JPS6286429A - Natural language processing system - Google Patents

Abstract

PURPOSE:To decide the modification of a verb by providing a means regarding the attribute relation of the world model as a case in which the said verb is conjugated other than the case described in a word dictionary. CONSTITUTION:Whether or not the description of case information of a verb exists in a dictionary is checked and when present, it is compared with that of an input sentence and when they are coincident, the modification processing is applied (forming of analysis tree). When no description exists in the case information of the verb in the dictionary or it is not coincident with the suffix information of the input sentence, the world model of the said verb is referenced and the model check the modification with other word. When in existing, the analysis tree is formed and when not in existence, the modification is regarded as disabled. The analysis tree is generated by a syntax analysis and the tree is formed by applying context processing taking the preceding input sentence into account. Thus, the verb case processing using the world model is sued is not applied and only the paragraph coincident with the case information of the word dictionary is modified.

Description

Detailed Description of the Invention (Field of Industrial Application) The present invention uses a natural language processing method, particularly a natural language for database search. Regarding processing method.

[Conventional technology]

Database searches are performed by inputting commands that specify the database name, the names of the major classifications and minor classifications used in the database, etc., but with this method, the searcher is familiar with the structure of the database and can search for the required information. It must be possible to create commands, and the scope is limited. If the input text is in natural language and the system can parse it, generate commands, and search the database, searching becomes much easier.

For example, the input text when searching a database in natural language is: What product did Fujiya sell 200 units on June 17th for 200 yen? That is what it is. When there is an input sentence, the system uses a word dictionary to perform morphological analysis, and uses a world model to perform syntactic and semantic analysis, context processing, and generates commands.

The world model describes the attribute relationships of various words to specific words used in this system, and is effective for syntactic/semantic analysis and context processing. The word dictionary contains various words, the corresponding class of each word, and if the word is a verb, the word associated with the verb and its particle (case information), etc.

[Problem that the invention seeks to solve]

The word dictionary also contains case information for verbs, so you can refer to the word dictionary for case processing of verbs (processing in which a verb and the noun clause that modifies it are treated as one verb clause). All the case information of the verb must be written in the dictionary, and if there is any omission, it may become impossible to analyze.

By the way, since the world model shows the attribute relationships of words, it is a powerful source of case information. Therefore, the present invention uses a world model for case processing of verbs, and uses attribute information of the world model as a default value for case information of verbs.

[Means for solving problems]

The present invention provides a word dictionary that describes various words and corresponding class names for each word, words that are associated with a verb if the word is a verb, and their particles, and a world model that summarizes specific words and various words that are associated with the word. In a method of processing input natural language into a parse tree using It is characterized by having a means for examining the attribute relationships of the world model and regarding the attribute relationships as possible cases of the verb.

[Effect]

If the attribute information of the world model is used for verb case processing, verb dependency can be determined even if the case information in the word dictionary is insufficient or not accurately described, making it extremely effective for natural language processing. be.

〔Example〕

To explain with reference to the drawings, FIG. 1 shows an overview of the entire system, and 1 is a terminal device with a CRT display la.
, a keyboard Ib, and a floppy disk IC.

11 is a word dictionary, 13 is a world model, and 15 is a database. 2 is interaction, 3 is morphological analysis, 4 is syntactic/semantic analysis, 5 is context processing, 6 is paraphrase generation, 7 is condo generation, 8 is data search, and 9 is response generation, and this part is software. 10 is an editor for a word dictionary 11, a world model 13, and a modeling system 12;
14 is a database management system. The modeling system 12 has system-defined classes, inchers, etc., under which the user-defined database resides.

As shown in FIG. 2, the word dictionary 11 includes words such as "sell" and "product" and their model information, that is, the word "sell".
The corresponding class is "sales", the attached word is quantity,
In the case of store and product, the quantity is n1l without a particle), the particle of store is [store is, or store is] and “ga” or “ha”, and the particle of “product” is “product”. ” and “wo”, and for the word “product”, the corresponding class is “product”, etc. There are various ways to write case information such as (seller), such as writing it directly or categorizing it into symbols. The world model of ``sales'' is as shown in Figure 3, where ``sales'' includes ``sales price'' such as ``sold for 200 yen,'' and ``sales stores'' such as ``sold by the Fuji layer.'' It shows that "products" such as "mechanical pencils were sold" and "sales dates" such as "sold on June 17" are linked.

These stores, sales unit prices, etc. are called instances.

Now, operate the terminal keyboard and click "Fuji layer is 6 for 200 yen.
Which product sold 200 units on the 17th of the month? ”, in the morphological analysis, the input sentence is broken down into words while referring to the word dictionary, and model information is attached to each word as follows.

Fuji layer will be available for 200 yen on June 17th (Seller)
(Price) (Date) 200 bottles sold
products are ? (Quantity) (Sales) (Products) In syntactic/semantic analysis, phrase integration is performed, and the worded input sentences are combined into phrases.

Next, the modification is analyzed (analysis tree generation processing).

Figure 4 shows the outline. First, it is checked whether there is a description of the case information of the verb in the dictionary, and if so, it is compared with that of the input sentence, and if they match, the dependency is processed (an analysis tree is created). When there is no case information for a verb in the dictionary, or even if there is, it does not match the particle information in the input sentence, the world model of the verb is referred to and the model is used to check connections with other words. . If there is a matching item, a parse tree is created using it, and if there is not, no modification is possible. Parse trees are created using syntactic and semantic analysis, and context processing takes into account the previous input sentence.

FIG. 5 shows a state in which the verb clause has been processed and a parse tree has been created. When the input sentence ``What product did Fuji Layer sell 200 units on June 17th for 200 yen?'' is morphologically analyzed with reference to the word dictionary and separated as shown by dotted lines, and the model information obtained from the dictionary is generated. , and the model information obtained by referring to the world model, a tree is created as shown in the figure through syntax-semantic analysis and context processing, and the mutual relationships are made clear. In this diagram,
The words "for 200 yen" and "200 pieces" were used as "Mo".
indicates that the model information was not found in the dictionary, so the relationship was sought using the world model of "sales". A sub-model is a world model that includes specific words, such as "sold" for sales in Figure 2, "Fuji layer" for dealers, and "Fuji layer" for sales price in Figure 2.
200 yen'', ``June 17th'' as the sales date, and ``200 pieces'' as the sales quantity, the first sub-model in Figure 5 is obtained. If the process reaches a dead end in the process of creating such an analysis tree, the searcher performs a paraphrase, displays the paraphrased words on the display, and asks the searcher whether this is acceptable. If OK is input, processing proceeds accordingly. This is the paraphrase generation routine.

Once the parse tree is created in this way, the next step is to generate commands. Assuming that the database has the following product table and sales table, the command to generate the product table and sales table is as follows.

get product name from product table 1 vhere product table, price - 200〃, product name - sales table, product name sales table, store - Fuji layer〃 01 sun - June 17th〃 sales volume = 200 The analysis tree is in the database If the analytic tree shown in FIG. 5 is obtained, it is possible to create the above command from this tree. Once the command is obtained, the table is searched using it, and when the desired product name is obtained, it is displayed on the display.

〔Effect of the invention〕

As explained above, according to the present invention, the attribute relationships of the world model are used to proceed with the case processing of verbs, so even if there are omissions in the word dictionary, it can be covered, and the word dictionary can be The advantage is that there is no need to write case information. For example, in the word dictionary in Figure 2, the word "sell"
Quantity ni! Since there are only three types of case information: , store, and product, no case information can be obtained for (sales) on month and day, but this can be determined by examining the world model of sales. Information can be obtained. Of course, even if the world model is referred to, the particles themselves cannot be obtained, and this point remains unclear, but if the attribute relationships are known, it is possible to create an analysis tree. In addition, the input sentence may contain particles that are not found in the word dictionary (in this case, ``no''), such as ``sold by a store,'' but in this case as well, by looking at the world model and knowing the attribute relationships, can be processed.

Note that the present invention can be applied to various applications by changing the criteria for allowing modification in the algorithm shown in FIG. Furthermore, in the input sentence for database update, errors in particles are not allowed because this requires strict semantic interpretation. That is, verb case processing using the world model is not performed, and only clauses that match the case information in the word dictionary can be modified.

[Brief explanation of drawings]

Fig. 1 is a block diagram showing an embodiment of the present invention, Fig. 2 is an explanatory diagram of a word dictionary, Fig. 3 is an explanatory diagram of a world model, Fig. 4 is an explanatory diagram of modification processing, and Fig. 5 is an explanatory diagram of a world model. It is an explanatory diagram of an analysis tree.

Claims (1)

[Claims] A word dictionary that describes various words and corresponding class names of each word, words that are associated with a verb if the word is a verb, and their particles, and a list of specific words and various words that are associated with the word. In a method of processing input natural language into a parse tree using a world model, when case processing is performed to convert a verb and a noun clause modifying it into a single verb clause, cases other than those described in the word dictionary are used. , a natural language processing method characterized by having means for examining the attribute relationships of the world model and regarding the attribute relationships as possible cases of the verb.