation device may perform error estimation.

  As described above, according to the first embodiment, not only the forward direction but also the reverse order is learned with respect to the arrangement order of the words in the sentence. Thereby, the learning apparatus 1 generates the language model 141 that can identify whether the word is an appropriate word regardless of the position in the sentence.

  Further, according to the first embodiment, by generating an RNN as the language model 141, it is possible to generate a language model 141 with higher accuracy than other learning algorithms such as an N-gram model. .

  Further, according to the first embodiment, by calculating the occurrence probability of each word / phrase in the target sentence using the forward and backward language models 141, it is possible to appropriately point out an error location in the target sentence. it can.

  Further, according to the first embodiment, it is also possible to output a correction candidate for a word / phrase estimated to be an error part using the language model 141. Thereby, the convenience can be improved.

(Embodiment 2)
In the present embodiment, a mode in which the character type and / or the number of characters of each word is used as a criterion for error estimation in addition to the occurrence probability of each word. In addition, about the content which overlaps with Embodiment 1, the same code | symbol is attached | subjected and description is abbreviate | omitted.
FIG. 9 is an explanatory diagram of error estimation processing according to the second embodiment. FIG. 9 illustrates the occurrence probability calculated from the forward and backward language models 141 for each word included in the target sentence for error estimation, as well as the character type and the number of characters of each word.

  In the present embodiment, the learning device 1 determines that when the occurrence probability in each of the forward and backward directions of the corresponding phrase is equal to or less than the threshold and the character type and / or the number of characters of the corresponding phrase satisfy a specific condition, Presumed to be an error location. For example, if the corresponding phrase is a character or a character string with only a hiragana character as a condition regarding the character type, the learning device 1 determines that it is an error part because there is a high possibility that it has not been converted. In addition, for example, the learning device 1 determines that an error location is present because there is a high possibility of erroneous input when the corresponding phrase is equal to or less than a predetermined number of characters as a condition regarding the number of characters.

  In addition to the character type and the number of characters, the learning device 1 may add other conditions to the error estimation determination condition. For example, the learning device 1 may estimate a location where a word having an occurrence probability equal to or less than a threshold continues for a predetermined number of times as an error. Thereby, for example, when a character string to be converted into Kanji is mistakenly unconverted, it is possible to appropriately estimate a portion where words or phrases having an occurrence probability equal to or less than a threshold value appear continuously.

FIG. 10 is a flowchart illustrating an example of a processing procedure of error estimation processing according to the second embodiment. Based on FIG. 10, the processing content of the error estimation processing according to the present embodiment will be described.
After executing the process of step S35, the control unit 11 of the learning device 1 estimates an error location based on the character type and / or the number of characters of each word in addition to the occurrence probability of each word calculated for each of the forward direction and the backward direction. (Step S201). For example, the control unit 11 determines whether the occurrence probability is equal to or less than a threshold value in both the forward direction and the backward direction, and whether the corresponding phrase is a specific character type or the number of characters. For example, the control unit 11 determines whether or not the corresponding phrase is hiragana as a condition regarding the character type. For example, the control unit 11 determines whether or not the corresponding phrase is equal to or less than a predetermined number of characters as a condition regarding the number of characters. In addition to the above conditions, the control unit 11 may further estimate a place where a predetermined number or more of words having occurrence probabilities relating to the forward direction and the backward direction are equal to or less than a threshold value as an error place. The control part 11 transfers a process to step S37.

  As described above, according to the second embodiment, not only the occurrence probability calculated using the language model 141 but also the character type and / or the number of characters of the corresponding phrase are used as the determination criteria, so that the error part can be estimated more appropriately. can do.

  Further, according to the second embodiment, it is possible to estimate an error more appropriately by estimating a place where words having an occurrence probability equal to or less than a threshold value continue as an error place.

(Embodiment 3)
In the present embodiment, by using a unit called a subword as a unit for dividing a sentence, even when dealing with a sentence having many unknown words that are not recorded in a general dictionary, such as a technical book, it is appropriately handled. The form which can be done is described.
FIG. 11 is a block diagram illustrating a configuration example of the learning device 1 according to the third embodiment. The auxiliary storage unit 14 of the learning device 1 according to the present embodiment stores a phrase list 142. The phrase list 142 is a database that stores phrases called subwords extracted from a predetermined sample document. In the present embodiment, a mode is described in which a medical certificate is assumed as a document (sentence) handled by the learning device 1 and error estimation is performed by dividing the certificate in subword units.

  FIG. 12 is an explanatory diagram illustrating an example of a record layout of the phrase list 142. The phrase list 142 has a phrase string and a score string. The phrase string stores phrases (characters or character strings) that are subwords extracted (divided) from the sample document. The score column stores a score (parameter) calculated based on the frequency of occurrence of each subword in the sample document in association with the subword.

  FIG. 13 is an explanatory diagram for explaining the subword learning process. The learning apparatus 1 acquires a large number of predetermined sample documents, extracts subwords from each sample document, and generates a phrase list 142. FIG. 13 illustrates a process in which subwords are extracted from a sample document and registered in the phrase list 142.

  A subword (partial word) is a unit of a word (character or character string) in which a sentence is divided according to the frequency of appearance in the sentence, unlike normal segmentation. Generally, `` word '' used as the minimum constituent unit of a sentence is a unit that is minimized from the viewpoint of grammar, etc. meaning a character or character string in a sentence, but a subword is not a unit based on meaning, grammar, etc. It is a unit that is minimized according to the frequency used in the text. According to the concept of subwords, low-frequency words (character strings) are grouped in units shorter than the words themselves, such as characters and partial character strings that constitute the words.

  Hereinafter, processing for learning subwords from a sample document will be described. In the present embodiment, the learning device 1 extracts subwords from a sample document using a BPE (Byte Pair Encoding) technique.

  First, the learning apparatus 1 divides the sample document in character units. In the example shown at the top of FIG. 4, the learning device 1 divides the sentence “no enlarged lymph nodes are recognized” into the characters “tumor”, “large”, “shi”, “ta”. doing.

  As shown in the second row of FIG. 13, the learning device 1 registers all the divided characters as subwords in the phrase list 142. In this case, the learning device 1 calculates the subword score (parameter) based on the appearance frequency of each subword (character) in the sample document, and registers the calculated score in the score column of the phrase list 142. The score is calculated by normalizing the appearance frequency, for example. In the example of FIG. 4, since “large” appears more frequently in the text than “tumor”, the score of “large” is 0.05, which is higher than the score of “tumor” 0.01. Yes.

  Note that values such as weights used for normalization are appropriately changed depending on circumstances. In the following description, a series of processing is performed based on a score (parameter) whose appearance frequency is normalized. For example, the learning device 1 may use an appearance frequency itself that is not normalized as a score. It suffices if a series of processing can be performed based on parameters according to the above.

  Next, the learning apparatus 1 registers a two-character character string obtained by connecting adjacent characters in the sample document in the phrase list 142 according to the appearance frequency of the character string. Specifically, the learning device 1 registers two characters having the highest score in the sentence in the phrase list 142.

  For example, the learning device 1 takes out character strings by two characters from the beginning to the end of the sentence, and calculates the score of each character string. In the example of FIG. 13, the learning device 1 first calculates a “swelling” score, then calculates a “great” score, a “do” score, and so on. The learning device 1 registers the character string having the highest score among the character strings composed of two characters in the phrase list 142. In the example shown in the third row of FIG. 13, since the score of the character string “transfer” is the highest, the learning apparatus 1 registers “transfer” as a subword in the phrase list 142. The learning device 1 also registers the score calculated based on the appearance frequency of the character string in the phrase list 142.

  Subsequently, the learning device 1 searches the sample document again and registers the two characters having the highest score in the phrase list 142. In this case, the learning device 1 considers the character string already registered in the phrase list 142 as a subword as one character and searches for a new subword. In the above example, since “transition” has already been registered in the word list 142, the character string of “transition” is regarded as one character. As described above, the learning device 1 compresses adjacent characters into one piece of information (character string) using the BPE method. When the learning device 1 calculates a score for a portion extending over “metastasis”, the “bone metastasis” obtained by connecting “bone” positioned before “metastasis” and “metastasis” is positioned after “metastasis”. A score is calculated by regarding “Transition”, which is a concatenation of “Ga” and “Transition”, as a two-character string.

  As described above, the learning device 1 connects two subwords (characters or character strings) registered in the phrase list 142 to specify a new subword (character string), and the new subword is determined according to the appearance frequency. Add to the phrase list 142. In the example shown in the fourth row of FIG. 13, since the score of the character string “bone metastasis” obtained by concatenating one character subword “bone” and two characters subword “transfer” is the highest, the learning apparatus 1 determines that the character string “ “Bone metastasis” is newly added to the word list 142 as a subword.

  Similarly, the learning apparatus 1 connects two adjacent subwords (characters or character strings) in the sample document, and registers a character string composed of the two subwords in the phrase list 142 according to the appearance frequency. Processing is performed in order. The learning apparatus 1 performs the process on a plurality of sample documents, and repeats the process until the number of subwords registered in the phrase list 142 reaches a predetermined maximum number (for example, 8000 words). Thereby, the learning apparatus 1 generates the phrase list 142 illustrated in FIG. In this way, the learning device 1 learns a character string pattern (subword) that is likely to appear in a sentence.

  In the above description, the learning device 1 registers the character string having the highest score according to the appearance frequency in the sentence in the phrase list 142. For example, the character string having a score equal to or higher than the threshold value is set. May all be registered as subwords. That is, the learning device 1 only needs to be able to register subwords in accordance with the appearance frequency, and the determination criterion related to the appearance frequency is not particularly limited.

  The learning device 1 divides a learning sentence (existing medical certificate in the above example) in units of subwords with reference to the word list 142, and generates a language model 141. For example, the learning device 1 divides the sentence so that the total value of the scores of all the subwords after the division is maximized. Similarly, the learning device 1 divides the target sentence (diagnosis document) to be error-estimated in units of subwords, and estimates the error location using the language model 141.

  As described above, the learning device 1 detects (estimates) an error part by dividing the target sentence by subwords classified according to the appearance frequency instead of a general standard such as meaning and grammar. As a result, it is not necessary to use a manually created dictionary, and even a sentence with many unknown words can be handled.

FIG. 14 is a flowchart illustrating an example of a processing procedure of subword learning processing. Based on FIG. 14, the processing content of the error estimation processing which the learning apparatus 1 performs is demonstrated.
The control unit 11 of the learning device 1 acquires a plurality of sample documents for learning (step S301). The control unit 11 divides the acquired sample document into character units (step S302). The control unit 11 registers all the divided characters as subwords in the phrase list 142 (step S303).

  The control unit 11 sets a score (parameter) according to the appearance frequency of the character string obtained by connecting two adjacent subwords in the sample document among the subwords (characters or character strings) registered in the phrase list 142. ) Is calculated (step S304). For example, the control unit 11 takes out subwords two by two from the beginning to the end of the sentence to form one character string, and sequentially calculates a score based on the appearance frequency of each character string.

  The control unit 11 registers a character string obtained by connecting two subwords as a new subword in the phrase list 142 according to the score calculated in step S304 (step S305). Specifically, the control unit 11 registers the character string having the highest score in the phrase list 142 among all the character strings whose scores are calculated in step S304. In this case, the control unit 11 registers a score based on the appearance frequency calculated in step S304 together with the phrase list 142.

  The control unit 11 determines whether or not a predetermined maximum number of subwords has been registered in the phrase list 142 (step S306). When it is determined that the maximum number of subwords is not registered (S306: NO), the control unit 11 returns the process to step S304. When it is determined that the maximum number of subwords has been registered (S306: YES), the control unit 11 ends the series of processes.

  As described above, according to the third embodiment, as an element (phrase) that divides a sentence, not a word or phrase based on meaning, grammar, or the like, but a subword that is classified according to the appearance frequency is used as a division unit. Use. As a result, it is not necessary to use a manually created dictionary, and even a sentence with many unknown words can be handled. In particular, by dividing in units of subwords, low-frequency words are divided by a short number of characters, so the second embodiment (error estimation based on character type and / or number of characters, error based on continuity of words whose occurrence probability is below a threshold value) A further improvement in accuracy can be expected by combining the estimation) and this embodiment.

(Embodiment 4)
FIG. 15 is a functional block diagram illustrating the operation of the learning device 1 having the above-described form. When the control unit 11 executes the program P, the learning device 1 operates as follows.
The acquisition unit 151 acquires a plurality of sentences. The dividing unit 152 divides each acquired sentence into a plurality of elements that are characters or character strings of a predetermined unit. The first learning unit 153 learns the element that appears next to each element in the order of the acquired sentences. The order conversion unit 154 rearranges the arrangement order of the plurality of elements in reverse order from the beginning to the end of the sentence. The second learning unit 155 learns the elements that appear next to the elements in the order of the rearranged sentences.

  The fourth embodiment is as described above, and the other parts are the same as those of the first to third embodiments. Therefore, the corresponding parts are denoted by the same reference numerals, and detailed description thereof is omitted.

  The embodiment disclosed this time is to be considered as illustrative in all points and not restrictive. The scope of the present invention is defined by the terms of the claims, rather than the meanings described above, and is intended to include any modifications within the scope and meaning equivalent to the terms of the claims.

DESCRIPTION OF SYMBOLS 1 Learning apparatus 11 Control part 12 Main memory part 13 Communication part 14 Auxiliary memory part P Program 141 Language model 142 Phrase list

Claims (9)

An acquisition unit for acquiring sentences;
A dividing unit that divides the acquired sentence into a plurality of elements that are characters or character strings of a predetermined unit;
A first learning unit that learns the elements that appear next to each element in the order of the acquired sentences;
An order conversion unit that rearranges the order of the plurality of elements in reverse order from the beginning to the end of the sentence;
A learning device comprising: a second learning unit that learns the element that appears next to each element in the order of the rearranged sentences. The learning device according to claim 1, wherein the first and second learning units generate a recursive neural network. A document dividing unit that divides each of a plurality of documents into predetermined units of characters or character strings;
A registration unit that registers the character or character string in a phrase list according to the appearance frequency of the character or character string in the document,
The learning device according to claim 1, wherein the dividing unit divides each character or character string registered in the word list. Get multiple sentences,
The acquired sentence is divided into a plurality of elements that are characters or character strings of a predetermined unit,
Learn the elements that appear next to each element in the order of the acquired sentences,
Rearranging the order of the plurality of elements from the beginning to the end of the sentence in reverse order,
A learning method, comprising: causing a computer to execute a process of learning the elements that appear next to the elements in the order of the rearranged sentences. Get multiple sentences,
The acquired sentence is divided into a plurality of elements that are characters or character strings of a predetermined unit,
Learn the elements that appear next to each element in the order of the acquired sentences,
Rearranging the order of the plurality of elements from the beginning to the end of the sentence in reverse order,
A program for causing a computer to execute a process of learning the element that appears next to each element in the order of the rearranged sentences. An acquisition unit for acquiring a target sentence that is an error estimation target;
A dividing unit that divides the acquired target sentence into a plurality of elements that are characters or character strings of a predetermined unit;
Based on the learning sentence, the occurrence probability of each of the elements included in the target sentence by referring to the first language model that has learned the elements next to the elements included in the sentence in the order of the sentences A first calculation unit for calculating
An order conversion unit that rearranges the order of the plurality of elements in the target sentence from the beginning to the end of the target sentence in reverse order;
Based on the learning sentence, the occurrence probability of each element of the rearranged target sentence is calculated by referring to the second language model that has learned the element next to each element in reverse order. A second calculation unit;
An estimation apparatus comprising: an estimation unit that estimates an element as an error location based on occurrence probabilities of the element calculated by the first and second calculation units, respectively. The estimation device according to claim 6, wherein the estimation unit estimates the element as an error location based on the occurrence probability of the element and the character type and / or the number of characters of the element. The estimation apparatus according to claim 6 or 7, wherein the estimation unit estimates a plurality of continuous elements as error locations when a predetermined number or more of the elements whose occurrence probabilities are equal to or less than a threshold value continue. An output unit that outputs, as a correction candidate, an element having the highest occurrence probability with respect to the element estimated as an error location with reference to the first or second language model. The estimation apparatus of any one of -8.

Images