JP3082889B2 - Topic structure recognition method and apparatus for monolog data - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method for recognizing a topic structure in natural language analysis.

[0002]

2. Description of the Related Art Conventionally, models relating to topics and their structures have been proposed. For this, for example, BJGrosz
and CLSidner: “Attention, intention and the str
uctureof discourse ”, Computational Lintuistics
Journal, volume 12, number 3, pp. 175-204 (1986). Since topics have a nested structure, the development of topics is modeled using a stack. The change of the nested structure of topics, that is, the operation of pushing or popping a topic on the stack, is determined by the transition of the intention of the speaker. In addition, what kind of topic develops,
Common sense called domain knowledge is involved.

[0003] Here, the domain knowledge is, for example, a superordinate-subordinate relationship of a concept such as "" Company A "is a kind of" telephone company "" or "" Company A "provides a service" Service A ". , For which it advertises. "

[0004]

However, in the above-described model relating to topics and structures, a method of recognizing intention is not provided, so that the structure of topics cannot be actually recognized. In addition, in terms of topic development, not only is there no way to provide what domain knowledge is needed and how to use it, but even if they are given, it is necessary for topic structure recognition. It is impossible to prepare in advance domain knowledge.

[0005] The present invention has been made in view of the above points,
For monolog data, not domain knowledge,
The purpose is to recognize topics by using topic development style and linguistic knowledge.

[0006]

According to the present invention, there is provided a topic structure recognition method for monolog data, which includes a dictionary storage unit for topic structure recognition preprocessing, a topic structure recognition preprocessing unit, a base expansion processing rule storage unit, and a base expansion processing unit. A topic structure recognizer for monolog data using a topic structure recognizer having a storage means for semantic expansion processing rules, a storage means for semantic expansion processing, a storage means for integrated processing rules and an integration processing means. First, morphological analysis processing is performed on the monolog data input using the topic structure recognition preprocessing dictionary stored in the topic structure recognition preprocessing dictionary storage means by the topic structure recognition preprocessing means. Performs pre-topic structure recognition processing consisting of simple sentence separation processing, salient noun phrase extraction processing, and block recognition processing, and then performs the base expansion processing by the base expansion processing means. From the result of the topic structure recognition preprocessing using the base development rule stored in the rule storage means, a topic about the base development that is explicitly indicated by a clue phrase or the like in the development of the topic in the monolog data is presented. Performing a base development process of sequentially performing identification processing of a topic establishment section to be established, topic word identification processing in a topic establishment section, and topic word nesting level and continuation section identification processing, and then the semantic The expansion processing means uses the semantic expansion processing rules stored in the semantic expansion processing rule storage storage means to determine the base expansion from the result of the topic structure recognition preprocessing and the result of each processing of the base expansion processing. The topic is presented about the semantic development that the topic develops in
Perform semantic expansion processing to sequentially identify the topic establishment section to be established, identify the topic word in the topic establishment section, and identify the nesting level of the topic word and the continuation section.
Then, the integration processing means performs integration processing from the result of the base expansion processing and the result of the semantic expansion processing using the integration processing rule stored in the integration processing rule storage means, thereby performing a monologue processing. A language that recognizes the topic structure of the entire data and extracts salient noun phrases for each single sentence in the monologue data as a result of simple sentence delimitation processing, as a prominent noun phrase extraction process in topic structure recognition preprocessing for monologue data By classifying expressions into explicit markers that only have the function of presenting salient noun phrases and other non-explicit markers, and registering their types and priorities, matching them with salient noun phrase marker priority rules, Extract and prioritize phrase candidates, select the candidate with the highest priority as a prominent noun phrase, and identify the topic structure for the monolog data. As a block recognition process in the preprocessing, a topic continuation phrase, which is a linguistic expression in which the topic continues, is registered, and information of the topic continuation phrase included in the monolog data and a unit of one sentence indicated by a period, etc. By using this information, a block that is a sentence set having a meaningful unity is recognized.

In this case, as a base development process of the topic structure recognition for the monolog data, the clue phrase extracted from the monolog data and the information of the type and the block information, which are the result of the topic structure recognition preprocessing, are obtained. By using, the topic establishment section in which the topic of the base development is presented / established is identified, and in the topic establishment section in the identified base development, from the prominent noun phrases according to the base candidate topic priorities for the base development, By selecting the topic candidate with the highest priority, if there is only one selected candidate, the candidate is taken as a topic, and if there are multiple selected candidates, the candidate that appeared earliest in time is selected, The topic word in the base development is identified, the topic level of the first topic word in the identified base development is set to 1, and the other topics are Determine the topic level according to the topic leveling rules for infrastructure development, set the beginning of the topic establishment section to which each topic belongs as the start point of the continuation section of the topic, and immediately before the start of the topic below the topic level and the end point of the monolog data By determining the temporally earlier one of the two as the end point of the topic establishment section, the topic word level and the continuation section in the base development may be identified.

As a semantic development process of topic structure recognition for monolog data, the title of the whole sentence, the block information as a result of the topic structure pre-recognition process, the salient noun phrase, The topic establishment section in which topics about the semantic development of topic structure recognition for monolog data are presented and established by using the topic topic candidate priorities and the topic establishment section in the base development as a result of the base development processing. In the topic establishment section in the identified semantic expansion, according to the topic candidate priority for semantic expansion, from the salient noun phrases, the topic candidate with the highest priority is selected, and the selected candidate is selected. If there is only one, the candidate is discussed as a topic, and if there are multiple selected candidates, the candidate that appeared earliest in terms of time is selected, The topic words in the expansion are identified, the provisional topic level is set to 1 for all the topic words in the identified semantic expansion, and the head of the topic establishment section to which each topic belongs is set as the start point of the continuation section of the topic. The end point of the topic establishment section is the earlier one of the three immediately before the start of the topic following the topic in the semantic development, immediately before the start point of the topic establishment section in the base development, and the end point of the language data. Thereby, the provisional level and the continuation section of the topic word in the semantic development may be determined.

A topic structure recognition apparatus according to the present invention includes an input unit for inputting monolog data, a unit for extracting a dictionary content of a preprocessing dictionary for topic structure recognition, and performing preprocessing using the dictionary content. The means for storing in the preprocessing storage unit, the development of the topic is divided into a base development explicitly indicated by clue phrases and the like, and a semantic development developed in the base development. Expansion processing storage section and semantic expansion processing storage section for storing the processing results of, and means for extracting each of the base expansion processing rules, and topic establishment section determination processing and topic word determination using the base expansion processing rules Means for performing processing and topic level section determination processing and storing them in the base development processing storage unit, means for extracting each rule of the semantic development processing rules, and topic establishment using the semantic development processing rules Means for performing interval determination processing, topic word determination processing, and topic level section determination processing and storing them in the semantic development processing storage unit, means for extracting each rule of the integration processing rules, and integration processing using the integration processing rule rules And storing the integrated processing storage unit in the integrated processing storage unit, and a display unit for displaying the contents of the integrated processing storage unit.

[0010]

[0011]

[0012]

[0013]

[0014]

[0015]

[0016]

The present invention recognizes semantically coherent blocks by topic continuation phrases and the like, and divides topic development into basic development explicitly indicated by clue phrases and the like, and semantic development developed in them. The topic is presented and established using block information, etc., and the topic establishment section in which the topic is established is determined. The topic word in each topic establishment section is selected from the candidates indicated by the topic marker, and the topics in the basic development and the semantic development are integrated. Recognize topic structure for monolog data. In this way, topics are recognized using only topic development styles and linguistic knowledge without requiring domain knowledge for monolog data.

[0017]

Next, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 is a diagram showing an outline of a topic structure recognition process according to one embodiment of the present invention, and FIG. 2 is a block diagram of a topic structure recognition device according to one embodiment of the present invention. The outline of the processing and the topic structure recognition device will be described.

[0019]

The topic structure recognition processing is performed according to the following procedure shown in FIG.

The input monolog data 110 is subjected to topic structure recognition pre-processing 120 to form blocks. Thereafter, the block is subjected to a base development process 130 and a semantic development process 140 to determine a topic level section independently. Subsequently, based on the topic level determined in the base development processing 130 and the semantic development processing 140, respectively, the integration processing 150 of the base development and the semantic development is performed.
To determine the topic structure 160.

The topic structure pre-recognition process 120 is configured to sequentially perform a morphological analysis 121, a single sentence separation process 122, a salient noun phrase extraction 123, and a block recognition 124. The base development processing 130 sequentially determines a topic establishment section determination 131, a topic word determination, and a topic level section determination , and the semantic development processing 140 determines a topic establishment section determination 141, a topic word determination 142, and a topic level section determination. 143 are performed in order.

In the block diagram shown in FIG. 2, the morphological analysis 121, the single sentence separation process 122, the salient noun phrase extraction 123, and the block recognition 124 performed in the topic structure recognition preprocessing 120 are performed by the topic structure recognition preprocessing unit 203 by the data input unit. 2
Pre-processing storage unit 2 for dialog data input from 01
02 according to the processing procedure stored in the dictionary management unit 21.
6 and the preprocessing dictionary 204. First, the input mode
A morphological analysis process is performed on the norolog data 110.
U. Morphological analysis processing is the input monologue data 1
The character string of 10 is divided into words to form a word string.
Identify the parts of speech of words and the inflected forms of inflected words.

After the morphological analysis 121 is performed, a single sentence segmentation process 122 is subsequently performed on the result of the morphological analysis. The simple sentence separator process 122 sentence including a plurality of predicates as embedded text and compound sentence, one that divides the simple sentence containing only one predicate, the block diagram of FIG. 2, the topic structure recognition preprocessing unit 203 pretreatment This is performed using the single sentence rule management and the word segmentation rules stored in the storage unit 202.

Next, in the salient noun phrase extraction 123, the noun phrase that is most emphasized in each simple sentence with respect to the input simple sentence delimitation processing is extracted.

Next, a block that is a unit that is semantically united is recognized. A block corresponds to a paragraph in the monolog data. Note that the processing here is performed by the topic structure recognition preprocessing unit 203 using the topic management rules and the topic structure recognition rules stored in the preprocessing storage unit 202 in the block diagram of FIG.

Next, the base expansion processing 130 and the semantic expansion processing 140 are performed on the recognized block, respectively, to determine topic establishment sections 131 and 141, topic words 132 and 142, and topic level section 133. , 143
Are sequentially performed. Here, the topic establishment section is a section in which a topic is presented and established.
By these three processes, the topic structure in each of the base expansion process 130 and the semantic expansion process 140 can be obtained.

As for the base development processing 130, only the result of the processing immediately before the base development is required as an input of each processing. On the other hand, regarding the semantic expansion processing 140, in order to determine a topic word, the semantic expansion processing 1
The result of the process immediately before 40 and the result of the same type of process in the base development process 130 are required. That is, as the input for performing the determination 141 of the topic establishment section in the semantic development processing 140, both the result of the block recognition 124 and the result of the determination 131 of the topic establishment section in the base development processing 130 are required.

Similarly, the input of the topic word determination 142 in the semantic expansion processing 140 is as follows.
0 and the result of the topic word determination 132 in the base development processing 130 are required. The input of the topic level section determination 143 in the semantic expansion processing 140 includes the semantic expansion processing 1
The result of the topic word determination 142 in 40 and the result of the topic level section determination 133 in the base development process 130 are required.

Finally, the respective topic structures obtained by the base development processing 130 and the semantic development processing 140 are input, and the integration processing 150 of the base development processing 130 and the semantic development processing 140 is performed. The topic structure 160 of the entire data is output.

In the block diagram shown in FIG. 2, the above-described infrastructure development processing 130 is a topic establishment section determination processing section 231, a topic word determination processing section 232, and a topic level section determination processing section 23.
3 is performed by referring to the base development processing rule management unit 205 and the base development processing rule 206 according to the storage contents of the base development processing storage unit 207 that stores the procedure of the base development processing. In the semantic development processing 140, the semantic development processing unit 240 including the topic establishment section determination processing unit 241, the topic word determination processing unit 242, and the topic level section determination processing unit 243 stores the procedure of the semantic development processing. In accordance with the contents stored in the semantic development processing storage unit 210, the processing is performed by referring to the semantic development processing rule management unit 208 and the semantic development processing rule 209.

Next, the specific contents of each process for performing topic structure recognition in the present invention will be described.

Topological Structure Recognition Preprocessing 120 Morphological Analysis 121 The morphological analysis 121 receives a Japanese character string, and outputs a result obtained by dividing the character string for each word and the part of speech of each word. For example, when a morphological analysis is performed using a Japanese character string “writing a patent” as an input, a Japanese character string divided into three words such as “patent”, “wo”, and “writing” is output, Part-of-speech information of each word such as "patent = noun", "wo = case particle", and "writing = termination of verb" is output. However, since the verb is a conjugation word, conjugation information such as "end form" is also added.

In order to perform the morphological analysis 121, a word dictionary that describes the parts of speech of each word and a concatenation dictionary that describes the ease of arrangement of parts of speech in a Japanese character string are required.
For example, in a conjunctive dictionary, a case particle is likely to follow a noun like "patent" or "wo", but a case particle is hard to follow after a verb like "write" or "wo". Is written.

When a Japanese character string is divided into words, for example, whether the character string "patent" is composed of one noun "patent" or whether the character string "patent" and two characters "permit" are also composed of words However, in the morphological analysis, the most appropriate analysis result is selected by using a word dictionary and a concatenation dictionary. The detailed method for morphological analysis is
"Yoshimura, Hidaka, Yoshida:" Morphological analysis of solid Japanese sentences using the minimum number of clauses method "Transactions of Information Processing Society of Japan, Vol. 2
4, No. 1, pp. 40-46 (1983)].

Single sentence delimiting process 122 The single sentence delimiting process 122 includes only one predicate by using a sentence containing a plurality of predicates such as an embedded sentence or a multiple sentence by using a prepared single sentence delimiting rule as shown in FIG. Break it into simple sentences. For example, this is a simple sentence because the only predicate in the sentence "I write a patent" is the verb "write". On the other hand, "If you invent,
Since the sentence "write a patent" includes two predicates, a verb "invent" and a verb "write", the sentence is divided into two simple sentences "if you invent" and "write a patent".

The single sentence separation rule shown in FIG. 3 is as follows. (1) Cut at a period (2) Except in the following cases, in principle, cut immediately after a relation (2-1) When the relation is an adjective or an adjunct form of an adjective (2-2) The relation is an adjective or an adjective verb A field of continuous use
If (3) it is not separated by a comma. However, if the relation is contained in the simple sentence before the reading point, it is separated after the reading point. (4) If the final particle is followed by the case particle "to", the morpheme is separated before the case particle "to". The Japanese sentence including a plurality of predicates is divided into simple sentences according to the type of word, the type of part of speech, and the inflected form obtained by the analysis. It checks whether each rule can be applied to a given Japanese sentence, and if possible, applies single rules to perform single sentence delimitation processing.

Extraction of salient noun phrases 123 The salient noun phrases that are most emphasized in each simple sentence are extracted. In Japanese, salient noun phrases are indicated by markers such as particles. Markers include “about”, “about”, “ha”
An explicit marker that only has the function of presenting a phrase, such as, and a marker that indicates a grammatical role, such as subject or object, such as "ga" or "wo", are used to present a phrase. An "explicit marker" exists. These are given by humans as rules in advance along with the priorities.

FIG. 4 shows an example of marker priority of salient noun phrases.

The highest priority is given to (1) explicit markers other than "ha", and then (2) explicit markers "ha",
(3) Followed by an unspecified marker.

By matching between the monolog data and these markers, candidates for salient noun phrases are extracted. However, in the case where the phrase indicated by the marker is a Dyxis expression that does not have a specific meaning by itself, such as a pronoun or “koto” or “mono,” it is not considered as a prominent noun phrase candidate.

When a plurality of candidates are extracted from one simple sentence, the one having the highest priority is selected as a prominent noun phrase according to the priority shown in FIG. 4, and there are a plurality of those having the highest priority. In this case, the one that appears earliest in time is selected as the salient noun phrase.

Block Recognition 124 A block which is a semantic unit in the monolog data is recognized. A block corresponds to a paragraph in the monologueta data.

The blanking lock recognition rule is as shown below, an example of which is shown in FIG. (A) Since topics continue in one sentence, the same block is set in one sentence. (B) If the topic continuation phrase indicates that the topic continues, the current sentence is included in the same block as the immediately preceding sentence. (C) to be recognized to belong to block the current sentence already by the rules described above, it is assumed that the new block starts from the current sentence.

The unit of one sentence indicated by a period etc. also indicates the continuation of a topic, so that one sentence is the same block. If a topic continuation phrase such as “depends on” or “to this” explicitly indicates that the topic continues from the previous sentence, the current sentence is It shall be included in the block to which it belongs. In other cases, it is assumed that a new block has started from the current sentence.

Base development processing 130 Determination of topic establishment section 131 In the base development in which topics are explicitly developed by clue phrases or the like, topic establishment sections in which topics are presented and established are identified. FIG. 6 shows a flow of a topic establishment section determination process in the base development.

First, the start of the topic establishment section is set at the start of the monolog data and at the beginning of the block containing the clue phrase (step S601). Here, the clue phrases include, as shown in FIG. 6, a nesting start type such as “first” and “first”, a topic conversion type such as “next” and “second”, and a “last” ”And“ at the end ”.

When the starting points of the topic establishment sections are collected, the topic presentation type in each topic establishment section is identified (step S6).
02). Here, the topic presentation type identification processing will be described later. Next, it is checked whether the topic presentation type is a gradual type (step S603). If the topic presentation type is a gradual type that presents a topic little by little, the topic establishment section end point has been recognized in the topic presentation type identification process, so that it is set as the topic establishment section end point (step S604). Of the original block, a part other than the topic establishment section is recognized as a “pseudo block” (step S605). If the topic presentation type is not the gradual type, that is, if the topic presentation type is a batch type, the range of the first sentence from the start point of the topic establishment section is recognized as the topic establishment section (step S).
606).

Next, each of the above steps S605 and S60
The topic establishment section obtained in 6 is recognized as a summary section (step S607). These series of processes are performed on the entire text, and then the process is terminated (step S60).
8).

Next, the topic presentation type identification processing performed in step S602 will be described.

As shown in FIG. 7, this process identifies whether the topic presentation type is a gradual type that presents topics little by little or a batch type that presents topics at once.

First, for one sentence from the start point of the topic establishment section in the base development, the top two salient noun phrases are extracted in accordance with the base development topic candidate priority shown in FIG. 8 (step S701). Of these two, the prominent noun phrase appearing earlier in time is p, and the prominent noun phrase appearing later in time is q
(Step S702).

Here, as criteria for extracting the top two salient noun phrases, (a) salient noun phrases included in the title, (b) salient noun phrases including proper nouns, and (c) explicit markers Are treated as the same priority, and prominent noun phrases presented with the implicit markers are treated as lower ranks of these.

Next, the topic candidate for base development is p,
It is checked whether both q are 1 (step S703).
When the priority of the topic candidate for base development is 1 for both p and q, that is, when it is included in the title, contains a proper noun, or is indicated by an explicit marker, the topic is presented. The type is assumed to be a gradual type (step S704), and the end of the single sentence immediately before the single sentence including q is set as the end point of the topic establishment section (step S70).
5). If the base development topic candidate priority is not 1 for both q and p, that is, if at least one of them is indicated by an implicit marker, or if there is only one that satisfies the base development topic candidate priority Assume that the topic presentation type is a batch type (step S706).

Determining Topic Words 132 Recognizing what topics are presented in a topic establishment section in a basic development where a topic is explicitly developed using clue phrases or the like. FIG. 9 shows an example of the flow of the topic word determination processing in the base development.

Each topic establishment section in the base development will be described first, and the section having the highest priority is extracted based on “topic candidates for base development” shown in FIG. If there is only one extracted candidate, it is identified as a topic (step S901). Thereafter, it is checked whether the number of extracted candidates is one (step S90).
2) If a plurality of candidates are extracted, the one that appears earliest in time is selected (step S903). This is recognized as a topic (step S904).

Determining Topic Level Section 133 For the topic in the base development, the topic level and the section in which the topic continues are determined. Here, it is assumed that the topic level of the outermost topic is 1, and each time the nesting increases by one, the topic level also increases by one.

FIG. 10 shows an example of a rule for assigning levels for board development. First, the topic level of the first topic of the monolog data is set to 1. Second, regarding the change in topic level due to a clue phrase, the topic level is increased or decreased according to a transition pattern from the immediately preceding clue phrase to the current clue phrase.

If the current clue phrase is the nesting start type, the topic level is increased by 1 regardless of the type of the clue phrase immediately before. Also, if the current clue phrase is the topic conversion type or the nesting end type, and the immediately preceding clue phrase is the nesting start type or the topic conversion type, the topic level does not change. If the current clue phrase is the topic conversion type or the nested type and the immediately preceding clue phrase is the nested type, the topic level is decreased by one.

FIG. 11 shows an example of a rule of topic continuation section determination processing in base development. Assuming that the topic currently being processed is A and its topic level is m, the beginning of the topic establishment section to which A belongs is the start point of the topic continuation section, and immediately before the start of the topic whose topic level is m or less, and the end of the monolog data. The one that appears earlier in time is the end point of the topic continuation section.

Semantic development processing 140 Determination of topic establishment section 141 In semantic development in which topics are developed in base development,
A topic establishment section in which a topic is presented and established is identified. FIG.
FIG. 2 shows an example of the flow of the topic establishment section determination process in the semantic development.

First, a topic establishment section in the base development is not included.
Including the topic establishment section in each block or foundation expansion, have
For each pseudo block in each block, it is checked whether the number of single sentences included in the block or the pseudo block is four or more (step S1201), and it is checked whether the first sentence in the block includes two or more single sentences. (Step S1202).
If the number of single sentences is 4 or more and the first sentence in the block or pseudo block includes two or more single sentences, a topic establishment section candidate is extracted from the block or pseudo block (step S1203).

Here, this topic establishment section candidate starts from the head of the block or the pseudo block, and ends at the end of the first sentence in the block or the pseudo block and at the end of the fourth single sentence from the end of the monologue. The earlier of the two is the end point.

If the number of simple sentences included in the block or the pseudo block is less than 4 , or if the first sentence in the block or the pseudo block includes only one simple sentence,
It is assumed that there is no topic in this block or pseudo block (step S1206). However, it is assumed that a human gives an appropriate value in advance to the value of "4 simple sentences" or "1 simple sentence" according to the property of the monolog data.

If there is a topic establishment section candidate, it is checked whether there is a prominent noun phrase that includes the topic candidate priority for semantic development (step S1204).

FIG. 13 shows an example of the priority of the topic candidates for semantic development. In this example, the priority is divided into two levels. The candidates with the higher priority are (1) salient noun phrases accompanied by question expressions such as “ask” and “ask” in the same sentence. (2) The candidate with the lower priority is indicated by (a) a prominent noun phrase included in the title or the immediately preceding summary section, (b) a prominent noun phrase including a proper noun, and (c) an explicit marker. Prominent noun phrases, which have the same priority.

If the topic establishment section candidate includes one or more semantic development topic candidate priorities, the topic establishment section candidate is identified as a topic establishment section (step S1205). If no topic candidate priority for semantic development is included, the topic establishment section candidate is rejected, and it is assumed that no topic establishment section exists in the block (step S1206).

The processing according to each of the above operations is performed on the entire text, and the processing is terminated (step S1207).

Determination of Topic Word 142 In the topic establishment section in the semantic development in which the topic is developed in the basic development, what kind of topic is presented is identified. FIG. 14 shows an example of the flow of topic word determination processing in semantic development.

First, for each topic establishment section in the semantic development, the section having the highest priority is extracted based on “topic candidates for semantic development priority” shown in FIG. 13 (step S1401). Next, the extracted candidate is 1
(Step S1402), and if there is only one extracted candidate, it is identified as a topic (Step S1404). If a plurality of candidates have been extracted, the candidate that appears earliest in time is selected (step S1403) and identified as a topic (step S140).
1404).

Determination of topic level section 143 For the topic in the semantic development, the topic level and the section where the topic continues are determined. FIG. 15 shows an example of the flow of topic level section determination processing in semantic development. Here, the topic currently being processed is A. First, the topic level of topic A is set to 1 (step S1501). Next, the head of the topic establishment section to which A belongs is set as the start point of the continuation section of A. In addition, immediately before the topic that appears next to topic A in the semantic development, immediately before the start point of the topic establishment section in the base development, and at the end of the monolog data, The end point of the continuation section of the topic A is set (step S1502).

Integration processing of base development and semantic development 150 The topic structure in the base development and the topic structure in the semantic development determined so far are integrated. FIG. 16 shows an example of the flow of the integration processing of the base development and the semantic development.

First, it is confirmed whether the topic is a topic of the base development (step S1601). As a result of the confirmation, for topics in the base development, the topic structure is integrated as it is (steps S1603 and S1604), and for topics in the semantic development, the topic level is corrected (step S1602). Perform

The topic level correction of the semantic development performed in step S1602 is performed by adding the maximum value of the topic level of the basic development at that time to the original topic level. The topic structure obtained after the integration is the final topic structure.

Description Using Monolog Data Example Next, a topic structure recognition method when the present invention is applied will be described in detail using a specific monolog data example.

[0076] Examples 17 topic structure recognition preprocessing 120 sentence separator process 122 is a diagram showing a monologue data example of an embodiment of the present invention. FIG. 1 shows an example in which the monolog data example is divided into simple sentences using the simple sentence separation rule example shown in FIG.
FIG.

Specific Example of Extraction of Salient Noun Phrases 123 A salient noun phrase is extracted from each of the divided simple sentences. FIG.
FIG. 19 shows the result of extracting salient noun phrases from the eight single sentence segmentation results using salient noun phrase markers as shown in FIG. Here, in FIG. 19, salient noun phrases are indicated by underlining, and a single sentence number for explanation is (1-
1), (1-2),...

In the simple sentence (1-1), the prominent noun phrase “communication service of company A” is extracted by the explicit marker “ha”. In the simple sentence (1-2), the implicit marker "wo" exists, but the marked word is a dichsis expression "them" having no specific meaning. Is not extracted. In addition, simple sentence (3-1)
Include “service A” indicated by the explicit marker “ha” and “competitor” indicated by the non-explicit marker “ni”. According to the priority shown in FIG. Since the priority is higher, “service A” is selected as a salient noun phrase. Similarly in other simple sentences,
Salient noun phrases are extracted by salient noun phrase markers.

Specific Example of Block Recognition 124 A block in monolog data is recognized. FIG.
FIG. 20 shows an example of block recognition with respect to the monolog data example shown in FIG. According to the block recognition rule example (a) shown in FIG. 5, one sentence belongs to the same block, so the sentences of (1) to (7) in FIG. 20 belong to the same block. Further, an example of the block recognition rule (b) in FIG.
As shown in (1), sentences (5) and (6) belong to the same block by the topic continuation phrase "this" of sentence (6).

Base Development Process 130 Specific Example of Topic Establishment Section Determination 131 in the Base Development Process 130 A topic establishment section in the base development is identified. FIG. 21 shows the identification result of the topic establishment section of the base development in the example of monolog data.

First, according to the topic establishment section determination process shown in FIG. 6, the start of the block including the clue phrase at the start of the monologue and the clue phrase is set. In FIG. 21, the head of each block of a, b, and c is the start point of the topic establishment section.

Second, topic presentation type identification processing is performed for each topic establishment section. As shown in the example of topic presentation type identification processing in FIG. 7, when the top two candidates for the base development topic candidates included in the first sentence from the start point of each topic establishment section are extracted, the topic establishment starting from block a is performed. From section, "Company A
Since only one communication service is detected, the topic establishment section is a collective type. Also, “various new services” and “service A, service B, service C, etc.” are detected from the topic establishment section starting from block b, but “various new services” have priority (2). Therefore, this topic establishment section is also a collective type. Further, since only the “conventional service” is detected from the topic establishment section starting from the block e, the topic establishment section is also a collective type.

Third, since the topic presentation type in all the topic establishment sections in the monologue example is the collective type, FIG.
The range of the first sentence from the start point of the topic establishment section is set as the topic establishment section according to the topic establishment section determination processing example. The results are shown in FIG. The topic establishment section recognized by the above processing is also recognized as a summary section.

Topic word determination 1 in base development processing 130
32 Specific Examples A topic word in a topic establishment section of the base development is determined. As shown in FIG. 21, three topic establishment sections were recognized in the base development. The “communication service of company A” in block a satisfies the topic candidate priority for platform development shown in FIG. 8, and is extracted as a topic word candidate according to the topic word determination process in the platform development of FIG. 11. Since this is the only candidate extracted from the topic establishment section included in block a, “communication service of company A” is determined as the topic word in the topic establishment section included in block a. Similarly, “various new services” are extracted as topic words in the topic establishment section included in the block b, and “conventional services” are extracted as topic words in the topic establishment section included in the block e.

Topic level section determination 13 in base development
Specific Example 3 The topic level section of the topic word in the base development is determined.
According to the rule of FIG. 10, the topic level of the first topic of the monologue “communication service of company A” is set to 1. Since the nesting-type clue phrase “first” is presented in the block in which the next topic “various new services” is presented, the topic level is increased by 1 according to the rule of FIG. .

The block in which the next topic "conventional service" is presented has a topic conversion start type clue phrase "next", and the clue phrase presented before it is "first". Nested start type,
According to the rule shown in FIG. 10, the topic level is set to 2, which is the same as before.

The continuation section of each topic is determined by using the rule shown in FIG. The continuation section of “Communication service of company A” is from the start of monologue data to the end of monologue data, and the continuation section of “various new services” is blocks b, c, and d in FIG.
Are the blocks e and f in FIG. FIG. 22 shows the recognition results of these topic level sections.

[0088] topic established interval determined in semantic expansion processing 140 semantic expansion processing 140 141
Specific example of topic Identify the topic establishment section in the semantic expansion. As shown in the example of the topic establishment section determination process in FIG. 12, here, only the blocks that do not include the topic establishment section in the base development or the pseudo blocks in the blocks that include the topic establishment section in the base development are targeted. Therefore, as shown in FIG.
Blocks c, d, and f are processing targets.

As for the block d, “the number of single sentences included in the block is 4
Or more. " Therefore, it is assumed that the topic establishment section is not included in the block d.

For blocks c and f, both the conditions of “the number of single sentences included in the block is 4 or more” and “the first sentence in the block includes two or more single sentences” are satisfied in the topic establishment section determination processing of FIG. . Therefore, in both blocks, the section from the end of the first sentence in the block and the end of the fourth single sentence from the end of the monologue to the one that appears first temporally is set as a topic establishment section candidate.

Each of these topic establishment section candidates has at least one candidate included in the “semantic development topic candidate priority” shown in FIG. Therefore, these topic establishment section candidates are recognized as topic establishment sections.

Specific Example of Topic Word Determination 142 in Semantic Expansion Processing 140 A topic word is determined in a topic establishment section of the semantic expansion. In the semantic development, two topic establishment sections were identified as shown in FIG. Since "service A" in block c satisfies the topic candidate priority for semantic development shown in FIG. 13, it is extracted as a topic word candidate according to the topic word determination process in the semantic development of FIG. In the topic candidate priority for semantic development shown in FIG.
It is assumed that a higher priority than that, that is, “prominent noun phrase with question expression” is not included in the topic establishment section of block c. Then, even if a candidate other than “Service A” is detected in this topic establishment section, the one that appears first earliest in time will be selected. Therefore, the first appearing “Service A” is recognized as a topic. . Similarly, in block f, “payment of directory guidance” is recognized as a topic.

Specific Example of Topic Level Section Determination 143 in Semantic Expansion Processing 140 The topic level section of the topic word in the semantic expansion is determined. According to the rule of FIG. 15, the topic level of all topic words is temporarily set to 1. In addition, the continuation section of each topic is shown in FIG.
Determined according to the rules. The continuation section of "service A" is blocks c and d in FIG. 22, and the continuation section of "payment of directory assistance" is block f in FIG. FIG. 24 shows the recognition results of these topic level sections.

Specific Example of Integration Process 150 of Base Expansion and Semantic Expansion The topic structures in the base expansion and the semantic expansion are integrated. The topic level of the topic in the semantic development is corrected according to the integration rule of the base development and the semantic development shown in FIG. FIG. 25 shows the integration result. This is the monologue
This is a topic structure obtained by applying the topic structure recognition method of the present invention to a data example.

Experimental Data The results of an evaluation experiment in which the topic structure recognition method of the present invention is applied to actual monolog data are shown. As the evaluation, a method of calculating the recall and the precision by comparing the topic structure recognized by humans with the topic structure recognized by the computer was adopted. Here, the recall is a measure of how much of the topic structure recognized by humans is also recognized by the computer, and the precision is how much of the topic structure recognized by the computer is human. Is also a measure of whether it is recognized. If the topic structures recognized by the human and the computer match, both the recall and the precision are 100%.

The monologue data used in the experiment was a total of 10 TV news manuscripts, which could be 3 words per sentence.
00. As a result of evaluation, the recall was 74.0%.
And the precision was 64.0%.

[0097]

As described above, according to the present invention, a topic and its structure can be recognized without using knowledge depending on a specific domain.

By presenting such a recognized topic and topic structure to the user, it is possible to assist the user in roughly grasping the contents of the monolog data. The topic structure can also be used as a table of contents.

[Brief description of the drawings]

FIG. 1 is a flowchart showing a process for recognizing a monologue data topic structure according to an embodiment of the present invention.

FIG. 2 is a block diagram of a monologue data topic structure recognition device according to an embodiment of the present invention.

FIG. 3 is a diagram showing an example of a single sentence separation rule used in one embodiment of the present invention.

FIG. 4 is a diagram showing an example of a marker priority order rule for salient noun phrases used in one embodiment of the present invention.

FIG. 5 is a diagram showing an example of a block recognition rule used in one embodiment of the present invention.

FIG. 6 is a flowchart showing a process of determining a topic establishment section in a base development according to an embodiment of the present invention.

FIG. 7 is a flowchart illustrating a topic presentation type identification process according to an embodiment of the present invention.

FIG. 8 is a diagram showing the priority order of topic candidates for base development according to one embodiment of the present invention.

FIG. 9 is a flowchart illustrating a topic word determination process in a base development according to an embodiment of the present invention.

FIG. 10 is a diagram showing an example of a leveling rule for board development used in one embodiment of the present invention.

FIG. 11 is a flowchart illustrating a topic establishment section determination process in a semantic development according to an embodiment of the present invention.

FIG. 12 is a diagram showing priority of topic candidates for semantic development according to one embodiment of the present invention.

FIG. 13 is a flowchart showing a topic word determination process in the semantic expansion of one embodiment of the present invention.

FIG. 14 is a flowchart showing a topic word continuation section determination process in base development according to an embodiment of the present invention.

FIG. 15 is a flowchart showing a provisional topic level section determination process in the semantic development of one embodiment of the present invention.

FIG. 16 is a flowchart showing an integration process of base development and semantic development according to one embodiment of the present invention.

FIG. 17 is a diagram showing an example of monolog data according to an embodiment of the present invention.

FIG. 18 is a diagram illustrating an example of a single sentence segmentation result with respect to an example of monolog data according to an embodiment of the present invention.

FIG. 19 is a diagram showing an example of salient noun phrases in an example of monolog data according to an embodiment of the present invention.

FIG. 20 is a diagram showing blocks in an example of monolog data according to an embodiment of the present invention.

FIG. 21 is a diagram showing a topic establishment section in a base development in a monolog data example according to an embodiment of the present invention.

FIG. 22 is a diagram illustrating a topic structure in a base development in a monolog data example according to an embodiment of the present invention.

FIG. 23 is a diagram illustrating a topic establishment section in a semantic expansion in a monolog data example according to an embodiment of the present invention.

FIG. 24 is a diagram showing a topic structure in a semantic expansion in an example of monolog data according to an embodiment of the present invention.

FIG. 25 is a diagram showing a topic structure in a monolog data example according to an embodiment of the present invention.

[Explanation of symbols]

 110 Monologue Data 120 Topic Structure Recognition Preprocessing 121 Morphological Analysis 122 Simple Sentence Separation Processing 123 Salient Noun Phrase Extraction 124 Block Recognition 130 Base Expansion Processing 131,141 Determination of Topic Establishment Section 132,142 Determination of Topic Word 133,143 Topic Level Section Determination 141 Semantic expansion processing 150 Integration processing of base expansion and semantic expansion 160 Topic structure 201 Data input unit 202 Preprocessing storage unit 203 Topic structure recognition preprocessing unit 204 Preprocessing dictionary 205 Base expansion processing rule management unit 206 Base Expansion processing rule 207 Base expansion processing storage unit 208 Semantic expansion processing rule management unit 209 Semantic expansion processing rule 210 Semantic expansion processing storage unit 211 Integration processing storage unit 212 Integration processing unit 213 Integration processing rule management unit 214 Integration processing rule 215 Display 2 0 base expansion processing unit 231, 241 topic section establishment determination processing unit 232 and 242 the topic word determination processing unit 233, 243 the topic level section determination processing unit 240 semantic development processing unit

────────────────────────────────────────────────── ─── Continuation of front page (56) References JP-A-4-306768 (JP, A) JP-A-4-332084 (JP, A) JP-A-5-266607 (JP, A) Atsushi Takeshita, “4H -3 Identification of Topics Using Dialogue Structure ”, Proc. Of the 43rd IPSJ (Second 1991) National Convention (3), September 24, 1991, p. 3-229 to 3-230 (58) Fields investigated (Int. Cl. ⁷ , DB name) G06F 17/27 G06F 17/30

Claims (4)

(57) [Claims] 1. A topic structure recognition preprocessing dictionary storage unit, a topic structure recognition preprocessing unit, a base expansion processing rule storage unit, a base expansion processing rule storage unit, a semantic expansion processing rule storage unit, a semantic expansion processing unit, and an integrated process. A method for recognizing a topic structure for monolog data using a topic structure recognizing device having a rule storage means and an integrated processing means, comprising: A topic consisting of morphological analysis processing, simple sentence separation processing, salient noun phrase extraction processing, and block recognition processing for monolog data input using the topic structure recognition preprocessing dictionary stored in the processing dictionary storage means. Performing a pre-recognition process, and then using the pre-expansion rule stored in the pre-expansion process rule storage means in the pre-expansion process means to perform the topic structure pre-recognition process. From the results of the above, the process of identifying a topic establishment section in which topics are presented and established and the process of identifying topic words in a topic establishment section are described with respect to a basic development in which the development of topics in the monolog data is explicitly indicated by clue phrases or the like. And the base word expansion processing for sequentially performing the nesting level of the topic word and the identification processing of the continuation section. Then, the semantic expansion processing means stores the meaning stored in the semantic expansion processing rule storage means. From the results of the topic structure pre-recognition processing and the results of each processing of the base development processing using the dynamic development processing rules, the topics for which the topics are presented / established about the semantic development in which the topics are developed in the base development A semantic expansion process is performed in which the identification process of the established section, the identification process of the topic word in the topic establishment interval, and the identification process of the nesting level of the topic word and the continuation section are sequentially performed. After that, the integration processing means performs integration processing from the result of the base expansion processing and the result of the semantic expansion processing using the integration processing rules stored in the integration processing rule storage means, thereby performing a monologue processing. A language that recognizes the topic structure of the entire data and extracts salient noun phrases for each single sentence in the monologue data as a result of simple sentence demarcation processing, as a prominent noun phrase extraction process in topic structure recognition preprocessing for monologue data By classifying expressions into explicit markers that only have the function of presenting salient noun phrases and other non-explicit markers, and registering their types and priorities, matching them with salient noun phrase marker priority rules, Extract and prioritize phrase candidates, select the candidate with the highest priority as a prominent noun phrase, and construct a topic structure for monolog data As a block recognition process in the knowledge processing, a topic continuation phrase which is a linguistic expression in which a topic continues is registered, and information of the topic continuation phrase included in the monolog data and one sentence indicated by a punctuation mark or the like are registered. A topic structure recognition method for monolog data characterized by recognizing a block, which is a sentence set having a meaningful unity, by using unit information. 2. The topic structure recognition method for monolog data according to claim 1, wherein the topic structure recognition processing for the monolog data is extracted from monolog data that is a result of the topic structure recognition preprocessing. By using the clue phrase and the information of the type and the block information, a topic establishment section in which a topic of the base development is presented / established is identified, and in the topic establishment section in the identified base development, According to the development topic candidate priority, the topic candidate with the highest priority is selected from the prominent noun phrases. If there is only one selected candidate, the candidate is set as a topic. By selecting a candidate that appears earliest in time, a topic word in the base development is identified, and a topic word in the identified base development is identified. Sets the topic level of the first topic word to 1 and determines the topic level for other topics according to the topic leveling rules for infrastructure development, and sets the beginning of the topic establishment section to which each topic belongs to the start of the continuation section of that topic By setting the end point of the topic establishment section as the end point of the topic establishment section between the time immediately before the start of the topic below the topic level and the end point of the monologue data and the end point of the monolog data, the level and the continuation section of the topic word in the base development are set. A topic structure recognition method for monolog data characterized by performing identification processing. 3. The topic structure recognition method for monolog data according to claim 1, wherein the semantic expansion processing of the topic structure recognition for the monolog data is a title of the whole sentence and a result of the topic structure recognition preprocessing. By using the block information, the salient noun phrase, the topic candidate priority for semantic development, and the topic establishment section in the base development that is the result of the base development processing,
Identify a topic establishment section in which a topic about the semantic development of topic structure recognition for monolog data is presented and established, and in the topic establishment section in the identified semantic development,
According to the topic candidate priority for semantic development, from the salient noun phrases, select the topic candidate with the highest priority,
If there is only one selected candidate, the candidate is discussed as a topic. If there are multiple selected candidates, the topic word in the semantic expansion is identified by selecting the candidate that appears earliest in time, The provisional topic level is set to 1 for all the topic words in the identified semantic expansion, the head of the topic establishment section to which each topic belongs is set as the start point of the continuation section of the topic, and the topic is identified by the semantic expansion. By setting the end point of the topic establishment section as the end point of the topic establishment section immediately before the start of the next topic, immediately before the start point of the topic establishment section in the base development, and the end point of the language data, the topic in the semantic development is obtained. A topic structure recognition method for monolog data, comprising determining a provisional word level and a continuation section. 4. An input unit for inputting monolog data, means for extracting dictionary contents of a pre-processing dictionary for topic structure recognition, pre-processing using the dictionary contents, and storing them in a pre-processing storage unit. Separation of means and topic development into base development explicitly indicated by clue phrases, etc., and semantic development deployed in the base development, and storing the processing results for each of the base development and semantic development Base expansion processing storage unit and semantic expansion processing storage unit, means for extracting each rule of the base expansion processing rules, topic establishment section determination processing, topic word determination processing, and topic level section determination processing using the base expansion processing rules Means for storing the information in the base development processing storage unit, means for extracting each rule of the semantic development processing rules, and topic establishment section determination processing and topic word determination using the semantic development processing rules. Means for performing processing and topic level section determination processing and storing them in a semantic expansion processing storage unit; means for extracting each rule of the integrated processing rules; and performing integrated processing using the integrated processing rule rules to perform integrated processing storage unit. And a display unit for displaying the contents of the integrated processing storage unit.