JP6021079B2 - Document summarization apparatus, method, and program - Google Patents

Description

  The present invention relates to a document summarization apparatus, method, and program for summarizing a given document.

  In conventional computer document summarization methods, a grammatical element in a document, for example, a continuous word string such as a clause or phrase included in a sentence, is used for each element to maximize a certain objective function. Extract so that the sum of degrees is maximized. In addition, when extracting grammatical elements, it is known that the quality of summarization is improved by considering parent-child relationships between grammatical elements, that is, rhetorical structure, rather than just considering them as a set. (For example, refer nonpatent literature 1).

In the technique described in Non-Patent Document 1, a document is expressed as a rhetorical structure tree in which a grammatical element in a document is “section” and a section is a node. Then, the document summarization is formulated as a combinatorial optimization problem in which a rooted subtree having the maximum sum of the importance of clauses and the summarization length of L _max or less is extracted as a summarization. The summary length is, for example, the number of words or characters included in the summary.

Tsutomu Hirao, Yasuhisa Yoshida, Masaaki Nishino, Norihito Yasudaand MasaakiNagata, "Single-Document Summarization as a Tree Knapsack Problem", Proc. Of the 2013 Conference on Empirical Methods in Natural Language Processing, pp. 1515-1520, 2013.

In conventional summarization techniques, grammatical elements such as sentences and clauses are determined in advance, and the document is represented as a rhetorical structure tree representing the parent-child relationship between those elements, thereby capturing the logical structure of the original document. Summary can be generated. In this case, a summary is generated by extracting continuous word strings in a document such as a predefined sentence or section. When a grammatical element is a sentence, the summary is generated by extracting a sentence to be included in the summary from a sentence included in the original document. However, if the summary length constraint L _max is severe (small), the number of sentences that can be extracted becomes extremely small, and there is a problem that the coverage of information as a summary decreases. On the other hand, when a grammatical element is a section, a summary is generated by extracting a section included in the summary from a section included in the original document. In this case, since a section is a unit smaller than a sentence, the coverage of information as a summary is higher than that when a grammatical element is a sentence, but there is a problem that readability is impaired.

  The present invention has been made in view of the above circumstances, and an object thereof is to provide a document summarization apparatus, method, and program capable of improving the coverage of information as a summary while ensuring readability. And

  In order to achieve the above object, a document summarization apparatus according to the present invention includes a rhetorical structure analysis unit that generates a rhetorical structure tree representing a document based on a rhetorical relationship between sentences in a document including a plurality of sentences. A dependency analysis unit that generates each dependency tree representing each of the sentences based on dependency relationships between words in each of a plurality of sentences, a maximum length of a predetermined summary, and an importance of each of the words At least one subtree is extracted from each of the dependency trees so as to satisfy a predetermined condition based on the rhetorical structure tree and the dependency tree, and based on the extracted subtree And a generation unit for generating a summary of the document.

  According to the document summarizing apparatus according to the present invention, the rhetorical structure analyzing unit generates a rhetorical structure tree representing a document based on a rhetorical relationship between sentences in a document including a plurality of sentences. The dependency analysis unit generates each dependency tree representing each sentence based on the dependency relationship between words in each of the plurality of sentences. Then, based on each of the dependency trees, the generation unit satisfies the predetermined condition based on the predetermined maximum length of the summary, the importance of each word, the rhetorical structure tree, and each of the dependency trees. At most one subtree is extracted, and a document summary is generated based on the extracted subtree.

  In this way, a document is represented by a nested structure of rhetorical structure trees and dependency trees, and at most one subtree is extracted from each dependency tree to generate a summary, thereby ensuring readability. However, the coverage of information as a summary can be improved.

  In addition, the predetermined condition is that the summary length is less than or equal to the maximum length, the rhetorical relationship represented by the rhetorical structure tree and the dependency relationship represented by the dependency tree are not impaired, and the importance of the words included in the summary It can be determined so that the sum of degrees is maximized. As a result, the logical structure of the document and the grammatical nature of the sentence are not impaired, so that a summary that ensures more readability can be generated.

  Further, the predetermined condition can be determined so as to maximize the objective function shown in the following equation (13) under the constraints shown in the following equations (1) to (12). As a result, the problem of document summarization can be formulated.

Where i is the sentence identification number, N is the total number of sentences included in the document, j is the word identification number, M (i) is the total number of words included in the sentence with the identification number i, and x _i is the identification number i. A decision variable that becomes 1 when the sentence is included in the summary, z _ij is a decision variable that becomes 1 when the word with identification number j of the sentence with identification number i is included in the summary, and w _ij is an identification number of the sentence with identification number i The importance of the word j, r _ij is a decision variable that becomes 1 when the word with the identification number j of the sentence with the identification number i is the root of the subtree, L _max is the maximum length, and parent (i) is the rhetorical structure tree A function that returns the identification number of the parent sentence of the sentence with the identification number i, parent (i, j) is a function that returns the identification number of the parent word of the word with the identification number j in the dependency tree representing the sentence with the identification number i. , A is a predetermined coefficient, and R _c (i) is a dependency tree representing a sentence with identification number i. The function root (i) that returns a set of identification numbers of words that are root candidates is a function that returns the identification number of a word that is a true root in a dependency tree representing a sentence with the identification number i, sub ( i) is a function that returns a set of identification numbers of the word that is the subject from the sentence of the identification number i, and obj (i) is a function that returns a set of identification numbers of the word that is the object from the sentence of the identification number i. is there.

  The document summarization method according to the present invention is a document summarization method in a document summarization apparatus including a rhetorical structure analysis unit, a dependency analysis unit, and a generation unit, wherein the rhetorical structure analysis unit reads a plurality of sentences. A step of generating a rhetorical structure tree representing the document based on a rhetorical relationship between sentences in a document including the dependency analysis unit, based on the dependency relationship between words in each of the plurality of sentences, A step of generating each dependency tree representing each sentence; and the generation unit determines a maximum length of a predetermined summary, importance of each word, the rhetorical structure tree, and each dependency tree. And extracting at most one subtree from each of the dependency trees so as to satisfy a predetermined condition, and generating a summary of the document based on the extracted subtrees. It is.

  A document summarization program according to the present invention is a program for causing a computer to function as each unit constituting the document summarization apparatus.

  As described above, according to the document summarization apparatus, method, and program of the present invention, a rhetorical structure representing a document based on a predetermined maximum length of summary, importance of each word, and rhetorical relationship between sentences. Extract at most one subtree from each dependency tree to satisfy a predetermined condition based on each dependency tree that represents each sentence based on the dependency relationship between the tree and the word To generate a summary. Thereby, the effect that the coverage of the information as a summary can be improved while ensuring readability is obtained.

It is a functional block diagram of the document summarization apparatus concerning this Embodiment. It is the schematic which shows an example of a rhetorical structure tree and a dependency tree. It is a flowchart which shows an example of the document summary process routine in this Embodiment. It is the schematic which shows an example of the document summary in this Embodiment.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  FIG. 1 is a block diagram showing a document summarizing apparatus 10 according to an embodiment of the present invention. The document summarization apparatus 10 can be constituted by a computer including a CPU, a RAM, and a ROM that stores a program for executing a document summarization processing routine described later.

  As shown in FIG. 1, the computer constituting the document summarizing apparatus 10 functionally has a sentence dividing unit 12 that divides an input document into sentences, and a word that divides each divided sentence into words. A dividing unit 14, a word importance assigning unit 16 that assigns importance to each word, a rhetorical structure analyzing unit 18 that analyzes a rhetorical relationship between sentences and generates a rhetorical structure tree representing a document, A dependency analysis unit 20 that analyzes each dependency relationship to generate each dependency tree that represents each sentence, a summary length, an importance of each word, a rhetorical structure tree, and a dependency tree Based on this, it can be expressed by a configuration including a generation unit 22 that generates a summary of a document. Hereinafter, each part is explained in full detail.

  The sentence dividing unit 12 receives a document (text data) including a plurality of sentences input to the document summarizing apparatus 10, gives a sentence break to the document, and divides the document for each sentence. An existing sentence divider can be used to identify sentence breaks. In addition, sentence breaks may be given simply using punctuation marks.

  The word dividing unit 14 receives each sentence divided by the sentence dividing unit 12 as an input, gives a word break to each sentence, and divides each sentence into words. An existing morphological analyzer can be used for recognition of word breaks. Further, when a document such as an English sentence in which a word break is explicitly given is input, word division may be performed according to the explicit word break.

  The word importance assigning unit 16 receives each word divided by the word dividing unit 14 as input, refers to the word importance database (DB) 30, and assigns importance to each word included in the input document. . In the word importance DB 30, each of a plurality of words and the importance of the word are stored in association with each other. The importance of each word stored in the word importance DB 30 can be defined using, for example, a tf-idf (term frequency-inverse document frequency) method generally used in an information search system or the like. .

  The rhetorical structure analyzing unit 18 receives each sentence divided by the sentence dividing unit 12 as input, and analyzes the rhetorical relationship between the sentences. For example, using a rhetorical structure analyzer (references “duVerle, D. and Prendinger, H,“ A Novel Discourse Parser Based on Support Vector Machine Classication ”, Proc of the 47th ACL, pp. 665-675, 2009.) Then, after generating an RST (Rhetorical Structure Theory) tree, for example, the rhetorical relationship between sentences can be analyzed by applying the rules described in Non-Patent Document 1. Moreover, it is not always necessary to generate an RST tree, and the rhetorical relationship between sentences may be analyzed using an analyzer that has been learned using data of a rhetorical structure tree that represents the rhetorical relation between sentences.

The rhetorical structure analysis unit 18 represents each sentence as a node based on the analyzed rhetorical relationship between sentences, and generates a rhetorical structure tree in which nodes corresponding to each other in the rhetorical relation are connected. Since two sentences in a rhetorical relationship have one parent and the other are children, the node corresponding to the parent sentence is the parent node and the node corresponding to the child sentence is the child node. Connect. An example of the rhetorical structure tree is shown in the left diagram of FIG. 2 left shows, sentence _{s 1} included in the input document, sentence _s 2, · · ·, each sentence _{s 8,} node _{s 1,} node _s 2, · · ·, node _{s 8} (figure (○) In addition, nodes corresponding to a sentence having a parent-child relationship are connected by arrows (edges) so that the start point side is a child node and the end point side parent node.

The dependency analysis unit 20 receives each word divided by the word division unit 14 as an input, and analyzes the dependency relationship between the words. An existing dependency analyzer can be used to analyze the dependency relationship. The dependency analysis unit 20 represents each word as a node based on the analyzed dependency relationship between the words, and generates a dependency tree in which nodes corresponding to each other in the dependency relationship are connected. Since the two words in the dependency relationship are in a parent relationship and the other is in a child relationship, the node corresponding to the parent word is a parent node, and the node corresponding to the child word is a child node. Connect between them. An example of a dependency tree is shown in the right side of FIG. The right figure of FIG. 2 shows each of the word w ₁ , the word w ₂ ,..., And the word w ₁₂ included in the sentence s ₈ corresponding to the node s ₈ included in the rhetorical structure tree shown in the left figure of FIG. w _1, node _w 2, ···, is represented by a node _{w 12} (figure ○ mark). Further, the nodes corresponding to the words having the parent-child relationship are connected by arrows (edges) so that the start point side is the child node and the end point side is the parent node.

Generator 22 receives a summary length _{L max} input. The summary length is a parameter that restricts the length of the summary. Here, the summary length L _max is the maximum number of words included in the summary. The summary length may be the number of characters. Further, not limited to the case of receiving the summary length L _max input may read the summary length L _max stored in advance such as in the ROM. The generation unit 22 receives each of the rhetorical structure tree generated by the rhetorical structure analysis unit 18 and the dependency tree generated by the dependency analysis unit 20. Further, the generation unit 22 receives the importance level given to each word by the word importance level giving unit 16.

Based on the summary length L _max , the importance of each word, the rhetorical structure tree, and the dependency tree, the generation unit 22 has a number of words included in the summary that is less than or equal to L _max , and the parent-child relationship between words and the sentence spacing The subtrees are extracted from each dependency tree so that the parent-child relationship is not impaired and the sum of the importance of the words included in the summary is maximized. Then, the generation unit 22 generates a summary from the word set corresponding to the nodes constituting each of the extracted partial trees.

  Here, the parent-child relationship between words does not impair that the parent-child relationship between words included in the summary maintains the structure of the corresponding part of the dependency tree generated by the dependency analysis unit 20. Means. Moreover, the parent-child relationship between sentences does not impair that the parent-child relationship between sentences included in the summary maintains the structure of the corresponding part of the rhetorical structure tree generated by the rhetorical structure analyzing unit 18. means. That is, when the summary is viewed focusing on the sentence, the structure between the sentences appears in a part of the rhetorical structure tree generated by the rhetorical structure analyzing unit 18. When the summary is viewed focusing on the words, the structure between the words appears in a part of the dependency tree generated by the dependency analysis unit 20.

In addition, when the generation unit 22 extracts a partial tree from each dependency tree, the generation unit 22 extracts at most one partial tree from each dependency tree. Extracting at most one subtree means that two or more subtrees are not extracted from one sentence, and there is a dependency tree that does not extract a subtree. For example, when a subtree (node w ₁ , node w ₂ , node w ₃ ) is extracted from the dependency tree representing the sentence s ₈ shown in the right diagram of FIG. 2, from the dependency tree representing the sentence s ₈ , No more subtrees are extracted. This is because multiple subtrees can be extracted from dependency trees representing compound sentences and heavy sentences, but one sentence is treated as having one meaning and multiple subtrees are extracted from one dependency tree. It restricts that. As a result, subtrees can be extracted from more sentences under the limitation of the summary length _Lmax , and the coverage of information as a summary can be improved.

  In addition, the generation unit 22 extracts a partial tree from the dependency tree corresponding to the parent node of the node corresponding to the dependency tree from which the partial tree is extracted. When the parent node of each node included in the rhetorical structure tree is traced, the root node is finally reached. Therefore, the node set corresponding to the dependency tree from which the subtree is extracted is the rooted subtree in the rhetorical structure tree. Become. In this way, by making a sentence set including words included in the summary a rooted subtree in the rhetorical structure tree, the logical structure of the document (for example, a structure such as a turnover) is reflected in the summary. This improves the coverage and accuracy of information as a summary.

  That is, in the summary generated by the generation unit 22, a sentence set constituting the summary is represented by a rooted subtree of the rhetorical structure tree, and a word string included in each sentence constituting the summary is a dependency tree subtree. expressed.

  The above summary generation problem is formulated as an integer programming problem that maximizes the objective function shown in the following equation (13) under the constraints shown in the following equations (1) to (12).

Here, i is a variable representing the position of the sentence in the input document (the number of the sentence from the beginning of the document), and is an identification number (sentence id) of each sentence. N is the total number of sentences included in the document. j is a variable representing the position of the word in the i-th sentence (the number of the word from the beginning of the sentence), and is a word identification number (word id). M (i) is the total number of words included in the i-th sentence. x _i is a decision variable that becomes 1 when the i-th sentence is included in the summary. z _ij is a decision variable that becomes 1 when the j-th word of the i-th sentence is included in the summary. w _ij is the importance level of the j-th word of the i-th sentence, and is a value given by the word importance level assigning unit 16. r _ij is a decision variable that becomes 1 when the node corresponding to the j-th word of the i-th sentence is the root node of the subtree extracted from the dependency tree.

The expression (1) is a constraint expression that guarantees that the number of words included in the summary is equal to or less than _Lmax . The expression (2) is a constraint on the parent-child relationship between sentences, and when extracting the i-th sentence as a summary, it is a constraint expression that guarantees that the parent sentence is also included in the summary. parent (i) is a function that returns the sentence id of the sentence corresponding to the parent node of the node corresponding to the i-th sentence in the rhetorical structure tree.

Equation (3) is a constraint on the parent-child relationship between words. When the j-th word of the i-th sentence is included in the summary, the parent word is guaranteed to be included in the summary. parent (i, j) is a function that returns the word id of the word corresponding to the parent node of the node corresponding to the j-th word in the dependency tree representing the i-th sentence. However, when the term r _ij is described together with the constraint expression shown in equation (9), a subtree whose root node is the node corresponding to the j-th word of the i-th sentence is extracted. Only if the word corresponding to its parent node is not included in the summary.

  Expression (4) is a constraint expression for ensuring that the i-th sentence is included in the summary when the j-th word of the i-th sentence is included in the summary. Equation (5) is a constraint equation for preventing only a sentence from being extracted without extracting a word. Here, a (i) is defined by the following equation.

  len (i) is a function that returns the number of words in the i-th sentence. By introducing a, a short sentence (here, 10 words or less) is not partially extracted but is extracted so that the original sentence is included in the summary. This is in order to prevent this because a word is partially extracted from a short sentence and the readability is significantly reduced.

Expressions (6) to (9) are constraint expressions for enabling extraction of a subtree having an arbitrary node as a root node. Expression (6) is a constraint expression that guarantees that at most one subtree is extracted from one dependency tree representing one sentence. The expression (7) is a constraint expression that guarantees that a subtree having a node other than a predetermined root node candidate as a root node is not extracted. Here, R _c (i) is a function that returns a set of word ids of words corresponding to nodes that are candidates for root nodes in the dependency tree representing the i-th sentence. The candidate for the root node can be, for example, a node corresponding to a word whose part of speech in the sentence is a verb. The expression (8) is a constraint expression that ensures that a word corresponding to a word is included in the summary whenever the node corresponding to the word is extracted as a root node of the subtree. Expression (9) is a constraint expression that guarantees that the parent word of the word corresponding to the root node of the extracted subtree is not included in the summary.

Expression (10) is a constraint expression that guarantees that the word corresponding to the true root node is not included in the summary when the true root node of the dependency tree is not extracted as the root node of the subtree. The true root node is a root node in the entire dependency tree, and is referred to as a “true root node” in order to distinguish it from the root node in the subtree. In the example of FIG. 2 right figure, node w ₁₂ is a true root node. Here, root (i) is a function that returns the word id of the word corresponding to the true root node in the dependency tree representing the i-th sentence.

  The expression (11) is a constraint expression in which one or more of the subjects included in the sentence are included in the summary when the subject includes a subject (SUB). Here, sub (i) is a function that returns a set of word ids of the word that is the subject from the i-th sentence based on the parent-child relationship between words. The expression (12) is a constraint expression in which one or more of the objects included in the sentence are included in the summary when the object (OBJ) is included in the sentence. Here, obj (i) is a function that returns a set of word ids of the word that is the object based on the parent-child relationship between words from the i-th sentence.

Next, the operation of the document summarizing apparatus 10 according to the present embodiment will be described. When a summary generation target document and a summary length L _{max as} a parameter are input to the document summarization apparatus 10, the document summarization apparatus 10 executes a document summarization processing routine shown in FIG.

  In step S10, the sentence dividing unit 12 receives the input document, gives a sentence break to the document, and divides the document into sentences. Next, in step S12, the word dividing unit 14 receives each sentence divided by the sentence dividing unit 12 as an input, gives a word break to each sentence, and divides each sentence into words. Next, in step S14, the word importance assigning unit 16 receives each word divided by the word dividing unit 14 as an input, refers to the word importance DB 30, and assigns importance to each word.

  Next, in step S16, the rhetorical structure analyzing unit 18 receives each sentence divided by the sentence dividing unit 12 as input, and analyzes the rhetorical relationship between sentences using, for example, a known rhetorical structure analyzer. Then, the rhetorical structure analysis unit 18 represents each sentence as a node, and based on the rhetorical relation between sentences, connects the nodes corresponding to the sentences in the parent-child relation to generate a rhetorical structure tree representing the document. .

  Next, in step S18, the dependency analysis unit 20 receives each word divided by the word division unit 14 as an input, and analyzes the dependency relationship between words using, for example, a known dependency analyzer. . Then, the dependency analysis unit 20 represents each word as a node, and based on the dependency relationship between the words, connects the corresponding nodes between the words in the parent-child relationship, and represents a dependency tree representing each of the sentences. Generate each of

Next, in step S20, the generation unit 22 inputs the summary length _Lmax , the importance of each word given by the word importance assigning unit 16, the rhetorical structure tree generated by the rhetorical structure analysis unit 18, and based on each of dependency receiving dependency generated by the analysis unit 20 trees, the number of words included in the summary below L _max, without impairing the parent-child relationship and sentences parent-child relationships between words, and are included in the summary At most one subtree is extracted from each dependency tree so that the sum of the importance of words is maximized. Next, in step S22, the generation unit 22 generates a summary from the subtree set extracted from each dependency tree, outputs the generated summary, and ends the summary generation processing routine.

  Here, an example of a summary generated by the document summarizing apparatus 10 according to the present embodiment will be described for a document including the following sentences 1 to 4.

Sentence 1: A Japanese apple is cropping up in orchards the way Hondas did on US roads.
Sentence 2: It is called the Fuji.
Sentence 3: Some fruit visionaries say the Fuji could someday tumble the Red Delicious.
Sentence 4: But the apple industry is ripe for change.

  As shown in FIG. 4, the document includes a node corresponding to sentence 1 as a root node, a node corresponding to each of sentences 2 and 3 as a child node of a node corresponding to sentence 1, and a node corresponding to sentence 4 as a sentence. It is represented by a rhetorical structure tree that is a child node of the node corresponding to 3. Further, each sentence is represented by a dependency tree having each word included in the sentence as a node. In the example of FIG. 4, a dependency tree representing a sentence is shown in a node corresponding to each sentence.

  From this, subtrees are extracted so that the parent-child relationship between sentences and the parent-child relationship between words are not impaired, and the sum of the importance of words is maximized under the restriction of the word length. In FIG. 4, nodes constituting the extracted subtree and nodes corresponding to the sentence including the extracted subtree are indicated by bold lines. Then, a summary as shown below is generated from the extracted subtree.

      A Japanese apple is cropping up in orchards.The Fuji could someday tumble the Red Delicious.But the apple industry is ripe for change.

As described above, according to the document summarizing apparatus according to the present embodiment, a document is expressed as a rhetorical structure tree representing a rhetorical relationship between sentences, and a sentence is represented as a dependency tree representing a dependency relation between words. By doing so, the document is understood as a nested structure of a tree having sentences as nodes and a tree having words as nodes. Then, from the dependency tree, the number of words included in the summary is equal to or less than L _max and the rhetorical structure of the document represented by the rhetorical structure tree is not impaired so that the sum of the importance of the words included in the summary is maximized. At most one subtree is extracted and a summary is generated from the extracted subtree. As a result, the logical structure of the document and the grammatical nature of the sentence are not impaired, so that readability can be ensured. Also, by extracting at most one subtree from one dependency tree, the coverage of information as a summary can be improved.

  Note that when a document provided with sentence breaks and word breaks is input to the document summarization apparatus, the components of the sentence divider 12 and the word divider 14 may be omitted.

  The document summarization apparatus described above has a computer system inside, but the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used.

  Further, in the present specification, the embodiment has been described in which the program is installed in advance. However, the program can be provided by being stored in a computer-readable recording medium.

DESCRIPTION OF SYMBOLS 10 Document summarization device 12 Sentence division part 14 Word division part 16 Word importance assignment part 18 Rhetorical structure analysis part 20 Dependency analysis part 22 Generation part 30 Word importance DB

Claims (5)

A rhetorical structure analysis unit that generates a rhetorical structure tree representing the document based on a rhetorical relationship between sentences in a document including a plurality of sentences;
A dependency analysis unit that generates each dependency tree representing each of the sentences based on a dependency relationship between words in each of the plurality of sentences;
Based on a predetermined maximum length of each summary, the importance of each of the words, the rhetorical structure tree, and each of the dependency trees, at most from each of the dependency trees to satisfy a predetermined condition. A generation unit that extracts one subtree and generates a summary of the document based on the extracted subtree;
A document summarization device.   The predetermined condition is that the summary length is less than or equal to the maximum length, the rhetorical relationship represented by the rhetorical structure tree and the dependency relationship represented by the dependency tree are not impaired, and the importance of words included in the summary is determined. 2. The document summarizing apparatus according to claim 1, wherein the sum is maximized. 2. The document summarization apparatus according to claim 1, wherein the predetermined condition is to maximize an objective function represented by the following equation (13) under the constraints represented by the following equations (1) to (12).

Where i is the sentence identification number, N is the total number of sentences included in the document, j is the word identification number, M (i) is the total number of words included in the sentence with the identification number i, and x _i is the identification number i. A decision variable that becomes 1 when the sentence is included in the summary, z _ij is a decision variable that becomes 1 when the word with identification number j of the sentence with identification number i is included in the summary, and w _ij is an identification number of the sentence with identification number i The importance of the word j, r _ij is a decision variable that becomes 1 when the word with the identification number j of the sentence with the identification number i is the root of the subtree, L _max is the maximum length, and parent (i) is the rhetorical structure tree A function that returns the identification number of the parent sentence of the sentence with the identification number i, parent (i, j) is a function that returns the identification number of the parent word of the word with the identification number j in the dependency tree representing the sentence with the identification number i. , A is a predetermined coefficient, and R _c (i) is a dependency tree representing a sentence with identification number i. The function root (i) that returns a set of identification numbers of words that are root candidates is a function that returns the identification number of a word that is a true root in a dependency tree representing a sentence with the identification number i, sub ( i) is a function that returns a set of identification numbers of the word that is the subject from the sentence of the identification number i, and obj (i) is a function that returns a set of identification numbers of the word that is the object from the sentence of the identification number i. is there. A document summarization method in a document summarization apparatus including a rhetorical structure analysis unit, a dependency analysis unit, and a generation unit,
The rhetorical structure analysis unit generates a rhetorical structure tree representing the document based on a rhetorical relationship between sentences in a document including a plurality of sentences;
The dependency analysis unit generating each dependency tree representing each of the sentences based on a dependency relationship between words in each of the plurality of sentences;
The dependency tree is configured so that the generation unit satisfies a predetermined condition on the basis of a predetermined maximum length of summary, importance of each word, the rhetorical structure tree, and the dependency tree. Extracting at most one subtree from each of the above and generating a summary of the document based on the extracted subtrees;
Document summarization method including   A document summarization program for causing a computer to function as each part constituting the document summarization apparatus according to any one of claims 1 to 3.

Images