JP5305971B2 - Abbreviation estimation apparatus and method - Google Patents

Description

  The present invention relates to an abbreviation estimation apparatus and method for estimating an abbreviation for an input word or set of words.

  There is a method in which after generating an abbreviation according to an abbreviation generation rule based on a character string of an original word (official name), the utterance probability of the generated abbreviation is calculated from the likelihood of the abbreviation generation rule (for example, see Patent Document 1). ). Also, it is determined whether the input word is an abbreviation or a formal name. If it is determined that the word is a formal name, an abbreviation is generated from the input word according to an abbreviation generation rule and used as a search word. On the other hand, if it is determined that the candidate is an abbreviation, the database is searched, and if there is an abbreviation / official name pair that matches the word, a search term is searched and a search term is interpolated to prevent a search omission (See, for example, Patent Document 2).

Japanese Patent No. 0324649 Japanese Patent Laid-Open No. 11-25117

As described above, conventionally, the original word is required to calculate the abbreviation of the word. However, abbreviations for words that can be converted in English or Katakana, such as Katakana, or words that have a formal name such as a TV program name, are complicated to generate. Has a problem that it is difficult to determine the abbreviation from the likelihood of the generation rule. For example, in the case of the abbreviation “M Smash” for the name “Music Smash”, the pronunciation “M Smash” of “M Smash” is not an abbreviation of the original word “Music Smash”, so it is necessary to consider the conversion of the notation, “Music Smash” Are generated by combining the notation conversion rule and the abbreviated conversion rule for each of the basic words “music” and “smash” divided at the division positions. At this time, the element word “M” obtained by converting the notation of “Music” to the English notation “Music” and abbreviated conversion and the original notation without converting the notation of the “smash” to the English notation “Smash” The word “M suma” is formed by concatenating the element word “suma”, and it is necessary to generate a combination of notation conversion rules and abbreviated conversion rules as appropriate, which is difficult to calculate.
Further, in a method that requires an original word to calculate the abbreviation of a word, the abbreviation of an abbreviation candidate cannot be calculated without the original word. For example, when only one word is input, such as a word that is clearly specified in parentheses in the text or a word that is input as a search character string, the abbreviation cannot be determined because there is no original word. There is a problem that generation rules and abbreviation estimation cannot be separated. Furthermore, if the rules for generating abbreviations are diversified, the probability is low for a rule to which the rule is rarely applied, and a lot of learning data for probability calculation is required. In the estimation method based on the probability for the abbreviation generation rule determined in advance, the generation rule is used when measuring the abbreviation likelihood. Therefore, the feature used to estimate the abbreviation likelihood is the abbreviation generation rule itself, and this is done manually in advance. Because it is necessary to prepare, it is not possible to learn the characteristics used for abbreviation estimation.

  On the other hand, abbreviation estimation methods that require a basic word database are difficult to deal with coined words and new words. In addition, in the method of abbreviation generation using the basic word database, an abbreviation is generated if a partial word of an official name before the abbreviation generation is registered in the database as a basic word, but the abbreviation of the generated abbreviation itself cannot be determined. .

  The present invention has been made to solve the above-described problems, and an abbreviation estimation device capable of estimating abbreviations based on features of abbreviation candidates and division positions into element words without depending on abbreviation generation rules, and It aims to provide a method.

  In order to solve the above-described problem, an abbreviation estimation apparatus according to the present invention includes a first acquisition unit that acquires a character string that can be an abbreviation candidate, and slicable position information that divides the character string, the character string, A calculation means for calculating a plurality of first features indicating the characteristics of the character string derived from the position information that can be divided from the character string; a second feature belonging to a desired category among categories that are indices indicating the concept of the characteristics; The storage means for storing the likelihood corresponding to the second feature, the plurality of first features and the second feature for each category are compared, and the likelihood corresponding to the same feature is obtained from the storage means. Second calculating means for calculating, and calculating means for calculating an abbreviation candidate rank, which is a rank indicating the correctness as an abbreviation for the formal name of the character string, from the likelihood acquired by the second acquiring means. It is characterized by.

  Also, the abbreviation estimation apparatus according to the present invention obtains a slicable position information for dividing the character string by dividing the character string between a first acquisition unit that acquires a character string that can be a candidate for the abbreviation, and the character string. Division means, calculation means for calculating a plurality of first features indicating the characteristics of the character string derived from the character string and the severable position information, and a category that is an index indicating the concept of the feature belongs to a desired category The storage means for storing the second feature and the likelihood corresponding to the second feature, the plurality of first features and the second feature for each category are compared, and the likelihood corresponding to the same feature is obtained. Second calculating means for acquiring from the storage means; and calculating means for calculating an abbreviation candidate rank that is a rank indicating the correctness of the character string as an abbreviation for the formal name from the likelihood acquired by the second acquiring means; , The features.

  In addition, the abbreviation estimation apparatus according to the present invention generates at least one abbreviation candidate that is a candidate for an abbreviation from the first acquisition unit that acquires a character string as an official name, and divides the abbreviation candidate. Generating means for generating at least one piece of position information that can be divided, calculation means for calculating a plurality of first features that indicate the abbreviation candidates and features of character strings derived from the pieces of position information of the abbreviation candidates; Storage means for storing a second feature belonging to a desired category among categories that are indices indicating the concept, likelihood corresponding to the second feature, the plurality of first features, and the second feature for each category And the second acquisition means for acquiring the likelihood corresponding to the same feature from the storage means, and the correctness as an abbreviation for the formal name of the abbreviation candidate from the likelihood acquired by the second acquisition means The A calculating means for calculating the abbreviation candidate ranking is to rank, characterized by comprising a.

  According to the abbreviation estimation apparatus and method of the present invention, it is possible to estimate the likelihood of abbreviations based on the characteristics of abbreviation candidates and the division positions into element words without depending on the abbreviation generation rules. For example, abbreviations that require katakana (alphabet) can be compared in the same way as abbreviations that do not use a notation conversion rule. In addition, even if abbreviation generation rules are complex or abbreviation candidates to which low-frequency rules are applied, estimation of abbreviations is unlikely to be affected, and abbreviation estimation can be performed accurately even with a relatively small amount of learning data. Is possible. Furthermore, even when the abbreviation generation rules are determined in advance, it is possible to learn features used for abbreviation estimation by estimating abbreviations from features different from those.

The block diagram which shows the abbreviation estimation apparatus which concerns on 1st Embodiment. The flowchart which shows operation | movement of the abbreviation estimation apparatus which concerns on 1st Embodiment. The figure which shows an example of the data stored in the characteristic storage part. The flowchart which shows operation | movement of a feature extraction part. The flowchart which shows operation | movement of a weight derivation | leading-out part. The flowchart which shows operation | movement of an abbreviation estimation calculation part. The block diagram which shows the 1st modification of the abbreviation estimation apparatus which concerns on 1st Embodiment. The block diagram which shows the 2nd modification of the abbreviation estimation apparatus which concerns on 1st Embodiment. The block diagram which shows the 3rd modification of the abbreviation estimation apparatus which concerns on 1st Embodiment. The block diagram which shows the 4th modification of the abbreviation estimation apparatus which concerns on 1st Embodiment.

Hereinafter, an abbreviation estimation apparatus and method according to embodiments of the present invention will be described in detail with reference to the drawings. In the following embodiments, the same reference numerals are assigned to the same numbered parts, and repeated description is omitted. Furthermore, in the following, basic words are words that are words as they are. An element word is a word that is divided at a position where a character string that is an official name corresponding to an abbreviation can be divided. The element word also includes a character string converted from the character string by applying one or more conversion rules. The original word is defined as the formal name that is the original word before being abbreviated to an abbreviation. Moreover, the position which can be divided | segmented points out the position judged to be divided | segmented by the method of dividing | segmenting a character string, and does not necessarily need to divide | segment into a word, You may divide between the characters of a character string. The conversion rule is, for example, a rule for converting a character string into Kanji, Katakana, Hiragana, or English.
(First embodiment)
The configuration of the abbreviation estimation apparatus according to this embodiment will be described in detail with reference to FIG.
The abbreviation estimation apparatus 100 according to the present embodiment includes an input unit 101, a feature extraction unit 102, a weight derivation unit 103, an abbreviation estimation calculation unit 104, an output unit 105, and a feature storage unit 106.
The feature extraction unit 102, the weight derivation unit 103, and the abbreviation estimation calculation unit 104 are collectively referred to as an abbreviation estimation unit 107.
The input unit 101 accepts input of a character string that can be one or more candidate abbreviations as a character string input from the outside. The abbreviation candidate indicates an abbreviation candidate that abbreviates the official name of the character string, and is hereinafter referred to as an abbreviation candidate. In addition, splittable position information for splitting the character string is added to the character string. The character string includes the case of one character. Here, the severable position information is a position in front of a character string unit indicating a group of meanings in which words used for abbreviations are extracted when an abbreviation is generated from an official name in an input character string, and The back position is shown. For example, it is assumed that the abbreviation “M Smart” is generated from the official name “Music Smash”. At that time, the character string unit indicating a group of meanings when “M”, which is a word used as an abbreviation, is extracted is “music”, and similarly, “suma”, which is a word used as an abbreviation, is used. A character string unit indicating a group of meanings at the time of extraction is “smash”. Information indicating the position between “music” and “smash”, the position before “music”, and the position after “smash”, which are character string units, is the position information that can be divided. Since the position after “Music” and the position before “Smash” are the same position, they are represented by either position. An element word is also defined as a character string unit divided by position information that can be divided.
Note that information that each element word in the abbreviation candidate was the beginning / end of the official name may be added to the position information that can be divided. This abbreviation candidate and the position information that can be divided are sent to the feature extraction unit 102.
The feature extraction unit 102 extracts the features of the abbreviation candidates from the abbreviation candidates and the splittable position information received from the input unit 101, and sends them to the weight derivation unit 103. Here, the feature is information unique to the character string derived from the character string and the position information where the character string can be divided. For example, information indicating that the character string includes one alphabetic character and two katakana characters (see FIG. 3) (corresponding to “1 English letter + 2 katakana”), and a specific example is shown in FIG. A plurality of features are generally grouped into a certain category. In other words, the category is an index indicating a common concept of a plurality of features, and is, for example, a combination of the number of characters and a character type (“C1: combination of character number + character type” in FIG. 3). However, the category may indicate only one feature, and specifically, “C5: number of elements” in FIG. 3 is an index indicating only one feature. The character type indicates the character type of the element word, and is, for example, an alphabetic character, hiragana or katakana. The feature extraction process will be described later in detail with reference to FIG. The operation of the feature extraction unit 102 will be described later in detail with reference to FIG.

The feature storage unit 106 stores features associated with character strings and their weights in association with each other. If necessary, a feature and a weight corresponding to the feature are sent to the search from the weight deriving unit 103. The data held in the feature storage unit 106 will be described in detail later with reference to FIG.
The weight deriving unit 103 first searches whether or not the feature extracted from the abbreviation candidate received from the feature extraction unit 102 is in the feature storage unit 106. If the feature corresponding to the feature storage unit 106 is stored among the stored features, the feature and the weight corresponding to the feature are sent to the abbreviation estimation calculation unit 104. In addition, when there is no corresponding feature among the stored features, a predetermined constant weight may be given as a feature not included in the stored feature. Alternatively, a weight may be given assuming that a certain category (for example, the number of characters) has a feature not included in the stored features. The operation of the weight deriving unit 103 will be described in detail later with reference to FIG.
The abbreviation estimation calculation unit 104 calculates the abbreviation likelihood of the abbreviation candidate from the feature received from the weight derivation unit 103 and its weight, and sends an abbreviation index to the output unit 105. The abbreviation is a rank indicating correctness as an abbreviation candidate, and is also referred to as an abbreviation candidate rank. In the calculation of abbreviations, for example, the abbreviation is estimated by calculating the sum of the weights. However, the function is not limited to the sum of powers, and any function that increases the value of abbreviations as the value based on the weight increases. You may calculate according to a function. For example, the feature weight may be positive or negative, and the abbreviation may be calculated by summation.
The output unit 105 outputs the abbreviation candidates and the abbreviations calculated by the abbreviation estimation calculation unit 104 in association with each other. When there are a plurality of abbreviation candidates, the abbreviation estimation unit 107 performs processing for any abbreviation candidate, and the output unit 105 associates each abbreviation candidate with the abbreviation calculated for each abbreviation candidate. Then output.

Next, details of the operation of the abbreviation estimation apparatus 100 according to the present embodiment will be described in detail with reference to the flowchart of FIG. Here, it is assumed that “M sum” is input as an example of an abbreviation candidate.
In S <b> 201, the character string “M summa” and its division position “/ M / summer /” are input in the input unit 101. However, here, “/” indicates the position where division is possible, and the element words are “M” and “smart”. Similarly, if the abbreviation candidate is “morning pita”, the division position “/ morning / pita /” of “morning pita” is input, and the element words are “morning” and “pita”.
In S202, the feature extraction unit 102 extracts a feature obtained from the abbreviation candidate and its divisionable position information. The feature is, for example, information such as “one alphabetic character + two katakana”, “three characters”, “two elements”.

  An example of data stored in the feature storage unit 106 is shown in FIG. In the table 1 in the feature storage unit 106, features belonging to a desired category and weights corresponding to the features are prepared in advance. The categories listed in FIG. 3 are combinations of the number of characters in the element word and the character type (C1: combination of the number of characters + character type), the combination of the character type of the element word (C2: combination of the character type), and the combination of the number of characters in the element word (C3 : Combination of the number of characters), the number of characters in the character string ("C4: number of characters"), and the number of element words ("C5: number of elements").

Also, the weights of features other than those listed in Table 1 may be held. This corresponds to features other than those listed in Table 1 of FIG. For example, in the category C4 of the number of characters, when there are only “3 characters” and “4 characters” in the table 1 in the feature storage unit 106, the other number of characters belongs to the “category C4 but in the table 1. Features such as “no characters” may be retained. Alternatively, in category C2 such as a combination of character types, when there are “Katakana” and “English characters” in table 1 in feature storage unit 106 and the input abbreviation candidate is “Hiragana”, table 1 in feature storage unit 106 Since “Hiragana” does not exist, the character type “character type belonging to category C2 but not in table 1” may be retained. In addition, a predetermined weight is held for any of the features in the table 1 and the features in the table 2.
The characteristics of the abbreviation candidate “/ M / smart /” are “1 letter + 2 katakana”, “English + katakana”, “3 letters” and “2 elements”. In addition, when the abbreviation candidate is “/ morning / pita /”, the characteristics are “1 kanji + 2 katakana” “kanji + katakana” “3 characters” “2 elements”.

Here, the details of the feature extraction processing in the feature extraction unit 102 will be described in detail with reference to the flowchart of FIG.
In S401, first, a character string S that is an abbreviation candidate is input. Here, it is assumed that the character string S is “M Smart”. Subsequently, in S402, the number of characters of the character string S is calculated. In the case of “M summa”, the number of characters is three. In S403, one of all element words E _i (i = 1 to N) constituting the character string S is extracted. Here, since “M” and “suma” are element words, the first element word “M” is extracted. Note that the order of extracting element words may be extracted randomly rather than from the top. That is, it is only necessary that all element words can be finally extracted. In S404, the number of characters of the element word is calculated for the element word extracted in S403. In the case of “M”, it can be seen that the number of characters is one. Subsequently, in S405, the character type of the element word is acquired. In the case of “M”, it is understood that the character type is an alphabetic character. In step S406, the processing in steps S403 to S405 is repeated to perform processing for all element words E _{i in} the character string S. In this example, “suma”, which is the last element word, is taken out, and the characteristics that the number of characters is two and the character type is katakana can be obtained.
Finally, in S407, a category string S obtained by the above processing, the number of characters, a combination of character types in each element word E _i constituting the string S, a combination of characters, a combination of characters and character type M features F _j (j = 1 to M, M is the number of features) are generated, and the feature extraction process in the feature extraction unit 102 is terminated. In the case of “/ M / smart /”, for example, it is possible to obtain the characteristics of “1 letter + 2 katakana”, “English + katakana”, “1 + 2 letters”, “3 letters”, and “2 elements”. Further, if “/ morning / pita /” is a character string S, characteristics such as “1 kanji + 2 katakana”, “kanji + katakana”, “1 +2 characters”, “3 characters”, “2 elements” are obtained. be able to. When the features included in each category in Tables 1 and 2 are all different, at most one character string feature is generated for each category. Therefore, a maximum number of character string features are generated for each category. Become.

Next, in S203, the weight derivation unit 103 searches the feature storage unit 106 for the same feature as the character string S. In S204, the weight derivation unit 103 also corresponds to each feature. Calculate the weight.
Here, the weight deriving process in the weight deriving unit 103 in S203 and S204 will be described in detail with reference to the flowchart of FIG.
First, in S501, M features _Fj of the character string S obtained by the feature extraction processing in the feature extraction unit 102 in FIG. 4 are input. Next, in S503, it is searched whether or not each inputted feature F _j is in the table 1 of the feature storage unit 106. In the condition determination of S504, wherein _{F j} is the flow proceeds to S505 if present table 1, to obtain the weight of its features. On the other hand, if the feature F _j does not exist in the table 1, the process proceeds to S506, the feature is changed to the feature of the table 2 of the same category, and the weight is acquired from the table 2 in S505. Then, in S507, and repeats the weight derivation process from S502 to S506 until the completion of all of the features _{F j.}
For example, for “/ M / smart /”, there are five characteristics F _{j of} “1 alphabet + 2 katakana” “English + katakana” “1 + 2” “3 characters” “2 elements” It is searched whether or not the feature F _j is in the table 1 of the feature storage unit 106. Since each feature F _{j of} “1 letter + 2 katakana”, “English + katakana”, “1 + 2 letters”, “3 letters” and “2 elements” exists in the table 1, it corresponds to these characteristics F _j The weights of “1 letter + 2 katakana” are 0.30, “English + katakana” is 0.20, “1 + 2 letters” is 0.21, “3 letters” is 0.25, “2 elements” is obtained from Table 1 as 0.40.
In addition, when “/ morning / pita /” is the input character string S, “1 character kanji + 2 characters katakana” “kanji + katakana” “1 character + 2 characters” “3 characters” “2 elements” Similarly, it is searched whether there are these features in the table 1 of the feature storage unit 106. In this case, the characteristics of “1 character + 2 characters”, “3 characters”, and “2 elements” exist in Table 1, but the characteristics of “1 character Kanji + 2 characters Katakana” and “Kanji + Katakana” do not exist in Table 1. Therefore, in S506, the characteristics are changed to the characteristics of the table 2 of the same category, and the “combination of the number of characters and character types belonging to the category C1 but not in the table 1” “the combination of character types belonging to the category C2 but not in the table 1” The weight is acquired from Table 2. Therefore, “a combination of the number of characters and character types belonging to category C1 but not in table 1” is 0.1, “a combination of character types belonging to category C2 but not in table 1” is 0.0005, “1 character + 2 characters” Is 0.21, “3 characters” is 0.25, and “2 elements” is 0.40.
Finally, in S508, and it outputs the acquired weights of M feature F _j of a string S by the above process. Thus, the weight derivation process in the weight derivation unit 103 ends.

After the processing of S204 and S205 described above is completed, in S205, abbreviation estimation processing is performed from the weight acquired in S204.
An example of the abbreviation estimation process in the abbreviation estimation calculation unit 104 will be described in detail with reference to the flowchart of FIG.
First, in S601, the weights for the M features _Fj of the character string S obtained by the weight derivation process of FIG. 5 are input. Next, in S602, the abbreviation of the character string S is calculated from the weight. As a method for calculating the abbreviation, for example, the abbreviation is obtained by taking the sum of weights. Finally, in S603, the abbreviation of the character string S obtained in S602 is output.
For example, for “M summa”, for example, the weights of five features F _j are 0.30, 0.20, 0.21, 0.25, and 0.40. “00126” is output as an abbreviation of “M-sum”. If it is “morning pita”, the weight of five features F _j , which is the sum of 0.1, 0.0005, 0.21, 0.25, and 0.40, 0.00000105 is “morning pita” Output as an abbreviation.

  In S206, the output unit 105 associates the abbreviation candidates with the abbreviations of the abbreviation candidates and outputs them to the outside. This is the end of the operation of the abbreviation estimation apparatus 100.

  According to the first embodiment described above, the characteristics of the input character string are extracted without using the abbreviation generation rules and the original language information, and the predetermined characteristics and the weights corresponding to the characteristics are obtained. The abbreviation can be estimated by referring to it. Also, abbreviations that require notation conversion rules (eg, Katakana ⇔ alphabet) in abbreviation generation rules can be compared in the same way as abbreviations that do not use notation conversion rules. Furthermore, even if abbreviation generation rules are complex or abbreviation candidates to which low-frequency rules are applied, estimation of abbreviation is unlikely to be affected, and abbreviation estimation can be performed accurately even with a relatively small amount of learning data. Is possible.

(First modification)
In the first embodiment, when the position information that can be divided is not added, the number of element words is unknown, so the feature of the number of elements cannot be extracted. However, in this modification, the abbreviation estimation can be performed by adding the severable position information to the abbreviation candidate that is the character string given to the input unit 101 when the severable position information is not added.
The configuration of the abbreviation estimation apparatus according to the first modification will be described in detail with reference to FIG.
The abbreviation estimation apparatus 300 according to the first modification includes a word division unit 701 after the input unit 101 in addition to the abbreviation estimation apparatus 100 according to the first embodiment. The word division unit 701 performs word division on the abbreviation candidates received from the input unit 101 and obtains at least one word division result. Thereafter, the word division result is added to the abbreviation candidate as dividable position information, and the abbreviation candidate and the dividable position information are associated with each other and sent to the feature extraction unit 102. The output unit 105 in this modification outputs the abbreviation candidates, the abbreviations, and the division position information in association with each other.
For example, if the abbreviation candidate input to the input unit 101 is “M summa”, the word segmentation unit 701 selects an arbitrary division position “/ M summa /” “/ M / summer /” “/ M sum / Divide by “/” to obtain the result of division. However, when the inputted abbreviation candidate is one character, it is not divided. Thereafter, the abbreviation estimation unit 107 performs processing for each of “/ M Smart /”, “/ M / Smart /”, and “/ M Smart / Ma /”. When “/ M Smart /” is input to the feature extraction unit 102, features F _j such as “3 characters” and “1 element” are extracted, and in the case of “/ M Smart / Ma /”, “2 characters + 1 character”. The feature F _j such as “2 elements” is extracted. Therefore, by performing the same processing as that of the first embodiment on each division result, the abbreviation likelihood in each of the three division positions for “M summa” is output from the output unit 105. When there are a plurality of input abbreviation candidate division results, all of them may be sent to the feature extraction unit 102, or one division result is selected from the plurality of division results and sent to the feature extraction unit 102. May be.

  According to the first modified example described above, even when the dividable position information is not added to the input abbreviation candidate, the dividable position information is generated by dividing the abbreviation candidate into words, and based on this As in the first embodiment, abbreviation estimation can be performed.

(Second modification)
In the first embodiment and the first modification described above, it is assumed that abbreviation candidates are input as input character strings. In this modification, when an official name (original word) is input, a plurality of abbreviation candidates that become abbreviations can be generated and the abbreviation corresponding to each can be output.
The configuration of the abbreviation estimation apparatus according to the second modification will be described in detail with reference to FIG.
The abbreviation estimation device 800 according to the second modification includes an abbreviation generation unit 801 after the input unit 101 in addition to the abbreviation estimation device 100 according to the first embodiment. The abbreviation generation unit 801 generates at least one abbreviation candidate for the formal name received from the input unit 101, and sends the generated abbreviation to the feature extraction unit 102. The output unit 105 in this modification outputs at least one abbreviation candidate and abbreviations corresponding to the abbreviation candidates.

  For example, it is assumed that the official name input to the input unit 101 is “music smash” and the division possible position is “/ music / smash /”. For this formal name, a plurality of abbreviation candidates generated by the abbreviation generation unit 801 and their divisionable positions are “/ Mu / Suma /”, “/ M / Suma /”, “/ Mu / S /”, “/ M / When obtained as “S /”, the abbreviation estimation unit 107 performs abbreviation estimation processing on each abbreviation candidate, and outputs from the output unit 105 abbreviation candidates and abbreviations corresponding to the respective abbreviation candidates. When a formal name (original language) is input, the division position of the abbreviation candidate may be determined from morphological analysis.

  According to the second modified example described above, even when a formal name (original word) is input as an input character string, a plurality of abbreviation candidates as abbreviations are generated and the abbreviation corresponding to each is output. And abbreviation estimation can be performed as in the first embodiment.

(Third Modification)
In the third modified example, when a plurality of abbreviation candidates are input, only those having high abbreviations can be selected and output from the plurality of abbreviation candidates. In addition, referring to external document data or the like It is possible to narrow down abbreviation candidates that are frequently used.
The configuration of the abbreviation estimation apparatus according to the third modification will be described in detail with reference to FIG.
The abbreviation estimation device 900 according to the third modification includes an abbreviation candidate selection unit 901 after the abbreviation estimation calculation unit 104 in addition to the abbreviation estimation device 100 according to the first embodiment. Furthermore, the abbreviation candidate selection unit 901 is connected to an external frequency calculation unit 902. The abbreviation candidate selection unit 901 selects abbreviation candidates that are more likely to be abbreviations from a plurality of abbreviation candidates obtained from the abbreviation estimation calculation unit 104, that is, abbreviated word candidates in descending order of abbreviation candidate ranking, and sends them to the output unit 105. Alternatively, frequency information in the document data is acquired from a frequency calculation unit 902 to be described later, abbreviation candidates are selected in the descending order of frequency information, and are sent to the output unit 105. The frequency information is information indicating the frequency of occurrence of abbreviation candidates in the document data. Finally, the output unit 105 in the present modification outputs abbreviation candidates and abbreviations corresponding to the abbreviation candidates.
The frequency calculation unit 902 measures the frequency with which abbreviation candidates are used. The frequency information may be the appearance frequency of external document data, or the number of character string search hits in an external database such as a web may be referred to as frequency information. The calculated frequency information is sent to the abbreviation candidate selection unit 901.

  As an example, in the abbreviation estimation calculation unit 104, abbreviations corresponding to abbreviation candidates are “/ Mu / Suma /”, “/ M / Suma /”, “/ Mu / S /”, “/ M / S /”. Assume that the likelihoods are calculated as 0.1, 0.2, 0.01, and 0.15, respectively. When the upper two are selected by the abbreviation candidate selection unit 901, “M summa” and “MS” are output. Alternatively, with respect to “M summa” and “MS”, the frequency calculation unit 902 selects “M summa” and “MS” having the highest number of appearances in the document (for example, the top 30). Furthermore, the character pattern “(*) Appearance” is added before and after, and the number of appearances in the abbreviation candidate document such as “Appearance of M Smartphone” or “Appearance of MS” is used, or “(*) is ( A character pattern such as “*)” is searched for a possible subject such as “M Smart is” or “MS is”, and an abbreviation candidate satisfying a condition (for example, the number of appearances is 1 or more) is selected.

  According to the third modification example described above, it is possible to select and output an abbreviation-like one from a plurality of abbreviation candidates, and further, the frequency in the external document data and the number of web hits Etc. can be referred to as frequency information, and abbreviation candidates that are used more frequently can be narrowed down, and the accuracy of abbreviations can be improved.

(Fourth modification)
In the embodiment and the modification described above, it is assumed that a predetermined value is used as the weight for the feature stored in the feature storage unit 106, but in the fourth modification, the feature storage unit 106 includes By updating the weight for the stored feature with the learning data, the accuracy of the weight for the feature can be improved.
The configuration of the abbreviation estimation apparatus according to the fourth modification will be described in detail with reference to FIG.
The abbreviation estimation apparatus 1000 according to the fourth modification includes a learning data input unit 1001, an abbreviation generation unit 801, a learning data feature extraction unit 1002, and a weight learning unit 1003 in addition to the abbreviation estimation apparatus 100 according to the first embodiment. Including. The learning data input unit 1001 to the weight learning unit 1003 are collectively referred to as a learning unit 1004 and connected to the feature storage unit 106. Moreover, since the abbreviation generation unit 801 performs the same operation as the abbreviation generation unit 801 shown in the second modification, description thereof is omitted here.
The learning data input unit 1001 receives an input of a known correct abbreviation and its formal name, and sends it to the abbreviation generation unit 801.
The learning data feature extraction unit 1002 extracts features from each abbreviation candidate generated by the abbreviation generation unit 801, and when there is a correct abbreviation corresponding to the official name input by the learning data input unit 1001. The characteristics of the abbreviation are stored in the characteristic storage unit 106. In addition, after storing all abbreviation candidates in the feature storage unit 106, information on all the abbreviation candidate features and which one is the correct abbreviation is sent to the weight learning unit 1003.

  The weight learning unit 1003 uses, as learning data, the features of each abbreviation candidate received from the learning data feature extraction unit 1002 and the known correct abbreviations, and the features included in the feature storage unit 106 and the features of the learning data Are matched, the weights for the features included in the feature storage unit 106 are learned (updated). For example, the weight of each feature is learned by Log-Linear Model.

  According to the fourth modified example described above, the feature for the feature stored in the feature storage unit 106 is learned (updated) by learning (updating) the feature of the abbreviation candidate using the known abbreviation and the formal name. It is possible to improve the accuracy of the weight for, and to learn the features used for abbreviation estimation by estimating abbreviations from different features even when abbreviation generation rules are determined in advance.

  Note that the present invention is not limited to the above-described embodiment as it is, and can be embodied by modifying the constituent elements without departing from the scope of the invention in the implementation stage. In addition, various inventions can be formed by appropriately combining a plurality of components disclosed in the embodiment. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, constituent elements over different embodiments may be appropriately combined.

100, 700, 800, 900, 1000, 1100, 1200 ... abbreviation estimation device, 101 ... input unit, 102 ... feature extraction unit, 103 ... weight derivation unit, 104 ... abbreviation estimation calculation 105, output unit, 106 ... feature storage unit, 107 ... abbreviation estimation unit, 701 ... word division unit, 801 ... abbreviation generation unit, 901 ... abbreviation candidate selection unit, 902 ... Frequency calculation unit, 1001 ... Learning data input unit, 1002 ... Learning data feature extraction unit, 1003 ... Weight learning unit, 1004 ... Learning unit.

Claims (12)

A first acquisition means for acquiring a character string that can be a candidate for an abbreviation and splittable position information for dividing the character string;
Calculating means for calculating a plurality of first characteristics indicating characteristics of the character string derived from the character string and position information of the character string that can be divided;
Storage means for storing a second feature belonging to a desired category among categories that are indices indicating the concept of the feature, and a likelihood corresponding to the second feature;
A second obtaining unit that compares the plurality of first features with the second feature for each category and obtains likelihood corresponding to the same feature from the storage unit;
An abbreviation estimation apparatus comprising: a calculation unit that calculates an abbreviation candidate rank that is a rank indicating a correctness as an abbreviation for an official name of the character string from the likelihood acquired by the second acquisition unit . First acquisition means for acquiring a character string that can be a candidate for an abbreviation;
Division means for dividing the character string between characters and obtaining splittable position information for dividing the character string;
Calculating means for calculating a plurality of first characteristics indicating characteristics of the character string derived from the character string and the splittable position information;
Storage means for storing a second feature belonging to a desired category among categories that are indices indicating the concept of the feature, and a likelihood corresponding to the second feature;
A second obtaining unit that compares the plurality of first features with the second feature for each category and obtains likelihood corresponding to the same feature from the storage unit;
An abbreviation estimation apparatus comprising: a calculation unit that calculates an abbreviation candidate rank that is a rank indicating a correctness as an abbreviation for an official name of the character string from the likelihood acquired by the second acquisition unit . First acquisition means for acquiring a character string as an official name;
Generating means for generating at least one abbreviation candidate that is an abbreviation candidate from the character string, and generating at least one splittable position information for dividing the abbreviation candidate;
Calculating means for calculating a plurality of first features indicating features of the character string derived from the abbreviation candidates and the position information of the abbreviation candidates that can be divided;
Storage means for storing a second feature belonging to a desired category among categories that are indices indicating the concept of the feature, and a likelihood corresponding to the second feature;
A second obtaining unit that compares the plurality of first features with the second feature for each category and obtains likelihood corresponding to the same feature from the storage unit;
An abbreviation estimation apparatus comprising: a calculation unit that calculates an abbreviation candidate rank that is a rank indicating the correctness of the abbreviation candidate as an abbreviation from the likelihood acquired by the second acquisition unit. .   The category includes the number of characters of the abbreviation candidate, the number of element words that are character string units divided by the position information that can be divided, the combination of the character type that is the character type of the element word and the number of characters of the element word, the element word The abbreviation estimation apparatus according to any one of claims 1 to 3, wherein the abbreviation estimation device is at least one of a combination of character types and a combination of the number of characters and the character type of the element word.   The abbreviation estimation apparatus according to any one of claims 1 to 4, further comprising a selection unit that selects a plurality of abbreviation candidates in descending order of the abbreviation candidate rank. A calculation means for calculating frequency information that is information indicating the frequency of occurrence of the abbreviation candidate from document data in an external document or in an external database;
The abbreviation estimation apparatus according to claim 5, wherein the selection unit selects the abbreviation candidates in descending order of the frequency information.   The abbreviation estimation apparatus according to any one of claims 1 to 6, wherein the likelihood is a weight indicating a degree of whether the feature is a correct element as an abbreviation of a formal name.   The second acquisition means compares the first feature and the second feature, and acquires a predetermined likelihood if there is no corresponding identical feature. The abbreviation estimation apparatus according to any one of the above. Obtaining means for obtaining a formal name and a correct abbreviation corresponding to the formal name;
Abbreviation candidate generating means for generating a plurality of abbreviation candidates from the official name;
When a plurality of abbreviation candidate features are extracted and the correct abbreviation feature corresponding to the official name is compared with the abbreviation candidate feature and there is an identical feature, the same feature is extracted and the storage means Extracting means for storing the same feature in
The characteristics of the abbreviation candidates are compared with the characteristics included in the storage means, the characteristics of the abbreviation candidates are rewritten to the characteristics included in the storage means, and which of the characteristics of the plurality of abbreviation candidates is correct? Learning means that updates the likelihood of the feature included in the storage means when the feature included in the storage means matches the feature of the learning data using the learning data that is data associated with The abbreviation estimation device according to any one of claims 1 to 8, further comprising: The first acquisition means acquires a character string that can be a candidate for an abbreviation and splittable position information for dividing the character string,
A first calculating means for calculating a plurality of first features indicating the characteristics of the character string derived from the character string and the severable position information of the character string;
Storage means stores a second characteristic which belongs to the desired category among the categories is an index showing the concept of features, and a likelihood corresponding to the second characteristic,
A second acquisition unit that compares the plurality of first features with the second feature for each category and acquires a likelihood corresponding to the same feature from the storage unit;
An abbreviation estimation method, wherein the second calculation means calculates an abbreviation candidate rank, which is a rank indicating the correctness as an abbreviation for the formal name of the character string, from the acquired likelihood. The first acquisition means acquires a character string that can be a candidate for an abbreviation,
The dividing means divides between the characters of the character string to obtain splittable position information for dividing the character string,
A first calculating means for calculating a plurality of first characteristics indicating characteristics of the character string derived from the character string and the splittable position information;
Storage means stores a second characteristic which belongs to the desired category among the categories is an index showing the concept of features, and a likelihood corresponding to the second characteristic,
A second acquisition unit that compares the plurality of first features with the second feature for each category and acquires a likelihood corresponding to the same feature from the storage unit;
An abbreviation estimation method, wherein the second calculation means calculates an abbreviation candidate rank, which is a rank indicating the correctness as an abbreviation for the formal name of the character string, from the acquired likelihood. The first acquisition means acquires the character string as a formal name,
Generating means generates at least one abbreviation candidate that is a candidate for an abbreviation from the character string, and generates at least one or more splittable position information for dividing the abbreviation candidate;
The first calculation means calculates a plurality of first features indicating the characteristics of the character string derived from the abbreviation candidates and the dividable position information of the abbreviation candidates;
Storage means stores a second characteristic which belongs to the desired category among the categories is an index showing the concept of features, and a likelihood corresponding to the second characteristic,
A second acquisition unit that compares the plurality of first features with the second feature for each category and acquires a likelihood corresponding to the same feature from the storage unit;
An abbreviation estimation method, wherein the second calculation means calculates an abbreviation candidate rank, which is a rank indicating the correctness of the abbreviation candidate as an abbreviation with respect to the formal name of the abbreviation candidate.

Images