JPH0367367A - Document answering method - Google Patents

Abstract

PURPOSE:To simplify the constitution and reduce its size by storing a group argument point sentence obtained by converting a document, retrieving a new argument point sentence according to the storage contents, and performing similar retrieval by using a conversion area when the retrieval result is not an answer sentence. CONSTITUTION:The input document is converted into argument point sentences through a RAM B for conversion under process control by a microcomputer and the group of the argument point sentence is stored in a RAM A. The microprocessor retrieves the new argument point sentence in the argument point sentence group and when it is decided that the new argument point sentence describes the fact, it is outputted as an answer sentence. When the sentence is not the answer, plural argument point sentences are stored hierachically from the RAM B for conversion and the RAM A and when there are plural argument point sentences in the same layer, similar processing is performed with the individual argument point sentences; and the layer of the argument point sentence which becomes the answer is tracked back and opened and this free area is utilized similarly. Therefore, the document answering method of simple, small-sized constitution is obtained by using a combination of the small- capacity storage part and microcomputer.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention provides a text input/sentence response system that operates on, for example, a personal computer or a microcomputer system using an artificial intelligence programming technique. The object of the present invention is to enable interaction with text data input into the system by inputting and outputting the text responses in natural language commonly used by humans.

What are the requirements for the text response invention explained in this invention example?

Unlike the methods using large computer systems that have been studied in this field, it is possible to produce powerful text responses even in programs running on very small computer systems, such as personal computers, with very small storage space and information processing capacity. In order to obtain output, the arrangement of the storage space on the storage device, the program structure, and the same path structure are devised in the structure of the present invention.

Conventionally, this kind of system has been used to unilaterally output only the information stored in one sentence, for example, by pre-recording a digital signal on an audio tape.

There are examples of systems that store information corresponding to a sentence, simply search for that part, and output necessary information 1, for example, answer information as a learning system.

Furthermore, the electronic device 9 for creating sentences 9, such as a Japanese word processor system, simply converts input such as katakana into kanji, outputs and displays the sentence, and inserts or deletes 9 characters in response to typos in a sentence. However, it has not yet been realized.

The learning system, word processor, etc. both have the function of simply searching for or changing pre-stored sentences and outputting them.

The present invention differs from the above-mentioned conventional methods in that when a question sentence is input for an input sentence or a group of sentences stored in advance, the group of sentences is broken down into thesis corresponding to the meaning of the sentence, and the answer is logically determined. It provides a system that infers a sentence that holds true based on the above-mentioned thesis, reshapes it, outputs it, and makes an announcement.

Below, we will proceed with the explanation by considering anything that expresses some meaning as a string of words to be a sentence. The objects of the present invention are as follows.

1) As a language learning machine, it inputs two groups of sentences, outputs automatically formed sentences that are semantically the same, and uses it to learn how to compose a single sentence.

2) Various information generated within the company (1), such as the meeting schedule behavior of multiple people, is input in natural language, questions are asked using natural language input, and text about the schedule behavior is output.

In this case, you can specify only the person who takes a specific action and output it in natural language.

3) By connecting to a database program or database system in a programmatic or software manner, information that could previously only be retrieved in terms of words or units within the database information can be logically analyzed and inferred from the meaning of the question writer's ability. It is possible to output two-word sentence data with a wide range of logical meaning.

For example, ``Mr. A will go to Tokyo on the 10th.'' ``Mr. 3 will go to Osaka on the 10th. Mr. J rC will go to Tokyo on the 11th.''

”, “If it is Tokyo, it is the head of Japan”, etc. are input.

The next question was, ``Who will go to the head of Japan on the 10th?

Even when you enter "Who is going to Tokyo on the 10th?"
For example, when inputting ``Mr.
Mr. A will go to Tokyo on the 10th. " is displayed in the output.

This has been difficult to achieve with conventional database search systems. By realizing this, the existing database can be utilized more effectively.

The structure and operation of one embodiment of the present invention will be explained below using FIGS. 1 to 14.

FIG. 1 shows a schematic configuration of the present invention, and FIG. 2 is a block circuit diagram thereof. The present invention provides an electronic device (1) such as a personal computer or word processor having a microprocessor (7), e.g. 8 or 16 bits, and a working area storage capacity of e.g. 256 bytes or less. 2) Start the Japanese input/output program. Please note that the language is not limited to Japanese, but English.

Although the present invention is also effective in German, the following explanation uses Japanese as a natural language.

Next, in step (3), the text response program, which is the main element of the present invention, is activated. When the program starts, multiple sentences are input in step (4). This is given in advance to the device (1) from, for example, the keyboard (6) as knowledge information.

For example, it may be provided from an external storage means such as a floppy disk (5).

Next, when the question text input step (8) is performed, an answer text in the form of text is output in step (9) according to the configuration and operation described below. In this case, the output is CRT (
10) shows an example of displaying sentences.As an example of a practical sentence response, an example of Japanese logic composition learning for elementary school students will be explained below.

In FIG. 3, for example, a group of sentences (31) and (32) as shown in the screen are entered into the CRT (10) using the keyboard (6) and displayed. Among the sentence groups, the sentence group (31) such as ``A lion is a mammal.'' is a factual sentence with a one-to-one correspondence between the nominative part and the predicate, and the sentence group ``If it is an animal, it is a living thing.'' etc. 32) are conditional sentences necessary to draw inferences, and one sentence group (32) has features connected by words stored in the conditional word dictionary (106), which will be described later, such as "nara" and "tatara". be. (33) in the diagram is C
This is the RT display cursor.

Next, the question text will be explained. Figure 4 shows the question text that is input after the sentence groups (31) and (32) in Figure 3 are input, and the question text (34) is ``What are your friends?'' It shows that For this question, the semantic answer to the question is reconstructed from a group of sentences (31) (321) using the method structure explained below with reference to FIG.
Penguins are friends. ” CR for “Elephants are friends”
T (output is displayed on 101. In this case, the output is CRT
[Not limited to 101, output audio sentences or actions corresponding to the sentences.

For example, it is possible to connect it to a robot motion drive signal.

Furthermore, input text groups (31) to (32), etc. are not limited to keyboard input, and various inputs 9 such as voice recognition input devices may be connected.

FIG. 14 shows output means according to another embodiment of the present invention.
An example of driving a movable part is shown. That is, in the other embodiment shown in FIG. 14, data is connected to the electronic device (1), which is a component of the present invention, by an external interface (2011, for example, a common bus such as R5232C, and +2011 is for example a lion, a penguin, etc.). Painted display (20
2) The corresponding lamps (2031 (204)
) (2051 is electrically connected to the interface (201). Therefore, in this example, "Penguins are friends." is used as the answer response text (35) at the answer text output step (9). Extract the word data of penguin and elephant from the answer essay "Elephants are friends," start the output interface, and light up the penguin lamp (204) and the elephant lamp (205).In this case, as described above It is also possible to output the sounds of elephants and penguins by generating the sounds of By showing it on a display, it can play a very important role in education, such as having a great learning effect and allowing you to experience the realism of writing a single word sentence.

Hereinafter, FIG. 6 shows text processing according to an embodiment of the present invention.

Inference and detailed processing methods will be explained.

FIG. 6 shows an input natural language sentence (101).

For example, it explains how to extract the main argument from a sentence such as ``If it's a quiet creature, it's definitely a friend.''

The input sentences (101) are "yabari", "apparently",
The character-digital data and the character string of one character-digital data corresponding to the words of the Excision Word Dictionary (102), which is a dictionary in which words that are clearly unrelated to the purpose of the argument, such as "I bet", are stored in advance. By comparing, the string is deleted.

In addition, the words in the question sentence ending dictionary (104) are stored in advance as dictionary contents, including endings that a question sentence should normally have (for example, "~ka.", "~?", "~kke", etc.). The digital data of the input sentence and the digital data of the character string at the end of the input sentence are compared, and if they match, it is deleted and the sentence is processed as a question sentence. In addition, there is a regular sentence ending dictionary (103) that stores in advance the endings that normal sentences would have (for example, "~da.", "~ deru.", etc.), and the endings of input sentences. By comparing , you can delete it and treat this sentence as a normal sentence from now on.

By applying each of the above dictionaries (1021 (1031 (104)) to the input sentence, the input sentence is converted into the sentence ``If you are a quiet creature, you are my friend.''

From the converted sentence (105) converted in this way, first,
"Words that have the role of connecting the conclusion to be discussed (A) and the sentence to be discussed to derive it (B)" (hereinafter referred to as conditional words). For example, "Nara", "Nara", "Dara", ..., etc.) in a dictionary stored in advance, that is, a conditional word dictionary (106). Conditional word dictionary (
If a word registered in 106) is detected, the input sentence is divided around that word (in this example, "nara"),
The first part (``a quiet creature'') is treated as B (1101), and the second part is treated as A (1091).If not detected, the entire input sentence is treated as A and the previous part. .

Also, in the previous part above, there is a compound sentence dictionary (107) (its contents are, for example, "~de,""~kute,""~noni"...
By using a word that plays the role of connecting sentences such as , etc., it is possible to detect and classify how many sentences of the above B type are connected (G pieces), memorize the count of G pieces, and write the following. It becomes possible to utilize this information as effective information to control inference through processing.

Furthermore, the (case) particle dictionary (1081 (contents are ``ha'', ``ga'', ``datte'', ``datte'',
(words with a particle role such as "also") are used. (As an example of a comparison method in the case of Japanese, if you want to compare strings sequentially, it is preferable to compare them sequentially from the end of the string.) Dictionary (If a word in 1081 is detected, use it as the center. It is also divided into a front and a back, with the front being treated as the subject and the back being treated as the predicate.

In this example, the subject and predicate are treated as -words, but the subject and predicate can be used multiple times as needed ("wo", "ni", "he", "kara")
It is possible to break down a more complex sentence into a gist sentence by using the ``no'' etc.), and also to identify what pattern of particle appearance the input sentence has. This information can be effectively utilized through the following logical development.

Now, if the word in the particle dictionary (108) is not detected, the system recognizes that the subject has been omitted, and the sentence in which ``O'' appears in the first argument of a complex sentence (this is the case in this example). In this case, the system automatically supplements a special main part (112) indicating main part omission such as rXJ. Sentences in which ``O'' does not appear in the first argument of a complex sentence (for example, ``You're an idiot and an idiot.''
), supplement it with a pronoun such as "it". Furthermore, in the case of the second and subsequent sentences in a complex sentence, the subject of the previous statement is supplemented as the subject.

By using such a means, a single-person sentence is decomposed into a group (1111) of thesis sentences necessary for the following logical development.

As mentioned above, the reason for having separate multiple dictionaries is 1 due to the characteristics of Japanese9. For example, phrases that are unique to the 9th ending only appear at the end of the argument, or excision words such as ``yabbari'' are not included in the sentence. This is because it is expected that it would be more efficient to cluster them based on differences in character, such as phrases that may appear anywhere in the world but do not affect logical development, and have them separately as a means of realizing the 9 functions. .

Furthermore, most of the dictionaries for conventional word processors have phrases that contain so-called stem words, such as ``animal'' or ``today,'' but the present invention, on the contrary, has phrases that are not stem words as a dictionary, and can be used effectively. It is intended to be applied to processing. The reason for this is 9. Due to its characteristics, the Japanese language uses a type of operator, typically a particle, to connect word stems that correspond to some kind of image, action, or concept, and then finally adds a ending that characterizes the entire sentence. This is because it is a type of language in which sentences are constructed, and so-called adjuncts, typically particles, play a role in influencing the logic of sentences.

Now, 2) in Fig. 6 is an example of a question (1131 inputs).

In the case of example 2), by passing it through the same processing means as in 1) in Figure 6, in this case, the ending of the word "~ka" is detected and recognized as a question sentence, and the ``C'' which indicates a question sentence is detected. 116) and is stored in a different location from A and B.

In addition, in the case of question sentences, unique phrases such as ``what'' and ``who'' appear. In the following processing, these words express the procedure that what = any word can be included, and who = any word can be included, but in the case of non-humans, the nuance will be biased. It is processed. This is also an essential process for natural Japanese processing.

In this way, the input text is decomposed into thesis sentences with only the thesis as the core, and each sentence is stored together with information necessary for controlling the process.

Therefore, FIG. 7 shows a part of the result of the text input example shown in FIG. 3 being decomposed into a thesis statement.

In Figure 3, the sentence ``Lions are mammals. Elephants are also very docile.If they are mammals, then they are living things.
Sentence number (120) that counts the number of sentences and information indicating the presence or absence and number of Otsu type thesis (123) (1
221 (For example, “The lion is ~” in sentence number l in Figure 7)
``0'' (124) for sentence number 11, ``2'' (126) for sentence number 11, ``Be quiet,...''), and the thesis statement A and B (if the number of conditions is 1 or more) are each Memorized and stored.

Therefore, as a means for realizing logical expansion processing, which is the core of the function of the present invention, the following describes how to specifically allocate this information in the memory of an electronic device and how to apply specific processing means.

This is what I will explain.

FIG. 8 is a diagram showing an example of the storage arrangement of the thesis and necessary information and its contents.

RAM A is a diagram showing an example of arrangement of the gist of an input text and necessary information.

RAM B is the first. FIG. 9 is a diagram showing an example of the storage arrangement of the second pointer and the newly converted thesis statement number.

As shown in FIG. 8, two storage means, RAM A and RAM B, which are functionally different, are provided as means for realizing one function. In this case, the RAM has the same functions and means even if it is an ROIJ in which information is written in advance, or whether it is held as an equivalent memory on various external storage devices via an interface.

In other words, RAMA is a means to realize the function of allocating and storing the thesis statement and necessary information of the input text, and a means to realize the function of logically developing the texts using the information in RAM B & iRAM A. The memory allocation for concrete progress shows the means.

Now, referring to FIG. 8, one embodiment of RAM A will be explained first.

The RAM ^ has three areas (Fig. 8, 139, 140, 141) corresponding to A (Fig. 7, 121), O (Fig. 7, 123), and C (Fig. 6, 116), respectively.
) is a thing.

A is the address ADD corresponding to the sentence number (Figure 7 120)
(131) Separately, 1 sentence number (130), thesis statement (132), and the number of 9 conditions (hereinafter referred to as G) (133)
, and if G is 1 or more, it is composed of a pointer (134) indicating the address of Party B.

(135) is a pointer (134) indicating Party B's address
This is a line illustrating a state in which "B" indicates the address of Party B.

In addition, Otsu (140) in Figure 8 is composed of a group of arguments assigned to the address ADD Otsu (1361) for each of Otsu's single purports, and the above (134) indicates the corresponding ADD Otsu of these Otsu. By pointing, it is connected to the instep.

In addition, C (1411, that is, the question essay) is stored as a thesis statement or a group thereof, similar to B, and is stored at a different address from A and B.

In the example, if a part of RAM A in FIG. 8 is 1, for example, if it is a sentence such as ``I am a quiet creature, I would definitely be friends with it.'' It passes through the above method and becomes 1 sentence number 10. A's thesis statement is r FXJ, Friend', and B's number of sentences (G) is 2.
．． The pointer points to ADD B, and to that point, G is 2
Therefore, ADD Otsu and ADD Otsu+1 r IrXJ
, docile'' and r 1rXJ , organism''.

Next, RAM B in FIG. 8 will be explained.

RAM B is used as a means to implement the logical expansion function.

It is characterized by having three areas Dq, PT and N.

First, the area TQ (142) is allocated as an address, for example, TQI-TQx~, and the content of each address is "Sentence number creation of the matching thesis statement" as detailed in the flow below. ~m, (1451J area for storing thesis statement newly converted and generated internally
146).

Furthermore, area PT (143) is an area for storing the first pointer (147) having the above address TQx as its contents, and the number and allocation method and means of converting the contents will be explained in the flow below. However, each PT, ~
It is characterized by having one or more storage locations assigned addresses PTADD~(143).

Further, in RAM B, in addition to nine or more areas, there is a storage area for a second pointer (144) whose content is the address prAoゎ (=N) of the storage location of the first pointer.

The content of (148) is PTAn. This is a line representing a state pointing to (=N).

In addition to these RAMs A and B, storage tables for work and answer output are required for the following processing.

They are omitted because they can be fully understood without illustrations. Function 9 of this memory table will be discussed in the flow below.

Example of thesis statement shown in Figure 8 of RAM B (146)
9 is an example showing the stored contents at a certain point in time (the procedure will be described later) in this embodiment.

In this embodiment, RAMB is singular, but RAMB is
There are multiple L-BNs, each with a communication means (e.g.
An independent CP with an IBM model 3270 terminal
Controlling by U is also one of the easy extensions of the present invention.

Hereinafter, specific means for realizing one function will be explained in detail using flowcharts according to FIGS. 9 to 12.

The following functional implementation means specifically shows how the RAM H allocation shown in FIG. 8 and the logic expansion means are implemented on the electronic device.

As described above, the procedures shown in FIGS. 9 to 12 are started after RAM allocation is completed. (1
511 Although not illustrated in this embodiment, processing for changing, deleting, or adding the contents of RAMA during or at the end of the processing shown in Figs. 9 to 12 can be easily added to the two procedures. By doing so, it becomes possible to develop logic that is more complex than the one example.

Now, starting from step (151), in step f1521, first, an initial value is set in the second pointer (contents are N) (1441) of RAMB. Normally, the initial value 1 is desirable.

Further, the contents of the variable HP that records the maximum value of the pointer and the first pointer (143) are set to zero as initial values.

After that, in step (1531), the question thesis statement (C) (138) of RAM A is loaded into the storage location of address TQI in area TQ (142).

The steps up to this point are performed once for each question input sentence. The following process is repeated as many times as necessary (specific means will be shown below) depending on the content of the question text.

Now, when the procedure of step (253) is completed, the procedure of step (155) is performed. In step (1551), 1 is added to the content N of the second pointer and the above HP is assigned to the content of the first pointer (143) whose primary address is PTIN+.

Then, out of (1451 areas (hereinafter referred to as IIR) of area TQ, *mptt+i-++*+~f
Set f1py++u to zero.

Before the processing in step (155), there is a return exit from the back (FIGS. 11 and 12) called wa f154).

Also, after the processing in step (155) (see Figure 12)
Note that there is another return loca (156) from .

Next, step (as 157+, create with the following process,
The thesis conversion table (T) to be converted is returned to its initial state in which nothing is stored.

Then, as in procedure A shown in step (158), TO
Search the thesis text from py+s-u*+ to TQprts+, and use words that appear more than once (for example, (1) in Figure 8).
46) The word "what" corresponds to this) is memorized in r! l (address α) respectively in table T
to be memorized.

The flow description will now turn to FIG. 1O.

After step (1581), first, in step (160), loop counter J is set to the contents of PT+N-0+1.

Then, the return lore (17) from the back (from Figure 12)
0), and in step (161) it is compared with whether J≦“contents of PT IN+”.

If J≦PT+s+, branch to step (162).

If J > PTIN+ content, branch to mouth (beginning of Figure 11).

Hereinafter, steps (1621 to (169)) are looped as a group, so steps (162) to (169)
) will be explained first, and then the processing below (FIG. 11) will be explained.

First, in step (162), the contents of mJ in the area TO (initial value is 0) are assigned to the second loop counter K.

In step (163), 1 is added to K.

Then, in step (1641), a comparison is made between K and the maximum value of the sentence number (1301nx) in the RAM ^ (referred to as MAX(nx)).

If ≦MAX (nxl, step (165
), and if K > MAX (nx), it branches to evening (Figure 12).

In step (165), TQ's thesis statement and RAM
The thesis statement of A's A whose address is stored in the position of K (referred to as A's A) is compared.

In this case, it is natural if the digital data of the two wheel names match completely, but it is allowed to consider cases where the digital data are different even if they are considered to be a match. Setting the degree to which the two are considered to be a match is a matter of discretion depending on the system design. In this example, as just one example, the word "X" (a special subject (or predicate) supplemented by the system) in Party A's thesis statement can be used with any word in the corresponding thesis statement on TQJ's side. Match. Also, “
It is set that the digital data of the word "What" is recognized as a match regardless of the digital data of the corresponding word in Party A.

In the flowchart, the symbol 4 is used to indicate that when there is a semantic or logical match, the match is recognized even if the digital data is different.

Now, in this way, the two are compared in step (165), and if they match, the process proceeds to step (166), and if they do not match, the process returns to step (163). Therefore.

In step (166), it is tested whether there is a contradiction between the words TQ and A according to the "position information (α)" stored in the table in procedure A of FIG. In procedure ^, the position where a common word must appear is memorized, so in procedure B of 2 steps f166), by performing the test using this position information, the number of conditions (B) (G ) even if there is more than one.

It is possible to prevent the conditions from causing problems in logical development.

Then, step (1671), if there is a contradiction, return to step knob (1631, if there is no contradiction, step (16)
8).

In step (168), the table T corresponding to the above α is added to the table T which has stored only one position information so far.
Make qJ and A memorize the words. Furthermore.

Detect whether the answer can be assigned to a table prepared for outputting the answer, which is different from T (for example, in the case of the example, T
If the word in QJ is "what" and the word in A is other than "x", the word in A is the answer, and this is substituted into the answer table] and substituted into the answer table.

Next, in step (169), K is substituted as the content of +1llJ, 1 is added to J, and step f1611
Return to

In FIG. 11, in step f161) of FIG. 1O, J>P
If the content is T4.41, processing starts from the branched point.

In step (1711), zero is assigned to the variable IDX that controls whether or not to output an answer in the following process, the contents of PT(H+ are assigned to the aforementioned HP, and PTIN is assigned to J.
-Substitute the contents of I+.

Step (Add 1 to J at 1721.

In step (173), it is determined whether the content of J≦rPTts+ is satisfied.

, 1≦PT+Nl, proceed to step (174).

If J>PTnu contents, branch to step (1791).

Now, if you proceed to step (1741), substitute mJ to K and step (1741).Next, in step (175), find out how many sentences in A that have K as the sentence number are stored in RAM A. Check whether the number of conditions (that is, the number of conditions GKI) is 0.

In step (175), if GK=0, return to step (172); if GK=Q, step (
176) and add 1 to IDX.

Then, as in procedure D shown in step (177), the word TQJ corresponding to the word A is stored (memory r).

At that time, if the word TQ matches the word in the table created in procedure C above, memorize the word A corresponding to TQ in that table instead of the word TQ. (Procedure D).

Then, in the ninth step (178), R is applied to TQHp-+ to TGHp, ax in the TQ area.
Load the thesis text of AMA Chuotsu ADDK ~ Otsu ADDK+ GK.

Then, according to the memory r in procedure D above, TQ□□
1~TQ□ya. Convert the word K.

At this point, the input question thesis statement has been converted into a new thesis statement and stored.

For example, in this example, "What is a friend?" thesis "What""Friend"
At this point, "what", "quiet" and "what"
It has been converted and memorized into two new thesis sentences for "Biology".

Then, GK is added to 1(P) and the process returns to step (172).

In addition, if the content of J>PT+s+ is found in step (173) and the process branches to step [1791], first.

Determine whether IDX=O.

If TDX≠0, branch to W (Figure 12).

If IDX=O, output the answer (see Figure 13) (180)
After that, add 1 to the contents of l1lpy +w+ in the TQ area and return to force (Fig. 9, 156).

Now, in FIG. 12, step (16) of FIG.
In 4), if K > MAX (nx), processing is performed from the branched point.

First, it is checked whether J>1.

If it is Jul, branch to step (1871).

If J≦l, branch to step (1831).

If it branches to step (187), J is PT+N
Compare whether it is equal to the content of -0+l.

If they are equal, subtract 1 from N, substitute l for IIIJ,
m.

- Add 1 to l and return to (1881, wa (Figure 9)).

If they are not equal, in step (189), first substitute 1 for the cup, add 1 to +mJ-1, subtract 1 from J, and then return to Figure 9).

When branching to step (11113), check whether the flag ANS indicating whether the answer has been output is "0" or not, and if NS>0, end the series of processing (186) AN
If S≦0, generate a random number and display 9 selections and 1 of the sentences expressing the meaning of "I don't know" prepared in advance in Step (184) (185), and then perform the processing. End (+861°) Figure 13 shows a method for generating response sentences from the thesis sentences of the answer table into which the answers are sequentially substituted in step (1681) in Figure 1O and output in step (180) in Figure 11. show.

Step (190) is an example of the thesis statement of the answer stored in the answer table. Particles, word endings, etc. are added to this thesis statement and output as a sentence.9 In this embodiment, as a means of adding case particles and word endings in step (191), a random number is first generated, and then a random number is generated. According to the random number, an arbitrary word is selected from the dictionary of case particles and word endings in step (1921 (1931), and is added to the thesis statement and output in step (+941).

In the case of more complex sentences, this method can be used to remember which words in the dictionary, such as particle 1 conditional words, appeared in which order during the sentence decomposition process shown in Figure 6 above. (means can be easily added to this embodiment), and can be handled by adopting a method of controlling the addition of particles, etc. based on that information.The details of one embodiment of the present invention have been explained above. However, the present invention has the following features in order to be able to operate with a small microcomputer system.

(1) As explained in Figure 6, it has the minimum information dictionary necessary for syntactically classifying one sentence in language syntactic analysis. Also, the dictionary can be of any size. As explained above, the syntax analysis procedure can be simplified, and the number of languages to be registered in the dictionary can be extremely small.

(2) As explained in Figure 7, normal sentences are divided into factual sentences and conditional sentences, and the number of two conditional sentences, the second N pointer, and the memory location of the sentence decomposed into thesis as shown in Figure 8 are calculated. It is possible to easily control the depth of inference of the thesis sentences efficiently arranged in the memory of RAM A by the first pointer PT indicating the address, and the present invention is realized with extremely few processing steps and the number of memories. I have the means to do so.

(3) As shown in Figure 13, the answer output sentences are 2 various particles 1
It is possible to easily output sentences with a variety of endings, and it is possible to compose emotional sentences equivalent to those written and expressed by a single person.

(4) As explained in Fig. 14, the digital data of the answer text output as external outputs and the i! It then compares it with predetermined digital data and when it matches, it can issue an external output signal.It can also analyze the logical meaning of questions input in natural language and take some action as an answer. I can make you say that. We can provide robots with so-called language and intelligence.

(5) I would like to add a further note: 9 The basic elements of the present invention are compactly constructed by sophisticated hardware and software as shown in Fig. 6.
The entire device of the present invention can be housed in a single case, can be constructed on the same substrate, or can even be constructed within an 1-chip microcomputer.

As an application of this invention, if you connect voice recognition to the input and voice 1 action output to the output in the embodiment of the present invention.

This means that a computer can come as close as possible to a human figure, which is unprecedented.9 It is also possible to provide a robot that is extremely compact, has a wide range of motion, and is close to a human figure, with artificial intelligence. Furthermore, if applied to the captain system used as a terminal in an information company today, it will be possible to extract and output sentences and words with similar logical meanings from six inputs, which will greatly improve the ability of rowers.

It can also be applied to controlling household information, for example, via a telephone line. For example, as prior knowledge, input one sentence into this system as follows. ``Gas can catch fire.'' ``Dryers can overheat.'' ``Electric drills are dangerous.''``It's dangerous to catch fire.'' ``Things that overheat are dangerous.'' ``If it's dangerous, turn it off while you're away.''
do. ” input what everyday people think in natural language.

For example, if you connect a voice recognition system from outside the device to the device of the present invention in which these are stored and ask a question in natural language, ``What turns off the switch?'' It is also possible to obtain an answer voice output saying, ``It turns off the gas, hair dryer, and electric drill.'' It is also possible to automatically switch them off according to the inference results.

(6) It can also be applied to word processing systems to create new sentences through semantic inference.

(7) The input/output of the present invention also applies to personal computers.
Since the response is in completely natural language, it can be applied to software for data extraction and input, program development, etc.

Next, we will discuss the feasibility of the present invention and how it has unprecedented features, and which will be used in future computer-related electronic devices (e.g., Comp).
uter assisted 1nstructton
device.

Word processors, database computer systems, personal computers with language output programs such as BASIC-CPM and Japanese, handy quive pocket computers, handheld electronic memo devices, etc.)
To explain how it has a significant impact on the system, the actual memory space is 128 bytes x 16 bits, microprocessor - 16 bits, internal ROM 128 bytes, RAM
64 bytes, built-in floppy disk tU byte x 2 x 16 bits as external memory (however, only a portion of the internal and external memory is used in this example).CRT
A detailed example of intelligent linguistic responses to sentences used for Japanese language learning by Japanese elementary school students, realized using today's so-called personal computers equipped with keyboard input devices, is shown below.

This example also shows that in the thesis analysis of a sentence, the fact (A) (121) corresponding to each sentence number in Figure 'M7 could be realized even if the sentence was not singular, but rather plural.

Regarding the example in which the teacher below inputs sentences as knowledge to be input into the response device while starting the component program of the present invention and interfacing with, for example, the Japanese input/output program, the following is the manual writing of this example. show.

Examples of human-powered sentences> Dogs are Japanese animals, can swim, are omnivores, and bark loudly.

Lions are animals, carnivores, and are of African origin.

I'm scared, but I can't swim.

The lion is the king of beasts after all, and it's sure to make a roar.

Monkey 6 Born in Japan, he is an animal, omnivorous, and quick.
I can climb trees, but I can't swim.

Giraffes have long necks, and elephants have long trunks.

Giraffes are herbivorous animals, native to Africa, and cannot swim.

They are elephants, herbivores, and were born in Africa, so they make loud noises.

It's an animal.

Elephants are big, and so are whales.

Whales are born in the sea and can absolutely swim.

Penguins are birds, and whales are apparently animals.

Penguins can swim, but elephants can't.

Penguins were born in Antarctica and now live in Africa.

Penguins cannot live in Africa, lions cannot live in Antarctica, and giraffes cannot live in Antarctica.

Lions are found throughout Antarctica. The dog is in Japan. There is a monkey in the sea. There are elephants in Africa. There are giraffes in Japan. And there are whales in the sea.

I really feel sorry for them if they can't live "somewhere" but are staying "somewhere".

Even if you can't swim, you'll drown if you're in the ocean.

I feel sorry for him if he drowns.

I was born in a ``certain place,'' and if I'm currently in a ``certain place,''I'm sure I'm happy.

If you feel sorry for someone, you should help them.

If an ``animal'' is born ``somewhere,'' then ``somewhere'' is the ``animal'' home.

Since "animals" are born "somewhere," the hometown of "animals" is "
Somewhere.”

We are all alive.

If you're alive, sing.

It's fun to be alive.

Even Okera is alive. Even frogs are alive. Even Amenbo is alive.

If everyone is still alive, they're your friends.

Animals are alive. If it's a bird, it's alive...
I typed... This entire sentence can be said to be a program that is completely similar to natural language, and can be called a super high-level language.

Below are examples of question writer skills and response/answer output. A teacher inputs a question to teach elementary school students. Here, the human ability to write questions will be abbreviated as a teacher, and the automatic response output will be referred to as a robot that composes a composition.

Teacher: What is happening in Africa now?

Robots: Elephants, giraffes, and penguins.

Teacher: What do you feel sorry for?

Robot: I feel sorry for the penguin.

Teacher: Why do you feel sorry for the penguin?

Robot: Penguins can't live in Africa.

Because penguins live in Africa.

Teacher, please send him back to his hometown.

Robot: Yes, I will.

At this time, the external output of the responsive electronic device of the present invention may drive an external behavioral object as explained in FIG.

For example, in this example, scenes are set in Africa and Antarctica, and a penguin doll placed in the African location is moved to the Antarctic location by a robot that imitates human hand and arm movements.

Additionally, we were able to achieve the following dialogue.

Teacher: Do you think the penguin is happy?

Robot: Yes, I think so.

Teacher: Why is the penguin happy?

Robot: Penguins are from Antarctica, because penguins are in Antarctica.

For example, even if new knowledge is additionally input here, it is only necessary to analyze the gist of the explanation from FIG. 6, add it to RAM A and H in FIG. 8, and reprocess it.

Here, if one of the teacher's students is Tanaka-kun, the teacher can add text input as shown below and obtain the following output group from the composition robot.

Teacher: Nitanaka-kun is alive.

Robot: Yes, I understand.

Then, I entered another question. I got the following response output.

Teacher: Tanaka-kun is my friend.

Robot: Yes, that's right.

Teacher: Who has fun?

Robot: I have fun. Okera and frog.

Amenbo, dogs, monkeys, etc.

Tanaka-kun is having fun too.

Teacher: That's great.

Robot: No, thank you. (This response text was prepared in advance in response to the question text, but output 6) Teacher: How about you, Nitanaka-kun?

Robot: Tanaka-kun is fun.

Tanaka-kun is my friend.

Tanaka-kun is alive.

The above is the content of the example taught to the students. Next, we will show an example in which students are asked to input their answers and judge whether they are correct or incorrect.

For example, 9. input a question in advance and have it randomly output and displayed from the response electronic device. for example.

Robot: Who can't swim?

In response, the student determines the meaning from the input sentence example and inputs the following.

Student: Monkeys, elephants, lions, and giraffes can't swim.

Since this answer input is exactly the answer/response output from the robot described above, it is sufficient to compare and match them and output the correct/incorrect determination.

for example.

Robot: Correct.

The answer is self-evident. Here, we have shown an example of a language sentence response learning device with full-scale intelligence.

I have so far shown the example of Japanese, but inferring the meaning from the gist of the language is common to almost all languages in the world.

When breaking down the general sentence (101) in Figure 6 into its thesis in English, for example, if it is a question sentence in English or American, most of the sentences will end with a "?", so put it in the dictionary (104). It is sufficient to keep it, and if the 1-condition word dictionary (106) is in English, rif, 1fthen, -.J, etc. apply. Furthermore, the fact in Figure 7 (A) (121).

Condition (B) (1231 is similarly decomposed into a group of English words and memorized, and is then processed in almost the same way as the Japanese case described above to obtain the same response output and effect. I can do it.

Even in the example of text input shown in FIG. 3, it can be processed without any problem even if it is translated into English. For example, the sentence group in (31) is.

Ａ　1ion　is　a　mam habitual.

^1ion is a frayed.

A penguin is a bird.

If it is quiet and ali
ving thing, 1tis just our
It is translated into English as "friends" and processed in the same way as in Japanese, and a response is requested.

As explained above, the device of the present invention has the following effects.

(1) Humans can interact with the device of the present invention in a language that is close to natural (Japanese, English, etc.).

(2) Knowledge and wisdom can be taught to the device of the present invention in language (hereinafter referred to as Japanese).

(3) The device of the present invention makes inferences and judgments by itself according to its knowledge and wisdom.

As mentioned above, the basic elements of the present invention are to realize the matters that will have a significant impact on the contribution of future computer application systems to society.
It can be realized with only 500 steps or less in a microcomputer programming language such as 16-bit or 16-bit, and can be called a truly revolutionary invention.

[Brief explanation of drawings]

FIG. 1 is a diagram showing a schematic configuration of an embodiment of the present invention. FIG. 2 is a circuit block diagram showing one embodiment of the present invention. 3 to 5 are explanatory diagrams for explaining the operation of an embodiment of the present invention. FIG. 3 is a diagram showing an input sentence. FIG. 4 is a diagram showing the input question text. FIG. 5 is a diagram showing an answer text as an output example. FIG. 6 is a diagram for explaining the means for converting an input sentence into a summary sentence, where ■) shows an example of a normal sentence, and 2) shows an example of a question sentence. FIG. 7 is an explanatory diagram showing the result of classifying sentences (A) that are discussed as a conclusion, conditions leading to them (B), and the number of sentences (B) according to an embodiment of the present invention. FIG. 8 is a diagram for explaining the gist and storage arrangement of necessary information and its contents according to an embodiment of the present invention. 9 to 13 are flowcharts for explaining one embodiment of the present invention. FIG. 14 is a diagram showing a schematic configuration according to another embodiment of the present invention. In the figure, (1) is an electronic device, (6) is a keyboard,
(1 (11 is CRT, (31) (321 is input text details, (34) is question text, (35] is answer text, (
102) is an excised word dictionary, (103) is a regular sentence ending dictionary, (104) is a question sentence ending dictionary, (1061 is a conditional word dictionary, (107) is a reconstitution dictionary. (1081 is a particle dictionary. The same reference numerals indicate the same or equivalent parts. Engraving of the drawings (Contents (no changes) 9< Figure Figure Figure Figure 1/7I ↓ 2) In contrast to the use of Ka2, 'f Ka・ /, /lI3 /15 //6 Figure 9 Figure 0 Figure 13 Figure procedural amendment (voluntary) Patent application A3 filed on January 1, 1990 2. Method of responding to the title of the invention 3゜ Person making the amendment Name of address related to the incident (6011

Claims (2)

[Claims] (1) A text response method that has the following elements: (a) A step of transforming a text into a thesis statement and storing multiple thesis statements as a thesis statement group; (b) A text response method that has the following steps and uses a conversion area. (b1) logically searching for a new thesis statement from a group of thesis statements based on the thesis statement; (b2) determining whether the searched thesis statement is a thesis statement stating a fact; (b3) Necessary for deriving the searched thesis statement when the searched thesis statement does not become an answer. A step of hierarchically storing one or more thesis sentences from a group of thesis sentences in a conversion area, (b4) If there are multiple thesis sentences stored in the same hierarchy, the above (b1) and (b2) are performed for each thesis sentence. (b5) After the step (b4) above, backtrack the hierarchy of the thesis statement stored in (b2) above, and use the conversion area released by backtracking to perform the above (b1). ) to (b4). (2) A text response method having the following elements: (a) having the following steps, transforming the text into various thesis sentences and storing a plurality of thesis sentences as a group of thesis sentences; (a1) converting the sentences into (a2) dividing ordinary sentences into factual sentences and conditional sentences, and generating factual thesis sentences from factual sentences; Step of generating and storing it in a group of thesis sentences, (a3) Divide the conditional sentences into a conclusion sentence that should be discussed as a conclusion, etc., and an argument sentence that is argued to derive this conclusion sentence, One or more essay thesis sentences are generated and stored in a thesis sentence group, a conclusion thesis sentence is generated from the conclusion sentence, and the one or more essay thesis sentences mentioned above for deriving this conclusion sentence can be discriminated, and the thesis sentence group is created. (b) A process of outputting an answer text from a question text using a conversion area, which has the following steps, (b1) A process of arranging a question thesis statement in a conversion area, (b2) A process of storing an answer text in a conversion area. (b3) a step of searching a group of thesis sentences based on the thesis sentence most recently placed in the conversion area and searching for a logically matching thesis sentence; (b4) When the thesis sentence retrieved in the step (b3) above is a conclusion thesis sentence, placing one or more thesis sentences determined from this conclusion thesis sentence from the thesis sentence group in the conversion area, (b5) When the thesis sentence retrieved in step (b3) above is a factual thesis sentence, this factual thesis sentence is logically similar to the thesis sentence already placed in the conversion area. (b6) After the steps (b2) to (b5) above, there are other essays that have not yet gone through the step (b3). (b7) After the above steps (b2) to (b6), the order of the thesis sentences stored in the above step (b2) is returned by one, and the conversion is performed. The process of opening the area and returning to the process of (b3).