JPH06149872A - Text input device - Google Patents

Abstract

PURPOSE:To provide a text input device capable of effectively performing the post-processing of an input text with computer resources much less than ever. CONSTITUTION:Only a word with high word extraction likelihood is extracted from an input unit string at a high likelihood word extraction part 104. A morpheme analysis part 106 performs morpheme analysis processing setting a partial string divided by the word with high extraction likelihood as the unit of processing. Therefore, an error can be reduced when the input unit string is divided into the partial strings, and the morpheme analysis processing can be performed at high speed. Also, it is possible to select an input candidate and to perform error detection by effectively using the continuity of the text at the front and rear terminals of a processing unit.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a sentence input device having a mechanism for inputting a sentence as electronic data and detecting an error when the inputted sentence contains an error.

[0002]

2. Description of the Related Art In recent years, there is an increasing need to convert documents into electronic data regardless of existing or new documents. However, input by pattern recognition of characters, voices, etc. may cause erroneous recognition in the recognition unit, and even when the operator inputs a document using the keyboard, erroneous input may be mixed in the input sentence due to erroneous conversion of kana-kanji conversion. Cannot be denied. In order to find such an error contained in a text converted into electronic data, and to automatically correct or to assist the correction work by the operator, there are various text input devices provided with a post-processing unit for inputting text. Proposed.

In such a text input post-processing unit, a partial sequence that may be a word is extracted by comparing the input unit sequence that has been input with the word dictionary, and the word sequence is composed of word candidates that do not overlap with each other. It is common to perform morphological analysis processing that determines the word sequence that seems to be correct depending on the grammatical connectivity between them and the statistical co-occurrence frequency. Then, in the conventional device of this type, the morphological analysis process is performed on all the input unit candidates that have been input. According to this processing method, in order to process a character string having a length of N characters and having M candidate characters at each character position, it is necessary to match the N power combinations of M with the word dictionary. Therefore, it takes a long time to perform matching, and an incorrect combination of candidate characters may be cut out as a word candidate at the time of matching, which may increase the load of the process of checking the possibility of connection between words thereafter.

[0004]

For this reason, the input character string is delimited at the transition of punctuation marks and character types (Kanji, Hiragana, Katakana, symbols, etc.), and a short partial character string becomes a candidate for processing. Techniques have been devised to prevent combinatorial explosion of the number of strings. However, if you separate the input string with punctuation, the length of the substring will be tens of characters,
Sometimes you can't reduce the amount of processing enough,
In the case of delimiting at a character type transition, there is a possibility that the character string may be erroneously cut in the middle of one word, and that part may be rejected by the morphological analysis processing in the subsequent stage.
Also, according to the method of separating the input character string at the punctuation mark and the change of the character type, it is difficult to select a correct candidate from the input character candidates at the end because the context of the substring that is the unit of processing is lost. There was also.

Further, when word candidates are extracted from the input character strings for all the words included in the word dictionary, hiragana words with particularly short word lengths are often mistakenly cut out as word candidates, which also takes time and time required for matching. This will increase the load of the subsequent processing.

The present invention solves the problem of the above-described conventional text input post-processing technique in that there is a high possibility that a word candidate is erroneously cut out from an input character string during morphological analysis, and that it takes a lot of time for processing. It is an object of the present invention to provide a text input device that can effectively post-process an input text with less computer resources and can be comfortably used by an operator.

[0007]

A text input device of the present invention is a text input device for inputting a text as a sequence of set input units, and an input device for obtaining one or more input candidates for each input unit. , A high-likelihood word extraction unit that adds word information to each word candidate, with a partial string that matches the notation data of a word having a higher word extraction likelihood than the input unit string set as a word candidate, and an input unit In a row
From the word information of the word extracted by the high-likelihood word extraction unit and the input candidates of the subsequence that is not recognized as a word, at least one of the word dictionary, the input unit connection rule, the word connection rule, the grammar rule, and the semantic rule. A morphological analysis unit that calculates the likelihood of a combination of input candidates using or, and a subsequence in the input sentence in which the likelihood obtained by the combination likelihood calculation unit is lower than a certain value as an input error. An erroneous input detection unit for detecting is provided.

A text input device according to claim 1 of the present invention comprises:
Using a word likelihood dictionary including at least word description data and word extraction likelihood data described as a string of input units, and using a word likelihood dictionary, from a value for which the word extraction likelihood is set from the input unit string It is characterized in that a partial sequence that matches the notation data of a high word is set as a word candidate, and a high-likelihood word extraction unit that adds word information to each word candidate is provided.

A text input device according to claim 2 of the present invention is
A high-likelihood word dictionary configured by extracting only words having a pre-calculated word extraction likelihood larger than a set value from a word dictionary containing general words to be processed by the sentence input device. A sub-string that matches the notation data of a word included in the high-likelihood word dictionary than the input unit string is a word candidate, and a high-likelihood word extraction unit that adds word information to each word candidate is provided. .

A text input device according to claim 3 of the present invention is
English words with a word length of K characters (K = 2,3, ...) or more, hiragana words with a word length of L characters (L = 2,3, ...) or more, and M characters (M = M) A high-likelihood word dictionary containing one or more kana-kanji words of 2, 3, ...) or more, and kana-kanji mixed words having a word length of N characters (N = 2, 3, ...) or more, and an input unit It is characterized in that a partial sequence that matches the notation data of a word included in the high-likelihood word dictionary rather than a column is set as a word candidate, and a high-likelihood word extracting unit that adds word information to each word candidate is provided.

A text input device according to claim 4 of the present invention is
Utilization information is attached to the entry of the utilization word included in the word likelihood dictionary or the high-likelihood word dictionary, and the utilization word generation is used to refer to this utilization information and generate a shape in which the utilization stem is added to the stem of the utilization word. And a high-likelihood word extraction unit is a word likelihood dictionary or a non-conjugated word written in the high-likelihood word dictionary and the conjugation forms of the conjugation words generated by the conjugation form generation unit are targeted for word extraction. Is characterized by.

A text input device according to claim 5 of the present invention is
When the character input means outputs a plurality of candidate characters, only the character string formed by the first candidate character is set as the processing target of the high likelihood word extraction unit.

[0013]

In the sentence input device of the present invention, the high-likelihood word extracting unit extracts words having a high likelihood of word extraction, separates the input unit string by the high-likelihood words extracted here, and separates each of the partial strings. On the other hand, the processing of the morphological analysis unit is executed.

[0014]

DESCRIPTION OF THE PREFERRED EMBODIMENTS FIG. 1 is an overall configuration diagram of an embodiment of the present invention.

Reference numeral 101 denotes an input means for converting an input unit into a code, which may be an OCR, an online handwritten character input device, a character input device such as a keyboard, a voice recognition device or the like. These input devices are at least the score at the time of recognition,
It is assumed that one or more ordered input candidates can be obtained by any of the similar character tables in consideration of the characteristics of the input means.

Reference numeral 102 is a word likelihood dictionary.

Reference numeral 103 denotes an inflectional form generating section for producing possible inflectional forms for inflectional words included in the word likelihood dictionary.

Reference numeral 104 is a high likelihood word extraction unit.

Reference numeral 105 is a part-of-speech connection rule table.

A morphological analysis unit 106 uses a word dictionary and a part-of-speech concatenation rule for the combination of the word information added to the word candidates by the high likelihood word extraction unit and the input character candidates that are not recognized as words. Perform morphological analysis processing to calculate the likelihood.

An erroneous input detection unit 107 detects a subsequence whose likelihood calculated by the combination likelihood calculation unit is lower than a certain value as an input error.

FIG. 2 is a processing flow of the high likelihood word extraction unit, FIG. 3 is a processing flow of the combination likelihood calculation unit, FIG. 4 is a part of the word likelihood dictionary, and FIG. 5 is a part of the part-of-speech concatenated likelihood table. FIG.

The operation of the text input device of the present invention will be described below with reference to the case of inputting a Japanese sentence shown in FIG.

In this embodiment, an OCR that outputs up to the third highest candidate character is assumed as the input means, and in this case, the input unit is a character. In addition to this, it is also possible to use an online handwritten character input device, a character input device such as a keyboard, a voice recognition device, etc. as the input means. The input unit in the case of the character input device is a character, and the input unit in the case of a voice recognition device. The unit is a recognition unit such as a phoneme or a syllable unique to the voice recognition device. FIG. 7 shows the result of reading the Japanese sentence of FIG. 6 by the OCR that outputs up to the third candidate character.

The high-likelihood word extraction unit collates a dictionary item having each character of the input character string as the first character with the input character string, and if they match, repeats the process of registering the word information in the word candidate table. When this process is completed for all the characters of the input character string, the information of the word candidate table and the information of the unprocessed character string are output and the process is completed. As a specific example, the processing in the high-likelihood word extracting unit when processing the character candidate string shown in FIG. 7 will be described according to the processing flow in FIG. In the high likelihood word extraction unit, the internal state is first initialized (steps 201 and 202). In the initial state, the pointer n = 1 for the input character string and the character C (1) = “light”, so the first character is “light” from the word likelihood dictionary and the inflectional form generation unit, and the word extraction likelihood is set. Get words that are greater than or equal to the value. The word extraction likelihood is a value indicating the proportion of a word included in the correct morphological analysis result when a word candidate is generated simply by pattern matching the input character string and the notation data written in the dictionary. Further, in the present embodiment, a word with a value of the word extraction likelihood of the word likelihood dictionary shown in FIG. 4 exceeding 0.5 is targeted for processing by the high likelihood word extraction unit. Therefore, ““ light ”described in the dictionary of FIG.
“(Noun) i = 1, p = 0.3” is rejected by the word reference in the high likelihood word extraction unit.

The next dictionary item "" light "(verb) i = 1, p
= 0.6, line 5 dan "is an inflection word, and thus, in the inflectional form generation unit," light "," light "," light "," light ", and" light "are generated and input to the high likelihood word extraction unit. . In the inflectional form generation unit, the inflectional word is added to the notation data described in the dictionary, the word length i is replaced with the generated inflectional word length, and the word extraction likelihood p is processed, for example, according to the inflectional form. Processing such as changing the word extraction likelihood is also conceivable. The word obtained in this way (steps 203 and 204) is a partial character string of the input character string {C (n), C (n + 1)}.
= Since the comparison with “light de” (step 205) does not match (step 206), if a word whose first character is “light” (for example, “optical” “light speed”) is registered in the dictionary, The steps after step 203 are repeated for the word. When this process is completed for all words whose first character is "light", the pointer n for the input character string
Is incremented by 1 (step 208). When n = 2, that is, when the above process is performed for the character C (2) = "de", the word "disk" i = 4, p =
"0.7" matches the input string, so "" disk "
(Noun) start position = 2 and end position = 5 ”are registered in the word candidate table (step 207). In this way n
= 51, and when the pointer n is advanced next, the position indicated by n exceeds the end of the input character string (step 209). Therefore, the information of the word candidate table and the information of the unprocessed character string are output ( (Step 210) The process ends.

FIG. 8 schematically shows the output of the high likelihood word extraction unit. In the figure, words indicate the word candidates and their part-of-speech names extracted by the high-likelihood word extraction unit in a frame. String length is 5
As a result, the length of the longest unprocessed character string is 6 with respect to the input character string of 1, which shows that the input character string is effectively divided.

In this embodiment, in the process of obtaining a word from the dictionary in step 203, the inflectional form is generated in the inflectional form generation unit from the stem of the word recorded in the dictionary and the utilization information of the inflectional word, Although all conjugation forms are input to the high likelihood word extraction unit, a processing form in which the stem and the conjugation ending are extracted and the connection test is performed in the morphological analysis unit in the subsequent stage is also conceivable.

Further, in the present embodiment, the character string consisting of only the first-ranked candidate of the input character is the processing target of the high-likelihood word extracting section, but the candidate characters up to the rank set in the high-likelihood word extracting section are selected. How to extract words for combinations,
Alternatively, a method of extracting the high-likelihood word from the character string of the first-ranked candidate and then extracting the high-likelihood word for the combination of the candidate characters of the unprocessed portion is also conceivable.

In the morphological analysis unit, the high-likelihood word extraction unit extracts word candidates for the unprocessed subsequence, and then calculates the connection likelihood between the word candidates based on the concatenation likelihood of the part-of-speech. The word candidate combinations obtained from the candidate combinations are ranked. As a specific example, the character position n = 9 to 1 in the output of the high-accuracy word extraction unit shown in FIG.
5, ie, the case of processing the “mass storage” portion, the processing of the morphological analysis unit will be described according to the processing flow of FIG. For the portion of n = 9 to 11, no word matching the character string was found in the high likelihood word extraction unit, so the word candidate is extracted in the morphological analysis unit (step 301). At this time, all the words in the word likelihood dictionary (FIG. 4) are targeted for word candidate extraction.

In this example, “large” (prefixed noun) start position = 9, end position = 9 “dog” (noun) start position = 9, end position = 9 “capacity” (noun) start position = 10, end position = 11 "customer" (noun) start position = 10, end position = 10 "each" (prefixed noun) start position = 10, end position = 10 "chapter" (noun / suffix) start position = 11, end position = 11 "quantity" (noun) 7 word candidates with start position = 11 and end position = 11 are extracted (Fig. 9). For the portions of n = 12 to 13 and n = 14 to 15, since ““ memory ”(noun)” and ““ device ”(noun)” are extracted by the high likelihood word extraction unit, these are word candidates. And Next, a connection likelihood is calculated for a combination of concatenated word candidates of these word candidates (step 30).
2). There are various conceivable methods for calculating the connection likelihood, but for example, for each combination of word candidates in a specific section, the method of using the product of the concatenation likelihood between all word candidates as the connection likelihood of the combination of word candidates is used. Shall be used. Figure 5
FIG. 10 shows an example in which this calculation method is applied to the combination of word candidates described above using the part-of-speech concatenated likelihood table shown in FIG. As a result, the word sequence “large / capacity / memory / device” is the maximum likelihood combination, and the result matches the input character string.

In this example, a correct combination of candidate characters is selected from the input candidate character strings by the above processing, but the morphological analysis fails, that is, a substring that cannot be identified as a word occurs, or all word candidate It is possible that the connection likelihood of the combination is less than the allowable range. As one of the causes of the failure of the morphological analysis, it is considered that erroneous extraction of word candidates occurs in the high likelihood word extraction unit, and as a result, a failure occurs in the morpheme analysis unit. In this case, if the morpheme analysis fails and there is a word candidate of a high likelihood word adjacent to the failure part, the word candidate is rejected, and the word candidate is again selected from the input character candidates for the low likelihood word. A correct result can be obtained by extracting and performing morphological analysis (step 303).

If a plurality of high-likelihood word candidates are redundantly extracted by the high-likelihood word extracting unit as in the section of n = 33 to 37 in FIG. 8, "online" is taken as the word candidate. The connection likelihood is calculated for both the case and the case where "line" is used as the word candidate (in this case, the word candidate is extracted from the combination of the input character candidates for the portions of n = 33 and 34), and the ranking is performed. I shall.

In the actual processing, the high-likelihood word extracting unit has unprocessed subsequences and high-likelihood words before and after it, that is, n = 2 to 13 and n = 17 to in FIG.
By using the portion 27 and the like as the processing unit of the morpheme analysis unit, the processing unit of the morpheme analysis unit can be shortened and the combination of character candidates / word candidates can be suppressed to a small number. Also,
The substring cut out in this way may include punctuation marks and symbols, but by calculating the connection likelihood for these special characters as a special part of speech, it is possible to select special character candidates that were difficult with the conventional method. And error detection becomes possible.

With the above processing, the partial character string which cannot be further subjected to morphological analysis is detected by the erroneous input detection unit and output. This output is supposed to obtain the correct character string after being corrected by the operator, but during this correction processing, information such as word candidates generated in the process of the high likelihood word extraction unit and the morphological analysis unit is appropriately used. It is possible to provide a user interface with excellent operability by displaying it.

The text input device according to claim 2 of the present invention is configured by extracting words having a higher word extraction likelihood than a word dictionary storing general words to be processed by the text input device. The high-likelihood word dictionary described above is provided, and by performing the high-likelihood word extraction process using this high-likelihood word dictionary, the same effect as that of the text input device of the present embodiment can be obtained while reducing the calculation amount at the time of execution. I will get it.

The text input device according to claim 3 of the present invention comprises a high-likelihood word dictionary composed of words having a preliminarily calculated word extraction likelihood larger than a preset value, and is described in this dictionary. High-likelihood word extraction processing is performed by pattern matching between words and input character strings.

Further, in the sentence input device according to claim 4 of the present invention, a word group consisting of English words having a word length of K characters (K ≧ 2, preferably K ≧ 3) or more, and a word length of L characters (L ≧ 2, preferably L ≧ 3) or more hiragana words, a word group of katakana words with a word length of M characters (M ≧ 2) or more, and a kana-kanji mixture of word characters of N characters (N ≧ 2) or more A high-likelihood word dictionary configured by a word group including one or more of the word groups based on the word is provided, and high-likelihood word extraction processing is performed by pattern matching between the words described in this dictionary and the input character string. Generally, a word having a long word length or a word composed of a character type such as a kanji that includes a large number of characters is rarely mistakenly extracted as a word candidate.
In this way, the shortest word length is determined for each character type, and by constructing a high-likelihood word dictionary with longer words, the same effect as that of the text input device of the present embodiment is obtained.

[0039]

As described above, in the sentence input device of the present invention, the high-likelihood word extracting unit extracts words with high likelihood of word extraction, and the high-likelihood words extracted here are used as input unit strings. Since the morpheme analysis unit executes the process for each substring by dividing the input unit string, the input unit string is effectively divided into substrings with few errors such as splitting in the middle of a word. Therefore, the morpheme analysis processing using this subsequence as a processing unit can be performed at high speed. In addition, since it is possible to utilize the part-of-speech information of high-likelihood words that become breaks in processing units for morphological analysis processing, the continuity of sentences is effectively used even at the front and rear ends of processing units to select and correct input candidates. It is possible to perform detection, and it is possible to effectively perform processing even on a portion such as a punctuation mark or a symbol in a sentence that has been conventionally difficult to select a candidate or detect an error.

[Brief description of drawings]

FIG. 1 is an overall configuration diagram of an embodiment of the present invention.

FIG. 2 is a diagram showing a processing flow of a high likelihood word extraction unit.

FIG. 3 is a diagram showing a processing flow of a combination likelihood calculation unit.

FIG. 4 is a diagram showing a part of a word likelihood dictionary.

FIG. 5 is a diagram showing a part of a POS concatenation likelihood table.

FIG. 6 is a diagram showing an example of a Japanese sentence used to explain the operation of the present invention.

7 is a diagram showing the result of reading the Japanese sentence of FIG. 6 by OCR that outputs up to the third candidate character.

FIG. 8 is a diagram schematically showing an output of a high likelihood word extraction unit.

FIG. 9 is a diagram showing an example in which word candidates are extracted from a combination of input character candidates in the morpheme analysis unit.

FIG. 10 is a diagram showing a calculation example of connection likelihood of a combination of word candidates.

[Explanation of symbols]

 101: Input means for converting an input unit into a code 102: Word likelihood dictionary 103: Inflectional generation unit 104: High likelihood word extraction unit 105: Part-of-speech concatenation rule table 106: Combination likelihood calculation unit 107: False input detection unit

Claims (7)

[Claims] 1. A sentence input device for inputting a sentence as a sequence of set input units, wherein input means for obtaining one or more input candidates for each input unit, and words described as at least a sequence of input units The word likelihood dictionary including the notation data and the word extraction likelihood data, and the substring of the input unit sequence that matches the notation data of the word in the word likelihood dictionary that has a higher word extraction likelihood than the set value. As a candidate, at least from a high likelihood word extraction unit that adds word information for each word candidate, word information of the word candidate extracted by the high likelihood word extraction unit, and a candidate string of the input unit of other parts, Word dictionary, input unit connection rules, part-of-speech connection rules,
A sentence input device, comprising: a morphological analyzer that calculates the likelihood of a combination of input unit candidates using any of a word concatenation rule, a grammar rule, and a semantic rule. 2. A sentence input device for inputting a sentence as a string of set input units, wherein the input unit obtains one or more input candidates for each input unit, and a high likelihood that is a subset of a general word dictionary. A word dictionary, a high-likelihood word extraction unit that performs a process of adding word information for each word candidate, using a substring of the input unit string that matches the word description data in the high-likelihood word dictionary as a word candidate, From the word information of the word candidates extracted by the high-likelihood word extraction unit and the candidate string of the input unit of other parts, at least the word dictionary, the connection rule of the input unit, the connection rule of the part of speech,
A sentence input device, comprising: a morphological analyzer that calculates the likelihood of a combination of input unit candidates using any of a word concatenation rule, a grammar rule, and a semantic rule. 3. The sentence input device according to claim 2, further comprising a high-likelihood word dictionary configured only with words having a word extraction likelihood larger than a set value. 4. A word group consisting of English words having a word length of K characters (K ≧ 2) or more, a word group consisting of hiragana words having a word length of L characters (L ≧ 2) or more, and a word length of M characters (M ≧ 2). ) A high likelihood word dictionary configured by a word group including the above katakana words and a word group including at least one word group including kana-kanji mixed words having a word length of N characters (N ≧ 2) or more The text input device according to claim 2. 5. Utilization information is attached to an entry of a utilization word included in a word likelihood dictionary or a high-likelihood word dictionary, and the utilization word is referred to by using a stem with a utilization ending. A high-likelihood word extraction unit is provided with an inflectional-form generation unit for generating, and the high-likelihood word extraction unit extracts the inflectional form of the non-individualized word described in the word likelihood dictionary or the high-likelihood word dictionary and the inflectional word generated by the inflectional-word generation unit The object of claim 1
Alternatively, the text input device according to claim 2. 6. When the input device outputs a plurality of input candidates, only the input unit string composed of the first-ranked input candidates is processed by the high likelihood word extraction unit. The text input device according to claim 1, claim 2, or claim 5. 7. An erroneous input detection unit for detecting, as an input error, a subsequence of which the likelihood obtained by the morphological analysis unit is lower than a certain value in the input sentence. The sentence input device according to claim 2, claim 5, or claim 6.