JPS61294591A - Information input device - Google Patents

Abstract

PURPOSE:To reduce the quantity of processing and to shorten the processing time when input information is recognized, by limiting statistically the candidate subjects to be recognized according to their emerging frequencies, etc. and recognizing only those candidate subjects that have high emerging probability. CONSTITUTION:Word voices are supplied to a pattern input part 1 and a feature vector is extracted. This vector is compared with the standard feature vector of the information on a subject to be recognized which is previously registered in a recognition dictionary 2 through a pattern recognizing part 3. The results of recognition are obtained and are arrayed in the order of candidate subjects to be recognized having higher degrees of resemblance. This array is delivered to a result selection part 4 in the form of the pattern recognition results. The patterns are recognized by the part 3 by limiting the information on the recognition subjects stored in a recognition dictionary 2 and to be recognized according to the information on the statistical frequency like each emerging frequency information, etc. related to the information on the recognition subjects stored in a frequency table 6. Then the part 4 delivers only the recognition subject candidates that have high emerging frequencies.

Description

[Detailed Description of the Invention] [Technical Field of the Invention] The present invention recognizes an input pattern given as speech or text, and inputs input target information such as linguistic information indicated by the input pattern into an information processing system. The present invention relates to an information input device.

[Technical background of the invention and its problems]

With the development of information processing technology, various language processing systems such as Japanese word processors have been developed.

This type of information processing system is generally configured to input a kana character sequence from a keyboard, process the input character sequence for original language information to obtain linguistic information such as words and phrases, and input this into the system.

However, inputting character strings using a keyboard requires so-called familiarity, making it difficult for general users to use, the input format is unnatural, it is easy to get tired, and information input speed cannot be increased. There is a problem.

Against this background, various attempts have been made to research information input devices that utilize character and voice recognition technology. This device uses handwritten characters themselves as input information, or inputs information by voice. According to such an information input device, it is possible to input information easily, naturally, and at high speed.

However, pattern recognition technology still has some imperfections, and it is not always possible to correctly recognize input patterns. Therefore, if a recognition error occurs, it is necessary to correct the error. Moreover, the pattern recognition technology is a very advanced technology, and there is also the disadvantage that the equipment cost is very high.

For this reason, there is a strong desire to develop an inexpensive and high-performance pattern recognition device, and to develop a high-performance information input device using this pattern recognition technology.

[Purpose of the invention]

The present invention has been made in consideration of these circumstances, and its purpose is to provide an inexpensive and inexpensive method that can accurately recognize input patterns and effectively input information indicated by the input patterns. The purpose of the present invention is to provide an information input device.

[Summary of the invention]

The present invention is based on high-speed semiconductor technology and pattern recognition technology.
In view of advances in performance, and the fact that the type (field) of information desired to input and the frequency of input of that information differ depending on the information inputter (LI), input consisting of syllables, kana characters, etc. is performed by referring to the input target information. When recognizing a pattern and inputting information consisting of words, phrases, etc. indicated by the input pattern or a series of input patterns into an information processing system, statistical Recognize the input pattern by limiting the input target information to be subjected to ML processing for the input pattern according to the frequency information, for example, by excluding those whose statistical frequency information is below a predetermined threshold from the recognition target information. , and the statistical frequency information regarding the input target information is sequentially updated according to the recognition processing result.

〔Effect of the invention〕

Thus, according to the present invention, since the input target information for the information that the user attempts to input is limited based on the information on the appearance frequency and the connection frequency, it is possible to improve the recognition processing speed and recognition accuracy for input patterns. .

In other words, depending on the user, information that appears infrequently is excluded,
Since the input pattern is recognized as a recognition processing result of only information with a high frequency of occurrence, it is possible to improve the processing speed, and since no unwanted recognition candidates appear, it is possible to improve the recognition accuracy. .

Therefore, an economical and high-performance information input device can be realized, and the user can have great practical effects, such as being able to easily input desired information.

[Embodiments of the invention]

Hereinafter, one embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a schematic diagram of an embodiment of an apparatus for recognizing word sounds and inputting the linguistic information into a computer.

This device basically consists of a pattern input section 1 that inputs a voice pattern, a pattern input section 3 that recognizes the input pattern by referring to a recognition word I2, and a result selection section that selects and outputs the result of this pattern a'i. 4. In addition, 5
is a system control unit that controls the operation of the entire device.

The pattern input unit 1 analyzes the input word sounds,
After performing preprocessing such as determining the beginning and end of the word sound,
Feature vectors such as phoneme are extracted. The pattern recognition unit 3 compares the feature vectors of such input speech with standard feature vectors of recognition target information (categories such as words) registered in advance in the recognition dictionary 2 to obtain recognition results (categories). For example, the composite similarity between the vectors is calculated. Then, the recognition target candidates (categories) are arranged in descending order of similarity value and outputted to the result selection unit 4 as a pattern recognition result.

In addition, the pattern recognition process by the pattern recognition unit 3 is performed based on statistical frequency information such as each appearance frequency information regarding the recognition target information accumulated in a frequency table 6, which will be described later. This is done by restricting categories (categories).

That is, among the recognition target information registered in recognition dictionary ■2,
Generally, a lot of information other than the field (range of linguistic information) that the user is inputting is also included.

Therefore, by referring to the user's information input history, it can be seen which information is frequently input and which information is rarely input. vA degree, both sides and 6, such information on the frequency of appearance, etc. is aggregated for each piece of recognition target information registered in the recognition word I2. According to the frequency information accumulated in the frequency table 6, the recognition target information used for the pattern matching is limited, for example by excluding recognition target information with a low appearance frequency from being matched with the input pattern. .

In this way, the pattern recognition unit 3 performs matching processing only with the limited recognition target information, and obtains a recognition result for the input pattern.

Therefore, the result selection unit 4 selects and outputs only recognition target candidates that appear frequently under the control of the frequency management unit 7 from among the plurality of VT recognition target candidates obtained by the pattern recognition unit 2. are doing.

That is, the result recognition unit 4 compares, for example, the frequency equivalent set by the frequency management unit 7 with the frequency information regarding each recognition target information stored in the frequency table 6, and determines the frequency of the information extracted as the recognition target candidate. It is determined whether the information is higher than the frequency threshold. Then, recognition target candidates having a frequency value lower than the frequency equivalency are excluded, and only the remaining recognition target candidates are output.

I! selected in this way! ! Recognition object candidates (categories) are presented to the user, and if necessary, the user is left to select the correct recognition result. W selected in this way
A candidate object is given as input information to a computer system.

At this time, the frequency management unit 7 obtains information on the recognition target candidates selected by the user and inputted into the computer system, and receives frequency information regarding each W1w and target candidate stored in the frequency table 6. is being updated.

As a result, the frequency information of each recognition target information stored in the frequency table 6 is updated according to the history of the information input, and is used for subsequent information input.

Note that the input unit 8 can input information for setting the frequency threshold of the frequency management unit 7. That is, lowering the threshold value when performing threshold processing for frequency means relatively increasing the frequency information, and increasing the threshold value is equivalent to lowering the frequency information.

By inputting this frequency threshold value, the above-mentioned process results in 1! It is possible to forcibly set the range of recognition target information to be excluded from candidates. For example, if you set it to a value lower than the frequency value below, it will be excluded! ME target information is completely lost.

At the same time, the frequency management unit 7 appropriately updates the frequency equivalent value for setting the range of recognition target information to be excluded in accordance with the further IFr process.

Furthermore, the input unit 8 is used to input information for forcibly setting selection criteria for recognition target information. That is, when the above-mentioned frequency threshold processing is performed, information to be input may be excluded from recognition targets because the frequency value thereof is low. When inputting such information with low frequency of appearance, the input unit 8 increases the frequency information related to the information from the outside in a preset manner.

By inputting and setting this frequency information, the recognition target information excluded from recognition candidates by the above-described process is recovered as recognition candidates. Therefore, even if the information is normally excluded from input targets, it is possible to input it as needed.

In this manner, according to the present VtR, since input target information is selected according to its frequency information and used for input pattern recognition processing, it is possible to speed up the pattern recognition processing.

In addition, since unwanted input target information is excluded from the input pattern recognition candidates, it is possible to improve the HIEIII degree. Therefore, it is possible to input information accurately in a comprehensive manner, and to improve the efficiency of inputting information.

FIG. 2 shows an example in which the present invention is applied to an audio word processor.

The input voice pattern is processed by the monosyllable detection recognition unit 11.
, and predetermined preprocessing and feature vector extraction are performed. The input speech is then recognized with reference to the recognition word [12]. This input speech recognition process is performed in the same manner as in the apparatus shown in FIG. 1 above.

The monosyllable detection and recognition unit 11 outputs the recognition result of the input monosyllable as one or more recognition candidates.

The concatenation processing unit 13 uses the concatenation dictionary 14 to test the conjunctive relationship of each syllable in the syllable candidate series obtained in this manner. This test of the syllable connection relationship is performed by referring to the connection dictionary 14 to check whether or not the connection relationship holds true for all combinations of syllables in the syllable candidate series.

Through this concatenation process, syllable candidate sequences that have possible concatenation relationships are extracted.

For the syllable candidate series extracted in this way, the kana-kanji conversion processing unit 15 refers to the kana-kanji conversion dictionary 16 and performs kana-kanji conversion ItI4 processing to convert the characters mixed with kana-kanji corresponding to the syllable candidate series. We are looking for column candidates (bunsetsu candidates). These phrase candidates are displayed on the result selection display section 17, and under the instruction input from the input section 18, the phrase candidates whose input speech has been correctly recognized are selected.

Here, the feature of this device is that the kanki unconnection processing and the kana-kanji conversion processing are performed under the control of the frequency management section 19.
It is controlled in accordance with each frequency information registered in the concatenation frequency table 20 and the kana-kanji conversion dictionary frequency table 21. That is, the concatenation dictionary 14 used in the concatenation process is selected according to the frequency value of each concatenation relationship shown in the concatenation frequency table 20 and is provided to the concatenation process IIi section 13. Further, the conversion dictionary 16 used in the kana-kanji conversion process is selected according to the frequency value shown in the kana-kanji conversion dictionary frequency table 21 and is provided to the kana-kanji conversion unit 15.

Through such dictionary selection processing, the connection processing i1
3 and the kana-kanji conversion processing unit 15 are designed to perform concatenation processing and kana-kanji conversion processing at high speed and with high accuracy within the selected processing target information range.

Note that 22 in the figure is a system control unit that controls the operation of the entire apparatus.

Now, to explain the control of the concatenation process according to the frequency information, the frequency management section 19 that controls the concatenation process is configured as shown in FIG. 3, for example.

The g14 degree management processing unit 23 sets a threshold value as a selection criterion for a set of syllable combinations (conjunctive set) in the threshold register 24,
The comparison unit 25 compares the appearance frequency value of each connection group registered in the connection frequency table 20 with the threshold value for each connection coarse angle.

Note that the appearance frequency value of each connection group in the connection vA degree table 20 is initialized in advance as a general-purpose value. As the concatenation process progresses, each time the conjunctive relationship in the syllable series is tested, the V4 degree value for that concatenated group is updated under the control of the frequency management processing section 23. Similarly, the contents of the threshold register 24 are also updated as necessary under the control of the frequency management processing section 23.

In the conjunction 114, a selection flag is set for each conjunction group (category) registered in the conjunction 1114, according to the comparison result of the frequency values. With reference to this selection flag, it is determined whether or not the linked group is to be used in the above-described linking process.

In other words, the concatenation processing unit 13 tests the conjunctive relationship of the input syllable series using only the information of the conjunctive groups flagged based on the set threshold.

Specifically, as shown in Figure 4, the frequency for the connected group (A → A) is (0), and the frequency for the connected group (A → A) is (0).
1oo) etc., set this as the threshold value (5
0), the connected group (A → A) is excluded from the target of connection processing, and the flag (1
) is set up and certified as a target for linked processing. Similarly, since the frequency is (sooo) for the connected group (A → I), it is connected! l! It is recognized that it is subject to Nu.

Using only the information (conjunction I) of the E conjunctive group that has been identified as the subject of concatenation processing in this way, the concatenation processing section 13 tests the conjunctive relationship of the input syllable series.

Through such connection group selection processing, the connection processing unit 13
The concatenation process is performed quickly, efficiently, and accurately by effectively utilizing only the information of concatenated groups that appear frequently.

Note that it is possible to forcibly set the value of the threshold register 23 using the input unit 18, and thereby the range of recognition target information is forcibly set. Also, if you want to input a special connection relationship, enter the frequency for that connection group in the input section 1.
8 to exceed the same value, and the information on the linked group selected based on the statistical frequency information may be used in the linking process.

On the other hand, the frequency management section 19, which controls the kana-kanji conversion process according to the frequency information, is configured as shown in FIG. 5, for example.

The frequency management processing unit 26 performs kana-kanji conversion target (category)
A threshold value serving as a selection criterion is set in the equivalency register 27, and the comparison unit 28 compares the appearance frequency value of each kana-kanji conversion target registered in the frequency table 21 and the above WA value for each kana-kanji conversion target. ing. Note that the appearance frequency value of each kana-kanji conversion target in the frequency table 20 is initialized in advance as a general-purpose value. As the kana-kanji conversion process progresses, the frequency value for each kana-kanji conversion target that has been converted into kana-kanji is updated under the control of the frequency management processing section 26.

The kana-kanji dictionary 16 is actually composed of an independent word dictionary section and an attached word dictionary section, and accordingly, the frequency table 21 is also composed of an independent word frequency table and an attached word frequency table. It has become.

The kana-kanji conversion unit 15 inputs only the kana-kanji conversion targets limited according to the frequency information, and converts the syllable series subjected to the above-mentioned concatenation test into kana-kanji characters.

Specifically, as shown in Figure 6, in response to the input voice ``Niyu uri yoku ni nara niara'', the response ``Niyu uri yoku de (ga) a(ba) ru(bu)'' is shown in Figure 6. When the syllable sequence ``na(ma)raba'' is obtained as a recognition result, it is subjected to an independent 8B test and an adjunct word test to convert kana to kanji. In this case, since the independent word "input" is selected as a frequently appearing kana-kanji conversion candidate, a conversion result "input j" is obtained.

As for the suffix gA word part, since the two candidates are ``is'' and ``is J'', the conversion processing result is, for example, ``if it is an input'' and ``if there is an input''. sought as a candidate.

However, if we refer to the information on the frequency of appearance of the ma part of the lever, it can be seen that the frequency of appearance of the auxiliary word ``ga'' and ``aru'' is higher than that of the auxiliary word ``J''. According to this appearance frequency information, the result selection display section 17 displays whether this voice input person (user)
It is determined that the adjunct ``aru'' is preferred, and this is output as the first candidate of the recognition result.

In addition, in the processing for this adjunct word, if the processing threshold is set to, for example, (5OO), the conversion output of appendix III, which is "is", is prevented from being converted from the adjunct word conversion processing target. Therefore, in this case, only the kana-kanji conversion result ``if there is input'' will be output.

By previously limiting conversion candidates for independent words and attached words in such kana-kanji conversion processing according to their frequency information, wasteful kana-kanji conversion processing can be omitted and ambiguity in the conversion processing results can be removed. It turns out. Therefore,
It becomes possible to improve recognition processing accuracy by that much. Furthermore, it is possible to improve the processing speed.

In this case as well, it is possible to restore independent words and adjunct words that have been selected according to frequency information and excluded from conversion targets by presetting their frequency values, etc. as conversion targets. Furthermore, it is also possible to set the same frequency values as described above. Therefore, when inputting information that is different from normal information inputting, by using this threshold value setting function, it becomes possible to input the information accurately.

As explained above, according to the present invention, when performing recognition processing (conversion processing) on input information (pattern), the recognition target candidates (conversion processing target candidates) are statistically limited based on their appearance frequency, etc., and the occurrence probability is Recognize only the things that are
(conversion) processing, it is possible to reduce the processing amount and shorten (speed up) the processing time. And! t
! Since the recognition (conversion) result is obtained from among candidates with high probability as II (conversion) candidates, it is possible to improve the recognition It (conversion) accuracy.

In addition, if information that has been excluded from recognition (conversion) targets based on frequency information is to be included in recognition (conversion) targets as necessary, it is only necessary to provide frequency information for that information as a preset. can respond flexibly to input information. Therefore, great practical effects can be achieved, such as making it easier to input information and making it possible to easily and accurately input necessary information.

Note that the present invention is not limited to the embodiments described above. Here, we gave an example of voice input of Japanese sentences, but
The present invention can be similarly applied to inputting handwritten character information. In addition, here, we used frequency information to control the selection of processing candidates in input pattern recognition processing, concatenation processing on the recognition results, and kana-kanji conversion processing, but each of these processing controls can also be performed independently. It is possible. In addition, the present invention can be implemented with various modifications without departing from the gist thereof.

[Brief explanation of drawings]

FIG. 1 is a schematic configuration diagram of a basic embodiment device of the present invention, FIG. 2 is a schematic diagram of an embodiment device applied to a spoken word processor, and FIG. 3 is a configuration diagram of a frequency management section used for concatenation processing. The fourth factor is a diagram showing an example of the configuration of the concatenation frequency table. Figure 5 is a diagram showing an example of the configuration of the frequency management unit used in the kana-kanji conversion process. Figure 6 is an example of the kana-kanji conversion process. It is a diagram showing 1... pattern input section, 2... H1i dictionary, 3...
・Pattern recognition section, 4... Result selection section, 5... System υ31111 section, 6... Frequency table, 7... Frequency management section, 8... Input section, 11... Monosyllable detection recognition part,
12... Recognition dictionary, 13... Concatenation processing unit, 14...
- Concatenation dictionary, 15... Kana-kanji conversion section, 16... Dictionary for kana-kanji conversion, 17... Result selection display section, 18.
...Input section, 19...Frequency management section, 20...Conjunction frequency table, 21...Kana-Kanji conversion world! Frequency table, 22...
System control unit. Applicant's representative Patent attorney Takehiko Suzue Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 (a)-E-14 (:9 U IJJ
<@J> S rt '5 1
J"Figure 6

Claims (5)

[Claims] (1) Recognize the input pattern by referring to the input target information,
In an information input device that inputs information indicated by the input pattern into an information processing system, the input pattern is recognized and processed by restricting input target information to be subjected to recognition processing for the input pattern according to statistical frequency information regarding the input target information. , an information input device characterized in that statistical frequency information regarding the input target information is sequentially updated according to the recognition processing result. (2) The information input device according to claim 1, wherein the statistical frequency information regarding the input target information includes appearance frequency information, connection frequency information, and the like. (3) Restriction of the input target information to be subjected to recognition processing is performed by excluding input target information whose statistical frequency information is below a predetermined threshold value from the recognition target information for the input pattern. Information input device described in section. (4) The information input device according to claim 1, wherein the input pattern consists of syllables or kana characters, and the input target information consists of words or phrases consisting of the series of the syllables or kana characters. (5) The information input device according to claim 1, wherein the restriction on input target information to be subjected to recognition processing is lifted as necessary.