JPH03214259A - Information extracting method - Google Patents

Abstract

PURPOSE:To extract the information at a high speed by starting simultaneously the forward and backward retrieving actions centering on a retrieving start point given in a document. CONSTITUTION:An item value extracting part 11 receives the restrieving start point, the characters and their character expressing definition and sends the retrieving start point and the character expressing definition of each received item to a forward retrieving part 16 and a backward retrieving part 17 to receive the character string which satisfies the given character expressing definition and the position of the character string in the document. Then the part 11 calculates the distances between the retrieving start point given first and the position of the character strings sent back from both parts 16 and 17 in the document and regards the character string having the smaller distance as the first received one to define it as the value of a relevant item. Then a pair of each item and the extracted item value is outputted. Thus the information can be extracted at a high speed and this information extracting method can be applied to a wide range of fields.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to an information extraction method for extracting items related to specified general foam content from a given document.

[Conventional technology]

Conventional information extraction methods use ``Proceedings on the Fits National Conference on Artificial Intelligence'' as a technology for extracting specified items from natural language documents.
(Procedures of the fifth
national conference on ar
tificialntellgence) + Morgan Calfman Publications, Yshires Inc. (Morgan
Kaufmann Publishers, Inc.
), 1986, Vol. 2, p. +089, the technology adopted by the ATRANS system is known. The ATRANS system is a system that extracts items such as remittance optical bank and remittance amount from a bank's tele-socus text.The input sentence is syntactically and semantically analyzed and expressed semantically, and specified items are extracted from the semantic expression.

[Problem to be solved by the invention] However, conventional information extraction methods require large-scale dictionaries, complex analysis techniques, and large-capacity, high-speed computers in order to perform syntactic and semantic analysis of the entire document. There was a problem. Moreover, the items that can be extracted are fixed,
The problem was that the target fields were extremely limited.

An object of the present invention is to provide an information extraction method that is fast and applicable to a wide range of fields.

[Means to solve the problem]

The information extraction method of the first invention is an information extraction method for extracting information from a given document, in which a concept definition section stores general foam content, items related to the general foam content, and character expression definitions of the items. When a certain document and a certain general bubble content in the document are given, the item and character expression definition corresponding to the given general bubble content are retrieved by referring to the concept definition section, and each extracted item is , perform a forward and backward search from the occurrence location of a given approximate bubble in a given document, and extract the first searched character string that satisfies the character expression definition corresponding to each retrieved item. It is constructed by setting the item values of the specified items.

Further, the information extraction method of the second invention is a check that presents the extracted item values to the user and accepts a judgment as to whether the extracted item values are correct or incorrect in the information extraction method of the first invention. If the item value is determined to be incorrect by the checking section, the search is continued again from the position where the item value determined to be incorrect was retrieved.

On the other hand, the information extraction method of the third invention is configured by providing a concept addition section in which the user adds new general bubbles, items, and character expression definitions to the concept definition section in the information extraction method of the first invention. has been done.

Furthermore, in the information extraction method of the first invention, the information extraction method of the fourth invention includes a concept relationship definition section that defines a superordinate relationship between the concepts, and refers to the concept relationship definition section to identify the subordinate concept. The subconcept name is constructed by inheriting the items and character expression definitions of the lower level concept name, and defining only the items and character expression definitions specific to the lower level general concept name.

[Effect]

First, the operation of the first invention will be explained.

In general, information related to a certain concept is often written together.

FIG. 7 is a document-side explanatory diagram showing an example of the guide text for "Farewell Party." For example, as shown in Figure 7, an example of a guide for a "farewell party" includes information such as "date", "time", etc. related to the "farewell party".
Information such as "location" is written near where the character string "farewell party" appears in the document. Therefore, when extracting item information related to "Farewell Party" such as "Date", "Time", and "Place" from this guidance text, it is necessary to move forward from the position where the character string "Farewell Party" appears in the document. If you search for the required item in parallel and backwards, you can often find it quickly and correctly.

On the other hand, there are various ways to write information such as "date" and "time." In the first invention, multiple character expression definitions representing writing patterns are given in advance for each item such as "date" or "time", and a character string that satisfies the character expression definition is searched from the document. Item information is extracted by

Note that the character expression definition is described using a predetermined grammar for character expression definition. For example, "date" can be written as "April 1st", r12/31J, etc. These patterns can be expressed as “ku number〉−′月”′−
It is written as a character expression definition such as [number>-“day ゛′”, [number>-number>-“/”-<number>-number”, etc.

On the other hand, the concept definition section contains general terms such as "farewell party," related items such as "date,""time," and "cost," as well as the above definitions for each item. It stores multiple character expression definitions. Then, given a certain document and a certain general bubble content in that document, the item and character expression definition corresponding to the general bubble content are extracted from the concept definition part, and each extracted item is extracted using the character expression definition. Perform processing. The extraction process searches forward and backward from a given approximate bubble-containing appearance location in a given document.
Search for a character string that satisfies the character expression definition for each item, and use the first character string found as the item value for that item. In this way, item values are extracted by searching the document for character strings that satisfy the character expression definition of each item in parallel forward and backward from the appearance point of the general bubble, so it is necessary to perform syntactic and semantic analysis of the entire document. There is no.

Furthermore, general foam content, items related to general foam content, and their character expression definitions are stored in the concept definition section, and items to be extracted and character expression definitions are retrieved according to the specified general foam content. Since the extraction is performed using the following methods, it can be applied to a wide range of fields.

Next, the operation of the second invention will be explained. In the first invention, when an item value is extracted, the item value is output and the process ends, so if the extracted item value is incorrect, it cannot be corrected. In the second invention, the extracted item value is presented to the user, the user checks whether the extracted item value is correct or incorrect, and if the extracted item value is incorrect, the user is allowed to continue searching again. be able to.

Next, the operation of the third invention will be explained. The concept of extracting information from a document differs depending on the user. You may also want to add new concepts.

In the third invention, a concept addition section is provided, and the user can newly add general bubbles, items, and character expression definitions to the concept definition section, so that more flexible processing that is adapted to the user can be performed. I can do it.

Next, the operation of the fourth invention will be explained. There may be common items between concepts. For example, the items related to "banquet" are "date", "time", "place", and "cost", and the items related to "meeting" are "date", "time", "place", and "participation". ``Date'', ``1 hour'', ``
"Place" is common between the two concepts. In such a case, it is cumbersome to write the definitions of "date", "time", and "place" twice. In the fourth invention, by defining a superior-lower relationship between concepts and having the lower-level general foam-containing inherit the items and character expression definitions of the upper-level general foam-containing, the lower-level general foam-containing Define only the item and character expression definition.

For example, a general content such as "event" is defined as a general content above "banquet" or "meeting", and "date", "time", and "place" are defined as items related to "event". Then, only "cost" is defined for "banquet", which is a lower level bubble, and other items are inherited from "event".

Similarly, only "participant" is defined for "meeting", and other items are inherited from "event". In this way, by defining the superordinate relationship between concepts and having the lower-level general bubbles inherit the items and character expression definitions of the higher-level general bubble-containers,
Since it is only necessary to define items specific to the lower general bubble-containing level and character expression definitions for the lower general bubble-containing level, the complexity of the character expression definition can be reduced.

〔Example〕

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

FIG. 1 is a block diagram showing an embodiment of the information extraction method of the first invention. As shown in FIG. 1, 11 is an item value mountain part, 12 is a concept definition part, 15 is a concept definition extraction part, 1
6 is a forward search section, and 17 is a backward search section.

The concept definition unit 12 stores a triplet of a concept name, an item related to the concept name, and a character expression definition of the item.

FIG. 5 is an information configuration diagram showing an example of the information configuration stored in the concept definition section 12. As shown in FIG. 5, for example, m items are associated with concept 1, and item 1-1
is associated with character expression definitions.

Further, the forward search unit 16 receives a search starting point in a given document and one or more character expression definitions, scans the given document forward from the search starting point, and satisfies any of the character expression definitions. Find a string.

If there is a string that satisfies the character expression definition, that string and its position in the document are returned. If there is no string that satisfies the character representation definition, a negative value is returned as the position of the string in the document.

On the other hand, the backward search unit 17 receives a search start point in a given document and one or more character expression definitions, scans the given document backward from the search start point, and searches for characters that satisfy any of the character expression definitions. Find a column. If there is a string that satisfies the character expression definition, that string and its position in the document are returned. If there is no string that satisfies the character representation definition, a negative value is returned as the position of the string in the document.

Note that the forward search section 16 and the backward search section 17 perform processing in parallel.

In FIG. 1, when it is desired to extract information about a certain concept from a document, the document and the search start point are passed to the item value extraction unit 11, and the concept name is passed to the concept definition extraction unit 15. The concept definition extraction unit 15 receives the concept name, and the concept definition extraction unit 12
, the item and character expression definition corresponding to the received concept name are extracted and passed to the item value extraction unit 11.

Therefore, the item value extraction unit 11 receives the document, the search start point, the item, and its character expression definition, and passes the search start point and character expression definition for each received item to the forward search unit 16 and backward search unit 17. Receives a character string that satisfies a given character representation definition and the position of that character string in the document. The distance between the position in the document of the string returned from the forward search section and the backward search section and the first given search start point is calculated, the one with the smaller distance is regarded as the first received string, and the It is set as the item value of the item, and a pair of each item and the extracted item value is output. If the position of either character string in the document is a negative value, the non-negative character string is used as the item value for that item. If the position of either character string in the document is a negative value, a symbol indicating that it is unknown is used as the item value.

FIG. 2 is a block diagram showing an embodiment of the information extraction method of the second invention. As shown in FIG. 2, 21 is an item value extraction section, 22 is an essential item storage section, 23 is a check section, 25
2 is a concept definition retrieval unit, 26 is a forward search unit, 27 is a backward search unit, and 29 is a position storage unit. 22 shown in FIG.
25.26°27 are respectively 12.16. in FIG.
They respectively correspond to IE3.17 and have the same functions.

Then, the checking unit 23 presents the pair of item and item value received from the item value extraction unit 21 to the user, and receives a determination as to whether the pair is correct or incorrect.

If it is determined to be correct, the process ends at that point. If it is determined that the item is incorrect, the item is passed to the item value extraction unit 21. Repeat the above operations until it is determined to be correct.

Further, the position storage unit 29 is a place where items, extracted item values, and their positions in the document are stored.

Therefore, when the item value extraction unit 21 receives the document, search start point, item, and its character expression definition, it performs the extraction process in the same way as in the case of FIG. 1, and outputs a pair of each item and the extracted item value. At the same time, pairs of each item and the position of the extracted item value in the document are stored in the position storage unit 29. Next, when only the item is passed from the checking unit 23, the position of the previously extracted item value of the passed item is retrieved with reference to the position storage unit 29, and the distance between the retrieved position and the search start point is determined.

This will be called the correction distance.

Then, the position forward by the corrected distance from the search start point and the character expression definition of the passed item are passed to the forward search unit 26, and in parallel, the position backward by the corrected distance from the search start point and the passed item's character expression definition are passed to the forward search unit 26. The character expression definition is passed to the backward search section 27 to search for a character string that satisfies either of the given character expression definitions, and from both the forward search section 26 and the backward search section 27, the given character expression is searched. Receives a string that satisfies the definition and its position in the document. The distance between the position in the document of the character string returned from the forward search unit 26 and the backward search unit 27 and the first given search starting point is calculated, and the character string with the smaller distance is selected as the item value of the item. regarded as. If the position of either character string in the document is a negative value, the non-negative character string is used as the item value for that item. If the position of either character string in the document is a negative value, a symbol indicating that it is unknown is used as the item value.

Each item, the extracted item value, and the position of the item value in the document are stored in the position storage unit 29, and the pair of each item and item value is passed to the checking unit 23.

FIG. 3 is a block diagram showing an embodiment of the information extraction method of the third invention. As shown in FIG. 3, 31 is an item value extraction section, 32 is a concept definition section, 34 is a concept addition section, 35 is a concept definition extraction section, 36 is a forward search section, and 37 is a backward search section. 31, 32, 35, 36, 37 shown in Figure 3
11.12, 15, 16.17 in Figure 1, respectively.
They have the same functions corresponding to each other.

Then, the concept addition unit 34 receives the triplet of the general bubble-containing, the item related to the general bubble-containing, and the character expression definition of the item,
It is added to the concept definition section 32. If the received general bubble-containing definition and item already exist in the concept definition unit 32, the newly received character expression definition is added to the set of character expression definitions. If the general bubble content is already in the concept definition section 3,
2, and if the received item does not exist, the pair of the newly received item and the character expression definition is added to the set of items.

Note that the item value extraction process is the same as in the case of FIG.

FIG. 4 is a block diagram showing an embodiment of the information extraction method of the fourth invention. As shown in FIG. 4, 41 is an item value extraction section, 42 is a concept definition section, 45 is a concept definition extraction section,
46 is a forward search section, 47 is a backward search section, and 48 is a conceptual relationship definition section. Each of 41.46.47 shown in FIG. 4 corresponds to 11.16.17 of FIG. 1 and has the same function. In the concept relationship definition section 48, higher-level and lower-level relationships between concepts are defined.

FIG. 6 is an information configuration diagram showing an example of the upper-lower relationships between concepts defined in the concept relationship definition section 48.

As shown in FIG. 6, the concept relationship definition unit 49 defines each superordinate concept for each concept.

In addition, in the concept definition unit 42, the lower-level general bubble inherits the items and character expression definitions of the higher-level affiliated bubble-container, and if the higher-level concept also has a registered higher-level concept to inherit,
Furthermore, since the items and character expression definitions of the higher-level concept are inherited, only the items and character expression definitions specific to the general bubble-containing concept are defined for the lower-level general foam-containing concept.

Therefore, the concept definition retrieval unit 45 receives the general foam content, refers to the concept definition unit 42, retrieves the item and character expression definition corresponding to the received general foam content, and at the same time refers to the concept relationship definition unit 48. Then, take out the uppermost foam-containing one. If there is a corresponding high-level generally foam-containing item, the concept definition unit 42 is referred to and the item and character expression definition corresponding to the high-level generally foam-containing item are extracted. This operation is repeated until the upper layer contains almost no bubbles.

Note that the item value extraction process is the same as in the case of FIG.

〔Effect of the invention〕

As described above, the information extraction method of the present invention is given a search starting point in a document and searches forward and backward simultaneously from there, so there is no need to analyze the entire text.
In addition to being able to process at high speed, it is possible to define general foam content, items related to general foam content, and their character expression definitions regardless of the type of document, so it has the effect of being applicable to documents in a wide range of fields. There is.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an embodiment of the information extraction method of the first invention, FIG. 2 is a block diagram showing an embodiment of the information extraction method of the second invention, and FIG. 3 is a block diagram showing an embodiment of the information extraction method of the second invention. FIG. 4 is a block diagram showing an embodiment of the information extraction method of the fourth invention, and FIG. 5 shows an example of the information structure stored in the concept definition section. FIG. 6 is an information configuration diagram showing an example of the upper-lower relationship between concepts defined in the concept relationship definition section. FIG. 7 is a document-side explanatory diagram showing an example of the guide text for "Farewell Party" It is. 11... Item value extraction section, 12... Concept definition section, 15
...Concept definition retrieval section, 16...Forward search section, 1
7... Backward search section, 21... Item value extraction section, 22.
... Concept definition part, 23... Check part, 24... Concept definition extraction part, 26... Forward search part, 27...
Backward search unit, 29... Position storage unit, 31... Item value extraction unit, 32... Concept definition unit, 34... Concept addition unit, 35... Concept definition extraction unit, 36... Forward search unit, 37... Backward search unit, 41... Item value extraction unit,
42... Concept definition section, 45... Concept definition extraction section, 46... Forward search section, 47... Backward search section, 48
... Conceptual relationship definition part.

Claims (1)

[Claims] 1. An information extraction method for extracting information from a given document, comprising a concept definition unit storing a concept name, an item related to the concept name, and a character expression definition of the item, When a document and a certain concept name in the document are given, the item and character expression definition corresponding to the given concept name are retrieved by referring to the concept definition section, and for each retrieved item, the given document is Search forwards and backwards from the occurrence of a given concept name in , and use the first searched character string that satisfies the character expression definition corresponding to each retrieved item as the item value of the retrieved item. An information extraction method characterized by: 2. The information extraction method according to claim 1, further comprising a checking section that presents the extracted item values to the user and accepts a judgment as to whether the extracted item values are correct or incorrect, and the checking section An information extraction method characterized by continuing the search again from the position where the item value determined to be incorrect is retrieved when it is determined that the item value is incorrect. 3. The information extraction method according to claim 1, further comprising a concept addition section through which the user adds new concept names, items, and character expression definitions to the concept definition section. 4. The information extraction method according to claim 1, further comprising a concept relationship definition section that defines a superior-subordinate relationship between concepts, and refers to the concept relationship definition section to add items of a superordinate concept name and a character expression definition to a subordinate concept name. , and defining only items and character expression definitions specific to the subordinate concept name in the subordinate concept name.