JP2014081710A - Frame undefined handwritten character recognition system or method thereof - Google Patents

Abstract

PROBLEM TO BE SOLVED: To raise predictability and to reduce load of handwriting input even in a frame undefined handwritten character recognition.SOLUTION: A morphological analysis unit 6 divides a first place candidate generated by a character string candidate generation unit 5 in words. A stroke string division unit 10 divides a stroke string to a part which has been finished writing and a part which is unfinished as a word based on data of a stroke string recording unit 4, the morphological analysis unit 6 and a morphological analysis result recording unit 7. A predictive candidate generation unit 11 calculates a predictive candidate based on information from the stroke string division unit 10, a verbal resource unit 8 and an input history recording unit 9. A candidate display unit 12 presents calculated character string candidate group and predictive candidate group calculated by the character string candidate generation unit 5 and the predictive candidate generation unit 11 to a user.

Description

  The present invention relates to a frame-undefined handwritten character recognition method, and more particularly to improvement of prediction accuracy.

  Today, a frame-undefined handwritten character string recognition technique is known as one method of handwritten character string recognition. The frame undefined handwritten character string recognition refers to handwritten character recognition in which an area for handwriting one character is not defined, unlike frame-based handwritten character string recognition in which one character is written in a defined frame. In recognition of a frame-undefined handwritten character string, it is necessary to confirm the input of one character by some method.

  Patent Document 1 discloses a method of automatically dividing a portion to be recognized as a character and a portion in the middle of input in an input stroke sequence. Specifically, the character recognition result for the part in the middle of input is determined to be noise in the prediction calculation, and a prediction candidate is obtained by removing this noise.

JP 2004-258800 A

  The prediction method disclosed in Patent Document 1 has a problem that prediction accuracy is low because a character for a portion in the middle of input is handled as noise.

  For example, when the word “end of summer” is entered, when the user starts writing a thread immediately after “summer”, the thread part is still judged to be noise, so the co-occurrence dictionary is Even if it is used, “end” cannot be presented as a prediction candidate.

  In order to solve such a problem, it is conceivable to predict all input stroke sequences using DP matching. However, it is not realistic to perform DP matching on all words in the dictionary in real time.

  An object of the present invention is to provide a frame-undefined handwritten character determination method or apparatus capable of solving the above problems, having high prediction accuracy, and reducing the load of handwriting input.

  (1) A frame-undefined handwritten character recognition apparatus according to the present invention includes a handwritten character dictionary storage unit that stores a correspondence dictionary of stroke strings and corresponding character strings, and a co-occurring combination of independent word character strings. Dictionary dictionary storage means, confirmed character string storage means for storing a combination of a conversion character string and a stroke string determined by a user for each word, stroke string recording means for storing a stroke string given from a handwritten character input unit, the stroke A morpheme analysis is performed on a character recognition unit that performs character recognition using the handwritten character dictionary storage unit and a recognized character string candidate group obtained by the character recognition unit with respect to the stroke sequence stored in the column storage unit. The result of the morpheme analysis by the morpheme analysis unit is completed as a word by comparing it with the result before one stroke input. A separation means that separates a word into an incomplete part as a word, a prediction candidate calculation means for obtaining a prediction candidate for an incomplete part, and a stroke string of a converted character string stored in the confirmed character string storage means and the Prediction that determines the prediction candidate based on the degree of coincidence with the stroke sequence of the unfinished part and the degree of co-occurrence with the last independent word among the parts completed as words using the co-occurrence dictionary Candidate determining means, and presenting means for presenting prediction candidates predicted by the prediction candidate calculating means.

  In this way, by using the morphological analysis results to determine prediction candidates for incomplete parts as words, it is possible to reduce the number of prediction calculation targets and perform highly accurate prediction.

  (2) In the frame undefined handwritten character recognition device according to the present invention, the prediction candidate determination means selects candidates from those having a high degree of coincidence of stroke sequence histories, and also presents the presentation means in the order of co-occurrence frequencies. Sort and display. Accordingly, candidates can be selected from those having a high degree of coincidence in the stroke sequence history, and rearranged in the order of the co-occurrence frequencies.

  (3) In the frame undefined handwritten character recognition apparatus according to the present invention, for the stroke string given from the handwritten character input unit, a handwritten character dictionary storage means for storing a correspondence dictionary between the stroke string and the corresponding character string is used. The character recognition means for performing character recognition, the recognized character string candidate group obtained by the character recognition means, the morpheme analysis means for performing morpheme analysis, and the morpheme analysis result by the morpheme analysis means are completed as words A separation means for separating a part into an incomplete part as a word, and a prediction candidate calculation means for obtaining a prediction candidate for an incomplete part, the stroke string of the converted character string stored in the confirmed character string storage means and the unfinished part Based on the degree of coincidence with the stroke string of the completed part, the prediction candidate determining means for determining the prediction candidate and the prediction predicted by the prediction candidate calculating means And it includes a presentation means, for presenting a candidate.

  In this way, by using the morphological analysis results to determine prediction candidates for incomplete parts as words, it is possible to reduce the number of prediction calculation targets and perform highly accurate prediction.

  (4) A frame-undefined handwritten character recognition program according to the present invention is a handwritten character recognition program for causing a computer to execute as the following means, and corresponds to a stroke sequence for a stroke sequence given from a handwritten character input unit. Character recognition means for performing character recognition using handwritten character dictionary storage means for storing a correspondence dictionary with character strings, morpheme analysis means for performing morphological analysis on a recognized character string candidate group obtained by the character recognition means, Separation means for separating a completed part as a word and an incomplete part as a word based on a morpheme analysis result by a morpheme analysis means, a prediction candidate calculation means for obtaining a prediction candidate for an incomplete part, wherein the fixed character Based on the degree of coincidence between the stroke string of the converted character string stored in the string storage means and the stroke string of the incomplete part Predicting candidate determining means for determining the prediction candidate, and presenting means for presenting the prediction candidate predicted by the prediction candidate calculating means.

  In this way, by using the morphological analysis results to determine prediction candidates for incomplete parts as words, it is possible to reduce the number of prediction calculation targets and perform highly accurate prediction.

  (5) In the frame undefined handwritten character recognition device or the handwritten character recognition program according to the present invention, the separating means is completed as a word by comparing the morpheme analysis result with the result before one-stroke input. And the unfinished part as a word. Therefore, separation by morphological analysis is possible.

  (6) In the frame-undefined handwritten character recognition device or the handwritten character recognition program according to the present invention, the prediction candidate determination means is from a connected character string dictionary that stores combinations of independent word character strings used in connection with each other. A combination of independent word strings in a connected relationship is read, and a part that is highly connected to the last independent word among the parts completed as words is displayed as a candidate. Therefore, the independent words in the connection relationship can be displayed as candidates.

  (7) In the frame undefined handwritten character recognition apparatus or the handwritten character recognition program according to the present invention, the prediction candidate determining means determines the prediction candidate after narrowing down candidates by the connected character string dictionary. Therefore, the number of stroke history determination processing targets can be reduced.

  (8) In the frame-undefined handwritten character recognition apparatus or the handwritten character recognition program according to the present invention, the prediction candidate determining means sets a candidate having a high degree of connection based on a connected character string dictionary as the candidate to be narrowed down. Accordingly, the prediction candidates are determined for those having a high degree of connectivity. Therefore, the determination target of the prediction candidate can be narrowed down.

  (9) In the frame undefined handwritten character recognition device or the handwritten character recognition program according to the present invention, the fixed character string storage unit stores a combination of the converted character string and the stroke string determined by the user for each word. ing. Therefore, candidates based on the user's confirmed history can be specified.

  (10) In the frame-undefined handwritten character recognition apparatus or the handwritten character recognition program according to the present invention, the prediction candidate determination means determines a candidate when the user actually has a connected history. Therefore, it becomes possible to specify a candidate based on a confirmed history that the user has connected and converted.

  (11) In the frame undefined handwritten character recognition method according to the present invention, the following 1) to 3) are stored in the storage unit of the computer, and 1) the stroke is given to the stroke sequence given from the handwritten character input unit. 2) a combination of independent word strings to be used in combination; 3) a combination of conversion strings and stroke strings determined by the user for each word; Run. When a stroke string is given from the handwritten character input unit, a step of performing character recognition with reference to the correspondence dictionary, a step of performing morpheme analysis on the obtained recognized character string candidate group, and a morpheme analysis by the morpheme analysis unit Based on the result, a step of separating the completed part as a word into an incomplete part as a word, based on the degree of coincidence between the stroke string of the stored conversion character string and the stroke string of the incomplete part , Obtaining a prediction candidate of the unfinished part, and presenting the predicted prediction candidate.

  In this way, by using the morphological analysis results to determine prediction candidates for incomplete parts as words, it is possible to reduce the number of prediction calculation targets and perform highly accurate prediction.

  In the present specification, the term “concatenated use” refers to a case where independent words are used in conjunction with or without accompanying words. In the embodiment, independent words in a co-occurrence relationship are applicable. .

It is a functional block diagram of the handwritten character recognition method apparatus. It is an example of the hardware constitutions at the time of constituting handwritten character recognition method device 1 using CPU. It is an example of the data structure of a co-occurrence dictionary. It is the whole flowchart. It is a figure which shows the history of the time series of the stroke sequence input. It is a detailed flowchart of a prediction candidate production | generation process. It is a figure which shows the conversion history memorize | stored in the input history memory | storage part 26r.

  Embodiments of the present invention will be described below with reference to the drawings.

(1.1 Function block)
FIG. 1 shows a functional block diagram of a handwritten character recognition apparatus 1 according to an embodiment of the present invention. The handwritten character recognition device 1 includes a handwriting input unit 3, a stroke string recording unit 4, a character string candidate generation unit 5, a morpheme analysis unit 6, a morpheme analysis result recording unit 7, a linguistic resource unit 8, an input history recording unit 9, and a stroke. A column division unit 10, a prediction candidate generation unit 11, a candidate display unit 12, and a candidate selection unit 13 are provided.

  A handwritten character pattern is input to the handwriting input unit 3 as a stroke. The stroke sequence recording unit 4 records the input stroke sequence. The character string candidate generation unit 5 generates a character string candidate from the stroke sequence recorded in the stroke sequence recording unit 4. The morpheme analysis unit 6 divides the first candidate generated by the character string candidate generation unit 5 into words. The morpheme analysis result recording unit 7 temporarily holds the morpheme analysis result at the time of the previous stroke input. The linguistic resource unit 8 stores linguistic information such as the co-occurrence frequency of words used as a clue for prediction. The input history recording unit 9 stores the correspondence between the input word string and the stroke string. The stroke sequence dividing unit 10 divides the stroke sequence into a part that has been written as a word and an incomplete part as a word based on the data of the stroke sequence recording unit 4, the morpheme analysis unit 6, and the morpheme analysis result recording unit 7. . The prediction candidate generation unit 11 calculates a prediction candidate based on information from the stroke string division unit 10, the linguistic resource unit 8, and the input history recording unit 9. The candidate display unit 12 presents the character string candidate group / prediction candidate group calculated by the character string candidate generation unit 5 and the prediction candidate generation unit 11 to the user. The candidate selection unit 13 causes the user to select a candidate from the candidate group presented by the candidate display unit.

  The character string candidate generation unit 5 generates candidate character string groups by referring only to the input stroke string and dictionary, and determines whether the end of the stroke string forms a character. Absent. Therefore, a recognition error may be included at the end of each candidate.

  In addition, the character recognition candidate generation unit 5 also outputs the correspondence between the received stroke sequence and each character.

  The morpheme analysis unit 6 divides the given character string into words, and gives parts of speech to each word. Any definition of words and parts of speech may be used, but it is assumed that it is possible to distinguish whether a word is an independent word. It is not assumed that the input is a complete sentence. Therefore, an analysis error may be included at the end of the analysis result.

(1.2 Hardware configuration)
A hardware configuration of the handwritten character recognition apparatus 1 shown in FIG. 1 will be described with reference to FIG. The handwritten character recognition device 1 includes a CPU 23, a memory 27, a flash memory 26, a touch panel display unit 31, a communication unit 37, a speaker 33, a position information acquisition unit 35, and a bus line 29. The CPU 23 controls each unit via the bus line 29 according to each program stored in the flash memory 26.

  The flash memory 26 stores various data in addition to an operating system program 26o (hereinafter abbreviated as OS) and a main program 26p. The program is installed in the flash memory 26 in advance, but may be downloaded via the communication unit 37.

  Data stored in the flash memory 26 will be described. In the dictionary storage unit 26j, in addition to a stroke correspondence dictionary (not shown) that stores correspondence between input stroke sequences and characters for recognizing handwritten characters, morphological analysis for morphological analysis is performed. A dictionary and a co-occurrence dictionary are stored. In this co-occurrence dictionary, as shown in FIG. 3, combinations of preceding words and succeeding words are stored together with their frequencies. In this co-occurrence dictionary, when the preceding word and the succeeding word are used continuously as they are, of course, they are stored as being co-occurring when they are connected by an attached word. For example, “summer” and “cool” include a case where “summer is cool”. Further, the character string whose input has been confirmed in the past has an input history storage unit 26r in which morphological analysis is performed and stroke string correspondence data for each word is stored.

  The area of each character is not limited on the touch panel 36, and when a stroke is input, the CPU 23 performs handwritten character recognition processing according to the main program 26p as described later.

  The RAM 27 stores a stroke string storage unit 27s in which the stroke string received from the touch panel 36 is stored, and a morpheme analysis result storage unit 27k in which the result of morphological analysis of the first character recognition candidate is overwritten each time one stroke is input. have.

(1.3 Explanation of flowchart)
Processing by the program 26p shown in FIG. 2 will be described with reference to FIG. In the following, as shown in FIG. 5, there is a history of inputting the character string “in summer in Koshien” from the touch panel 36, and then predictive conversion in the case of handwriting input as “in summer in Koshien” will be described as an example. To do.

  The handwritten character recognition process can be broadly divided into the following: 1) completion part determination process (steps S101 to S111 in FIG. 4), 2) prediction candidate generation process (step S113), and 3) prediction candidate determination process (step S115). To step S119).

  The CPU 23 empties the stroke string storage unit 27s and the morpheme analysis result storage unit 27k (step S101). The CPU 23 determines whether or not a stroke is received from the touch panel 36 (step S103). When the CPU 23 receives the stroke, the CPU 23 stores it in the stroke string storage unit 27S. For example, when a stroke “horizontal bar” as shown in FIG. 5A is received, data indicating this shape is stored.

  The CPU 23 acquires a recognized character string candidate group using the stroke string correspondence dictionary in the dictionary storage unit 26j (step S107). Here, it is assumed that the recognized character string candidate “one” corresponding to the stroke “horizontal bar” has been acquired. The CPU 23 performs morphological analysis on the first candidate (step S109). In this case, the morphological analysis is performed on the character string “one”.

  The CPU 23 divides and stores the part that has been written as a word and the unfinished part as a word (step S111). In the present embodiment, the character string stored in the previous morphological analysis result storage unit 27k is compared with the character prefix stored in the morpheme analysis result storage unit 27k. Is an incomplete part. Therefore, in this case, since the common prefix does not exist, it is stored as no part that has been written as a word.

  The CPU 23 generates a prediction candidate for the character string to be input using the part that has been written as a word (step S113). In this case, since there is no portion that has been written as a word, a prediction candidate is not generated. CPU23 displays a prediction candidate (step S115). In this case, since no prediction candidate exists, no display is performed.

  The CPU 23 determines whether or not there is a candidate selection (step S117). If there is no candidate selection, the processing from step S103 is repeated. Hereinafter, similarly, it is assumed that the stroke sequence is input until “summer”. In this case, the character string “summer / no” is stored in the morphological analysis result storage unit 27k (see FIG. 5). In FIG. 5, “/” represents a word break.

  The CPU 23 determines whether or not a stroke is received from the touch panel 36 (step S103). When the CPU 23 receives the stroke, the CPU 23 stores it in the stroke string storage unit 27S. Here, the stroke “vertical bar” as shown in FIG. 5A is received, and the data itself indicating this shape is stored.

1) Completed part determination process The CPU 23 acquires a recognized character string candidate group using the stroke string correspondence dictionary in the dictionary storage unit 26j (step S107). Here, it is assumed that the character string candidate “1” is acquired for the stroke “vertical bar”. The CPU 23 performs morphological analysis on the first candidate (step S109). In this case, the morphological analysis is performed on the character string “summer 1”, and it is decomposed into “summer / no / 1”.

  The CPU 23 divides and stores the part that has been written as a word and the unfinished part as a word (step S111). In this case, the character string “summer / no” is stored in the morpheme analysis result storage unit 27k as the previous morpheme analysis result. Therefore, the CPU 23 determines that the character string “Natsu / no” as the common prefix has been written as a word. As a result, as shown in FIGS. 5C and 5D, the character string “summer / no / 1” is stored as the morphological analysis result.

2) Prediction candidate generation process CPU23 produces | generates a prediction candidate about the incomplete part identified by the morphological analysis of step S109 (step S113). The prediction candidate generation process of step S113 is demonstrated using FIG.

  The CPU 23 extracts the word W at the end of the word string of “the portion that has been written as a word” (step S201 in FIG. 6). In this case, the last word “no” is extracted as the word W from the character string “summer / no” which is the part that has been written. The CPU 23 determines whether or not the word W is an independent word (step S203). Whether or not it is an independent word is determined from, for example, the part of speech of the morphological analysis result.

  In this case, since the word W is the word “no”, it is determined that it is not an independent word, and the independent word W ′ that is closest to the left of W is extracted from the word sequence of “the part that has been written as a word”. 'Is a word W (step S205). In this case, the word “summer” is extracted. The CPU 23 extracts the word Wi, which is the closest independent word and located to the right of the word W, from the past input history recorded in the input history storage unit 26r (step S209). Here, it is assumed that there is an input history as shown in FIG. In this case, the words “rest”, “fireworks”, “travel”, “vegetables”, “na”, “sea”, “koshien”, and “hot” are extracted.

  Note that in step S209, the word Wi, which is located to the right of the word W and is the closest independent word, is extracted because the independent word is followed by a plurality of attached words. This is because there may be.

  The CPU 23 refers to the history of the input character string and the stroke string stored in the input history storage unit 26r, and refers to the stroke string (the first part) in the input history for the word Wi and “the word is incomplete. The “part” stroke sequence is geometrically collated, and the degree of matching Yi is calculated for each word extracted in step S207 in FIG. 6 (step S211). In this case, the word “rest” -0, the word “fireworks” -0, the word “travel” -0, the word “vegetable” -0.65, the word “na” -0, the word “sea” -0, the word “ It is assumed that Koshien -0.7 and the word "hot" -0.7 are obtained. In the present embodiment, the DP matching calculation method is used for such geometric matching.

  The CPU 23 calculates the co-occurrence frequency Xi between the word W and the word Wi using the co-occurrence dictionary. Specifically, the co-occurrence frequency Xi is calculated for the word co-occurring with the word “summer” from the co-occurrence dictionary shown in FIG. In this case, the co-occurrence frequency Xi includes the word “rest” -0, the word “fireworks” -0, the word “travel” -0, the word “na” -0, the word “sea” -0, and the word “hot”. -46533/46533, the word "vegetable" -44234/46533, and the word "Koshien" -41267/46533. In this embodiment, normalization is performed with the maximum value of the frequency, but this is arbitrary.

3) Prediction candidate confirmation process CPU23 displays a prediction candidate on the touch panel 36 (step S115). In the present embodiment, candidates are arranged and displayed in order of the co-occurrence frequency Xi for the top n values of the matching degree Yi. In this case, “vegetables”, “koshien”, and “hot” are selected as candidates, and are arranged in the order of “hot”, “vegetables”, and “koshien”.

  The user selects “Koshien”, which is an input request. The CPU 23 determines whether or not a candidate is selected (step S117). Since the selection is made, the result is stored in the input history storage unit 26r (step S119).

  In this way, the user can easily input “Koshien” that he / she wants to input. Note that the prediction accuracy is further improved as the input of the stroke sequence proceeds.

  In the present embodiment, linguistic information such as co-occurrence obtained from a part written as a word and the geometric similarity of stroke sequences obtained from an unfinished part as a word are used in combination. Yes. Thereby, highly accurate input prediction is attained. Therefore, the user can input text by selecting from the prediction candidates only by writing the first few strokes of each word to be input without writing all the text to be input.

  Also, in this case, the timing to start the prediction was once the morphological analysis was performed, and a part that was incomplete but was completed as a word, and stroke input was started for the part that was not completed as a word It is a middle stage. Thereby, the object which determines a stroke sequence can be reduced. On the other hand, in the case of general handwritten character recognition, it is the timing at which stroke input is started, and is different from general kana-kanji conversion in which one character is input as it is from a keyboard or the like.

(2. Other embodiments)
In the present embodiment, in step S111 in FIG. 4, “the part that has been written as a word” is extracted instead of “the part that has been written as a word”. The word or phrase is natural as the granularity, and the most effective when the prediction is correct is the beginning of writing, and 2) Kanji with many strokes often includes parts that are also used as characters, It is difficult to judge. Therefore, the morphological analysis result may not be appropriate as the input of the stroke sequence proceeds. In this case, the determination of “the part that has been written as a word” may not be appropriate.

  For example, in FIG. 5, the character recognition result for this part changes every time a stroke is input until “summer” is written until “summer” is written. It is determined that the part has been written as.

  On the other hand, there are two cases depending on the character recognition result and the morphological analysis result at the time when writing of “K” is completed and when the next “child” starts to be written.

  One is a case where the end part of the stroke sequence is recognized as “suggest”. In this case, since the last morpheme does not match the previous morpheme analysis result, “summer” is still “the part that has been written as a word”.

  The other is when the end of the stroke sequence is recognized as “Kou”, and the morphological analysis also matches the previous morphological analysis results up to “K”, such as “Summer / No / Koh / End”. is there. In this case, “upper” is regarded as “the part that has been written as a word”. Furthermore, when “Koshi” is written (if the character recognizer recognizes “Koshi” and the morphological analyzer analyzes “Koshi” as one morpheme), “Summer” will be written as “Word” in both cases. The finished part.

  In this embodiment, the candidates are rearranged in descending order of the co-occurrence frequencies Xi after narrowing down the top n geometrical matching degrees Yi of the stroke sequences. However, the present invention is not limited to this, and the degree of coincidence Yi is not an arbitrary number, and a threshold value may be provided so that candidates less than this may not be targeted.

  Further, it may be determined comprehensively from the geometric matching degree Yi and the co-occurrence frequency Xi of the stroke sequence. For example, the score may be determined by weighting and adding either or both.

  For example, by calculating the prediction candidate score as the linear sum S = αC + βM of the co-occurrence score C and the DP matching score M, and calculating the co-occurrence score at the history search stage, the score difference from the first rank is calculated. DP matching may not be performed for candidates that are larger than βM ′ (M ′ is “average DP matching score with correct answer”).

  Moreover, in this embodiment, what is displayed as a candidate from the history input by handwriting in the past is limited. This is because the character string is used once, so that the possibility is high.

  However, the present invention is not limited to this, and a candidate may be selected from the co-occurrence dictionary, and may be selected as a candidate when it matches the stroke sequence. In this case, at the stage where the word Wi is determined in step S209 in FIG. 6, one or more words that are likely to follow are extracted from the co-occurrence dictionary as candidates, and a stroke sequence is extracted for the extracted candidates. What is necessary is just to calculate the coincidence degree Yi.

  Thus, by limiting the candidates determined by DP matching to those in the dictionary storing the co-occurring candidates, the amount of DP matching calculation can be reduced.

  For example, FIG. 3 shows an independent word appearing immediately after “Summer” and its frequency in a corpus (including an adjunct appearing between “Summer” and an independent word). ) If it is recognized that “summary / no” in the input stroke sequence is “the part that has been written as a word”, the next expected word is the co-occurrence frequency of “summer”, the last independent word. Is likely to be a sufficiently high word such as “hot” or “festival”. Therefore, DP matching with “unfinished part as a word” in the (standard) stroke sequence and input stroke sequence for those words can be performed only for those with high frequency of occurrence with “summer”. That's fine.

  Thereby, even if it is not memorize | stored as a log | history, it can be predicted about the word which starts from the character memorize | stored in the stroke sequence corresponding dictionary.

  In this embodiment, the character string for morphological analysis is targeted only for the generated first candidate. However, if there is a surplus in computational resources, the candidate for the second or lower candidate is predicted in the same way. You may make it produce | generate.

  In addition, if there are more than a certain number of stroke strings having “unfinished parts as words”, the calculation cost of DP matching may increase, making it impossible to immediately respond to user input. Therefore, in such a case, the prediction may not be performed.

  In addition, instead of the input history as a prediction candidate in “Generation of prediction candidate”, a corpus or dictionary prepared in advance can be used.

  In addition, in the prediction candidate selection process, the co-occurrence frequency of independent words was used as the priority of the candidate, but various other linguistic resources such as character n-gram and semantic similarity can also be used. .

  In the present embodiment, DP matching is performed for “an incomplete part as a word”, but any matching of stroke sequences is possible as long as it matches the position and shape. It doesn't matter.

  Note that if the word W has never appeared in the input history, the prediction may not be performed.

  In the present embodiment, “parts written as words” and “unfinished parts as words” are separated, but in this embodiment, the character stored in the previous morpheme analysis result storage unit 27k. Compared with the column, the part that is the common prefix is set to “the part that has been written as a word”. However, the present invention is not limited to this. For example, recognition is performed in units of characters, and the “recognition result at the end is discarded” (characters are regarded as unfinished parts, or common to recognition result character strings without using morphological analysis. Take the prefix and divide it into “completed part” and “unfinished part”, or the last word of the morpheme analysis result when the word cost and the concatenation cost with the previous word are above the threshold A stroke sequence corresponding to a word may be incomplete.

  Note that “word cost” in morphological analysis represents the difficulty of appearance of the word, and “joint cost” represents the difficulty of connection between two (or more) adjacent words. For example, in Japanese, the word “se” often appears, but the word “丗” (30) rarely appears. In other words, the word cost of “Se” is smaller than the word cost of “丗”.

  In addition, “magazine”, “wo”, “ni”, etc. are likely to appear immediately after “newspaper”, but inflection endings such as “mas” and “t” are unlikely to appear. That is, the connection cost of “newspaper-magazine”, “newspaper-to”, and “newspaper-ni” is smaller than the connection cost of “newspaper-mass” and “newspaper-ta”.

  In the above description, in order to simplify the description, the case where the prediction candidate is not provided for the word “summer” has been described as an example. Actually, except for an attached word or the like, if there is a co-occurrence in relation to an independent word input before that, “summer” may be presented as a candidate for prediction. For example, in the previous sentence, if "... is a high school ball child" has been entered, and if the string "Summer of a high school ball child" has been entered in the past, "... is a high school ball child" After "is", when several strokes are input for "summer", "summer" is presented as a candidate.

  In the present embodiment, when a stroke is input even for a portion that has been written, the first character string candidate converted from the stroke string is determined. However, the present invention is not limited to this, and the determination from the stroke sequence may be omitted for a portion that has been repeatedly determined as a word.

  For example, even if you have never entered "Summer Koshien", if you have entered "Koshien", if "Summer" and "Koshien" are in the co-occurrence dictionary, you may select them as candidates. Good.

  When the handwritten character is confirmed, the input history storage unit 26r that stores the stroke string overwrites the correspondence with the stroke string for the predicted input word. Thereby, although it is not actually input, the converted word can be memorize | stored.

  In this embodiment, candidates are predicted by performing DP matching on incomplete parts. Such DP matching is computationally intensive. Therefore, by searching the user input history first, candidates for DP matching processing are narrowed down as much as possible.

  In the present embodiment, conversion from a stroke sequence is performed, morphological analysis is performed, and the history and co-occurrence of the input history storage unit 26r are divided into a completed part and an incomplete part as words. Prediction candidates are determined based on the dictionary.

  However, without using the history of the input history storage unit, the correspondence between the stroke string and the character string in the dictionary storage unit 26j may be referred to and the candidate may be determined using the co-occurrence dictionary.

  Also, in this embodiment, conversion from a stroke sequence is performed, this is morphologically analyzed, divided into a completed part and an incomplete part as words, and prediction candidates are determined by DP matching for the incomplete part. ing. This eliminates the need for DP matching for all input character strings, thus reducing the arithmetic processing.

  In this embodiment, although the case where the dictionary was memorize | stored in the handwritten character recognition apparatus was demonstrated, you may make it memorize | store a part or all in a server, and send a candidate to a handwritten character recognition apparatus.

  Further, if the number of strokes does not become a predetermined number (for example, 2 or more) for the undetermined portion, the above candidate calculation may not be performed.

  Also, instead of targeting the past input history, refer to the correspondence between the stroke sequence and the character in the dictionary storage unit, determine the candidate for one character, and select the candidate word that is at the top of the candidate. You may make it select from a co-occurrence dictionary.

  In the present embodiment, a predetermined number of candidates are selected from those having a high co-occurrence degree according to the co-occurrence dictionary. However, candidates to be narrowed by a threshold may be determined.

  In the present embodiment, the case where it is applied to Japanese has been described. However, the present invention can be similarly applied to other languages (such as Chinese and English).

  In the above embodiment, the CPU 23 is used to realize the function shown in FIG. 1, and this is realized by software. However, some or all of them may be realized by hardware such as a logic circuit. In addition, you may make it make an operating system (OS) process a part of program.

23 CPU
27 Memory 26 Flash Memory 36 Touch Panel

Claims (11)

A handwritten character dictionary storage means for storing a correspondence dictionary of stroke strings and corresponding character strings;
Co-occurrence dictionary storage means for storing a combination of independent word strings to be co-occurred,
A confirmed character string storage means for storing a combination of the converted character string and the stroke string determined by the user for each word;
Stroke sequence recording means for storing a stroke sequence given from a handwritten character input unit;
Character recognition means for performing character recognition using the handwritten character dictionary storage means for the stroke strings stored in the stroke string storage means,
Morphological analysis means for performing morphological analysis on the recognized character string candidate group obtained by the character recognition means,
Separating means for separating a morphological analysis result by the morphological analysis means into a part that is completed as a word and an incomplete part as a word by comparing with the result before one stroke input,
Prediction candidate calculation means for obtaining a prediction candidate of an incomplete part, the degree of coincidence between the stroke string of the converted character string stored in the fixed character string storage means and the stroke string of the incomplete part, and the co-occurrence dictionary Prediction candidate determination means for determining the prediction candidate based on the degree of co-occurrence with the last independent word of the part completed as a word,
Presenting means for presenting prediction candidates predicted by the prediction candidate calculation means;
A frame-undefined handwritten character recognition device characterized by comprising: In the frame undefined handwritten character recognition device according to claim 1,
The prediction candidate determining means selects candidates from those having a high degree of coincidence of stroke sequence histories, and causes the presenting means to rearrange and display in order of co-occurrence frequency,
A handwritten character recognition device. Character recognition means for performing character recognition using a handwritten character dictionary storage means for storing a correspondence dictionary of stroke strings and corresponding character strings for a stroke string given from a handwritten character input unit;
Morphological analysis means for performing morphological analysis on the recognized character string candidate group obtained by the character recognition means,
Separating means for separating a completed part as a word and an incomplete part as a word based on the morphological analysis result by the morpheme analyzing means,
Prediction candidate calculation means for obtaining a prediction candidate for an incomplete part, based on the degree of coincidence between the stroke string of the converted character string stored in the fixed character string storage means and the stroke string of the incomplete part, A prediction candidate determining means for determining a prediction candidate;
Presenting means for presenting prediction candidates predicted by the prediction candidate calculation means;
A frame-undefined handwritten character recognition device characterized by comprising: A frame-undefined handwritten character recognition program for causing a computer to execute as the following means.
Character recognition means for performing character recognition using a handwritten character dictionary storage means for storing a correspondence dictionary of stroke strings and corresponding character strings for a stroke string given from a handwritten character input unit;
Morphological analysis means for performing morphological analysis on the recognized character string candidate group obtained by the character recognition means,
Separating means for separating a completed part as a word and an incomplete part as a word based on the morphological analysis result by the morpheme analyzing means,
Prediction candidate calculation means for obtaining a prediction candidate for an incomplete part, based on the degree of coincidence between the stroke string of the converted character string stored in the fixed character string storage means and the stroke string of the incomplete part, A prediction candidate determining means for determining a prediction candidate;
Presenting means for presenting prediction candidates predicted by the prediction candidate calculation means. In the frame undefined handwritten character recognition apparatus of claim 3 or the frame undefined handwritten character recognition program of claim 4,
The separating means separates the completed morphological analysis result into a word and an unfinished part as a word by comparing with the result before one stroke input,
It is characterized by. In the frame undefined handwritten character recognition apparatus or the frame undefined handwritten character recognition program of claim 5,
The prediction candidate determining means reads out a combination of independent word character strings in a connected relationship from a connected character string dictionary storing combinations of independent word character strings used in connection, and Of these, the candidate that is highly connected to the last independent word is displayed as a candidate,
It is characterized by. In the frame undefined handwritten character recognition apparatus or the frame undefined handwritten character recognition program according to claim 6,
The prediction candidate determining means determines the prediction candidate after narrowing down candidates by the connected character string dictionary;
It is characterized by. In the frame undefined handwritten character recognition apparatus or the frame undefined handwritten character recognition program of claim 7,
The prediction candidate determination means sets a predetermined number of candidates as candidates for narrowing the candidates from those having a high degree of connection according to a connected character string dictionary;
It is characterized by. In the frame undefined handwritten character recognition device or the frame undefined handwritten character recognition program according to any one of claims 3 to 8,
In the confirmed character string storage means, a combination of a converted character string and a stroke string confirmed by the user is stored for each word,
It is characterized by. In the frame undefined handwritten character recognition apparatus or the frame undefined handwritten character recognition program according to any one of claims 3 to 9,
The prediction candidate determination means determines a candidate when the user actually has a connected history term.
It is characterized by. Store the following 1) to 3) in the storage section of the computer.
1) For stroke strings given from the handwritten character input part, correspondence dictionary of stroke strings and corresponding character strings, 2) combinations of independent word strings used in connection, 3) conversion determined by the user for each word A combination of a string and a stroke string,
That the computer performs the following steps:
When a stroke string is given from the handwritten character input unit, referring to the correspondence dictionary and performing character recognition;
A step of performing morphological analysis on the obtained recognition character string candidate group,
A step of separating a completed part as a word and an incomplete part as a word based on a morphological analysis result by the morpheme analyzing means;
Obtaining a prediction candidate for the incomplete part based on the degree of coincidence between the stroke string of the stored conversion character string and the stroke string of the incomplete part;
Presenting the predicted prediction candidates,
A frame-undefined handwritten character recognition method characterized by

Images