JPH07271774A - Document input device - Google Patents

Abstract

PURPOSE:To provide the document input device which has a function for lightening the burden of input by predicting and displaying a word appearing next by using information on the meaning that a word that is already inputted has. CONSTITUTION:This document input device includes an input means 1 for inputting a document, a document buffer 2 which stores the inputted document, a meaning vector generating means 3 which generates meaning vectors of the document from the document stored in the document buffer 2, a word predicting means 4 which predicts following words from the meaning vectors of the document, a meaning dictionary 5 which is used to predict the words, a predicted word display means 6 which displays the predicted word, and a word selecting means 7 which selects one word out of the predicted words. Then the word predicting means 4 is equipped with a function which compares the similarity between the meaning vectors of the document and the meaning vectors of the words described in the meaning dictionary and selects the next word. Further, the word predicting means 4 has a function for narrowing down the predicted words by using information on the character which is inputted after predicting the words.

Description

Detailed Description of the Invention

[0001]

The present invention relates to a personal computer / electronic notebook /
The present invention relates to a text input device such as a word processor.

[0002]

2. Description of the Related Art Conventionally, as a function (learning function) for reducing the input load by utilizing data that a user has input in the past, words already input in a sentence at the time of Kana-Kanji conversion are
There is a method of placing it higher in the conversion candidates. In addition, a method has been put into practical use in which a word that has already been input appears as a candidate when the first character is input and the conversion key is pressed.

A technique using statistical information is also known (“Statistical language model using kana / kanji character chain information”, Tomokazu Yamada and 3 others, Proceedings of the Acoustical Society of Japan (1993). March 1993), pp. 95-96), but the information used here is information in character units, not as words.

[0004]

When inputting a sentence in a word processor or the like, it is necessary to input a correct kana in each character. As a method of reducing this load, a method of setting a word already input in the sentence at the time of Kana-Kanji conversion as a higher rank as a conversion candidate, or a word already input by inputting the first character and then the conversion key In the method of pressing and displaying as a candidate, the minimum condition is that the same word has already been input. However, the probability that the exact same word appears multiple times is low.

On the other hand, some documents have patterns that are deeply connected and extremely easy to appear depending on the content. For example, if you have typed "Dear Sir, Masumasu Masumasu", the next word will often be "Goseiei no" or "Goseisho no". Also, "long-term research plan"
The word "planning / promotion / reviewing / submitting" is likely to appear relatively frequently.

Therefore, it is expected that the input load can be reduced by predicting / displaying the word that appears next using the semantic information of the words that have already been input.

An object of the present invention is to provide a text input device having a function of predicting / displaying a word that appears next by using the semantic information of a word that has already been input and reducing the addition of input. To do.

[0008]

The text input device of the present invention creates an input means for inputting a text, a text buffer for storing the input text, and a meaning vector of the text from the text stored in the text buffer. A meaning vector creating means, a word predicting means for predicting the next word from the meaning vector of the sentence, and a semantic dictionary stored for predicting the word,
The prediction word display means for displaying the predicted word and the word selection means for selecting one from the plurality of predicted words are included.

The word predicting means has a function of comparing the meaning vector of a sentence and the meaning vector of a word described in the meaning dictionary to select the next word.

Further, the word predicting means has a function of narrowing down the predicted word using the information of the character input after predicting the word.

[0011]

When the user inputs a sentence to some extent, the system then predicts the word and displays the prediction result. If the word desired by the user exists in the displayed predicted words, the user selects the word by using a selection method such as a key or a mouse. If it is not in the predicted word, enter it with the key as in a normal word processor. If it is in the predicted word, it is input only by selecting the word, so the input load is reduced. In addition, by performing prediction, words that are semantically connected are displayed, which helps the user in creating a sentence.

[0012]

DESCRIPTION OF THE PREFERRED EMBODIMENTS The text input device of the present invention will be described below.

FIG. 1 shows a block diagram of an embodiment of a text input device of the present invention.

In FIG. 1, reference numeral 1 is an input section for words or sentences. For example, it corresponds to a keyboard, a pen, a tablet, a character recognition device, a voice recognition device, and the like. Two
Is a sentence buffer that stores the input sentence. Normally, a part of the contents of this buffer is displayed on the display device, but it is omitted here. Reference numeral 3 is a semantic vector creation unit that creates the semantic features of the input sentence from the contents of the sentence buffer. The meaning vector of the sentence created here does not have to be one, and a plurality of semantic features may be created. Reference numeral 4 is a word prediction unit that uses the meaning vector of the sentence created in 3 to predict words (or word strings, clauses, sentences, etc.) that are closely related to the meaning vector of the sentence. The word prediction unit has a function of selecting the next word by comparing the degree of similarity between the meaning vector of the sentence and the meaning vector of the word described in the later-described meaning dictionary.
Reference numeral 5 is a semantic dictionary used in blocks 3 and 4.
Reference numeral 6 is a predicted word display unit that displays the word predicted in 4. Reference numeral 7 is a word selection unit for the user to select a word from the displayed words. If the word the user wants to input is not in this display, this selection is not made, and the word is input as usual (from the keyboard or the like). Further, the word selected by this selection unit is stored in the sentence buffer in the same manner as that input from one input unit and becomes a part of the sentence.

Next, an example of word prediction in the case where a sentence of "long-term research plan" is input will be described.

First, an example of the semantic dictionary shown in FIG. 2 will be described. Here, nine words are shown as a representative. In the dictionary here, each word is represented by a vector consisting of six elements in the horizontal direction. Creating a semantic dictionary is
It can be created by giving meaning to each element and manually giving the relationship with the element, and as described in the patent application (Japanese Patent Application No. 5-181265) filed by the same applicant, It can also be created automatically from the data.

In this example, the word "long-term research plan" has already been input, so the meaning vector of the sentence is created using this word.

FIG. 3 shows an example of the calculation of the meaning vector of the sentence when the sentence "Long-term research plan" is included.

From the semantic dictionary of FIG. 2, the vector of each word forming the word "long-term research plan" is long-term (8,
1, 2, 4, 3, 2), research (7, 9, 4, 7, 2,)
1), plan (2,5,7,3,1,1), and the semantic vector of the sentence is calculated by averaging these already input vectors (however, the decimal point is rounded off). That is, (8 + 7 + 2) /3=5.6, (1 + 9 +
5) ÷ 3 = 5, (2 + 4 + 7) ÷ 3 = 4.3, (4 + 7
+3) ÷ 3 = 4.6, (3 + 2 + 1) ÷ 3 = 2, (2+
1 + 1) /3=1.3. As a result of this calculation, the average (6, 5, 4, 5, 2, 1) of these three vectors becomes the meaning vector of the sentence.

Here, although the average is taken as the meaning vector of the sentence, various methods such as a weighted average can be considered as the method of creating the meaning vector of the sentence. It is also possible to create multiple meaning vectors, such as creating both the meaning vector of the entire sentence and the meaning vector of the sentence considering only the neighborhood (words close to the current input).

Next, the next word is predicted from the created meaning vector of the sentence. At the time of prediction, by calculating the similarity between the calculated meaning vector of the sentence and the meaning vector of each word described in the meaning dictionary, some are selected from those with a large similarity (small distance). realizable.
It is conceivable that the number of words to be selected is a fixed number in descending order of similarity. Further, it is also possible to select one having a (similarity of target words / maximum similarity) of a certain threshold value or more. In addition, the maximum similarity refers to the maximum value among the similarities of each word.

The similarity between the meaning vector of a sentence and the meaning vector of a word can be calculated using the Euclidean distance (square error), and as shown in FIG. 4, the meaning vector of the sentence (6, 5, 4, 5, 2, 1) in order of decreasing distance (planning: distance: 31), propulsion (same: 31), review (40), watermelon (74), science (89), list (10)
7) (words already in the input sentence are omitted).

If three words are displayed in the order of similarity among them (here, three words are set to be displayed in descending order of similarity), words of planning, promotion, and review are displayed as candidates. . If the user wants to input these three words, the user selects those candidates by using the keys or the mouse. For words other than these words, input is performed using conventional kana-kanji conversion or character recognition.

FIG. 10 shows a flow of the operation procedure of the text input device of the present invention.

In step 1, it is checked whether it is the beginning of a sentence. It is assumed that the topic is different for each sentence, and at the beginning of the sentence, the sentence vector is initialized at 2. This means that all the elements of the meaning vector indicating the topic of the sentence are 0.
It can be realized by At 3, it is checked whether a word has been input. When a word is entered and confirmed,
At 4, the meaning vector of the word is extracted from the meaning dictionary. In step 5, a new sentence vector is created by using the meaning vector of the current word, using the meaning vector of the sentence (sentence vector) obtained and extracted based on the meaning vector of the already input word. In other words, when the input of 3 words is confirmed,
The sentence vector is updated at 4 and 5. At 6, it is checked whether the word prediction key has been input. When the prediction key is input, the sentence vector obtained in 5 and
The degree of similarity between the word and the meaning vector stored in the meaning dictionary is calculated as 7. In 8, the results obtained in 7 are sorted in descending order. In 9, some are displayed as predicted words in descending order of similarity. At 10, it is checked whether the user has selected a candidate from the displayed words. When the displayed candidate is selected, the word is confirmed as an input character at 11. If a key other than the candidate selection is entered, the candidate word display is erased and the next entry is awaited.

In the above word prediction example, a method of displaying a word having a meaning vector close to the meaning vector of the input sentence is shown. However, in ordinary text input, more than 10,000 words are used. If the accuracy of the prediction is sufficiently high, there is no problem, but if the accuracy of the prediction is not high, the load is not so different from the load of the conventional non-prediction method.

Next, a method for improving this point will be described. In the above example, when the sentence “of the long-term research plan” is confirmed, three words are predicted and displayed. In this improvement method, when the next one character is input, the prediction candidate is limited using that one character. In other words, the previous 3 words (planning / promotion / review) are displayed as the prediction result when there is no input, but if "ri" is input to the next character, it will start with that character. Limited to words of things. In the example of FIG. 2, since the distance is small and the word starting with “ri” is only “planning”, only “planning” remains as a candidate display. If it is not a character-by-character unit, but Roman characters are input, each information is used even when "r" in the planning is input or when two or more characters such as "Ritsu" are input. It is possible to limit it.

Even if there are a large number of words with a small distance at the stage of no limitation, the candidates are limited by using the next input to perform the minimum input, and then the word of the prediction result is obtained. Can be used, and the input load is reduced.

A block diagram of the embodiment in this case is shown in FIG.

In FIG. 5, each of the input unit 1, the sentence buffer 2, the meaning vector creation unit 3, the word prediction unit 4, the meaning dictionary 5, the predicted word display unit 6, and the language selection unit 7 is shown in FIG. Although it has the same function as in the block diagram, the word prediction unit 4 has a function of narrowing down the predicted word by using the information of the character input after predicting the word in addition to the above. Therefore, after using the meaning vector of the sentence created by the meaning vector creating unit 3 to predict words having a close relationship with the meaning vector of the sentence by the word predicting unit 4, in order to narrow down this predicted word, the next character Indicates that you must enter.

FIG. 6 shows a block diagram of means for creating a semantic dictionary.

In FIG. 6, 1 is a sentence storage unit that stores a document input in advance, 2 is a sentence analysis unit that performs morphological analysis of the document stored in the sentence storage unit 1, and 3 is a sentence analysis unit 2.
Sentence analysis dictionary used to perform morphological analysis of 4
Is an independent word storage unit that stores an independent word as a result of analysis by the sentence analysis unit 2, and 5 is an independent word sequence stored in the independent word storage unit 4 and a shared co-occurrence registration dictionary that will be described later. A co-occurrence relation creating unit that creates a co-occurrence relation table, 6 is a co-occurrence registration dictionary that describes which independent word is used when creating a co-occurrence dictionary, and 7 is a co-occurrence relation creating unit. It is a co-occurrence dictionary storage unit for storing the created co-occurrence relation table (co-occurrence dictionary). Details are described in a patent application (Japanese Patent Application No. 5-181265) filed by the same applicant.

FIG. 7 shows a conceptual diagram for explaining the creation of the semantic dictionary.

Now, let us say that the inputted text is "The goal of the reorganization of the industry is to expand the management base by merger and reduce excessive facilities" (Fig. 7a). The input sentence is analyzed by the sentence analysis unit 2 and the independent word “industry, organization, goal,
Merger, management, foundation, expansion, excess, equipment, reduction "(Fig. 7
b) is extracted. The extracted independent word is the independent word storage unit 4
Memorized in. After inputting a sentence, the independent word storage unit 4
A co-occurrence relation is created by the independent words stored in. In the input example sentence of FIG. 7, the extracted independent word described below is obtained. From this extracted independent word, co-occurrence relation table (= semantic dictionary)
(FIG. 7c). In this case, two independent words are extracted from the extracted independent words, and the two independent words A are extracted.
And the independent word B. In the co-occurrence relation table, 1 is added to the square corresponding to the combination of AB. In other words, we assume that combinations of words that appear in the same sentence have some relationship.

For example, if "industry" and "composition" are selected, the horizontal axis (row direction) of the co-occurrence relationship is "industry" and the vertical axis (column direction) is "composition". (Figure 7c
1 is listed in). A co-occurrence relationship table is created by extracting all two independent word pairs from the extracted independent words and adding 1 to the corresponding co-occurrence relationship table element. In this way, the semantic dictionary is created.

FIG. 8 and FIG. 9 show an example of creating a sentence for creating a research plan discussion material.

In FIG. 8, "About the research plan of this year,
If you press the key for "Word Prediction" at the stage where "Development" has been entered, the word prediction result of the portion below it will be displayed. Currently, the number one forecast is "Schedule",
"Schedule" is also displayed in the text as a candidate for a word to be input. If the next candidate key is pressed, the next prediction candidate "plan" is displayed instead of "schedule". If the cancel key is pressed, all prediction results will be erased and will be the same as normal input. If it is canceled and a key other than the next candidate is input, the prediction word currently displayed is confirmed and the prediction result is erased. "Word prediction"
The keys such as "Next candidate" and "Predictive cancel" do not have to be independent, and may be used together with the conversion key for Kana-Kanji conversion or the cancel key.

As in FIG. 8, FIG. 9 shows a case where “development about the research plan of this year” is input and one character “MO” is further input. If the "word prediction" key is pressed after "mo" is input, only words starting with "mo" are displayed as the prediction result. Thereby, more accurate prediction can be performed. Of course, it is also possible to output the prediction result without pressing the word prediction key. In this case, the current prediction candidate is always displayed somewhere on the screen. Each time a character is input, the information is used to update the prediction candidate. If there is a word to be input in the candidates, the user can input the word by selecting it. Of course, instead of words, it is possible to use example sentences as prediction candidates.

[0039]

[Effect] According to the present invention, by predicting a word or a phrase or a sentence to be input next from an already input sentence and letting the user select it, only the selecting operation is required without inputting all the sentences. It is possible to input words and reduce the input load.

[Brief description of drawings]

FIG. 1 is a block diagram of an embodiment of the present invention.

FIG. 2 is a diagram showing an example of a semantic dictionary.

FIG. 3 is a diagram showing an example of calculation of a semantic vector of a sentence.

FIG. 4 is a diagram showing a distance according to a semantic vector.

FIG. 5 is a block diagram of another embodiment of the present invention.

FIG. 6 is a block diagram of means for creating a semantic dictionary.

FIG. 7 is a conceptual diagram illustrating creation of a semantic dictionary.

FIG. 8 is a diagram showing a display example according to an embodiment of the present invention.

FIG. 9 is a diagram showing a display example according to another embodiment of the present invention.

FIG. 10 is a diagram illustrating a flow of an operation procedure of the text input device of the present invention.

[Explanation of symbols]

 1 Input Section 2 Text Buffer 3 Semantic Vector Creation Section 4 Word Prediction Section 5 Semantic Dictionary 6 Predicted Word Display Section 7 Word Selection Section

Claims (3)

[Claims] 1. An input unit for inputting a sentence, a sentence buffer for storing the inputted sentence, a meaning vector creating unit for creating a meaning vector of the sentence from the sentences stored in the sentence buffer, and a meaning vector for the sentence. A word prediction means for predicting the next word, a semantic dictionary used for predicting the word, a prediction word display means for displaying the predicted word, and a word selection for selecting one from a plurality of predicted words A text input device comprising means. 2. The word predicting means has a function of selecting the next word by comparing the similarity between the meaning vector of the sentence and the meaning vector of the word described in the meaning dictionary. The text input device according to claim 1. 3. The text input according to claim 1 or 2, wherein the word predicting means has a function of narrowing down predicted words using information of characters input after predicting the words. apparatus.