JP2002073644A - Device and method for extracting and processing important statement and computer readable storage medium stored with important statement extraction processing program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To perform important statement extraction with high accuracy without needing a large number of calculation resources. SOLUTION: When important statement extraction processing that extracts a statement constituting sentence data as an important statement from inputted sentence data is performed, weighting (112) by subject chaining and weighting (113) by representation chaining are performed on the basis of the analyses of the subject chaining and the representation chaining of each statement about the statement constituting the sentence data, the weights of respective statements obtained in a weight area 122 are totalized, the weight of each statement is further adjusted (114) on the basis of the analysis of the connection relation of a plurality of statements, and a statement coinciding with a prescribed extraction condition is extracted as important statement. In the adjustment (114) of the weight of each statement, topological sort is utilized on the basis of the connection relation.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention stores an important sentence extraction processing apparatus, an important sentence extraction processing method, and an important sentence extraction processing program for extracting a sentence constituting sentence data as important sentences from input sentence data. The present invention relates to a computer-readable storage medium.

[0002]

2. Description of the Related Art Conventionally, in the field of natural language analysis using a computer, an automatic summarization system for extracting an important sentence from input sentence data and creating a summary sentence has been known.

In such a system, an important sentence is extracted from sentence data input based on a specific algorithm, and the extracted important sentence is used as it is, or appropriate automatic editing is performed to connect the extracted important sentence. The summary sentence is created by outputting through the above processing.

[0004]

The known important sentence extraction algorithm (method) can be mainly classified as follows according to the information to be used.

[0005] 1) Appearing keywords or their frequencies 2) Positional information in sentences 3) Information such as titles 4) Automatic learning 5) Information such as relationships between sentences 1) The methods of 1) are "in short" and "conclusion". By searching for a keyword that indicates an important sentence such as "" (including when the user specifies it as a search target) or a keyword specified from the frequency information of words that appear, the sentence containing that keyword is used as an important sentence. It is to extract.

The method 2) performs extraction based on the presumption that positional information of a sentence in a sentence, for example, the first sentence, the last sentence, etc., is important. Can also be classified into this category.

The method 4) is a method in which information such as keywords and position information used in the methods 1) to 3) is stored as a dictionary, and the dictionary is automatically updated.

[0008] The method 5) analyzes the relationship between a plurality of sentences constituting a sentence and uses the analysis result. For example, the relationship between sentences such as concatenation and chain is analyzed, the weight of each sentence is determined from the relationship between the main and subordinate sentences, and a sentence having a large weight value is extracted.

Conventionally, the existence of a concatenation theory and a chain theory for analyzing the relationship between sentences has been known, but in the conventional summarization system, weights are determined mainly by using the concatenation theory.
The concatenation refers to a relationship between a plurality of sentences connected by a connective such as "consequence".
As an example, two sentences connected by such a conjunction will naturally have a higher weight in the later sentence.

[0010] However, analysis using the conventional connection theory does not provide sufficient accuracy. The reason is that conjunctive theory only states the relationship between two sentences, and does not detect the role of the sentence throughout the sentence. It is difficult to imagine that a method relying only on the connection relation has detected sufficient information for extracting an important sentence.

[0011] On the other hand, the chain theory uses the elements of a sentence such as a subject, a predicate, and a particle to clarify the function of the sentence in the whole sentence. There are two types of statement chains. Linkage theory can be thought of as trying to identify the role of a sentence in the whole sentence.

Conventionally, there has been little known about a sentence analysis technique that makes use of the characteristics of the above-described connection relationship analysis and chain relationship analysis.

It is an object of the present invention to analyze an articulation and a linkage of sentences when extracting an important sentence from an input sentence, and to use such information to obtain high accuracy without requiring a large amount of computational resources. The purpose is to extract important sentences.

[0014]

In order to solve the above-mentioned problems, the present invention provides an important sentence extraction processing apparatus and an important sentence extraction process for extracting a sentence constituting sentence data from input sentence data as an important sentence. A method and, for each sentence constituting the sentence data in a computer-readable storage medium storing an important sentence extraction processing program, determining a sentence weight from a subject chain weighting rule based on an analysis of a subject chain of the sentence; For each sentence constituting the data, the weight of the sentence is determined from the statement chain weighting rule based on the analysis of the statement chain, and the weights of the respective statements obtained by the subject chain weight and the statement chain weight are summed up. Based on the analysis of the concatenation relation of a plurality of sentences among the sentences that constitute the data, each sentence obtained by the sum of the weights from the concatenation weighting rule Adjust the body, on the basis of the weights adjusted by the weight adjusting adopts the configuration that extracts a sentence that matches the predetermined extraction condition as an important statement.

[0015]

Embodiments of the present invention will be described below with reference to the accompanying drawings.

[System Configuration] FIG. 1 shows the configuration of a processing system capable of implementing the present invention. The configuration shown can be implemented on general computer hardware such as a personal computer or a workstation,
It comprises control means 11 and storage means 12.

The control means 11 includes, in terms of hardware, a CPU for executing each processing step indicated by reference numerals 111 to 115 in the block, and a semiconductor memory or an external storage device ( A storage means such as a hard disk, a floppy (registered trademark) disk, or another storage medium.

Although the storage means 12 can be originally considered as a part of the control means 11, the configuration for storage used in the present embodiment is separately illustrated, and the text data 121 to be processed, Processing steps 111 to 115
(A specific illustration is omitted), constant information necessary for processing (weighting rule described later) 1
24, 126, and 127, and a weight area 122 used for storing weight data of a sentence as output data.

The storage unit 12 is composed of a semiconductor memory such as a RAM (or a virtual storage unit configured using an external storage device).

The important sentence extraction process 111 of the present embodiment firstly inputs the sentence data 121, decomposes it into a known single sentence,
It performs decomposition processing into elements (morphemes) such as words or parts of speech, analyzes subject chains, declarative chains and concatenation relations for each sentence, and analyzes the subject chain analysis results 123 and the statement chain analysis results (however, at this stage, The storage unit 12 stores the analysis result “125” (described later) and the analysis result 128 of the connection relationship.

Subsequently, the important sentence extraction processing 111 includes a weighting processing 112 based on a subject chain based on the analysis results 123, 125, and 128 of the respective chains and connection relations, a weighting processing 113 based on a statement chain, and a weight adjustment based on the connection relations. The processing 114 is executed, and the weighted data is output to the weight area 122.

The weight is a score which is individually assigned to each sentence which is an element constituting the sentence data 121, and is used to store a weight value for each sentence as shown in the lower part of FIG. An area is prepared in the weight area 122. Then, as a result of the important sentence extraction processing 111, the weight area 1
The final important sentence can be extracted by selecting a sentence to be output by an appropriate algorithm from the weight data output to 22.

Note that the subject chain, the declarative chain, and the concatenation relationship may be executed collectively without separately performing the analysis and the weighting. For use, it is desirable that at least analysis of the concatenation relation of each sentence be completed before weighting of the statement chain.

[Whole Flow of Weighting Process] FIG. 2 shows a flow of the whole weighting process in the important sentence extraction process of the present embodiment. At this stage, the analysis of the subject chain, the analysis of the statement chain (however, at this stage, up to the analysis of the sentence unit), and the analysis of the concatenation relationship have been completed. Shall be stored in

The details of the analysis of the subject chain, the statement chain, and the concatenation relationship will be described later when the respective weighting processes are described in detail.

In this embodiment, first, the weight area 122 used for output is cleared (step S101).
The weight area 122 includes not only the weight of the final output but also a weight area for storing an intermediate result obtained in each processing step described later.

Subsequently, the text data 121 is read into the input area of the storage means 12 of FIG. 1 (step S102). The text data 121 may be read from other computers (including those on a LAN or the Internet) or data stored in a storage medium such as a floppy disk via an appropriate input means.

The sentence data 121 is passed to an important sentence extraction process 111 of the control means 11. In the important sentence extraction process 111, first, each sentence which is a component of the sentence data 121 is weighted by subject chain 112 (step S1
03 to S105), and weighting 11 by a statement chain
3 is performed (steps S106 and S107). The result of the weighting 112 based on the subject chain is temporarily
(Step S10)
5).

In the present invention, the sentence data 121 is
Decompose the sentence into its constituent sentences, and further disassemble the sentence into elements (morphemes) such as connectives, particles, and verbs in order to analyze the linkage and concatenation described later, or reconstruct them into phrases as necessary There is a need. Such decomposition into sentences and elements may be performed by an appropriate algorithm, for example, various known morphological analysis techniques. For example, decomposition into sentences can be performed by detecting delimiters such as punctuation marks, and processing such as decomposition of sentences into elements (morphemes) such as connectives, particles, and verbs, or processing such as reconstruction into phrases is performed using dictionary data prepared in advance. For example, by performing pattern matching.

The details of the subject and statement chain analysis processing will be described later, but here, the subject chain weighting rule 124 and the statement chain weighting rule 126 stored in the storage means 12 of FIG. 1 are read, and these rules are read. This is performed by executing a weighting process as described later on the basis of each.

Subsequently, the weights obtained by analyzing the subject and the statement chain are summed (step S108), and the sum of the weights is written to the corresponding area in the weight area 122 (step S109).

Further, in the present embodiment, in addition to the weights 112 and 113 based on the subject and the statement chain for each sentence constituting the sentence data 121, weight adjustment processing 114 and 115 based on the concatenation relation of each sentence are performed. (Step S110) Based on the obtained result, the weight of each sentence determined by the weights 112 and 113 based on the subject and the statement chain is adjusted.

In the weight adjustment process 114 by concatenation, the weight adjustment rule 127 is used. At this time, the weight adjustment is not performed simply in the order according to the appearance order of the sentences, but the order determined by the weight adjustment order determination process 115 is determined. Used.

In the weight adjustment order determination processing 115, a dependency graph corresponding to the relation between sentences is created as described later,
The dependency graph is topologically sorted to generate order data, and weight adjustment processing 1 by concatenation is performed according to the order data.
14 applies weight adjustment by concatenation to each sentence.

As described above, accurate weighting can be performed by performing weighting 112 and 113 based on the subject and the statement chain and further performing weight adjustment by concatenation.

The weight obtained for each sentence is the weight area 1
The sentence to be output is output to an area for storing the final output weight 22 (step S111), a sentence to be output is selected by an appropriate algorithm, and a final important sentence is extracted. The output algorithm based on this weight value can be performed by, for example, extracting those whose weight value is equal to or more than a predetermined point (step S112). The output of the important sentence (step S113) can be output as data in an arbitrary file format such as text or HTML to an output device such as another computer or printer.

Hereinafter, each part of the important sentence extraction process described with reference to FIG. 2 will be sequentially described. First, the overall flow of the weighting process using the subject chain and the weighting process using the statement chain will be described.

[Weighting Process by Chain Theory and Subject / Declaration Chain] FIG. 3 shows a flow of the subject chain weighting process 112 in the important sentence extraction process. In the subject chain weighting process, the sentence data 121 is received (step S201), and the subject chain weighting rule 12 is received.
4 is read (step S202), and a loop (steps S203 to S207) is executed. That is, for each sentence constituting the sentence data 121 (step S203), the sentence is read (step S204), and the subject chain analysis result 123 of each sentence is read (step S2).
05), the weight of the sentence is determined based on the subject chain type of the sentence (described later) and the subject chain weighting rule 124 (step S206), and the obtained weight is stored in the weight area 122 for each sentence (step S206). Step S207), for the upper-level important sentence extraction processing routine that called the subject chain weighting processing of FIG. 3, in the form of a pointer to the weight area of each sentence in the weight area 122, the obtained sentence Returns the value of the subject chain weight.

FIG. 4 shows a weighting process 113 based on a statement chain in the important sentence extraction process. In this weighting process, the text data 121 is received (step S30).
1) Read the statement chain weighting rule 126 (step S302), loop (steps S303 to S308)
Execute

The loop of steps S303 to S308 is as follows:
This is executed for each sentence. First, the analysis result 128 of the connection relation is read (step S304), and all the sentences having the connection relation with the relevant sentence are read (step S3).
05), the analysis result of the statement chain of each sentence (in this stage, only the analysis result of the “end” described later) 125 is read (step S306), and the sentence and the sentence having a connection relationship with the sentence are read. Is determined from the statement chain type (described later) and the statement chain weighting rule 126 (step S307), and the obtained weight is stored in the weight area 122 for each sentence (step S308). In the upper-level important sentence extraction processing routine that called the weighting process by the statement chain in FIG. 4, the statement chain of each obtained statement is used in the form of a pointer to the weight area of each statement in the weight area 122. Returns the weight value.

Here, FIGS. 7 and 11 show examples of assignment of weight values actually used in the weighting by the subject chain and the statement chain, and the details of the weighting of the subject chain and the statement chain will be described based on the assignment examples. In the following, how to analyze the type of the subject chain / declaration chain will be described together, and the details of the above flowchart will be supplemented as necessary.

As described above, the chain in the chain theory is
When viewed throughout the sentence, it indicates the function of a sentence. There are two types of chains: subject chains and statement chains. Subject chains focus on the subject of the entire sentence. Is a statement focusing on the statement (predicate) of the entire sentence.

In the analysis by subject chain in this embodiment, sentences are classified into four categories as shown in FIG. That is, there are four phenomena, a judgment, a predicate, and a quasi-judgment.

Of these, the phenomenon sentence is a sentence whose subject is indicated by something similar to the case particle "ga". Similarly, the judgment sentence refers to a sentence indicated by something similar to the case particle “ha”. The quasi-judgment sentence is a form in which the subject of the judgment sentence is omitted.
A predicate sentence is a sentence that does not require a subject. For example, when analyzing the subject chain for the following sentences: One Sunday morning. (Predicate sentence) I got up early after a long time. (Decision sentence) I can hear the singing of the little bird called "Twipie". (Phenomenon statement) I opened the curtain. (Assistant sentence) It was a great tit. (Quasi-judgment sentence).

Therefore, as for the phenomena sentence and the judgment sentence, if the sentence has been previously decomposed into morphemes, for example, parts of speech, by a known morphological analysis, a dictionary in which case particles "ga" and "ha" are prepared in advance It can be detected by performing pattern matching using data. The predicate sentence is detected by judging that the subject does not exist based on the detection of the case particle, and the quasi-judgment sentence is judged by judging that the subject is a sentence pattern in which the subject is omitted based on the detection of the case particle. be able to.

The above sentence example and the sentence examples in FIGS. 9 to 13 which will be described later are described in Ken Nagano's "Sentence theory review" (ISB
N4-254-51011-X C3081).

In the weighting based on the subject chain according to the present embodiment, the process of classifying into four types of the phenomena sentence, the judgment sentence, the predicate sentence, and the quasi-judgment sentence is performed in the case of FIG.
Has already been obtained. In the weighting of FIG. 3, the subject chain analysis result 123 is read (step S20).
5) According to the result, weights are allocated as shown in FIG. 7 (step S206). At this time, the weight allocation may be changed according to the appearance position in the paragraph of the sentence as shown in the figure.

That is, in FIG. 7, “transcription” and “transcription” as the appearance positions of the sentence indicate whether the sentence is at the start position in the paragraph (the first sentence in the paragraph), respectively. Indicates whether the sentence is at the end position in the paragraph (is the last sentence in the paragraph). In the present embodiment, subject chain analysis is performed only on the sentences in the “transcription” and “transcription” in the paragraph, and weighting is performed only on these sentences. This is because an important sentence is likely to appear as a sentence at the start and end of a paragraph, and it is intended to save computational resources by processing only these sentences. Alternatively, subject chain analysis and weighting may be performed for all sentences using an appropriate weighting rule.

In the example of FIG. 7, the judgment sentence and the predicate sentence are irrespective of the appearance position of the sentence (regardless of whether the appearance position is "transcription" or "writing").
The weight value of +10 is similarly assigned to the quasi-predicate sentence regardless of the appearance position of the sentence.

For the phenomena sentence, a weight value of -8 is set at the transcript position, a weight value of +10 is set at the transcribed position if the preceding sentence is not a phenomena sentence, and-is set otherwise.
A weight value of 3 is assigned.

That is, in the example of FIG. 7, the weighting points of the judgment sentence and the predicate sentence are high, and for the phenomena sentence, when the preceding sentence is not a phenomena sentence at the writing position of the paragraph, the score is raised. At the position, the score is the same as the quasi-judgment sentence.

The weight value as shown in FIG. 7 indicates that the larger the numerical value is, the more likely it is to be an important sentence. However, the weight value allocation for each type of subject chain is appropriate according to the system design concept. Should be determined.

Incidentally, regarding the appearance position in the paragraph of the sentence,
Needless to say, it can be easily detected by detecting a paragraph (which can be detected by detecting the number of indents or line feeds) in advance.

The analysis method using subject chains shown in FIG. 7 is merely an example, and the weight value is adjusted according to the appearance position of a sentence in a paragraph (transcription or writing) as shown in FIG. Processing need not necessarily be performed.

On the other hand, a chain of statements is obtained by classifying a sentence into four types, a relational word, a unified word, a predicate word, and a communicative word by using a word, and then a plurality of sentences (usually two sentences) having a concatenated relationship. It can be analyzed by examining the combination of the words.

Whether a sentence is a relational word, a unified word, a predicate word, or a convey word can be distinguished by the form of the word at the end of the sentence.
For example, in the following sentence, since the auxiliary verb "ne" at the end of the sentence is a contagion, the sentence as a whole is determined to be a contagion sentence.

6. I wonder if the cherry blossoms have bloomed.

More specifically, the functions of relational words, unified words, predicates, and convey words are classified as shown in FIG. In particular,
These words are associated with sentence elements (parts of speech) as shown in FIG.

By the way, it can be analyzed in more detail that the above sentence "I wonder if the cherry blossoms have bloomed" is composed of a chain of "D" as shown in FIG. it can.

That is, from the statement chain, this sentence is composed of “sakura”, “flower”, “blooming” “lyric”, and “no”.
(Relation), "ga" (relation), "ta" (unification),
It is composed of "Kashira" (predicate and handwritten words) and "Ne" (handwritten words). When considering the function of a sentence, it is considered that the last (end of sentence) word represents the function of this sentence, so that usually only the last word of the sentence need be extracted as the word that determines the function of the sentence.

The part-of-speech detection as shown in FIG. 9 can be performed by pattern matching with a dictionary prepared in advance. As a result, the element at the end of the sentence that determines the function of the sentence is detected. It can detect whether the sentence is a unified word, predicate word, or convey word.

The statement chain analysis result 125 in FIG. 1 is the result of analyzing this sentence type of each sentence, and this analysis result has already been obtained when the weighting of the statement chain in FIG. 4 is started. Shall be.

In the weighting process of the statement chain in FIG. 4, by examining the order in which relational words, unified words, predicate words, and convey words appear in a plurality of sentences having a concatenated relation, The type of statement chain of the plurality of sentences can be specified and weighted (step S in FIG. 4).
305 to S308).

For example, as shown in FIG. 11, the type of statement chain can be classified as a structure of a chain of plural sentences, and weights can be assigned according to the type of each statement chain. More specifically, the specific sentence b is replaced by another sentence a.
When the chain relation corresponds to one of the chain relations in FIG. 11, the corresponding weight is set to the specific sentence b.
To be given.

In FIG. 11, the same weight value +15 is assigned to the type of statement chain shown in FIG. 11, but the relationship of statement chains of all two sentences is not listed in FIG. No weight value is assigned to the type of statement chain of (No. 0).

In FIG. 11, the same weight value +15 is assigned to the type of statement chain shown in the figure for simplicity, but it is of course possible to change the allocation more finely.

As described above, in order to classify the structure of the statement chain of a plurality of sentences as shown in FIG.
It is necessary to analyze the concatenation relation of each sentence that constitutes (the concatenation analysis result 128 of FIG. 1 is prepared). However,
In the weighting by the statement chain, it is sufficient that the connection relation can be used to specify at least a plurality of sentences to be weighted by the statement chain, and the type of the connection relation required for weight adjustment by the connection described later (see FIG. 13 and FIG. It is not necessary to use up to the analysis result of FIG.

Alternatively, as a simpler processing, weighting of the statement chain is performed without using the connection relation.
A configuration in which two adjacent sentences are sequentially taken out and weighting of the statement chain is conceived based on the classification of the lexicon may be considered. There is an advantage that a statement chain of a plurality of sentences having a connection relationship that is not directly adjacent can be reflected in the processing.

As described above, the subject chain and the statement chain are analyzed, and the weight can be assigned to each of the sentences constituting the sentence based on the respective results. The weighting rules for the subject chain and the statement chain shown in FIGS. 7 and 11 may be stored as the subject chain weighting rule data 124 and the statement chain weighting rule data 126 using an appropriate data format.

[Connection theory and weighting processing by connection]
In the present invention, weighting is performed by the subject chain and the statement chain as described above, and weight adjustment is performed based on the analysis result of the connection relation. In the weight adjustment by this connection relation, the connection relation of sentences (which sentence is related)
Not only the type of connection but also the type of connection (how they are connected), and weight adjustment is performed according to the result.

The flow of weight adjustment by concatenation is shown in FIG. 5 and FIG. 6. Before describing the details of this processing flow, concatenation of sentences will be described.

Concatenation is a concept indicating the relationship between two or more sentences, and the types are mainly classified into an equal type, a capture type, an expansion type, an opposite type, an additive type, a contrast type, and a conversion type. .

For example, in the following sentence: The daughter of Itoya in Hommachi, Osaka.

2. My sister is 16 and my sister is 15.

3. The national lords kill with bow and arrow.

4. The daughter of Itoya kills with his eyes.

Sentence 2 has a relationship of expansion with respect to 1.
Sentence 3 has a relation of conversion to 2. Sentence 4 has a relation of contrast to 3, and has a relation of expansion to the combination of 1 and 2.

The meanings of the above-mentioned types of connection are as shown in FIG. In FIG. 12, in addition to the above-described types, a stepping stone type in which two sentences are connected with a sentence therebetween, and two or more sentences
Although the stone type that is linked to the sentence is shown, the following focuses on the relationship between two sentences that is considered to be relatively easy to analyze by a computer. Will not be mentioned directly.

FIG. 13 shows a sentence example of each type of concatenation. As shown in the figure, the type of concatenation relation can be identified by detecting a conjunction in many cases. However, there are some types such as a cumulative type which need to detect continuous use of particles.

In the analysis of the connection relation according to the present embodiment, for each of two adjacent sentences constituting the sentence data 121,
The connection is analyzed by pattern matching of a dictionary prepared in advance, and the connection relation for each two sentences is identified by performing processing such as detecting the consecutive use of particles.

Then, as shown in FIG. 14, the weight of each sentence obtained by weighting with the subject chain and the descriptive chain is adjusted in units of two sentences corresponding to the connection relation. In FIG. 14, the weight adjustment is not performed for the conversion type. However, since the conversion type has a function of changing the topic, the weight should not be changed between the two sentences before and after the connection in the conversion type. Because there is no.

The weight adjustment based on the analysis of the type of concatenation shown in FIG. 14 cannot be directly applied in the order of appearance of sentences. This is because weight adjustment is performed in units of two sentences (or more than one sentence when a stone-stone type or the like is detected), and the connection can be regarded as a dependency between sentences. is there. That is, if weight adjustment is performed in an incorrect order, accurate weights cannot be obtained. Therefore, when weight adjustment is performed by concatenation, the order must be determined first.

Therefore, in the present embodiment, when weight adjustment by concatenation is performed as shown in FIG. 5 (process 114 in FIG. 1), the weight adjustment order determination process in FIG. 6 (process 1 in FIG. 1) is performed.
15. Step S403 in FIG. 5 is performed.

That is, in the weight adjustment by concatenation in FIG. 5, first, the weight of each sentence obtained by the subject chain and the statement chain as described above is input (step S40).
1) Then, a weight rule by concatenation (127 in FIG. 1) is read (step S402), and the weight adjustment order determination processing in FIG. 6 is called here.

In the present embodiment, the connection of sentences is considered to be a dependency, and the connection of each sentence (actually, sentence identification data such as a sentence number is used) in the weight adjustment order determination process of FIG. Register the relationship in the dependency graph in association with the dependency.

As is well known, a dependency graph is a concept used to analyze a set of events having a dependency, and a node (also called a vertex or a point) and a branch (an edge or the like) representing the dependency. ). In the field of the graph algorithm, a slightly different term may be used in some cases, but the “dependency graph” in the present embodiment is a so-called directed graph (Directed Graph) in which the relationship between nodes is directional.
h).

In determining the weight adjustment order shown in FIG. 6, the connection relation between sentences is converted into a dependency graph. First, all connection relations are read (step S501). Here, a unique sentence number is assigned to the read sentence, and the sentence number increases in the order of appearance in the sentence.

Then, a loop is executed for all the connected relations of the sentences constituting the sentence (S502), and when the connected relation of the two sentences is equal or captured, the sentence with the larger sentence number is set as the starting point, Generate a branch to be the end point (step S
503, S504). If the concatenation relationship between the two sentences is expanded, reversed, or cumulative, a branch is created with the sentence with the lower sentence number as the starting point and the larger sentence as the end point (step S505).
S506).

When the connection relation between the two sentences is a contrast type (step S507), it is considered that the two sentences depend on each other in a semantically equal relationship, and the relationship between the two sentences is reflected in the weight. First, a branch having a sentence with a high sentence number as a start point and a small sentence as an end point is created (Step S).
508), and create a dummy node connected by a branch in the opposite direction (steps S509 and S510). By performing such processing, the dependency graph is converted into a so-called D
AG (Directed Acyclic Group)
h) can be easily handled.

The dependency graph can be expressed on the storage means by a known arbitrary program method such as various lists, and can be handled by various known list operation methods.

The dependency graph generated based on the connection relation of sentences has a structure as shown in FIG. 15, for example. This figure 1
The dependency graph No. 5 is created from an example sentence of an example shown in FIG. In FIG. 15, six sentences (sentence numbers 1 to
6) (corresponding examples will be described later in detail), of which sentences 2 and 3 and sentences 5 and 6 are in a concatenated connection relationship, and are shown in step S5 in FIG.
Through the processing from 07 to S510, dummy nodes 2 ', 3', 5 'and 6' connected by branches in opposite directions are generated. This is because the contrast type connection relation is a relation in which two sentences are semantically equivalent. Sentence 4 has a branch between two nodes of sentence 3 and sentence 5 because sentence 4 has an opposite type of connection with sentence 5.
FIGS. 15 and 16 will be referred to again in the description of the examples after FIG.

In the process of determining the weight adjustment order of the present embodiment, a one-dimensional structure called a weight adjustment order must be generated from the dependency graph generated as shown in FIG. The topological sort is used to generate the weight adjustment order from the dependency graph expressing the concatenation relation (step S511).

Topological sorting is one of the methods for sorting ordered sets, and is a process of rearranging all points into a one-dimensional structure so as not to contradict partial order relations in the ordered sets. The method of topological sorting is known, and details of the program are omitted. For example, when the dependency graph of FIG. 15 is topologically sorted, a one-dimensional structure as shown in FIG. 16 is obtained. When the topological sort is performed, as shown in FIG. 16, the branches (arrows in the figure) indicating the dependencies are directed from left to right in the figure, and the branches are arranged so that those in the opposite direction do not appear.

In the topological sort, since only the alignment of the dependencies (directions) is a problem, the solution is not unique, and a plurality of correct solutions may exist. For example, FIG.
May be located on the rightmost side of the figure. By the processing based on such topological sort, the sentence data 1
21 can be reduced to only one time.

Further, by using a concatenation relationship, a chain of subjects, and a chain of statements together, more precise weighting and extraction of important sentences can be performed.

The weight adjustment order obtained as shown in FIG.
For example, FIG.
(Step S512).

Referring again to FIG. 5, the weight adjustment processing receives the weight adjustment order from the processing of FIG. 6 (step S40).
4), determine a sentence to be processed first based on the weight adjustment order (step S405), and execute a loop for processing each sentence according to the weight adjustment order (steps S406 to S).
411).

That is, in the order of the sentence number column in the weight adjustment order, for all sentences (steps S406 and S411),
One sentence is selected one by one, and further, for all the sentences that have a connection relationship with the selected sentence (step S40).
7) The weight is adjusted according to the adjustment rule read in step S402.

Then, only when the current weight <the adjusted weight, that is, when the weight value becomes larger (step S409), the adjusted weight is written into the weight area corresponding to the sentence (step S409). S410).

The sequence of weights of each sentence finally obtained by the above processing is returned to the higher-order important sentence extraction processing which called the processing of FIG.

[Output Processing Based on Weight] Weight area 1 in FIG.
The weight of each sentence finally obtained at 22 can be considered as a score indicating the importance of the sentence in the input sentence data 121.

Therefore, an important sentence can be extracted based on the weight obtained in the weight area 122. As a method of determining an important sentence, a method of extracting all sentences having a weight equal to or more than a certain appropriate value, a method of extracting only a sentence having the largest value, and the like can be considered. For example, it is difficult for an end user to determine the value, so it is appropriate to indicate some steps by a programmer or the like (see examples below). Further, at the time of output, a process of appropriately inserting elements such as connectives may be performed to adjust the connection of the extracted sentences as necessary.

[Specific Example of Important Sentence Extraction Processing] Here, the above-described important sentence extraction processing will be described with reference to specific examples. Here, a newspaper article (Tokyo Shimbun, November 15, 1981)
An excerpt from one paragraph of the “Journal Review” on the second page of the Japanese and Korean editions will be used to explain the above-described important sentence extraction processing of the present invention using this article. The following is the full paragraph of this article. "Because they are the representatives of Egypt and Israel who have made the most efforts to reconcile Israel, and thus to bring peace between Israel and the Arab States.
It is well known that Sadat's efforts and that he was assassinated. On the other hand, many readers are bizarre to name Dayan as a representative of Israel. Prime Minister Beguin signed the agreement on behalf of Israel. However, it is unlikely that Beguin had the same long-term peace plan as the late President Sadat. Dayang had more in common. "
Here, it is assumed that the sentence is divided by the reading point “.” (Of course, other division rules may be used).
As a result of the decomposition into sentences, each sentence is assigned a natural number that gradually increases in the order of appearance as shown in FIG. Furthermore, each sentence needs to be decomposed into morphemes in advance by morphological analysis in order to analyze the subject chain, the declarative chain, and the concatenation relationship, but the morphological analysis method is not the gist of the present invention, and Therefore, the detailed description is omitted here.

For each sentence in FIG. 17, first, the above-mentioned subject chain weighting is performed. Here, the weight of each sentence is determined using the analyzed subject chain type (shown in FIG. 17) and the subject chain weighting rule shown in FIG. Weighted area by subject chain weighting (122 in FIG. 1)
The weight of each sentence changes as shown in FIG.

Next, the statement chain is weighted for each sentence in FIG. Here, the weight of each sentence is determined using the type of the analyzed statement chain (shown in FIG. 17) and the rule of weighting the statement chain shown in FIG. A weight region (12 in FIG. 1) is obtained by weighting by a chain of statements.
The weight of each sentence in 2) changes as shown in FIG.

In FIG. 19, +15 is applied only to sentence 3 and sentence 4.
In the case of the sentence shown in FIG. 17, the relation between the words corresponding to the statement chain shown in FIG.
(Unification word)-> Chain relation of sentence 3 (Unification word / predicate word) (Fig. 11
113) and sentence 3 (unified phrase / predicate)-> sentence 4
This is because there is only the chain relation (chain relation 114 in FIG. 11) of (predicate).

In the above processing, the weight of each reference sentence can be determined based on the analysis of the subject chain and the statement chain. Finally, the weight of each sentence is adjusted based on the analysis of the connection relation.

In the weight adjustment processing (processing in FIG. 6), as described above, a dependency graph is created based on the connection relation, topological sorting is performed on the dependency graph, and weights of the sorted sentences are adjusted in order from the top. Here, first, a dependency graph as shown in FIG. 15 is generated based on the analysis of the connection relation (illustrated in FIG. 17) of the sentences (sentence numbers 1 to 6) in FIG.

At first, the dependency graph is empty, and sentences are read in order from sentence number 1. In the state where the sentence 1 is read, nothing is performed because the connection relation is not defined. Next, sentence 2 is read. At this time, since the connection relation is in a trapped relationship with sentence 1, a branch from 2 to 1 is added to the dependency graph.

Next, when sentence 3 is read, since sentence 3 is in a concatenated relation with sentence 2, in this case, first, node 2 of the dependency graph (the node in FIG. 15 corresponds to the sentence number)
Is added to the dependency graph, and dummy nodes 2 ′ and 3 ′ corresponding to nodes 2 and 3 are added.
(A dummy is indicated by the symbol "'"), and a branch from node 3' to 2 'is added to the dependency graph to satisfy the correspondence, and node 2'
A branch similar to that in the case of nodes 2 and 1 is added to the dependency graph between and. In this way, create a dummy node,
The reason why the branch in the opposite direction is generated is that two sentences are semantically equivalent in the contrast type connection relation. The generation of the dummy node is also performed for the sentences 5 and 6 which are also of the contrast type.

By repeating the above processing up to the last sentence in sequence, a dependency graph as shown in FIG. 15 can be generated based on the connection relation of the sentences.

The dependency graph is represented in the storage unit 12 in a data format such as a list. FIG. 20 shows how the dependency graph changes in the storage area during the generation process. In FIG. 20, ｛｝ represents the entire graph,
() Indicates a branch connecting two nodes, the number in () indicates a node (sentence number), and the appearance order indicates the direction (dependency) of the branch. For example, (2,1) is a branch from node 2 (sentence 2) to node 1 (sentence 1). φ indicates that the graph is empty. Of course, the state of the graph when the last sentence 6 in FIG. 20 has been processed is equivalent to the graph in FIG.

If the dependency graph generated as described above is topologically sorted, the order of nodes (sentence numbers) to be weight-adjusted can be obtained as shown in FIG. That is, the order of the sentence numbers to be weight-adjusted based on the connection relation determined based on the connection relation of the sentences is 2, 4,
6 ′, 5,3 ′, 3,5 ′, 6,2 ′, 1. Note that the topological sort may be performed by a known algorithm, and a detailed description thereof will be omitted here.

Once the weight adjustment order has been determined, the weight adjustment processing (FIG. 5) is performed in accordance with the order.
For sentences (or more), adjust the weight of each sentence according to the concatenation relation. The weight adjustment rule by concatenation is shown in FIG.
This is shown in FIG. At this time, the weight of the sentence in the weight area 122 is rewritten only when it is larger than the current weight as described above.

First, the sentence 2 corresponding to the node 2 is extracted in accordance with the above-described weight adjustment order. The connection relation of the sentence 2 is a capture type in relation to the sentence 1 and a contrast type in relation to the sentence 3. . Therefore, the weight of the sentence in the weight area 122 changes as shown in FIG. 21 based on the adjustment rule of FIG.

Next, according to the weight adjustment order, the node 4
, And performs the same processing. Node 4 (sentence 4) has an interlocking relationship between node 3 and dummy node 3 ', and has an opposite type connection relationship between node 4 and dummy node 5'. Therefore, the weight of the sentence in the weight area 122 changes as shown in FIG. 22 based on the adjustment rule of FIG.

At this time, the same weight area as that of the original node is used for the dummy node. Hereinafter, the weight adjustment processing based on the connection relation is performed on all the sentences as described above. As a result, a weight adjustment result as shown in FIG. 23 is obtained.

An important sentence can be extracted by using the weight obtained in the weight area 122 as described above. As a method of determining an important sentence, a method of extracting all sentences having a weight equal to or more than a certain appropriate value, a method of extracting only a sentence having the largest value, and the like can be considered.

For example, several levels (five levels in FIG. 24) of compression ratio levels as shown in FIG. 24 are presented as extraction conditions (extraction criteria) on a display or the like, and are selected by the user, and are associated with the compression ratio levels in advance. Using the threshold value of the weight determined in this way, only sentences having a weight value equal to or greater than the threshold value are output. This output processing can be performed based on screen display, printout, or any protocol (a protocol such as HTTP or SMTP) for a disk file or another computer.

[Effects of the Embodiment] As described above, by performing weighting based on the subject and the statement chain, and further performing weight adjustment by analyzing the concatenation relation, weighting of the sentence can be performed with high accuracy. An accurate important sentence extraction process can be performed. The important sentence extraction can be used, for example, in a process of automating the creation of a summary sentence, which has conventionally been performed manually, and can significantly reduce the labor required for the creation of a summary sentence.

In this embodiment, when weight adjustment is performed by analyzing the connection relation, the connection relation is regarded as a dependency relation, a dependency graph expressing the connection relation of a sentence is generated, and the weight is determined by using topological sorting. The adjustment order is determined, and the weight adjustment based on the connection relation can be performed easily and at high speed using a computer and its software. Fast and reliable extraction of important sentences.

Note that the subject chain, the statement chain,
The allocation of the weight values assigned to the relations such as concatenation and the size thereof are merely examples, and can be arbitrarily changed according to the design concept of the system.

The control program for extracting important sentences according to the present invention can be stored and supplied in any computer-readable storage medium such as a ROM, a hard disk, a floppy disk, an optical disk, a magneto-optical disk, and a memory card. .

[0124]

As is apparent from the above description, according to the present invention, an important sentence extraction processing apparatus, an important sentence extraction processing method, and a method for extracting sentences constituting text data as important sentences from input sentence data. For each sentence constituting the sentence data in a computer-readable storage medium storing the important sentence extraction processing program, the sentence weight is determined from the subject chain weighting rule based on the analysis of the subject chain of the sentence, and the sentence data is determined. For each of the constituent sentences, the weight of the sentence is determined from the statement chain weighting rule based on the analysis of the statement chain of the statement, and the weight of each statement obtained by the subject chain weight and the statement chain weight is summed, and the sentence data is calculated. Among the constituent sentences, based on the analysis of the concatenation relation of a plurality of sentences, the weight of each sentence obtained by the sum of the weights is adjusted from the concatenation weighting rule. Based on the weight adjusted by the weight adjustment, a sentence that matches a predetermined extraction condition is extracted as an important sentence. Therefore, the subject and the statement chain are weighted, and the connection is further analyzed. By performing the weight adjustment in the order obtained as a result of the topological sort, highly accurate sentence weighting can be performed with a small amount of computational resources. Based on this, accurate important sentence extraction processing can be performed. The use of important sentence extraction for the process of automating the creation of a summary sentence, which has been performed, has an excellent effect that the labor required for the creation of a summary sentence can be significantly reduced.

[Brief description of the drawings]

FIG. 1 is a block diagram showing the structure of a processing system capable of implementing the present invention.

FIG. 2 is a flowchart showing an entire flow of an important sentence extraction process according to the present invention.

FIG. 3 is a flowchart illustrating a flow of a weighting process based on a subject chain;

FIG. 4 is a flowchart showing a flow of a weighting process by a statement chain.

FIG. 5 is a flowchart showing a flow of weight adjustment processing by concatenation.

FIG. 6 is a flowchart illustrating a flow of a weight adjustment order by concatenation.

FIG. 7 is a table showing an example of a weighting rule based on a subject chain.

FIG. 8 is a table showing classifications in a statement chain format.

FIG. 9 is a table showing an example of a part of speech corresponding to a classification in a statement chain format.

FIG. 10 is a table showing a structure focusing on a statement chain of a certain sentence.

FIG. 11 is a table showing an example of a weighting rule based on a statement chain.

FIG. 12 is a table showing types of connection relationships.

FIG. 13 is a table showing an example of a connection relationship and a language form.

FIG. 14 is a table showing an example of a weight adjustment rule based on concatenation.

FIG. 15 is an explanatory diagram illustrating a dependency graph expressing a connection relationship.

FIG. 16 is an explanatory diagram showing a result of topological sorting of the dependency graph of FIG.

FIG. 17 is a table illustrating sentences of an actual example to be processed for important sentence extraction according to the present invention.

18 is a table showing the weights of sentences determined by subject chain analysis in the example of FIG. 17;

FIG. 19 is a table showing the weights of sentences determined by analyzing the statement chain in the example of FIG. 17;

20 is a table showing how a dependency graph is generated by analyzing concatenation in the example of FIG. 17;

FIG. 21 is a table showing how weight adjustment is performed in the order determined based on the dependency graph in the example of FIG. 17;

FIG. 22 is a table showing how weight adjustment is performed in the order determined based on the dependency graph in the example of FIG. 17;

FIG. 23 is a table showing how weight adjustment is performed in the order determined based on the dependency graph in the example of FIG. 17;

FIG. 24 is a table showing an example of extraction conditions (extraction criteria) for extracting an important sentence.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 11 Control means 111 Important sentence extraction processing 112 Weighting by subject chain 113 Weighting by statement chain 114 Weight adjustment processing 12 Storage means 121 Text data 123 Analysis result of subject chain 125 Analysis result of statement chain 128 Analysis result of connection relation 124 Subject chain weighting Rule 126 Statement Chain Weighting Rule 127 Weight Adjustment Rule

Claims (6)

[Claims] 1. An important sentence extraction processing device for extracting a sentence constituting sentence data as an important sentence from input sentence data, wherein each sentence constituting the sentence data is analyzed based on a subject chain analysis of the sentence. Subject chain weighting means for determining the weight from the subject chain weighting rule; statement chain weighting means for determining the weight of the sentence from the statement chain weighting rule based on the analysis of the statement chain for each sentence constituting the sentence data; Weight summing means for summing the weights of the respective sentences obtained by the subject chain weighting means and the statement chain weighting means; and Weight adjusting means for adjusting the weight of each sentence obtained by the weight summing means, and based on the weight adjusted by the weight adjusting means. Sentence extraction processing apparatus characterized by comprising an extraction means for extracting a sentence that matches the predetermined extraction condition as an important statement. 2. The weight adjusting means, prior to adjusting the weight of each sentence among sentences constituting the sentence data,
Generating a dependency graph expressing a connection relation between the plurality of sentences, and determining an order for adjusting the weight of each sentence based on order information obtained by topologically sorting the dependency graph. Item 1. An important sentence extraction processing device according to item 1. 3. An important sentence extraction processing method for extracting a sentence constituting sentence data as an important sentence from input sentence data, wherein for each sentence constituting the sentence data, the sentence A subject chain weighting step of determining the weight from the subject chain weighting rule, and a statement chain weighting step of determining the weight of the sentence from the statement chain weighting rule based on an analysis of the statement statement chain for each sentence constituting the text data, A weight summing step of summing the weights of the sentences obtained in the subject chain weighting step and the statement chain weighting step; and A weight adjustment step of adjusting the weight of each sentence obtained in the weight summation step from the above, based on the weight adjusted in the weight adjustment step Sentence extraction processing method characterized by comprising an extraction step of extracting a sentence that matches the predetermined extraction condition as an important statement. 4. The weight adjusting step includes, prior to adjusting the weight of each sentence among sentences constituting the sentence data,
Generating a dependency graph expressing a connection relation between the plurality of sentences, and determining an order for adjusting the weight of each sentence based on order information obtained by topologically sorting the dependency graph. Item 3. An important sentence extraction processing method according to Item 3. 5. A computer-readable storage medium storing an important sentence extraction processing program for extracting a sentence constituting sentence data from an inputted sentence data as an important sentence, wherein a sentence constituting each sentence data is provided. A subject chain weighting step of determining the weight of the sentence from the subject chain weighting rule based on the subject chain analysis of the sentence data; and for each sentence constituting the sentence data, the statement weighting of the sentence based on the statement chain analysis A statement chain weighting step of determining from the following: a weight summation step of summing weights of the respective sentences obtained by the subject chain weighting step and the statement chain weighting step; and a concatenation of a plurality of sentences among the sentences constituting the text data. A weight adjusting step for adjusting the weight of each sentence obtained in the weight summing step from the connection weighting rule based on the analysis of the relations A computer-readable storage of an important sentence extraction processing program, wherein an extraction step of extracting a sentence that matches a predetermined extraction condition as an important sentence based on the weight adjusted by the weight adjustment step is stored. Storage medium. 6. In the weight adjusting step, prior to adjusting the weight of each sentence among the sentences constituting the sentence data, a dependency graph expressing a connection relation of the plurality of sentences is generated. The important sentence extraction processing program according to claim 5, wherein a control procedure for determining an order for adjusting the weight of each sentence based on the order information obtained by topologically sorting the dependency graph is stored. A computer-readable storage medium that stores the information.