JP2849111B2 - Natural sentence semantic analysis processor with learning function - Google Patents

Description

The present invention analyzes the meaning of a natural sentence read from an input device, and outputs a semantic structure that can be processed by a response processing device or the like connected at a later stage. The present invention relates to a natural sentence semantic analysis processing device with a learning function.

[Conventional technology]

As a conventional semantic analysis processor, it performs morphological analysis and syntactic analysis using dictionaries, grammatical rules, and field-dependent knowledge, and converts it into a semantic structure based on the generated dependency relations and predefined rules. Therefore, a semantic analysis processing device that performs a semantic analysis process is known. However, in such a semantic analysis processing device, if the input sentence does not satisfy a grammar rule defined in advance, the dependency relationship cannot be obtained, and the analysis fails. Since ordinary users do not know the grammatical rules defined by the semantic analysis processing device, they often input sentences that do not satisfy the grammatical rules.

To solve this, as shown in Japanese Patent Application No. 63-132247, a sentence that the system can accept is registered in advance, a sentence similar to the input sentence is displayed, and the user is made to select the sentence. A device for processing a sentence has been proposed.

[Problems to be solved by the invention]

However, every time the user inputs a sentence that does not match the pre-registered sentence, the user has to perform an operation of selecting a display sentence. Must be done,
In some cases, the smooth flow of dialog between the user and the computer was hindered.

[Object of the invention]

An object of the present invention is to reduce the user's selection operation by maintaining the relationship between a pattern and a pre-registered pattern when a sentence of the same pattern is input many times. An object of the present invention is to provide a semantic analysis processing device for a dialog processing system that can realize a smooth dialog.

[Means for solving the problem]

A semantic analysis processing device according to the present invention includes at least a regular sentence pattern dictionary, a sentence pattern matching unit, a pattern storage unit, a similar sentence search unit, an inquiry unit, a learning processing unit, and a semantic structure generation unit. It has a configuration that

[Action]

The sentence that can respond appropriately and its semantic structure are registered in the regular sentence pattern dictionary. If the input sentence pattern matches the regular sentence pattern, the sentence is sent to the semantic structure generating means and the meaning is sent. For a sentence that does not match the regular sentence, a similar sentence pattern is extracted by a similar sentence search means, and a regular sentence pattern is extracted by a query means based on the extracted regular sentence pattern and the input sentence. A display sentence is generated and displayed for each sentence pattern, the user is allowed to select a sentence that suits his or her intention, and the learning processing means patterns the input sentence and the name of the regular sentence pattern corresponding to the selected display sentence. The regular sentence pattern is sent to the semantic structure generating means and the semantic structure is output. The above operation is repeated every time a sentence is input, and the regular sentence pattern is stored in the pattern accumulating unit. Each time the same pattern as the sentence pattern is input, the corresponding frequency is updated. If the frequency of the sentence pattern stored in the pattern storage unit exceeds a predetermined threshold, the frequency matches the sentence pattern. When a sentence is input, the sentence of the same pattern is input many times by sending the corresponding regular sentence pattern to the semantic structure generating means and outputting the semantic structure without passing through the similar sentence search means and inquiry means. In such a case, the burden of the user on the input sentence having the pattern is reduced.

〔Example〕

FIG. 1 is a block diagram showing an embodiment of the present invention. In FIG. 1, 1 is a regular sentence pattern dictionary, 2 is a pattern storage unit, 3 is a morphological analysis unit, 4 is a sentence pattern matching unit, 5 is a control unit, 6 is a similar sentence search unit, 7 is a query unit, and 8 is an inquiry unit. The learning processing means 9 is a semantic structure generating means.

Hereinafter, the operation of the semantic analysis processing device according to the present invention will be described with reference to FIG.

The regular sentence pattern dictionary 1 stores a regular sentence pattern defined by a permutation having semantic attributes and words as constituent elements, and a semantic structure defined for the regular sentence pattern. In the pattern storage unit 2, based on past input sentences,
The sentence pattern and the name and frequency of the regular sentence pattern corresponding to the sentence pattern are stored.

The input sentence is input to the morphological analysis means 3, the input sentence is divided into words, the semantic attributes of each word are extracted, and the input pattern IP having the semantic attributes and the words as constituent elements is generated, and the sentence pattern matching is performed. Send to means 4.

The sentence pattern matching means 4 extracts a regular sentence pattern that matches the input pattern IP from the regular sentence pattern dictionary 1 and sets the regular sentence pattern as an analysis target pattern. Also,
If a sentence pattern that matches the input pattern IP exists in the pattern storage unit 2 and the frequency of the sentence pattern exceeds a predetermined threshold, the sentence pattern is extracted and a regular sentence pattern ( (The name of the regular sentence pattern corresponding to the pattern storage unit 2 is defined.)
Is an analysis target pattern.

Next, in the control means 5, when the sentence pattern matching means 4 obtains the pattern to be analyzed, the semantic structure generating means 9
If not, the input pattern IP is sent to the similar sentence search means 6. That is, if the input sentence is the first pattern and does not match the regular sentence pattern, the sentence is sent to the similar sentence search means 6 and the sentence that matches the regular sentence pattern or has already been input in the past. If the sentence matches the pattern, and the frequency of the pattern exceeds a predetermined threshold, the pattern to be analyzed does not go through the similar sentence search means 6, inquiry means 7, and learning processing means It is sent to the structure generation means 9.

The similar sentence retrieval means 6 extracts a regular sentence pattern SPi (i = 1, 2,... N) similar to the input pattern IP from the regular sentence pattern dictionary 1 and sends it to the inquiry means 7.

The inquiry means 7 generates a display sentence for each regular sentence pattern SPi (i = 1, 2,... N =) based on the regular sentence pattern SPi obtained by the similar sentence search means 6 and the words of the input sentence. The sentence matching the user's intention is selected by the user, and the regular sentence pattern SPk corresponding to the selected display sentence is set as the analysis target pattern.

If the sentence pattern that matches the input pattern IP is not stored in the pattern storage unit 2, the learning processing unit 8 sets the name of the regular sentence pattern Sk to the sentence pattern using the input pattern IP as the sentence pattern of the pattern storage unit 2. As a name of a corresponding regular sentence pattern, a predetermined initial value is stored as the frequency, and if the frequency is already stored, the corresponding frequency is updated.

The semantic structure generating means 9 outputs the semantic structure of the pattern to be analyzed based on the semantic structure defined in the regular sentence pattern dictionary 1 and the words constituting the input sentence for the regular sentence pattern corresponding to the pattern to be analyzed. I do.

 Next, a description will be given based on a specific example.

FIG. 2 is a definition example of the semantic attribute of each word, and FIG. 3 is a definition example of the regular sentence pattern dictionary 1. In the regular sentence pattern in FIG. 3, the notation [x] represented by parentheses indicates that the semantic attribute [x] is used, and the notation (x) indicates that the word x is a constituent element.

First, a case where a sentence that matches a regular sentence pattern is input will be described. For example, "Fax with broadcast function
I want. Is input. The morphological analysis unit 3 generates an input pattern having words and semantic attributes as constituent elements as shown in FIG.

The sentence pattern matching means 4 compares the sentence pattern with the regular sentence pattern in the regular sentence pattern dictionary 1. This matching is performed, for example, by determining whether the semantic attributes or the words match in all of the constituent elements at the same position in the input pattern and the regular sentence pattern. as a result,
Regular sentence pattern in regular sentence pattern dictionary 1 in FIG.
It is determined that it matches P1, and the regular sentence pattern P1 is sent to the semantic structure generating means 9.

The semantic structure generating means 9 generates the semantic structure defined for the regular sentence pattern P1 based on the words of the input sentence shown in FIG. Generation occurs. Here, in order to make the explanation easy to understand, as described above, the semantic structure is represented by an expression similar to a natural sentence. [X] in the semantic structure is
It is assumed that the configuration is such that a semantic structure is generated by replacing the word in the input sentence having the semantic attribute [x].
As a result, [function name] is replaced with "broadcast" and the semantic structure "I want a fax with a broadcast function."
Is output.

Next, a description will be given of a case where an example sentence "I want a fax capable of broadcasting" is input, which does not match the regular sentence pattern.

The morphological analysis means 3 generates the input pattern shown in FIG.

In the sentence pattern matching means 4, there is no regular sentence pattern that matches this pattern, and no sentence pattern is registered in the pattern storage unit 2, so that a pattern to be analyzed cannot be obtained.

Therefore, the input pattern is sent to the similar sentence search means 6 by the control means 5.

For example, it is assumed here that the similar sentence search means 6 is configured to extract a regular sentence pattern having a semantic attribute required for generating a semantic structure as an input pattern. In this case, the regular sentence pattern P2 also includes the semantic attribute [function name], and the regular sentence patterns P1 and P2 are extracted.

Inquiry means 7 generates display sentences D1 and D2 based on these regular sentence patterns as shown in FIG. Here, it is assumed that the user removes the display sentence D2 and selects the display sentence D1. In that case, the analysis target pattern is P1.

The learning processing unit 8 registers the input pattern and the name P1 of the corresponding regular sentence pattern in the pattern storage unit 2 as shown in FIG. 7, and gives 1 to the frequency as an initial value.

In the semantic structure generation means 9, the pattern to be analyzed is P1
Since the word of the input sentence having the semantic attribute [function name] is "broadcast," the semantic structure "broadcast" is the same as when "I want a fax with a broadcast function." I want a fax with a communication function. "

Next, the seventh operation is performed in the pattern storage unit 2 by the above operation.
A case where "I want a fax that can be intercepted" after the sentence patterns are accumulated as shown in the figure will be described. Here, for the sake of simplicity, it is assumed that the frequency threshold in the pattern storage unit 2 is 0.

At this time, the input pattern obtained by the morphological analysis means 3 is such that the word “broadcast” in FIG.
It just replaced it.

Therefore, the sentence pattern matching unit 4 determines that the input pattern matches the sentence pattern U1 in the pattern storage unit 2. Since the frequency exceeds the threshold value 0, the control unit 5 sends the regular sentence pattern P1 corresponding to the sentence pattern U1 to the semantic structure generation unit 9 as a pattern to be analyzed.

The semantic structure generation means 9 replaces the semantic attribute [function name] with the interception, and replaces the semantic structure "FA with the interception function".
I want X. Is generated. In this way, after learning with "I want a fax that can broadcast", a non-regular sentence "I want a fax that can be intercepted." The semantic structure can be output without passing through the similar sentence search means 6, inquiry means 7, and learning processing means 8.

〔The invention's effect〕

As described above, according to the present invention, when a sentence having a pattern matching a pre-registered regular sentence pattern is input, the semantic structure defined in the regular sentence pattern dictionary is used. Can be output at high speed, and in the case of a pattern that does not match, a similar sentence pattern is extracted by a similar sentence search unit, a display sentence is generated by an inquiry unit, and the user is allowed to select a sentence that matches his or her intention. , The semantic structure is output based on the selection, so that various expressions of the user can be absorbed. In addition, if a sentence of the same pattern is input several times, the input sentence having the pattern is input by the user. The semantic structure can be output at high speed while reducing the burden of the selection operation.

In addition, the system designer registers the basic sentence accepted by the system as a regular sentence in the regular sentence pattern dictionary, tests the system with a large number of users, and checks the sentence pattern stored in the pattern storage unit. By registering a sentence pattern having no ambiguity in a target field as a new regular sentence, a dialog processing system capable of absorbing more various expressions can be constructed.

[Brief description of the drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of the present invention.
Fig. 3 is a definition example of the semantic attributes of words, Fig. 3 is an example of a regular sentence pattern dictionary, Figs. 4 and 5 are examples of input patterns, Fig. 6 is an example of display sentences, and Fig. 7 is a pattern storage unit. FIG. 6 is a diagram showing an example of sentence patterns stored in the embed. In FIG. 1, 1 is a regular sentence pattern dictionary, 2 is a pattern storage unit, 3 is a morphological analysis unit, 4 is a sentence pattern matching unit, 5 is a control unit, 6 is a similar sentence search unit, 7 is a query unit, and 8 is an inquiry unit. The learning processing means 9 is a semantic structure generating means.

──────────────────────────────────────────────────続 き Continuation of the front page (56) References JP-A-61-150068 (JP, A) JP-A-61-260367 (JP, A) JP-A-59-183469 (JP, A) Matsuo, Oyama, Nakagawa , "User input support for Japanese-language dialog processing", Proc. Of the 38th Annual Meeting of the Information Processing Society of Japan (Early 1988), p. 400-p. 401 (Mar. 17, 1989) (58) Field surveyed (Int. Cl. ⁶ , DB name) G06F 17/20-17/28 JICST file (JOIS)

Claims (1)

(57) [Claims] 1. A natural sentence semantic analysis processing device having means for inputting an input sentence and means for outputting a semantic analysis result of the input sentence, comprising: a regular sentence pattern defined by a permutation having semantic attributes and words as constituent elements. And a regular sentence pattern dictionary that stores the semantic structure defined for the regular sentence pattern, an input sentence divided into words, and the semantic attributes of each word are extracted. A morphological analysis unit for generating, a pattern storage unit for storing a sentence pattern and the name and frequency of a regular sentence pattern corresponding to the sentence pattern, and a regular sentence pattern that matches the input pattern obtained by the morphological analysis unit. The regular sentence pattern is extracted from the regular sentence pattern dictionary, and the regular sentence pattern is set as an analysis target pattern. A sentence pattern that matches the input pattern is searched from the pattern storage unit. If the frequency of the sentence pattern exceeds a predetermined threshold, the sentence pattern is extracted, and a sentence pattern matching unit that uses a regular sentence pattern corresponding to the sentence pattern as an analysis target pattern, and a normal pattern similar to the input pattern. A similar sentence search means for extracting a sentence pattern from the regular sentence pattern dictionary; and a display sentence for each of the regular sentence patterns based on the regular sentence pattern obtained by the similar sentence search means and words of the input sentence. Query means for displaying a sentence that matches the user's intention and allowing the user to select a regular sentence pattern corresponding to the selected display sentence as an analysis target pattern; Otherwise, the input pattern is used as the sentence pattern of the pattern storage unit, and the name of the regular sentence pattern corresponding to the pattern to be analyzed. As a name of a regular sentence pattern corresponding to the sentence pattern, accumulating a predetermined initial value as its frequency, and updating the corresponding frequency if it has already been accumulated; A semantic structure generating means for generating a semantic structure based on the sentence; and a sentence pattern generating means for sending the sentence pattern generating means to the semantic structure generating means if the pattern to be analyzed is obtained; And a control means for sending to the similar sentence search means.