JPH0743728B2 - Summary sentence generation method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Industrial Application] The present invention relates to a summary sentence generation method capable of effectively generating a high-quality summary sentence of a sentence composed of natural language.

(Prior Art) Conventionally, the generation of a summary sentence of a sentence composed of a natural language generally involves a sentence including a large number of frequently appearing words in the entire sentence,
For example, a sentence including a specific word or expression such as "important thing ..." is judged to be a sentence of high importance and extracted, and these sentences are connected.

However, since such a sentence extraction criterion is merely focused on the frequency of occurrence of a word or a specific word, a plurality of sentences having the same meaning in content are likely to be extracted in duplicate.
As a result, there has been a problem that redundant summary sentences are often generated. In addition, since the sentence extraction standard is not based on the flow of the whole sentence, it is difficult to determine the connection relation between the extracted plural sentences, and it is difficult to generate an appropriate summary sentence. In addition, the essential sentences generated by connecting a plurality of extracted sentences often include a necessary conjunction or an inappropriate conjunction, which makes the meaning of the sentence ambiguous. There was a problem that it was easy to become a thing.

On the other hand, recently, various attempts have been made to analyze the context structure of the whole sentence and examine the flow of the whole sentence and the relative relationship between the individual sentences constituting the sentence. The analysis of this context structure can be performed, for example, in “Information Processing Society of Japan, Research Report Vol.89 No.6 89-NL-70-2 pp1-8 1989.1.20”, “Information Processing Society of Japan Symposium Vol.89 No.5 pp125-”. 136 1989.11 ”, etc., basically, using a rhetorical expression such as a conjunction that appears in a sentence as a clue, a preferable sequence as a connection relation between a plurality of sentences from the rhetorical expression is referred to as a sequence between connection relations. It is done by looking up the rules.

However, even if the context structure of the entire sentence is obtained by such an analysis process, in general, only the flow of the argument and the connection relation between the multiple sentences forming the sentence are shown, and such information can be obtained. The method of generating a summary sentence using is not clarified.

(Problems to be Solved by the Invention) As described above, in the conventional summary sentence generation process, it is easy to duplicate sentences having the same meaning in terms of content, and the connection relation between the plurality of extracted sentences. Is difficult to determine,
When a plurality of extracted sentences are connected to each other to generate a summary sentence, there are problems such as a redundant summary sentence and the possibility that an inappropriate conjunction is often included in the summary sentence.

The present invention has been made in consideration of such circumstances, and an object thereof is to extract a high-quality summary sentence by extracting important sentences in a sentence while clearly determining the connection relation between sentences. It is to provide a summary sentence generation method that can be effectively generated.

[Structure of the Invention] (Means for Solving the Problem) A summary sentence generation method according to the present invention provides a means for analyzing a sentence in natural language to obtain a logical structure of the whole sentence and a connection relation between a plurality of sentences. The sentence structure is composed of two sentences based on a means for expressing the logical structure of the sentence in a tree structure using a co-categorical marker as a unit and a selection rule corresponding to the pre-stored co-categorical marker. The minimum unit structure that is a partial tree structure is extracted, the extracted minimum unit structure is classified according to the co-categorical marker included in the minimum unit structure, and the preceding sentence or the latter sentence or both sentences in two sentences are deleted according to the classification. Means for recursively repeating the rejection operation, and means for extracting a sentence remaining in the tree structure after the rejection operation as an important sentence,
And a means for obtaining a summary sentence by generating a connection expression between the important sentences using the co-categorical markers remaining in the tree structure after the rejection operation.

(Operation) According to the present invention, a context structure indicating a flow of the whole sentence is obtained by analysis processing, and a plurality of sentences forming the sentence according to this context structure are combined into a common category representing a connection relation between the plurality of sentences. The tree structure is represented using a dynamic marker, the minimum unit structure which is a partial tree structure composed of two sentences is extracted from the tree structure based on the selection rule corresponding to the co-categorical marker, and the extracted minimum unit structure is extracted. Are classified according to the co-categorical markers included in the minimum unit structure, and the rejection operation of deleting the preceding sentence, the latter sentence, or both sentences in the two sentences is recursively repeated according to the classification. Sentences remaining in the sentence can be extracted as important sentences, and after the rejection operation, the connected representation between the important sentences can be generated using the co-categorical markers remaining in the tree structure to obtain a summary sentence. it can.

(Embodiment) Hereinafter, a summary sentence generation method according to an embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a schematic configuration diagram of a document processing device configured by applying the embodiment method, and 1 is a sentence input unit for inputting a sentence (text data) in natural language. The context structure analysis unit 2 uses a rhetorical expression such as a conjunction that appears in the text with respect to the text input from the text input unit 1 by using, for example, the method described above, and a plurality of the texts are extracted from the rhetorical expression. The preferred sequence as the connection relation between the connections is checked with reference to the sequence rule between the connection relations, and the context structure indicating the argument structure is obtained.

According to the context structure of the input sentence analyzed as described above, the tree structure generating unit 3 uses a plurality of sentences forming the above-mentioned input sentence as a unit of sentence and a co-categorical marker indicating a connection relation between the plurality of sentences. It is a tree structure expression.

The connection relation judging unit 4 is registered in the rule unit 5 in advance,
One of the sentences is referred from the co-categorical indicator indicating the connection relation between a plurality of sentences in accordance with the tree structure indicating the context structure generated by the tree structure generation unit 3 with reference to the selection rule specific to the connection relation between sentences. Alternatively, it is determined whether or not both can be rejected. When a sentence to be rejected is found, the information about the sentence is rejected from the tree structure. The unnecessary sentence rejection operation by the connection relation determination unit 4 is recursively repeated while modifying the tree structure. When a predetermined number of reject operations are repeated, or when the number of sentences left by the above reject operation is less than or equal to a predetermined number of sentences, unnecessary reject operations are stopped, and the remaining sentences are removed from the tree. By connecting in accordance with the structure, a summary sentence for the input sentence is generated and output.

The flow of such summary sentence generation processing will be described in more detail.

For example, when a sentence as shown in FIG. 2 is input from the sentence input unit 1, the context structure analysis unit 2 detects, for example, a punctuation mark as a division of minutes, and a conjunction or a specific sentence expression appearing in the sentence. The connection relation between multiple sentences is examined by using the rhetorical expression of. The connection relation between the sentences is obtained by referring to a dictionary constructed as shown in FIG. 3, for example, and connecting the words (sentence expressions) with the connectives or specific sentence expressions appearing in the sentence as clues. The "connection type" and the "connection relationship name", which indicate the connection relationship between the sentences, are respectively obtained, and are represented by a co-category indicator symbol. At this time, a series preferable as a connection relation between a plurality of sentences is checked from the rhetorical expression over the entire sentence with reference to a series rule between the connection relations, and a context structure indicating a logical structure of the input sentence is determined.

The connection relation between the sentences thus obtained is information such as "overlap", "supplement", "reason", etc., as the connection relation of "statement type" as shown in FIG. "Forward connection,""reverseconnection,""samerow," etc. as the connection relation, and "parallel", "contrast", "conversion", etc. as the connection relation for "parallel type", and other connection relations It consists of information such as “plan” and “reference”.

The context structure analysis unit 2 analyzes the input sentence as shown in FIG. 2 in this way, and the input sentence is ...
It consists of 8 sentences of ((((1 → 2)-(((3 → 4) ap5) × 6)) →
7) * 8) It is found that there is a connection relationship. In addition, the co-categorical mark symbol shown here [→] indicates “forward connection”, [−] indicates “contrast”, [×] indicates “reverse connection”, and [*] indicates “reference”, [Ap] indicates "continuation", respectively.

The tree structure generation unit 3 follows the context structure showing the connection relation in sentence units of the input sentence obtained as described above.
For example, as shown in FIG. 5 (a), a tree structure is generated, and the subcategory indicator symbol described above is attached to the node part of each subtree indicating the connection relation between individual sentences, whereby the input sentence Expresses the argument structure of, that is, its contextual structure. In this way, according to the context structure of the input sentence represented by the tree structure using the co-category indicator symbol, the process of generating a summary sentence is executed under the connection relation judging unit 4 as shown in FIG. 4, for example. It

This processing procedure is started by first initializing an upper limit value I that defines the number of sentences of the summary sentence to be generated (step a). The upper limit value I defines how many sentences or less are to be generated as the essential translation sentences. Then, as a preprocessing, as described above, from the context structure of the input sentence expressed in a tree structure using the co-category indicator symbol, "reference type""scheduledtype"
It is determined that the partial structure (subtree) having the co-category indicator symbol of is a redundant sentence for constructing the summary sentence, and a process of removing the partial structure from the context structure of the input sentence is executed ( Step b). Specifically, in the context structure of the tree structure shown in FIG. 5 (a), since the partial structure having the "reference type" co-category indicator symbol [*] is a sentence, this sentence is included in the input sentence. , And the tree structure is deleted from FIG. 5 (b).
Modify and change as shown in.

After executing such a processing procedure, the number of remaining parts in the tree structure, that is, the number of sentences included in the context structure shown in FIG. Initialize (step c). Then, it is determined whether or not the control value J indicating the number of sentences has reached the allowable number of abstract sentences as the above-mentioned upper limit value I (step d). When it is detected that the following is true, the tree structure that has been modified by the deletion of the sentence as described above (irreduced context structure)
According to the above, they are connected and output as a summary sentence (step e).

However, in general, it is impossible to keep the number of sentences below the upper limit value I only by deleting the substructures (subtrees) having the "reference type" and "scheduled type" co-category indicator symbols described above. is there.

Therefore, in such a case, pay attention to the minimum unit part in which two sentences are directly structured as shown in the form [sentence connection sentence] from the tree structure showing the context structure of the whole sentence. Then, irreducible reduction is performed on the minimum unit part. This irreduction is performed based on a selection rule stored in advance in the rule section 5 and peculiar to the connection relation between sentences. For example, kth (k = 1, ... L; L is the total number of unit structures). The minimum unit structure of is between the sentence Nk and the sentence Mk,
When the connection relation is represented as [Rk] [Nk, Rk, Mk], (1) When the connection relation Rk is a series type, the whole unit is a sentence Mk.
Replace with.

[Nk, Rk, Mk] → [Mk] (2) When the connection relation Rk is assertive, the whole unit is sentence Nk
Replace with.

[Nk, Rk, Mk] → [Nk] (3) Connection relation When Rk is a parallel type, the entire unit is deleted.

It is performed according to the rules such as [Nk, Rk, Mk] → [Delete]. Such irreducible processing is recursively repeatedly executed from the portion of the minimum unit structure of the tree structure showing the context structure.

That is, in this irreducible processing, first, all of the minimum unit structures [Nk, Rk, Mk] are extracted from the tree structure showing the context structure (step f), and the total number L of the minimum unit structures is set in the control parameter K. (Step g). Then, for the minimum unit structure specified by the control parameter K, it is determined whether the connection relation Rk is "statement type" or "serial type" (steps h and i). When the connection relation Rk is "statement type", the minimum unit structure [Nk, Rk, Mk] is replaced with the sentence Nk (step j). When the connection relation Rk is "serial type", the minimum unit structure [Nk, Rk, Mk] is replaced with the sentence Mk (step k).
If the connection relation Rk is neither "series type" nor "statement type", it is shown that the remaining "parallel type" is apparent from the connection relation shown in FIG. , The entire minimum unit structure [Nk, Rk, Mk] is deleted (step m).

While decrementing the control parameter K by such processing (step n), until the value of the control parameter K becomes zero (0), that is, all the minimum unit structures [Nk, Rk, Mk extracted as described above are obtained. ] Is repeatedly executed until the processing for each of the above is completed (step o).

By such irreducible processing, sentences and sentences are deleted from the context structure as shown in FIG. 5 (b) according to the rules for the "serial type" connection relation, and the tree structure is changed to FIG. 5 (c). It is irreducible as shown. The input sentence shown in FIG. 2 is organized as shown in FIG. 6 (a).

When the above irreducible processing is completed, the number of sentences at that time is obtained, and the number of sentences is set as a new control value J (step p). Then, it is determined whether or not the number of sentences indicated by the control value J is equal to or less than the upper limit value I described above (step q), and when the allowable number of sentences I has not been reached, the irreducible processing described above is recursively performed. To repeat.

For example, when the number of sentences [2] is set as the upper limit value I, the irreducible processing described above is executed again. And the fifth
From the minimum unit structure [Nk, Rk, Mk] of the tree structure shown in Fig. (C), sentences having a "parallel type" connection relation are found,
By deleting each of these sentences, a tree structure as shown in FIG. 5 (d) is obtained, and the sentences are organized as shown in FIG. 6 (b).

Further, when the sentence number [1] is set as the upper limit value I, the irreducible processing is started again for the irreducible processing result as shown in FIG. 5 (d). And FIG. 5 (d)
From the minimum unit structure [Nk, Rk, Mk] of the tree structure shown in Fig. 5, sentences "in parallel connection" are found, and by deleting each sentence, the tree structure is changed to Fig. 5 (e )
Is corrected as shown in FIG. 6 and the sentence is arranged as shown in FIG.

In this way, when the sentences less than or equal to the upper limit value I are extracted from the input sentence, that is, the number of sentences less than or equal to the upper limit value I is reduced by the irreducible processing of the context structure shown in the tree structure. When left, these sentences are connected according to the connection relation between the sentences, and this is generated and output as a summary sentence (step e). In this summary sentence generation process, for example, the sentences belonging to the irreducible context structure (tree structure) are arranged in order from the left-most sentence in the earliest appearance order, and the sentences shown in the tree structure between them are displayed. It is performed according to the category sign symbol while inserting the connection expression corresponding to the symbol.

Thus, the irreducible processing of the context structure in this embodiment is
It is executed while saving the connected expressions that should be included as a summary sentence as co-categorical sign symbols in the tree structure. Then, the irreducible process for the context structure is executed while deleting the sentence determined to be unnecessary from the connection relation in the subtree which is the minimum unit structure. As a result, it is possible to effectively extract only important sentences as summary sentences and create summary sentences, and even if the number of sentences included in the summary sentences changes, Since the relation between them can be saved as a context structure, it is possible to always create a high-quality summary sentence with an appropriate connection expression.

When there are a plurality of sentences having the same meaning in the input sentence, they are eliminated as follows. For example, when the context structure is obtained as (((1 → 2) = (3 ← 4)) → 5) by the analysis process for the input sentence and the tree structure is obtained as shown in FIG. By irreducible processing, the structure of ((2 = 3) → 5) is obtained. Here, the co-categorical mark symbol [=] that connects sentences means "same sequence", and sentences having the same meaning are included in parallel in the summary sentence. However, in this case, if the above-mentioned upper limit value I is the number of sentences [2], the irreducible processing is performed again on this irreducible context structure, and the connection relation between the sentences and the “serial type The sentence will be deleted from the rule. Therefore, if there is a margin in the length of the summary sentence, there is no problem even if there is a sentence with the same meaning in the above-mentioned parallel relationship, but if there is no margin, set the upper limit value I. By fixing
One of the duplicate sentences can be effectively deleted. Therefore, by recursively repeating the irreducible processing in this manner, it becomes possible to easily and effectively generate a high-quality summary sentence of an appropriate expression.

The present invention is not limited to the above embodiment. In the embodiment, the upper limit value I of the number of sentences as the summary sentence is defined and the irreducible processing for the context structure is repeatedly performed, but the number of repetitions itself may be limited. Further, the sentence deletion regulation according to the connection relation may be defined more finely according to the connection relation name. In addition, the present invention can be variously modified and implemented without departing from the scope of the invention.

EFFECTS OF THE INVENTION As described above, according to the present invention, the minimum unit structure, which is a partial tree structure composed of two sentences, is extracted from the tree structure in which the logical structure of a natural language sentence is expressed in sentences. The extracted minimum unit structure is classified according to a co-categorical marker included in the minimum unit structure, and a rejection operation for deleting the preceding sentence, the latter sentence, or both sentences in two sentences is recursively repeated according to the classification, and the rejection operation is performed. After that, since the two sentences remaining in the tree structure are extracted as important sentences, the connection expression between the important sentences is generated and summarized by using the co-categorical markers remaining in the tree structure after the rejection operation described above. You can get a sentence. Therefore, the practically great effect such as the effective adjustment of the length of the summary sentence can be achieved.

[Brief description of drawings]

FIG. 1 shows an essential sentence generation method according to an embodiment of the present invention. FIG. 1 is a schematic configuration diagram of a natural language processing device configured by applying the embodiment method, and FIG. 2 is an input sentence. FIG. 3 is a diagram showing classified expressions of connection expressions used in context analysis, co-categorical sign symbols, and their properties, and FIG. 4 is a diagram showing an algorithm of essential sentence generation processing in the embodiment, FIG. 5 is a diagram showing a tree structure representation of a context structure using a co-categorical sign and changes in the tree structure due to irreducible processing, and FIG. 6 is a diagram showing an example of sentences organized by the irreducible processing. , FIG. 7 is a diagram for explaining a process for avoiding duplication of sentences having the same meaning. DESCRIPTION OF SYMBOLS 1 ... Text input part, 2 ... Context structure analysis part, 3 ... Tree structure generation part, 4 ... Connection relation determination part, 5 ... Rule part, b ... Deletion processing of sentences according to connection relation, e ... Necessary according to irreducible context structure Translation processing, f ... Extraction of unit structure part from tree structure, j, k,
m ... Sentence deletion processing according to connection relationships (irreducible processing of context structure).

Claims (1)

[Claims] 1. A method for analyzing a sentence of a natural language to obtain a logical structure of an entire sentence, and a co-categorical marker indicating a connection relation between a plurality of sentences, to form a tree with the logical structure of the sentence as a unit. Based on the means for expressing the structure and the selection rule corresponding to the pre-stored co-categorical marker, a minimum unit structure which is a partial tree structure composed of two sentences is extracted from the tree structure, and the extracted minimum unit structure is extracted. Means for recursively repeating a rejection operation for classifying according to a co-categorical marker included in the minimum unit structure and deleting the preceding sentence or the latter sentence or both sentences in the two sentences according to the classification; and the tree structure after the rejecting operation. Means for extracting the remaining sentences as important sentences, and means for obtaining a concatenated sentence by generating a connection expression between the important sentences using the co-categorical markers remaining in the tree structure after the rejection operation. Features that are equipped Statement generation scheme.