JP2012242966A - Knowledge acquisition device, knowledge acquisition method, and program - Google Patents

Abstract

To provide a knowledge acquisition device capable of outputting results more quickly.
A first storage unit storing a word pair table having a pair of words having a specific relationship and a morpheme structure related to a sentence including the pair of words, and the first storage unit A search key is created by adding a specific event that limits the search target to the extracted word pair and morpheme structure, and the second storage means storing the search target document group is searched using the search key The number of occurrences of the search key is obtained, the number of occurrences obtained is associated with the word pair and the structure of the morpheme and stored in the word pair table, and the word pair table is referred to. The relevance between each pair of words and the specific event was evaluated based on the degree of matching between the trend of the number of appearances of each morpheme related to the pair and the number of appearances of each morpheme related to the pair of words. An evaluation means for outputting an evaluation value; Knowledge acquisition apparatus comprising.
[Selection] Figure 3

Description

  The present invention relates to a technique for acquiring knowledge such as word pairs from a data group using a computer, and more particularly to a knowledge acquisition device, a knowledge acquisition method, and a program for acquiring knowledge about a relationship between a specific event and a word pair.

  2. Description of the Related Art Conventionally, data including a keyword (search key) is searched from an accessible data group using the Internet or the like. Data acquired as a result of the search is mainly an electronic document.

  Knowledge acquired through such data search can be applied to the creation of dictionaries, dictionaries, etc., and the provision of electronic information services. For example, if you search a data group using “* is effective for **”, “* is effective for **”, etc. as a search key, a pair of words such as “disease” and “food that has an effect on it” will be found. It is expected that multiple sets will be acquired. Here, the above “*” and “**” indicate a portion where data search is performed without specifying the contents.

  Hereinafter, a word pair having a specific relationship obtained in this manner is referred to as a word pair. In the above case, “cold” and “radish”, “cold” and “ginger”, “headache” and “plum pickled”, etc. can be acquired as word pairs. In addition to the above "* works for **", there are various things such as "place" and "cleaning tool suitable there", "season" and "cooking", "food ingredients" and "condiment" Are considered as word pairs. It is considered that useful information services can be provided by comprehensively collecting such word pairs.

  On the other hand, parts other than “*” and “**”, that is, “Ha” and “Effective”, and “Ha” and “Effective” are extraction rules for obtaining word pairs. It can be said. Hereinafter, such an extraction rule is referred to as a “context pattern”.

  A word pair and a context pattern can be acquired from a large-scale data by an automatic algorithm by a computer (see, for example, Patent Document 1 and Non-Patent Document 1).

  Specifically, first, a known word pair called a seed is given to the computer. The computer searches the large-scale data for context patterns including seeds. When the context pattern is obtained, a document including the obtained context pattern is searched, and an unknown word pair is obtained from the obtained document. By repeating these steps, the number of word pairs and context patterns increases. Eventually, a plurality of word pairs and a plurality of context patterns are acquired as knowledge, and a database that can be used for applications such as a dictionary and a dictionary is created.

  It should be noted that a process of excluding knowledge (noise, dust) that seems to be uncommon may be performed by giving a condition such as “the number of appearances is a predetermined number or more”. At this time, it is also assumed that the context pattern is weighted by inserting various setting inputs of the user into the computer operation.

US Pat. No. 7,146,308

Stijn De Saeger et al., “Acquisition of large-scale semantic relationships by pattern learning using word semantic classes”, 16th Annual Conference of the Language Processing Society, pp 932-935, 2010

  By the way, when acquiring knowledge as described above, there is a need to collect data limited to a specific event related to data (electronic document). Typical examples of specific events are categories such as region, industry, and field. Hereinafter, for the sake of simplicity, it is referred to as a category.

  Specifically, when acquiring word pairs such as the above-mentioned “disease” and “food effective for it”, (1) use only word pairs acquired from a medical-related authoritative document, (2 On the other hand, there may be a request to collect grass root word pairs comprehensively. Further, due to the user's residence, etc., there may be a request such as (3) collecting word pairs that are common in West Japan alone.

  However, the process in which a computer acquires word pairs and context patterns from large-scale data using an automatic algorithm includes many repetitive processes. Therefore, if a process is performed after a category is specified, the processing time becomes long. As a result, there is a high possibility that the desired result cannot be quickly provided to the user.

  On the other hand, it is conceivable to acquire word pairs and context patterns for each category in advance and store them in a database. However, in this case, the amount of data becomes enormous and the demand for resources increases. In addition, it is not realistic to predict all categories that can be specified by the user.

  The present invention is to solve such problems, and a main object of the present invention is to provide a knowledge acquisition device, a knowledge acquisition method, and a program that can output results more quickly.

One aspect for achieving the above object is as follows:
A first storage means storing a word pair table having a pair of words having a specific relationship and a structure of a morpheme related to a sentence including the pair of words;
A second search key is created by adding a specific event limiting the search target to the word pair and morpheme structure extracted from the first storage means, and storing a search target document group using the search key. Retrieving the number of occurrences of the search key by searching the storage means, the appearance number information acquisition means for storing the obtained number of occurrences in the word pair table in association with the word pair and the morpheme structure;
With reference to the word pair table, based on the degree of coincidence between the number of occurrences of each morpheme structure related to each word pair and the number of appearances of each morpheme structure related to all word pairs, An evaluation means for outputting an evaluation value obtained by evaluating an association with the specific event;
It is a knowledge acquisition device provided with.

  ADVANTAGE OF THE INVENTION According to this invention, the knowledge acquisition apparatus etc. which can output a result more rapidly can be provided.

1 is an overall view of an information system including a knowledge acquisition device 1 according to an embodiment of the present invention. It is a hardware structural example of the knowledge acquisition apparatus 1 of a present Example. It is a functional structural example of the knowledge acquisition apparatus 1 of a present Example. It is a figure which shows typically the whole image of the process performed by the knowledge acquisition apparatus 1 of a present Example. It is a figure which shows typically a mode that the operator 110 inputs a seed with respect to the client computer 100, and the knowledge acquisition apparatus 1 produces | generates the word pair table 40. FIG. It is an example of the word pair table 40 produced | generated by the word pair acquisition part 30. FIG. It is a figure which shows typically a mode that the user 120 inputs a query with respect to the client computer 100, and the knowledge acquisition apparatus 1 produces | generates the appearance number table 42. FIG. It is an example of the appearance number table generated by the search unit. It is a figure which shows typically a mode that the knowledge acquisition apparatus 1 produces | generates the word pair 44 with a score. 4 is an example of a scored word pair 44 output by the output of the score calculation unit 34; It is a schematic diagram for contrasting the process which can be performed with the conventional apparatus, and the process performed by the knowledge acquisition apparatus 1 of a present Example. 4 is a flowchart showing a flow of characteristic processing executed by a word pair acquisition unit 30. 5 is a flowchart showing a flow of context pattern extraction processing executed by a word pair acquisition unit 30. It is a figure which shows typically the structure of the morpheme used in the analysis of S410. 4 is a flowchart showing a flow of characteristic processing executed by a search unit 32. 4 is a flowchart showing a flow of characteristic processing executed by a score calculation unit.

  DESCRIPTION OF EMBODIMENTS Hereinafter, embodiments for carrying out the present invention will be described with reference to the accompanying drawings.

  Hereinafter, a knowledge acquisition device 1 according to an embodiment of the present invention will be described with reference to the drawings.

[Hardware configuration]
FIG. 1 is an overall view of an information system including a knowledge acquisition device 1 according to an embodiment of the present invention. As illustrated, the knowledge acquisition device 1 is a server device in which one or a plurality of client computers 100 are connected via a network 50, for example. The knowledge acquisition apparatus 1 can access the large-scale document 200 via the network 50. The network 50 includes the Internet, a mobile phone, a PHS (Personal Handy-phone System) radio network, a LAN (Local Area Network), and the like.

  FIG. 2 is a hardware configuration example of the knowledge acquisition apparatus 1 of the present embodiment. The knowledge acquisition device 1 includes, for example, a CPU (Central Processing Unit) 10, a drive device 12, an auxiliary storage device 16, a memory device 18, an interface device 20, an input device 22, and an output device 24. Information processing apparatus. These components are connected via a bus, a serial line, or the like.

  The CPU 10 is a processor having, for example, a program counter, an instruction decoder, various arithmetic units, an LSU (Load Store Unit), a general-purpose register, and the like.

  The drive device 12 is a device that can read a program and data from the storage medium 14. When the recording medium 14 on which the program is recorded is mounted on the drive device 12, the program is installed from the recording medium 14 to the auxiliary storage device 16 via the drive device 12. The recording medium 14 is a portable recording medium such as a CD-ROM, a DVD disk, or a USB memory. The auxiliary storage device 16 is, for example, an HDD (Hard Disk Drive) or a flash memory.

  In addition to using the storage medium 14 as described above, the program can also be installed by the interface device 20 being downloaded from another computer via the network 50 and installed in the auxiliary storage device 16. Further, it may be stored in advance in an auxiliary storage device 16 or a ROM (Read Only Memory) at the time of shipment of the information processing apparatus.

  When the CPU 10 executes the program installed or stored in advance as described above, the information processing apparatus having the mode shown in FIG. 1 can function as the knowledge acquisition apparatus 1 of the present embodiment.

  The memory device 18 is, for example, a RAM (Random Access Memory) or an EEPROM (Electrically Erasable and Programmable Read Only Memory). The interface device 20 controls connection with the network.

  The input device 22 is, for example, a keyboard, a mouse, a touch pad, a touch panel, a microphone, or the like. The output device 24 includes, for example, a display device such as an LCD (Liquid Crystal Display) and a CRT (Cathode Ray Tube), a printer, a speaker, and the like.

  The interface device 20 can access the client computer 100 and the large-scale document 200 via the network 50.

  The client computer 100 includes a CPU, a drive device, an auxiliary storage device, a memory device, an interface device, an input device, an output device, and the like, like the knowledge acquisition device 1 on the server side. Detailed description thereof will be omitted.

  The large-scale document 200 is an electronic document stored in any storage device accessible via the network 50.

[Function configuration]
FIG. 3 is a functional configuration example of the knowledge acquisition apparatus 1 of the present embodiment. The knowledge acquisition device 1 includes a word pair acquisition unit 30, a search unit 32, and a score calculation unit 34. These functional blocks function when the CPU 10 executes program software stored in the auxiliary storage device 16 or the like.

  In addition, the knowledge acquisition device 1 generates a word pair table 40, an appearance number table 42, and a scored word pair 44 in predetermined areas of the memory device 18 and the auxiliary storage device 16.

  FIG. 4 is a diagram schematically showing an overall image of processing executed by the knowledge acquisition device 1 of the present embodiment. As illustrated, the knowledge acquisition device 1 generates a word pair table 40 based on the large-scale document 200 and the seed 250.

  In addition, the knowledge acquisition device 1 performs a category search based on the word pair table 40 and the query 260 to generate the appearance number table 42, calculates a score based on the appearance number table 42, and outputs a scored word pair 44. . Here, the “category” is an example of “a specific event that limits a search target” in the claims.

  Among these processes, the generation of the word pair table 40 is performed as a pre-process for service provision (preliminary batch process), and the generation of the appearance number table 42 and the score calculation are performed when a service request is made by the user. It is suitable to be executed (after the application destination is determined). Hereinafter, these processes will be described.

{Acquire word pairs}
The word pair acquisition unit 30 searches the large-scale document 200 for a context pattern including a seed based on a given seed.

  Here, the seed is a context pattern that appears in a word pair (word pair) having a specific relationship or in a sentence including the word pair. Both of these may be given as seeds. Examples of word pairs include “cold” and “ginger” in the sentence “Ginger is effective for colds”, and the context pattern in this case is “Ha” and “Effective for”. The context pattern is a characteristic character string other than a word pair included in a sentence in the large-scale document 200 that is a data search target, and is understood as an extraction rule for acquiring a word pair from the large-scale document 200. be able to. Further, as will be described later, the context pattern in the present embodiment is an example of “a morpheme structure relating to a sentence including a word pair” in the claims. In this embodiment, a “sentence” with a delimiter “.” As a delimiter is a data search target, but a “sentence” in which a plurality of sentences are collected may be a data search target.

  The seed is provided to the CPU 10 via the network 50 when the operator inputs an arbitrary word pair and / or context pattern to the client computer 100, for example. FIG. 5 is a diagram schematically illustrating how the operator 110 inputs a seed to the client computer 100 and the knowledge acquisition device 1 generates the word pair table 40.

  When the seed is a word pair, the word pair acquisition unit 30 first searches the large-scale document 200 for a context pattern including the seed. When the context pattern is obtained, a document including the obtained context pattern is searched from the large-scale document 200, and an unknown word pair is obtained from the obtained document. By repeating these steps, the number of word pairs and context patterns increases. The word pair acquisition unit 30 holds the acquired word pairs and context patterns, for example, in the form of an empty data table using these as labels.

FIG. 6 is an example of the word pair table 40 generated by the word pair acquisition unit 30. As shown in the figure, the word pair table 40 uses word pairs and context patterns as labels in the row direction and column direction, and data content portions corresponding to combinations of word pairs and context patterns are blank. An appearance number table 42 is generated by storing the number of appearances (number of hits) output as a search result of the search unit 32 described later in this blank part.
Here, the data search from the large-scale document 200 is performed by a known search technique when the large-scale document is owned, and represented by Google (registered trademark) when the large-scale document is not owned. This is done using technology implemented by existing search service providers. The specific contents of such data search are well-known and commonly used techniques, and will not be described in detail. However, search conditions such as AND, OR, and NOT used in general document search should be set appropriately. Thus, a document including a word pair or a context pattern can be extracted.

{Search}
Based on the word pair table 40 generated by the word pair acquisition unit 30, the search unit 32 performs a data search that reflects a query input by an operator or user on the large-scale document 200. Then, the result is stored as an appearance number table 42.

  FIG. 7 is a diagram schematically illustrating how the operator 110 or the user 120 inputs a query to the client computer 100 and the knowledge acquisition device 1 generates the appearance number table 42. Note that the client computer 100 in this figure may be different from that in FIG.

  For example, the operator 110 or the user 120 inputs a category such as ““ National Hospital ”OR“ City Hospital ”” (an example of “specific event” in the scope of claims) as a query. Then, the search unit 32 performs a data search that satisfies the query ““ national hospital ”OR“ city hospital ”” for all combinations of word pairs and context patterns stored in the word pair table 40. That is, a search is performed from the large-scale document 200 using, as a search key, one word pair and one context pattern extracted from a plurality of word pairs and a plurality of context patterns stored in the word pair table 40 as a search key. This process is performed for all word pairs and context patterns.

  Then, the search unit 32 acquires the number of appearances in the search result satisfying the query ““ national hospital ”OR“ city hospital ”” for each word pair and context pattern combination, and stores it in the blank portion of the word pair table 40. Thus, the appearance number table 42 is obtained. For example, if the number of appearances when searching for AND (“national hospital” OR “city hospital”) is 21 (“cold” works for “cold”), this is expressed as “cold” in the word pair table 40. "-" Radish "and" A is effective for B ". The appearance number table 42 is generated as a dedicated table corresponding to the query (“national hospital” OR “city hospital”). That is, the search unit 32 generates a dedicated table corresponding to the input query. FIG. 8 is an example of the appearance number table 42 generated by the search unit 32.

  In this way, the search unit 32 acquires the number of appearances of a document that satisfies the input query and corresponds to all combinations of word pairs and context patterns stored in the word pair acquisition unit 30 in advance. A number table 42 is generated.

{Score calculation}
The score calculation unit 34 refers to the appearance number table 42 generated by the search unit 32, evaluates the relevance between each word pair and the query, and performs output based on the evaluation result.

  The output of the score calculation unit 34 is made, for example, in the form of a scored word pair 44 with a score added to the appearance number table 42. The score in the present embodiment is (1) relevance to the category input as a query, and (2) the number of occurrences of the word pair itself, that is, whether or not the word pair is generally permeated. It is the evaluation value shown.

  FIG. 9 is a diagram schematically illustrating how the knowledge acquisition device 1 generates a scored word pair 44. FIG. 10 is an example of a scored word pair 44 output by the score calculation unit 34. As shown in the figure, the score calculation unit 34 adds a score to each word pair stored in the appearance number table 42 and outputs it.

The score can be calculated by, for example, the following formula (1). In the formula, Score _i represents the score of the i-th word pair, f (w, p) represents the number of simultaneous occurrences of word pair w and context pattern p, and N represents the occurrence of all word pairs and context pattern combinations. Indicates a number. The subscript j is a context pattern identification parameter, and the subscript k is a word pair identification parameter.

  The above equation (1) calculates the number of occurrences for each context pattern (the numerical values in the “total” line in FIG. 10) for the entire word pair, and the value obtained by dividing this by the total number of occurrences is used as a weighting factor. Multiply by the number of occurrences for each context pattern. Therefore, the score in the present embodiment is based on the degree of matching between the appearance number tendency for each context pattern for each word pair and the appearance number tendency for each context pattern for all word pairs, and the relationship between each word pair and the query. It becomes an evaluation value that is evaluated appropriately. Referring to FIG. 10, for example, the score of the word pair (“cold” − “radish”) is expressed by the following equation (2) because N is 1047.

Score _{(“cold” − “radish”)} = 21 × (201/1047) + 8 × (155/1047) + 36 × (83/1047) +... = 0.21 (2).

  As a result, a pair of words having a high number of appearances does not obtain a good score, but a word pair showing a tendency close to the appearance tendency of context patterns corresponding to all word pairs satisfying the query obtains a high score. It will be. For example, in the category of “hospital relations”, when there is a tendency that the context pattern 1 is often used but the context pattern 3 is not used much, a word pair showing a tendency close to such a tendency is high. Give a score. As a result, a high score can be given to word pairs that are well suited to the category, and the user's request to “get knowledge of word pairs that match the category” can be met. Note that the score may be calculated by performing some normalization processing, for example, by multiplying the total number of word pairs.

  When the scored word pair 44 is generated as described above, it may be output to the user as it is, or a ranking (rearranged) in the order of score may be output. Moreover, you may provide to a user only in the top ranking. As described above, various outputs based on the score calculation result can be considered.

  Moreover, as described above, the generation of the word pair table 40 is performed as a pre-process for providing the service, and the generation of the appearance number table 42 and the score calculation are preferably performed when a service request is made by the user. It is. FIG. 11 is a schematic diagram for comparing a process that can be executed by a conventional apparatus with a process that is executed by the knowledge acquisition apparatus 1 of the present embodiment.

  In the figure, the upper part schematically shows processing executed by a conventional apparatus. As shown in the figure, in order to acquire word pairs for each category in the conventional apparatus, it is necessary to perform the same processing as the word pair acquisition unit 30 for each category. However, the processing of the word pair acquisition unit 30 is to repeatedly acquire word pairs and context patterns from the seed, and the required time is relatively long. As a result, the processing time from when the user or operator inputs a category until the word pair is acquired becomes longer.

  On the other hand, in the knowledge acquisition apparatus 1 of the present embodiment in which the outline of the process is represented in the latter stage of FIG. 11, the word pair table 40 is generated as a pre-batch process, and the appearance number table 42 after the application destination is determined. Only generation and score calculation are performed. The generation of the appearance number table 42 is a process for searching the large-scale document 200 by setting a search key reflecting the query, and is completed in a relatively short time. Also, the score calculation is a simple calculation process, so the processing time is short. As a result, the processing time from when the user inputs a query until the word pair is acquired can be shortened. That is, the result can be output more quickly.

[Processing flow]
Hereinafter, processing of each functional block included in the knowledge acquisition device 1 will be specifically described with reference to a flowchart. For the entire flow, refer to FIG. 4 and the illustration is omitted.

  FIG. 12 is a flowchart showing a flow of characteristic processing executed by the word pair acquisition unit 30.

  First, the word pair acquisition unit 30 adds the input seed 250 to the word pair list (S300). The word pair list 40 </ b> A and a context pattern list 40 </ b> B, which will be described later, serve as labels in the row direction and column direction of the word pair table 40, and are set in predetermined areas of the memory device 18 and the auxiliary storage device 16.

  Next, the word pair acquisition unit 30 searches the large-scale document 200 with the word pairs stored in the word pair list 40A, extracts a context pattern, and adds it to the context pattern list 40B (S302; details are shown in FIG. 12). Description).

  Next, the word pair acquisition unit 30 determines whether even one new context pattern has been extracted (S304). If no new context pattern can be extracted, this flow ends.

  On the other hand, if even one new context pattern can be extracted, the large-scale document 200 is searched with the context patterns stored in the context pattern list, and word pairs are extracted and added to the word pair list 40A (S306). ).

  Next, the word pair acquisition unit 30 determines whether even one new word pair has been extracted (S308). If no new word pair can be extracted, this flow ends.

  On the other hand, if even one new word pair can be extracted, the process returns to S302. In this way, word pairs and context patterns are repeatedly extracted and added to the word pair list 40A and the context pattern list 40B.

  When this flow ends, the word pair acquisition unit 30 generates the word pair table 40 based on the contents of the word pair list 40A and the context pattern list 40B, and stores the word pair table 40 in a predetermined area of the memory device 18 or the auxiliary storage device 16.

  FIG. 13 is a flowchart showing a flow of context pattern extraction processing executed by the word pair acquisition unit 30. This flow corresponds to S302 in FIG.

  First, the word pair acquisition unit 30 extracts one word pair in order from the top of the word pair list 40A (S400).

  Next, the word pair acquisition unit 30 determines whether or not the search has been completed up to the end of the word pair list 40A, that is, whether or not a new word pair has been extracted in S400 (S402). When the search is completed up to the end of the word pair list 40A, this flow ends.

  On the other hand, if the search has not been completed up to the end of the word pair list 40A, a large-scale document is searched using the extracted word pair as a search key (S404). The search result is temporarily stored in the memory device 18 or the auxiliary storage device 16.

  Next, the word pair acquisition unit 30 takes out one search result of S404 (S406). Then, it is determined whether or not a search result has been extracted to the end, that is, whether or not a new search result has been extracted in S406 (S408). If a new search result cannot be extracted, the process returns to S400.

  On the other hand, if a new search result can be extracted, it is determined whether or not the word pairs included in the search result form a predetermined syntax pattern (S410).

  FIG. 14 is a diagram schematically showing a morpheme structure used in the analysis of S410. The result of document search using a certain word pair ("cold"-"radish") as a search key includes words other than word pairs and context patterns, such as "The radish is most effective for colds". It is assumed to be a sentence.

  First, the word pair acquisition unit 30 performs a process called morpheme analysis that divides a Japanese sentence into minimum units called morphemes. In the morpheme unit, for example, the word “tea” is counted as “o (prefix)” and “tea (noun)”. Then, when the sentence “The radish is the most effective for colds” is analyzed, the sentence is divided into “cold / ni / ha / the radishes / most / effective / yo”.

  Next, the word pair acquisition unit 30 performs dependency analysis. In dependency analysis, first, morpheme strings are grouped into units called phrases. The above sentences are grouped into phrases such as “(cold / ni / ha), (after all), (radish / ga), (first), (effective / yo)”. Then, the relationship between clauses is defined. The relationship in Japanese is analyzed based on the principle that (1) there is one member and (2) it is from the front to the back. The morphological analysis and dependency analysis are well known in the existing technology, and are automatically processed with an accuracy of about 99% and 91%, respectively.

The relationship between the above sentences is as follows.
(Cold / ni / ha) → (works / yo)
(After all) → (Effective / Yo)
(Daikon / ga) → (Effective / yo)
(First) → (works / yo)
(Effective / Yo) → <End of sentence>

  When such a relation is represented by a tree structure and an empirical rule that “the morpheme in the phrase is related to the morpheme immediately after” is used, the structure illustrated in FIG. 14 is acquired.

  Further, a method for extracting the character portion of “he” from the tree structure of the morpheme solution will be described with reference to the example of FIG. First, it is assumed that “cold” and “radish” are given as word pairs. Then, a minimum partial structure including both of these two word pairs (morphemes) is extracted. In the example of FIG. 14, the shaded “he” portion corresponds to this minimum partial structure.

  When the minimum partial structure is obtained in this way, the word pair acquisition unit 30 determines that a predetermined syntax pattern is formed when the word pairs are within a predetermined distance. When it is determined that a predetermined syntax pattern is formed, a part obtained by removing a word pair from the obtained minimum partial structure is recognized as a context pattern.

Here, the “predetermined distance” refers to, for example, the number of links connecting morphemes in the morpheme tree structure shown in FIG. The number of arrows (5) below corresponds to this.
(Cold) → (To) → (Ha) → (Effective)
(Radish) → (ga) → (works)

  By limiting the distance in this way, it is possible to suppress extraction of insignificant context patterns from word pairs included in relatively long sentences. When this distance is not limited, or when the distance is allowed up to about 30, from a sentence such as "There is no natto on the list that came to the shopping at the request of a wife who had a cold." This is because “cold” and “natto” are extracted, and insignificant context patterns may be extracted from this sentence. There is no special rule for determining the distance threshold, and it may be determined empirically depending on the number of errors in the knowledge obtained.

  Note that the tree structure of the morphological solution is not necessarily a “he” as shown in FIG. 14, and can take various forms. For example, the result of the morphological analysis for the sentence “I bought a radish that works for a cold” would be “cold / effective / effective / radish / obtained / buy / taken”.

The result of dependency analysis is as follows.
(Cold /) → (Effective)
(Effective) → (Daikon /)
(Daikon /) → (Buy / Ta)

In this case, the tree structure of the morpheme is not a “f” character but is a straight line as follows, and the extracted pattern is also a straight line.
(Cold / ni) → (Effective) → (Daikon //) → (Bought / taken)
(Cold →) → (Effective) → (Daikon → →) → (Buy →)

  Return to the description of the flow. If the word pair forms a predetermined syntax pattern, the extracted context pattern is added to the context pattern list 40B (S412). If the word pair does not form a predetermined syntax pattern, the process returns to S406.

  FIG. 15 is a flowchart showing a flow of characteristic processing executed by the search unit 32.

  First, the search unit 32 extracts one word pair from the top of the word pair table 40 in order (S500).

  Next, the search unit 32 determines whether or not the word pair has been processed to the end of the word pair table 40, that is, whether or not a new word pair has been extracted in S500 (S502). When the process is completed up to the end of the word pair table 40, this flow is finished.

  On the other hand, if the processing is not performed to the end of the word pair table 40, one context pattern is extracted in order from the leftmost column of the word pair table 40 (S504).

  Next, the search unit 32 determines whether or not processing has been performed up to the end of the word pair table 40 regarding the context pattern, that is, whether or not a new context pattern has been extracted in S504 (S506). When the process is completed up to the end of the word pair table 40, the process returns to S500.

  On the other hand, if the processing is not performed to the end of the word pair table 40, the word pair and context pattern extracted in S500 and S504 and the input query are combined to create a search key (S508). The scale document 200 is searched (S510). The search key is created, for example, by combining a sentence in which a word pair is embedded at positions A and B of a context pattern (for example, “A works for B”) and an input query using an AND condition.

  Then, the number of appearances (number of hits) is acquired, stored in the location corresponding to the word pair and context pattern extracted in S500 and S504 in the appearance number table 42 (S512), and the process returns to S504.

  Note that the flow of FIG. 15 does not automatically proceed, but may output a search result to the user and perform a confirmation operation for each search execution.

  Here, in the flow of FIG. 15, a search using a character string is performed without handling the morpheme tree structure. Therefore, depending on the search key setting method and the search specification, “a radish is still effective for colds” If there is "After all" etc., the number of appearances may not be counted. However, since (1) such a search is performed from a large-scale document, it is expected that there will be many expressions that do not contain “After all”, etc. (2) In the search processing, an accurate number of appearances is required. It is only necessary to know the tendency in a specific category. (3) Assuming that the ratio of character strings such as "Yappari" is constant for each pattern, "Yappari" etc. The decline in the number of searches due to the expression is not a big problem because it can be ignored. Rather, since detailed search is not performed, processing can be simplified and high-speed processing can be realized.

  FIG. 16 is a flowchart showing a flow of characteristic processing executed by the score calculation unit 34.

  First, the score calculation unit 34 refers to the appearance number table 40 and calculates the number of appearances of each context pattern and the total number of appearances N (S600).

  Next, the score calculation unit 34 extracts one word pair in order from the top row of the word pair table 40 (S602).

  Next, the score calculation unit 34 determines whether or not the word pair has been processed to the end of the word pair table 40, that is, whether or not a new word pair has been extracted in S602 (S604). When the process is completed up to the end of the word pair table 40, this flow is finished.

  On the other hand, when the processing is not performed to the end of the word pair table 40, one context pattern is extracted in order from the leftmost column of the word pair table 40 (S606).

  Next, the score calculation unit 34 determines whether or not processing has been performed up to the end of the word pair table 40 regarding the context pattern, that is, whether or not a new context pattern has been extracted in S606 (S608). When the process is completed up to the end of the word pair table 40, the process returns to S602.

  On the other hand, when the processing is not performed up to the end of the word pair table 40, the total number of occurrences of the corresponding context pattern (in the figure (for example, (2) in the figure)) A value obtained by multiplying the value obtained by dividing (3)) by the total number of appearances N ((4) in the figure) is calculated. Then, this value is cumulatively added to the score of the corresponding word pair ((1) in the figure) (S610).

  By such processing, the score expressed by the above formula (1) is calculated for each word pair.

"Summary"
According to the knowledge acquisition apparatus 1 of the present embodiment described above, since the occurrence number table 42 can be generated and the score can be calculated after the application destination is determined, the word pair is acquired after the query is input. The processing time can be shortened. Therefore, an evaluation value that appropriately evaluates the relevance between each word pair and the query can be output more quickly.

  The best mode for carrying out the present invention has been described above with reference to the embodiments. However, the present invention is not limited to these embodiments, and various modifications can be made without departing from the scope of the present invention. And substitutions can be added.

  For example, data corresponding to the word pair table 40 may be input from the outside or stored in advance in the auxiliary storage device 16 or the like. In this case, the word pair acquisition unit 30 can be omitted.

  Further, as described above, the score calculation unit 34 performs the output based on the result of evaluating the relationship between the word pair and the query, but on the contrary, based on the result of evaluating the relationship between the context pattern and the query. Output may be performed. In this case, the score calculation unit 34 determines whether the context pattern and the category are based on the matching degree between the appearance ratio for each word pair for one context pattern and the appearance ratio for each word pair for all context patterns. Assess relevance. Such output can be suitably used when the user is not an end user. In this case, the score is calculated by completely replacing the word pair and the context pattern in the above formula (1). In this way, an evaluation value that appropriately evaluates the relationship between the context pattern and the query can be output more quickly.

  In addition, “category” is given as a representative example of the contents of the query entered by the user. However, the query is not limited to the concept of “category” in general as long as it limits the search target. You can accept it. For example, there may be a time limit such as “documents from what day, what month, what day”.

Regarding the above description, the following items are further disclosed.
(Appendix 1)
A first storage means storing a word pair table having a pair of words having a specific relationship and a structure of a morpheme related to a sentence including the pair of words;
A second search key is created by adding a specific event limiting the search target to the word pair and morpheme structure extracted from the first storage means, and storing a search target document group using the search key. Retrieving the number of occurrences of the search key by searching the storage means, the appearance number information acquisition means for storing the obtained number of occurrences in the word pair table in association with the word pair and the morpheme structure;
With reference to the word pair table, based on the degree of coincidence between the number of occurrences of each morpheme structure related to each word pair and the number of appearances of each morpheme structure related to all word pairs, An evaluation means for outputting an evaluation value obtained by evaluating an association with the specific event;
A knowledge acquisition device comprising:
(Appendix 2)
The knowledge acquisition device according to attachment 1, wherein
The evaluation means multiplies the number of appearances of each morpheme related to each word pair by the number of appearances of each morpheme related to all word pairs, and outputs the value divided by the total number of appearances as an evaluation value. Means,
Knowledge acquisition device.
(Appendix 3)
The knowledge acquisition device according to appendix 1 or 2,
Searching the second storage means using a given word pair or morpheme structure as a search key, adding the obtained search result to the search key and repeating the data search to obtain a plurality of Data collection means for storing the word pairs and the structure of a plurality of morphemes in the storage means,
Knowledge acquisition device.
(Appendix 4)
The knowledge acquisition device according to any one of appendices 1 to 3,
The morpheme structure related to the sentence including the word pair is the morpheme structure obtained by removing the morpheme of the word pair from the morpheme structure that is the minimum partial structure obtained by performing dependency analysis on the target sentence of the data search. Structure,
Knowledge acquisition device.
(Appendix 5)
A first storage means storing a word pair table having a pair of words having a specific relationship and a structure of a morpheme related to a sentence including the pair of words;
A second search key is created by adding a specific event limiting the search target to the word pair and morpheme structure extracted from the first storage means, and storing a search target document group using the search key. Retrieving the number of occurrences of the search key by searching the storage means, the appearance number information acquisition means for storing the obtained number of occurrences in the word pair table in association with the word pair and the morpheme structure;
With reference to the word pair table, the structure of each morpheme is determined based on the degree of match between the number of occurrences of each word pair related to the structure of each morpheme and the number of occurrences of each word pair related to the structure of all morphemes. An evaluation means for outputting an evaluation value obtained by evaluating an association with the specific event;
A knowledge acquisition device comprising:
(Appendix 6)
The word pair and morpheme structure extracted from the first storage means storing the word pair table having the word pair having a specific relationship and the morpheme structure related to the sentence including the word pair are searched. A search key to which a specific event for limiting the target is added is created, the second storage means storing the document group to be searched is searched with the search key, and the number of appearances of the search key is obtained. Storing in the word pair table in association with the word pair and the morpheme structure;
With reference to the word pair table, based on the degree of coincidence between the number of occurrences of each morpheme structure related to each word pair and the number of appearances of each morpheme structure related to all word pairs, A process of outputting an evaluation value that evaluates the association with the specific event;
A knowledge acquisition method in which a computer is executed.
(Appendix 7)
The word pair and morpheme structure extracted from the first storage means storing the word pair table having the word pair having a specific relationship and the morpheme structure related to the sentence including the word pair are searched. A search key to which a specific event for limiting the target is added is created, the second storage means storing the document group to be searched is searched with the search key, and the number of appearances of the search key is obtained. Storing in the word pair table in association with the word pair and the morpheme structure;
With reference to the word pair table, based on the degree of coincidence between the number of occurrences of each morpheme structure related to each word pair and the number of appearances of each morpheme structure related to all word pairs, A process of outputting an evaluation value that evaluates the association with the specific event;
A program that causes a computer to execute.

1 Knowledge acquisition device 10 CPU
DESCRIPTION OF SYMBOLS 12 Drive apparatus 14 Storage medium 16 Auxiliary storage apparatus 18 Memory apparatus 20 Interface apparatus 22 Input apparatus 24 Output apparatus 30 Word pair acquisition part 32 Search part 34 Score calculation part 40 Word pair table 40A Word pair list 40B Context pattern list 42 Appearance number table 44 Word Pair with Score 50 Network 100 Client Computer 110 Operator 120 User 200 Large Document 250 Seed 260 Query

Claims (6)

A first storage means storing a word pair table having a pair of words having a specific relationship and a structure of a morpheme related to a sentence including the pair of words;
A second search key is created by adding a specific event limiting the search target to the word pair and morpheme structure extracted from the first storage means, and storing a search target document group using the search key. Retrieving the number of occurrences of the search key by searching the storage means, the appearance number information acquisition means for storing the obtained number of occurrences in the word pair table in association with the word pair and the morpheme structure;
With reference to the word pair table, based on the degree of coincidence between the number of occurrences of each morpheme structure related to each word pair and the number of appearances of each morpheme structure related to all word pairs, An evaluation means for outputting an evaluation value obtained by evaluating an association with the specific event;
A knowledge acquisition device comprising: The knowledge acquisition device according to claim 1,
The evaluation means multiplies the number of appearances of each morpheme related to each word pair by the number of appearances of each morpheme related to all word pairs, and outputs the value divided by the total number of appearances as an evaluation value. Means,
Knowledge acquisition device. The knowledge acquisition device according to claim 1 or 2,
Searching the second storage means using a given word pair or morpheme structure as a search key, adding the obtained search result to the search key and repeating the data search to obtain a plurality of Data collection means for storing the word pairs and the structure of a plurality of morphemes in the storage means,
Knowledge acquisition device. The knowledge acquisition device according to any one of claims 1 to 3,
The morpheme structure related to the sentence including the word pair is the morpheme structure obtained by removing the morpheme of the word pair from the morpheme structure that is the minimum partial structure obtained by performing dependency analysis on the target sentence of the data search. Structure,
Knowledge acquisition device. The word pair and morpheme structure extracted from the first storage means storing the word pair table having the word pair having a specific relationship and the morpheme structure related to the sentence including the word pair are searched. A search key to which a specific event for limiting the target is added is created, the second storage means storing the document group to be searched is searched with the search key, and the number of appearances of the search key is obtained. Storing in the word pair table in association with the word pair and the morpheme structure;
With reference to the word pair table, based on the degree of coincidence between the number of occurrences of each morpheme structure related to each word pair and the number of appearances of each morpheme structure related to all word pairs, A process of outputting an evaluation value that evaluates the association with the specific event;
A knowledge acquisition method in which a computer is executed. The word pair and morpheme structure extracted from the first storage means storing the word pair table having the word pair having a specific relationship and the morpheme structure related to the sentence including the word pair are searched. A search key to which a specific event for limiting the target is added is created, the second storage means storing the document group to be searched is searched with the search key, and the number of appearances of the search key is obtained. Storing in the word pair table in association with the word pair and the morpheme structure;
With reference to the word pair table, based on the degree of coincidence between the number of occurrences of each morpheme structure related to each word pair and the number of appearances of each morpheme structure related to all word pairs, A process of outputting an evaluation value that evaluates the association with the specific event;
A program that causes a computer to execute.

Images