JP2007034871A - Character input apparatus and character input apparatus program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a document input apparatus and its program capable of presenting character strings a user wants to input, such as words and expressions, at a higher order of prediction candidates and enabling the user to select his or her desired words and the like from the prediction candidates without performing complicated operation. <P>SOLUTION: A context dictionary generation part 202 generates a context dictionary on the basis of the documents stored in a document set to output it to a context dictionary set 203. An input part 204 decides whether or not the character strings are inputted by the user. A context dictionary selection part 205 selects an optimal context dictionary from the context dictionary set 203 according to content or a context of the document being inputted. A character string conversion processing part 206 predicts a character string the user wants to input on the basis of the input character string from the user to extract the prediction candidates from the selected context dictionary. An output part 207 displays the prediction candidates of the character strings the user wants to input on a display 104. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a character input device and a character input device program for creating a document by inputting characters using a word processor function of a personal computer.

  In recent years, opportunities for document creation using a word processor function of a personal computer or the like have increased. For example, in addition to the conventional opportunity for creating documents such as business documents and papers, electronic medical records and the like have been newly introduced in medical institutions such as hospitals. Therefore, opportunities for creating documents by doctors and nurses are increasing.

  Under such circumstances, it is desired to reduce a burden related to information input of a user, such that even a person who is not skilled in keyboard operation can input information accurately with a few key touches.

  As one solution to such a request, when a “reading” at the beginning of a character string is input, it is used as a keyword, and a character string that the user thinks is desired is retrieved from a dictionary and presented. A method for reducing the burden has been proposed.

  However, in this method, for example, a word string starting from “reading” is simply extracted and presented from words or expressions registered in a simple dictionary. As the number of expressions and the like increase, a large number of prediction candidates are presented, and there is a problem that the operation of searching for and specifying a desired candidate from among them becomes complicated.

One method for solving such a problem is described in Patent Document 1 below. In this patent document 1, when creating a mail by a mail function of a mobile phone or performing a character input by another application, a plurality of dictionaries are provided and used by the mobile phone in order to reduce the burden of the user's character input. The method of switching the dictionary to be used according to the attribute of the application or character input position (for example, the address entry field in the case of mail creation) (the name dictionary has priority when filling in the mail address entry field) Proposed. However, with this method, for example, when a character is input on a single application by a word processor function of a personal computer and a document is created, the above problem cannot be solved yet.
JP 2001-325252 A

  Therefore, according to the present invention, when a document is created using the word processor function of a personal computer or the like, when the user inputs a character string, the user desires to input according to the contents of the document being created. Provided is a document input device that can present a character string such as a word or expression on the top of a prediction candidate, and as a result, allows a user to select a desired word or the like from the prediction candidate without complicated operations, and a program thereof For the purpose.

  The document input device according to the present invention is a character input device that performs character input by converting input information input by a user into a character string, and is one or more used when converting the input information into a character string. The context dictionary selection means selects a context dictionary to be used from the one or more context dictionaries based on the contents of the document being input. Then, the character string conversion means converts the input information into a character string based on the selected context dictionary.

  If there is only one context dictionary, the context dictionary selecting means inevitably selects such one context dictionary.

  According to the present invention, the context dictionary selection means selects a context dictionary to be used based on the contents of a character string that is confirmed to be input from among documents in which characters are being input. Therefore, since a context dictionary suitable for creating a document during character input is selected, the user can easily input a desired character string, and the user's burden of character input can be reduced.

  Furthermore, since an appropriate context dictionary is automatically selected from the contents of the document being input, the user does not need to perform an operation of selecting the context dictionary appropriate for the input document.

  In the document input device according to the present invention, the context dictionary generation unit generates one or more context dictionaries based on a document set including one or more documents.

  According to the present invention, the context dictionary generation unit generates one or more context dictionaries used when converting input information from the user into a character string based on the document set. Therefore, at the time of introduction of the document input device, it is not necessary to prepare a plurality of context dictionaries independently in advance. In addition, since the context dictionary is generated based on the document set created and stored so far, the context dictionary suitable for the characteristics of the document frequently created by the user is generated. Therefore, by using such a context dictionary, the user can easily input a desired character string, and the user's burden of character input can be further reduced.

  In the character input device according to the present invention, the character string conversion means extracts one or more character string candidates that the user wants to input based on the input information from the selected context dictionary and displays the characters. Column candidate display means is provided.

  According to the present invention, the user can select a desired character string from the character string candidates displayed by the character string candidate display means. Therefore, the user can easily input a desired character string, and the burden of character input on the user can be further reduced.

  In the document input device according to the present invention, the context dictionary generation unit generates one or two or more clusters by performing a clustering process on the document set, and each document is converted into the one or two or more. A document classification means for classifying into one of the clusters, a first language feature extraction means for extracting linguistic features reflecting the contents of each document from one or more documents belonging to each cluster, And a context dictionary output means for outputting a context dictionary for each cluster based on the characteristics.

  In the document input device according to the present invention, the context dictionary selecting unit includes a second linguistic feature extracting unit that extracts a linguistic feature reflecting the content of the document being input, and an extracted linguistic feature. A similarity calculation means for calculating the similarity between each feature and the linguistic feature of each context dictionary corresponding to each cluster, and the context dictionary of the cluster with the highest calculated similarity should be extracted and used Context dictionary extraction means for outputting as a context dictionary.

  According to the two aspects of the present invention, the document set is appropriately clustered according to the characteristics of the document, and a context dictionary suitable for each cluster is generated. When a user creates a document during character input, a context dictionary suitable for character input to the document is selected from the characteristics of the document. Therefore, the user can easily input a desired character string by the context dictionary selected in this way.

  In the character input device according to the present invention, the character string candidate display means extracts and displays character string candidates from the selected context dictionary according to the linguistic characteristics of the context dictionary.

  According to the present invention, the character string candidate display means extracts and displays character string candidates in accordance with the linguistic characteristics of the context dictionary, for example, the appearance frequency of the character string stored in the context dictionary. With this configuration, the character string desired by the user is displayed at the top of the character string candidates. Therefore, the user can easily select and input a desired character string, thereby further reducing the user's character input burden.

  In the document input device according to the present invention, in the context dictionary generation unit, the context dictionary update unit regenerates the context dictionary at a predetermined timing.

  According to the present invention, since the context dictionary is updated every predetermined period, the optimum context dictionary can be used.

  In addition, when the number of documents having new contents increases, a context dictionary reflecting the contents is generated. Therefore, for example, even when a document of a field or type that was not initially planned is created, it can be easily handled.

  A document input device program according to the present invention is a program for a character input device that gives a computer the function of each means described in any one of claims 1 to 5.

  According to the present invention, when a document is created using the word processor function of a personal computer or the like, when the user inputs characters, the user desires input according to the contents of the document being created. To provide a document input device and a program thereof that can present a character string such as a word or expression on the top of a prediction candidate, and as a result, allows a user to select a desired word or the like from the prediction candidate without complicated operations. Can do.

 The significance or effect of the present invention will become more apparent from the following description of embodiments.

  However, the following embodiment is merely one embodiment of the present invention, and the meaning of the term of the present invention or each constituent element is not limited to that described in the following embodiment. Absent.

Example 1
DESCRIPTION OF THE PREFERRED EMBODIMENTS Hereinafter, embodiments of the present invention applied to input prediction processing of an electronic medical record system used in medical institutions such as hospitals and clinics will be described with reference to the drawings.

  FIG. 1 shows a configuration diagram of an embodiment of a character input device according to the present invention.

  As shown in FIG. 1, the character input device 100 includes a storage device 101, a CPU 102, an input device 103, and a display device 104.

  The storage device 101 stores a program for executing the functions of the character input device 100 described later, and various tables and electronic files of documents (hereinafter simply referred to as documents) that are referred to when the functions are executed. Etc. are stored. The CPU 102 executes processing according to the character input device program stored in the storage device 101. The input device 103 is used for inputting information such as a character string necessary for the user to execute the character input device program. The display device 104 displays information such as a character string input by the user and the execution result of the character input device program by the CPU 102.

  Therefore, in the character input device 100 of the present invention, when information such as a character string is input from the input device 103 by the user, the CPU 102 executes the character input device program stored in the storage device 101 and displays the execution result. A document that is displayed on the display device 104 and created as a result of execution is stored in the storage device 101.

  FIG. 2 shows a functional block diagram constituting the document input device 100. The functional block diagram is a diagram in which the document input device 100 is classified for each function and expressed for each block for realizing each function.

  In FIG. 2, a character input device 100 includes a document set 201 composed of one or more documents created so far, and a context dictionary generation unit 202, 1 or 2 that generates a context dictionary described later based on the document set 201. A context dictionary that determines and selects which context dictionary to use: a context dictionary set 203 including two or more context dictionaries, an input unit 204 that acquires an input character string input from the input device 103, and a context dictionary set 203 A selection unit 205, a character string conversion unit 206 that predicts a character string desired by a user from an input character string with reference to the selected context dictionary, and displays or confirms a predicted user desired character string candidate on the display device 104 The output unit 207 stores the input character string as a document in the document set 201.

  In the present embodiment, it is assumed that the document set 201 is composed of electronic files (hereinafter referred to as medical record documents) of a plurality of patient charts created by a medical institution such as a hospital or clinic.

  The context dictionary generation unit 202 generates a context dictionary, which will be described later, based on a plurality of medical record documents.

  Specifically, for example, it is assumed that N (N = 1, 2, 3,...) Medical record documents initially exist in the document set 201. The context dictionary generation unit 202 first classifies these N medical record documents into, for example, C clusters (C ≦ N, C = 1, 2, 3,...). As a result, a cluster corresponding to the contents of the chart is formed.

  An existing document classification technique can be applied to the process of classifying into clusters, that is, the clustering process. For example, each document can be classified into clusters by generating document vectors after morphological analysis of each document and applying a clustering algorithm such as k-means method or Ward method to these document vectors.

  FIG. 3 shows a state in which N medical record documents are classified into C clusters for each disease group such as heart disease, digestive organ disease, respiratory disease, and the like according to the contents. Here, C, which is the number of clusters, may be set in advance, or may be automatically determined by the document dictionary generation unit 202 according to the contents of the medical record document.

  In fact, as described above, clusters are not strictly formed for each disease group, but a classification result is obtained in which most of the medical record documents belonging to each cluster have related contents. If there is no problem.

Next, the context dictionary generation unit 202 extracts a linguistic feature reflecting the contents of the document from the medical record document belonging to each cluster, and generates a later-described context dictionary for each cluster.
Here, the linguistic feature reflecting the document content is, for example,
(1) Appearance frequency or appearance probability of each word in each cluster (2) Co-occurrence frequency or co-occurrence probability between words in each cluster (3) Appearance frequency or appearance probability of each character in each cluster (4) In each cluster The number of co-occurrence between characters or the co-occurrence probability.

  In the present embodiment, the appearance frequency of each word in each cluster is used as a linguistic feature reflecting the document content.

In order to extract linguistic features, the context dictionary generation unit 202 first performs processing (morpheme analysis) for dividing each medical record document into words (provided that morpheme analysis is performed during cluster classification). Does not perform morphological analysis again.) next,
In order to obtain the appearance frequency of each word in the medical record document belonging to each cluster, first, referring to the morphological analysis result of each medical record document, the words that appear in the medical record document and the number of appearances thereof are counted. Then, for each of the C clusters, the number of appearances of words appearing in the medical record document belonging to each cluster is totaled for each word. Thereby, a word appearance frequency list for each cluster is obtained.

  For example, out of the C clusters in FIG. 3, a cluster mainly including a heart disease medical record document is referred to as cluster 1, and a cluster mainly including a digestive organ disease medical record document is referred to as cluster 2.

4 and 5 respectively show word appearance frequency lists for cluster 1 and cluster 2 generated by the above-described procedure. However, in this case, functional words such as particles, auxiliary verbs, and conjunctions that are considered to hardly reflect the contents of the medical record document are excluded from the list, and it is considered that the contents of the medical record document are most reflected in order to efficiently reduce the dictionary size. Only for nouns. As shown in FIG. 4 and FIG. 5, each list includes a “No” column indicating a word serial number, a “notation” column indicating a word notation, a “read” column indicating a word reading, It consists of a “part of speech” column indicating the part of speech and an “appearance frequency” column indicating the appearance frequency of the word. Each list is composed of only words whose appearance frequency is equal to or higher than a predetermined threshold (threshold 2 or higher in FIGS. 4 and 5), and are arranged in descending order of appearance frequency. As a result, a list of n ₁ words such as arrhythmia and electrocardiogram is obtained in cluster 1 as shown in FIG. 4, and a list of n ₂ words such as gastric mucosa and inflammation is obtained in cluster 2 as shown in FIG. .

  Similarly, the context dictionary generation unit 202 generates a word appearance frequency list for the remaining clusters.

The context dictionary generation unit 202 outputs the C word appearance frequency list generated as described above as a context dictionary corresponding to each cluster,
Saved in the context dictionary set 203.

  Next, input prediction processing by the context dictionary selection unit 205 and the character string conversion processing unit 206 will be described below.

  Note that the input prediction process is to select the context dictionary corresponding to the input confirmed character string of the medical record document being created, the user recognizes the character string being input, predicts the input desired character string, and is selected. This is a series of processes in which prediction candidates are extracted from the context dictionary and displayed on the display device 104.

  FIG. 6 shows an input screen of a medical record document displayed on the display device 104 during examination of a patient who has visited the hospital complaining of symptoms of abdominal pain and vomiting.

  According to FIG. 6, the user has a stomachache more than last night in the [main complaint / current medical history] field on the input screen of the medical record document. In the “history: H8 stomach ulcer” field, enter “Occupation: Company employee drinking history 25 years” in the column of “Lifestyle information”, and after confirming these inputs, the user will enter “Acute” At the time of inputting “suspected gastritis” from the keyboard (not shown) of the input device 103 to “Kyuse”, prediction candidates “Acute gastritis”, “Acute peritonitis”, “Acute appendicitis”, “Acute colon” It shows a state where six words “flame”, “acute hepatitis”, and “acute pancreatitis” are displayed.

Whenever the input character string from the user is confirmed, the context dictionary selection unit 205 performs morphological analysis on the character string that has already been confirmed, and extracts the noun and the number of appearances. In the case illustrated in FIG. 6, the context dictionary selection unit 205 selects “last night”, “abdominal pain” (twice), “this morning”, “breakfast”, from a character string input and confirmed before the unconfirmed character string “Kyuse”. “Vomiting”, “Past history”, “Gastric ulcer”, “Occupation”, “Company employee” and “Drinking history” are extracted. This is equivalent to generating a vector (hereinafter referred to as a document vector) having a weight of the appearance frequency of a word as follows from a medical record document during character input.

(Last night, abdominal pain, this morning, breakfast, vomiting, medical history, stomach ulcer, occupation, company employee, drinking history)
= (1, 2, 1, 1, 1, 1, 1, 1, 1, 1)
Next, the context dictionary selection unit 205 displays the extracted word and the number of times of the word stored in the “notation” column of the context dictionary corresponding to each cluster and the appearance frequency stored in the “appearance frequency” column. To match. Specifically, a cluster of document vectors having the appearance frequency of each word stored in the “occurrence frequency” column of each context dictionary as the weight of each corresponding word stored in the “notation” column Considering it as a representative document vector, the cosine similarity S between the document vector of the medical record document during character input and the document vector of each cluster is obtained. That is, as shown in the following equation (Equation 1), the appearance frequencies of words appearing in both are multiplied and added, and this is divided by the product of norms of two vectors.

d: Document vector representing the document currently being input

Di: Document vector representing the i-th cluster

(d∩Di): A set of words that appear in both d and Di

W (d, w): Weight of word w in vector d (appearance frequency)

W (Di, w): Weight of word w in vector Di (appearance frequency)

The context dictionary selection unit 205 calculates the similarity S with the C context dictionaries, extracts the context dictionary having the maximum similarity S, and selects the context dictionary having the most similar tendency of the appearing words. Select as context dictionary. In the case of FIG. 6, the cluster 2 mainly composed of the medical record document of the digestive system disease is selected as the most similar cluster.

  The selection of the context dictionary can be performed at any time within the range that can be processed by a computer such as a PC to be used whenever a character string is newly entered and confirmed. However, when the input confirmed character string is extremely small, such as immediately after creation of a new document, it is difficult to select the context dictionary with high accuracy. Therefore, if the input confirmed character string is less than a certain number of characters (or a certain number of words), the context dictionary may not be selected.

  The character string conversion processing unit 206 recognizes a reading character string during character input, and predicts a character string that the user intends to input. Then, prediction candidates are extracted from the context dictionary selected by the context dictionary determination unit 205 and displayed on the output unit 207.

  In the case of FIG. 6, the character string conversion processing unit 206 reads from the context dictionary of cluster 2 selected by the context dictionary selecting unit 205 the word “acute gastritis” whose reading matches the forward reading character string “Kyuse” being input. “Acute peritonitis”, “acute appendicitis”... Are extracted and displayed on the chart input screen via the output unit 207 as prediction candidates in descending order of appearance frequency.

  Details of a general input prediction method are disclosed in, for example, Japanese Patent Laid-Open Nos. 7-334499 and 09-274613.

  FIG. 7 shows a flowchart of the program of the character input device executed by the CPU 102.

  In step S 301, the context dictionary generation unit 202 generates a context dictionary based on the document chart stored in the document set as described above, and outputs the generated context dictionary to the context dictionary set 203.

  In step S302, the input unit 204 determines whether or not a character string has been input from the user. If there is a character input, the process proceeds to step S303. Otherwise, the process waits until there is a character input.

  In step S <b> 303, the context dictionary selection unit 205 selects an optimum context dictionary from the context dictionary set 203 according to the content and context of the document being input.

  In step S304, the character string conversion processing unit 206 predicts the user's desired character string based on the input character string from the user, and extracts prediction candidates from the selected context dictionary.

  In step S <b> 305, the output unit 207 causes the display device 104 to display a prediction candidate for the user-desired character string.

  As described above, the character input device 100 of the present invention can present words suitable for the context at the top as compared to the method of extracting words from a single context dictionary and constructing prediction candidates in order of frequency, The operability is greatly improved. For example, since there are more than 200 words whose readings are prefixed with “Kyusei” only for medical terms that start with “Acute”, the word that the user wants to input rarely appears at the top, and the user is in the prediction candidate list. Scroll down to find the desired word, or enter longer readings to narrow down the candidates.

  Also, in order to automatically determine the optimum context dictionary, a plurality of dictionaries are prepared for each field in advance, and compared with a method in which a dictionary is selectively used by the user indicating the field of the input document. This instruction operation becomes unnecessary.

  In addition, it is almost necessary to generate a context dictionary dynamically based on an existing medical record document so that all possible context dictionaries are prepared in the system. There is no problem of having to keep a context dictionary that does not.

  In the above example, the word appearance frequency TF (Term Frequency) is used as the word weighting scale in the context dictionary and the document vector. However, the word weighting scale is not limited to this, and various existing scales such as words Various existing measures such as IDF (Inverse Document Frequency) representing the degree of occurrence in a small number of specific documents and the product of TF and IDF can be used.

  Further, in addition to (or instead of) the appearance frequency in units of words, a co-occurrence probability between words or a connection probability can be employed. The co-occurrence probability between words is a numerical value representing the number of times two (or more) words appear together in the same document. Generally, the higher the co-occurrence probability is, Relevance is high. The concatenation probability is a probability that two (or three or more) words appear consecutively, and the larger this value is, the higher the possibility that these words appear consecutively. By using these statistics together to construct a context dictionary, it is possible to improve the accuracy of more suitable context dictionary determination, and to perform more accurate prediction processing that reflects the ease of connection between words It becomes possible.

Furthermore, when performing the prediction process using the selected context dictionary, in some cases, instead of using only one context dictionary having the maximum similarity in the selection of the context dictionary, the top k contexts (k <C) may be used. Dictionaries may be used in combination (for example, if there are a plurality of context dictionaries having a similarity with a medical record document during character input exceeding a threshold value, these are used in combination). Moreover, in order to prevent omission of a prediction candidate, you may comprise the prediction candidate by adding the word extracted from the context dictionary of all the clusters to the word extracted from the adapted context dictionary.

(Example 2)
A case where the context dictionary stored in the context dictionary set 203 is updated in the electronic medical record system described in the first embodiment will be described below.

  In Example 1, N medical record documents that already existed at the time of introduction of the electronic medical record system are classified into C clusters to deal with disease groups such as heart disease, digestive tract disease, respiratory disease, and blood disease. C context dictionaries were generated. Then, for example, it is assumed that the number of diabetic patients has increased due to the adoption of doctors specializing in diabetes treatment in this medical institution, and the tendency of the contents of medical record documents has changed. In such a case, the context dictionary at the time of initial introduction cannot perform sufficiently accurate prediction processing according to the contents and context of the document being created. Therefore, by re-clustering the medical record documents, C ′ clusters reflecting the trend of the latest medical record document are obtained, and C ′ context dictionaries corresponding to these clusters are reconstructed. If the medical record documents of diabetic patients are insignificant compared to other disease groups, a cluster composed mainly of diabetic medical record documents is obtained as one of the C ′ clusters, and the corresponding context is obtained. A dictionary is generated. As a result, the accuracy of the prediction process is improved even in the case of input to a medical record document of a diabetic patient who could not properly perform the prediction process in the context dictionary at the time of initial introduction.

In actual operation, in addition to the case described above, the context dictionary may be rebuilt at predetermined timings as described below.
(1) After a predetermined time (days) has elapsed since the previous context dictionary generation (2) After a user has accessed (created, edited, viewed, received, etc.) a predetermined number of chart documents Alternatively, it is possible to add only the updated medical record document to the classified cluster and generate the context dictionary at the first timing, and to perform all from the clustering at the second timing. This is due to the fact that clustering is computationally expensive.

  Furthermore, when a document sufficient for generating a context dictionary cannot be secured, a similar chart document may be acquired from the network using a clustered chart document as a query and used for generating the context dictionary.

  Furthermore, since clustering a large number of documents generally requires a large calculation cost, for example, clustering of medical record documents is performed only once a month, and once a week, any of the clusters in which a medical record document updated in the week was generated last time It is also possible to adopt a configuration in which only the context dictionary is regenerated to reflect the contents of the newly added chart document.

Furthermore, in this embodiment, since the context dictionary is dynamically reconfigured based on the medical record document, even when a context dictionary having new contents that cannot be assumed at the time of introduction is required (for example, Documents such as SARS (Severe Acute Respiratory Syndrome) several years ago, cases where new infectious diseases are prevalent, cases such as hepatitis C, where the possibility of infection due to phytotoxicity is revealed, and the number of examinations increases rapidly) It is possible to easily generate a new necessary context dictionary based on new contents and types of medical record documents stored in the set.

(Example 3)
A case where the present invention is used for context processing (post-character recognition processing) in a handwritten character input device will be described.

  In this case, the input character string input by the user is a handwritten character string, and the input prediction processing by the context dictionary selecting unit 205 and the character string conversion processing unit 206 recognizes the handwritten character string by the character recognition unit, and further results thereof. This corresponds to a process of outputting an appropriate character string candidate based on language knowledge from the obtained character recognition candidate group.

  Post-processing methods in character recognition include “JP 2000-90201 bigram dictionary and its miniaturization method and handwritten character recognition processing method and device (Tokyo Univ. Of Agriculture and Technology)” and “JP 9-282420 character pattern recognition device (Hitachi). As is described in “)” and the like, a method for obtaining an optimal combination of candidate characters by using a connection probability between characters is generally used, but an outline will be described below.

Character string evaluation value L (C | X for character string candidate C = {C ₁ , C ₂ , ..., C _N } for handwritten input character string X = {X ₁ , X ₂ , ..., X _N } ) Is obtained by the following equation (Equation 2), and the character string candidate that maximizes this is set as the first candidate.

Here, X ₁ , X ₂ , ..., X _N are individual handwritten character patterns constituting the handwritten input character string X, and C ₁ , C ₂ , ..., C _N constitute a certain character string candidate C S (X _i , C _i ) is the similarity (recognition score) of the recognition candidate character C _i obtained from the handwritten character pattern X _i , and P (C _{i + 1} | C _i ) is the character The connection probability from C _i to the letter C _{i + 1} , w represents a weight constant determined by experiment.

  FIG. 8 is a diagram showing a process for recognizing a character string input by handwriting based on the connection probability by the character recognition means.

  In FIG. 8, four characters “book”, “day”, “ha”, and “sunny” are input by handwriting, and “book”, “old”, “cloth”, and “date” as recognition candidate characters for “book”. Three candidate characters such as “eyes”, “day”, and “month” are obtained. The post-character recognition process here is a process for obtaining an optimal candidate character string from 81 candidate character strings obtained by combining these 3 characters × 4 candidate characters based on the connection probability between characters.

  In a general character recognition post-processing method, a character connection probability (character bigram) used for post-processing is obtained in advance from a large amount of text and used.

  On the other hand, in the present invention, as in the first embodiment, existing N documents are classified into C clusters at the time of introduction. Then, C character bigram dictionaries are generated from the documents belonging to each cluster, and this is used as a context dictionary corresponding to each cluster. Then, the context dictionary selection unit 205 selects an optimum context dictionary from the contents of the document being input. Subsequently, the character string conversion processing unit 206 performs post-character recognition processing using the selected context dictionary. As a result, a highly accurate post-processing result suitable for the context can be obtained rather than using a single general-purpose character bigram dictionary.

  Further, as in the second embodiment, since the context dictionary is reconfigured at a predetermined timing, it is possible to perform appropriate post-character recognition processing even if the tendency of the input document changes.

  In this case, in the context dictionary selection process in the context dictionary selection unit 205, instead of using the word unit appearance frequency as in the first embodiment, a character bigram is extracted from a document in which characters are being input, It is also possible to select an optimal dictionary from the degree of coincidence with the clustered character bigram dictionary. However, it is also possible to obtain the word appearance frequency of each cluster separately and select the context dictionary based on the word appearance tendency as in the first embodiment.

  The present invention is not limited to the above embodiment, and various modifications can be made within the technical scope described in the claims.

It is a figure which shows the structure of the document input device which is one of embodiment of this invention. It is a figure which shows the functional block of the document input device which is one Embodiment of this invention. It is a figure which shows the cluster classification | category of the medical record document by the context dictionary production | generation part 202. FIG. It is a figure which shows the example of the word appearance frequency list produced | generated by the context dictionary production | generation part 202. FIG. It is a figure which shows the example of the word appearance frequency list produced | generated by the context dictionary production | generation part 202. FIG. It is a figure which shows the input screen to a medical chart document. It is a figure which shows the flowchart of a character input device program. It is a figure which shows the recognition process of the handwritten character string by a character recognition means.

Explanation of symbols

DESCRIPTION OF SYMBOLS 100 Character input device 201 Document set 202 Context dictionary generation part 203 Context dictionary set 204 Input part 205 Context dictionary selection part 206 Character string conversion process part 207 Output part

Claims (8)

In a character input device for inputting characters by converting input information into a character string,
One or more context dictionaries to use when converting input information to a string;
A context dictionary selecting means for selecting a context dictionary to be used from the one or more context dictionaries based on the contents of a document during character input;
A character string conversion means for converting input information into a character string based on the selected context dictionary;
A character input device comprising: A context dictionary generating means for generating the one or more context dictionaries based on a document set including one or more documents;
The character input device according to claim 1, further comprising: The character string conversion means is a character string candidate display means for extracting and displaying one or more character string candidates that the user desires to input based on the input information from the selected context dictionary,
The character input device according to claim 1, wherein the character input device is provided. The context dictionary generating means includes:
A document classification means for generating one or more clusters by performing a clustering process on the document set, and classifying each document into one of the one or more clusters;
First linguistic feature extracting means for extracting linguistic features reflecting the contents of each document from one or more documents belonging to each cluster;
A context dictionary output means for outputting a context dictionary for each cluster based on the extracted features;
The character input device according to claim 2, further comprising: The context dictionary selecting means includes
Second linguistic feature extracting means for extracting linguistic features reflecting the contents of the document during character input;
Similarity calculation means for calculating the similarity between the extracted linguistic features and the linguistic features of each context dictionary corresponding to each cluster;
A context dictionary extracting means for extracting a context dictionary of a cluster whose calculated similarity is equal to or greater than a predetermined value and outputting as a context dictionary to be used;
The character input device according to claim 4, further comprising: 6. The character string according to claim 4, wherein the character string candidate display means extracts and displays the character string candidate from the selected context dictionary according to a linguistic feature of the context dictionary. Input device. 7. The character input device according to claim 2, wherein the context dictionary generation unit includes a context dictionary update unit that regenerates the one or more context dictionaries at a predetermined timing. . The character input device program which provides the function of each means as described in any one of Claim 1 thru | or 7 to a computer.

Images