JPS61221873A - Question and answer system using natural language - Google Patents

Abstract

PURPOSE:To attain a question/answer system using a natural language by extracting and storing the facts contained in sentences and at the same time collating the fact contained in a question sentence with the fact contained in the data base of the natural language produced previously. CONSTITUTION:A close input KANA (Japanese syllabary)/KANJI (Chinese character) converting part is provided together with a parser part and a question/answer part. A close input KANA string is converted into a KANA/KANJI mixture sentence by a KANA/KANJI converting part with use of a dictionary and a table 7. The parser part uses a parser grammar rune memory 9 and a parser dictionary memory 8 to analyze the sentence structure of each word and produces a sentence structure analysis tree for input sentences. Thus a question sentence using an input natural language undergoes the anlysis of form elements through the KANA/KANJI converting process and is tirned into a simple sentence through analysis of sentence structure and meaning and the simple sentence production processing. Then the facts contained in said simple sentence are extracted and stored. While the fact contained in the question sentence is allocated with the fact contained in the data base of a natural language produced previously. Then an answer is produced to the question sentence.

Description

[Detailed Description of the Invention] Technical Field> The present invention can be applied to an interactive device (computer) such as a diagnostic device, so that it is possible to interact with the device using natural language (for example, EI real language). This is related to the question-and-answer method.

<Prior Art> Conventional database manipulation languages mainly involve programming methods and table/manipulation languages, which are completely difficult for beginners to use. In addition, many computer operating languages are command-based and cannot be easily handled. Either way, it's an operation. It takes a lot of time to learn the operating language, and this has been a major problem in operating computers up until now.

Purpose> Therefore, the present invention provides knowledge (target databases and commands) to a device using natural language or a program, and once the knowledge is stored, questions, searches, and commands can be sent to the device using natural language. The present invention aims to provide a natural language question answering method in which the device can respond to questions, searches, and commands in natural language.

<Example> The present invention will be explained in detail below based on the drawings.

FIG. 1 is a system configuration diagram of an apparatus that employs the natural language question answering method of the present invention.

In the figure, 1 is a keyboard for inputting kana character strings, etc., 2 is a CRT display device, 3II'i is a central processing unit, and 4 is an internal memo! 7.5Fi external memory. Further, 26 is a command correspondence table memory, 7 is a kana-kanji system dictionary memory, 8 is a parser dictionary memory, and 9 is a parser grammar memory.

This device consists of a solid kakana-kanji conversion section, a parser section (sentence analysis section), and a question answering section. It is converted into a sentence containing kanji, and during this conversion process, the delimiter information (
Here, a word refers to a morpheme such as a noun, particle, or conjugated ending) and part-of-speech information. The parser section utilizes this morpheme information and parses each word using the parser grammar rule memory 9 and the parser dictionary memory 8 to generate a parse tree for the input sentence. Further, the parser section generates monocultures and relationships between simple sentences in a format that can be used by the question answering section later from the obtained parse tree. Figure 2 shows the relationship stack between monocultures and simple sentences in the case of ``If he buys a red Y, I will also buy Y'', and the relationship between this monoculture stag and simple sentences. The stack is used in the question answering section described later.

FIG. 3 shows the processing flow of the question answering section, which can be functionally divided into five primary parts.

(1) Expand the relationship stack between simple sentences into an and sentence or sentence.

(21 input sentences are created into a database according to question sentences, imperative sentences, declarative sentences, and hypothetical sentences.

(Find solutions by creating a database of 31 things and 1 rule (unification).

(4) Interpretation and execution of commands.

(5) Response in natural language.

Next, each of the above functions will be explained in order.

i) Regarding the development from the relationship stack between simple sentences to an and sentence or sentence, a simple sentence here refers to a sentence that contains only one predicate (verb, adjective, adjectival verb).

Generally, a sentence is made up of multiple simple sentences and, or
, if, and adjunctive relationships, and it is extremely difficult to process them in just one form.

However, each of A-D is a simple sentence.

Table 1 For example, in Table 1 (al, simple sentences A, B, C are and, o
The sentences connected by r and further composed of A, B, and C are connected by simple sentence ri and if. On the other hand, Table 1 (bl is an expansion of the sentence in Table 1 (al) into and, or, and corresponds to a binary expansion in mathematics.
The items to be connected are grouped together in this way, and later processing becomes easier.

Figure 4 is a diagram showing the actual stack configuration, and is 10F in the figure.
The i relation stack, 111fi and*or stack is shown. Here, the relation stack 10 shows the relations between simple sentences (Table 1 (a)). In this example, simple sentences B and C
1rior relationship, sentence #-t consisting of a simple person and or
The relationship of and, and a sentence composed of and is a simple sentence and an if relationship.

The and*or stack 11 is for extracting and, or, and if quantity relationship simple sentences from the relationship stack. For example, if if is a parent node, the left child node is determined to be the condition part, and the right child node is determined to be the conclusion part. As is clear from the figure, first, and, o of the left child node of if
Examine the r relationship and find Aand(BorC). This can be expressed as (A andB) or (A a
ndC), and condition part 1 of and/Or stack 11
31CA, B and A, C are stacked. Next, the right child node is a simple statement, so it is an and/or stack 11
The relationship shown in Table 1 (b) is derived.

ii) Concerning database creation, sentences analyzed by the parser section (text analysis section) are converted into simple sentences with deep structure, and the preprocessing section of the question answering section adds relationships such as and and or to each simple sentence. is extracted. Using these processing results, for example, declarative sentences (including hypothetical sentences) are compiled into a database as facts and rules, while interrogative sentences are compiled into a database and converted into a set of machine-specific commands, or facts and rules are compiled into a database. The solution is found by pattern matching and pack tracking against the database.

The structure of the database is 1) Addition, deletion, and expandability of information 2) Ease of pattern matching Taking this into account, we have adopted the following format.

Predicates (ID, situation l, deep case, case demonstrative, situational predicates are usually expressed in the final form of the predicate, but if there is an adnominal 5, etc. in the sentence, the following case demonstrative is used as the key. It is related to the preamble.

However, the ro~r3i constants, RO~R1, and XO~X1 are variables in Table 2. For example, in a) of Table 2, the case phrase following 'Sono' is '
° It is an ice ball, and since the sentence number of ro has the same maxim, it is associated with that sentence. At this time, “predicate J
I became a drefer and put #'trentai as a deep case. Predicates can also be expressed as variables. As shown in Table 2 a), rIDJ assigns Hfact to declarative sentences excluding hypothetical sentences and stores them as facts in a database. For hypothetical sentences, as shown in Table 2 C), Figoal is given to the conditional part of the sentence, and head is given to the conclusion part of the sentence.

Furthermore, interrogative sentences are given goa 1 as shown in Table 2 b). These are necessary for syllogism reasoning during the unification process after IDIr1.

"Situation 1" is a monocultured sentence number, in which rO to rn are assigned to the IDs of head and fact, and RO to Rn are assigned to Km of the turtle with the ID of goal. Further, the r o -r n tri constant, RO-Rn, is defined as a variable, and is also used in the later unification process.

"Deep case" is agent, 5ubj, obj,
We have prepared a few rounds of lO such as Rentaj. ``Case pronoun'' is a formal expression that falls into the deep case, and usually includes ``he'', ``school'', but interrogative pronouns ``who'', ``what'', ``°°X'',
If 'Y' appears in the sentence, Table 2 b).

XO, XI, and X2 are stored as variables as shown in C).

“Situation 2” contains a sentence number associated with a relation such as refer. If it has a normal predicate other than refer, it is nil.

Since the database will be subject to pattern matching later in the unification process, the structure of the database is devised to speed up the pattern matching process.

Note that l) Kanji data for predicates and case expressions are registered in the word area, and the registered addresses are stored in the predicates and case expressions, so when a pattern is matched, it is sufficient to compare address pointers. 2) Since the predicate, ID, and situation 1 are the same in each simple sentence, they are combined into one.

Regarding '+iD unification, the unification process is as shown in the schematic flow in Figure 5,
Its basic operation is pattern matching and backtracking between the goal side (interrogative sentence database) and the source side (declarative sentence database).

For example, when Table 2 a), c), and d) are already stacked on the source side, if you enter Table 2 b) as a question, the result of unification will be "ball",
"Apples" are obtained. As shown in Table 3, this operation is as follows: 5), 6) on the goal side and 1), 6) on the source side.
2) is unified, the ball is bound to the variable XO, and the solution is found. Furthermore, in order to obtain the next solution, the values of variables Ro and xo are unbound and unified with 8) and 9) on the source side. At this time.

RO-r 2 OwXt is bound. 8), 9) It head
Therefore, set 7) whose ID is goal as a sub-goal, and then try to unify with 10) on the source side.

At this time, R1 engineering r3 R1 = apple is bound.

XO gong Xl = apple ``Apple'' is found as another solution from the link.

Furthermore, even if you try to find the next solution, it will fail because there is no match on the source side.

Note that 1), 2), etc. correspond to Table 2.

Table 3 W) Interpretation and Execution of Commands Imperative sentences and interrogative sentences are once stored in a database, but if the predicate is not in a system command, unification with the database of declarative sentences is performed. Also, if it is a system command, unification with the declarative database is not performed.

FIG. 6 shows the argument table 14 for the command, I! The components are the command name, such as print or copy, and the type of argument for the command.

Sets the address of the memory where the command body program is stored.

Table 4 shows examples of command interpretation and execution,
For example, a sentence entered in Japanese is stored in a database, and its predicate is compared with the command name in the command table.

If they match, then the deep case is compared with the case of the argument described in the command table. In this example, it matches argument 1, and "I file 1w" is passed to argument 1.

If there are no arguments to fill in the command table, the arguments are passed to the main body of the program and executed.

If the minimum required arguments are not present or the case of the arguments does not match, the user is notified of this via the CRT 2.

■) Regarding responses using natural language.

Table 3 is.

This is an example of unification for the question ``What will he buy?'' 6) expresses what he will buy. Furthermore.

obj corresponds to 1 and 2 xO corresponds to "what", and as a result of unification, Xo = ball and apple are found as solutions, so there is 1 pole for "what",
If you replace question 1 with ``apple'' and remove the question 1, a response will be generated in Japanese. Namely.

"He buys the ball" "He buys apples." A sentence is generated and displayed on the CRT 2.

In this way, according to the method of the present invention, the database
For example, it can be accessed using a Japanese interface. Therefore, questions, searches, and commands can be made in Japanese to various question-and-answer type devices (eg, diagnostic equipment, etc.), and furthermore, the devices can respond to these questions, searches, and commands in Japanese.

As described above in detail, a sentence is morphologically analyzed using the kana-kanji conversion process of the present invention.

A monocultural text is created through syntactic analysis, semantic analysis, and monocultural processing, and the facts contained in this text are extracted and stored, and the facts contained in the question text are compared with the facts contained in the own database created in advance. By checking with the facts of
Since a response to the question text is created, it is possible to interact with the device (computer) using natural language.

Therefore, there is no need to learn a complicated operating language for the device, and even beginners can use the device extremely easily.

[Brief explanation of the drawings] Fig. 1 is a system configuration diagram of a device adopting the method of the present invention;
Figure 2 is a diagram showing the relationship stack between monocultures and simple sentences, Figure 3 is a diagram showing the processing flow in the question answering section, and Figure 4 is ll.
FIG. 5 is a diagram illustrating the expansion into a 1and statement+6r statement, FIG. 5 is a diagram illustrating a unification processing flow, and FIG. 6 is a diagram illustrating a command argument table. 1Iri keyboard, 2t/1CRT, 3t-1cPU
, 4 is internal memory, 5 is external memory, 6 IIi command correspondence table, 7 is Kana-Kanji system dictionary memory, 8 is parser dictionary memory, 9 Iri parser grammar rule memory Agent Patent attorney Aihiko Fuku (and 2 others) Figure 1 Figure 2 Figure 4 Rodoro 3 Figure 512 t9 Figure 6

Claims (1)

[Claims] 1. Natural language input means, morphological analysis means for obtaining word delimiter information and part-of-speech information, syntactic analysis means for determining dependencies between parts of speech, semantic analysis means for determining semantically correct dependencies, and embedded sentences. It has a monocultural processing means that breaks down into simple sentences and a database consisting of natural language, and uses kana-kanji conversion processing to morphologically analyze the input question sentences in natural language, and performs syntactic analysis, semantic analysis, and monocultural processing. After processing, it becomes a monocultural text, and the facts contained in the text are extracted and
The natural language is characterized in that the response to the question text is created by accumulating the questions and comparing the facts included in the question text with facts in a natural language database created in advance. Question-and-answer method.