JP2014191777A - Word meaning analysis device and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To perform ranking of a word having a plurality of meanings as to what meaning it is likely to be used according to a retrieval object.SOLUTION: A word meaning characterizing word extraction unit 11 of a word meaning analysis device 1 extracts a word meaning characterizing word which characterizes each word meaning out of word meaning description text data in which a plurality of word meanings are described about a word having a plurality of meanings. A related word extraction unit 12 extracts a related word of the word out of a collection of text data on the basis of a co-occurrence relation with the word having a plurality of meanings. A similarity calculation unit 13 calculates similarity between the word meaning characterizing word which is extracted by the word meaning characterizing word extraction unit 11 and the related word extracted by the related word extraction unit 12. A ranking processing unit 14 determines the order of likelihood to be used of the word meaning corresponding to each word meaning characterizing word on the basis of the similarity calculated by the similarity calculation unit 13.

Description

  The present invention relates to a semantic analysis device and a program.

  Conventionally, as an index of the meaning of an ambiguous word having a plurality of meanings that is easily used, information manually generated in a Japanese dictionary is used. In addition, WordNet, which is an English dictionary that classifies the meaning of words, is based on SemCor Corpus (http://www.gabormelli.com/RKB/SemCor_Corpus), where meanings are manually assigned to each word. In addition, the meaning of words is ranked (see Non-Patent Document 1).

"WordNet", [online], December 27, 2012, PRINCETON UNIVERSITY, [Search on March 14, 2013], Internet <URL: http://wordnet.princeton.edu/>

  Since it takes a lot of time to rank the ease of use of meaning by a person, it is difficult to generate and update a dictionary. Further, when using the meaning ranking result for the purpose of search, the ease of use of the meaning should be set for each search target, and it is very difficult to perform this setting work manually.

  The present invention has been made in consideration of such circumstances, and a semantic analysis device capable of ranking according to a search target what kind of meaning a word having a plurality of meanings is likely to be used, and Provide a program.

[1] One aspect of the present invention is a meaning-characterizing word extraction unit that extracts meaning-characterizing words that characterize each of the meanings from text description text data that describes a plurality of meanings of words having a plurality of meanings. , A related word extraction unit that extracts a related word of the word from a set of text data based on a co-occurrence relationship with the word, the semantic characterization word extracted by the semantic characterization word extraction unit, and the related word A similarity calculation unit that calculates the similarity with the related word extracted by the extraction unit, and a ranking in which the meaning corresponding to the meaning-characterizing word is likely to be used based on the similarity calculated by the similarity calculation unit. And a ranking processing unit for determination.
According to this invention, the meaning analysis device extracts from the meaning description text data the meaning-characterizing words that characterize each meaning of a word having a plurality of meanings, and creates a co-occurrence relationship with the word having a plurality of meanings. Based on the set of text data, related words are extracted. The semantic analysis device calculates the similarity between the extracted semantic characterization word and the related word, and based on the obtained similarity, ranks in which each semantic meaning described in the semantic description text data is easy to use are calculated. decide.
As a result, the semantic analysis apparatus can rank the meanings of words having a plurality of meanings. Moreover, the meaning analysis apparatus can determine the meaning ranking according to the search target by changing the text data from which the related word is extracted.

[2] One aspect of the present invention is the semantic analysis device described above, wherein the semantic characterization word extraction unit is a noun included in a final phrase of a definition sentence for each semantic meaning described in the semantic description text data. Is extracted as a word meaning characterization word.
According to the present invention, the semantic analysis device extracts a noun that becomes a semantically characterized word from the final clause of each semantic definition sentence described in the semantic description text data.
Thereby, the meaning analysis apparatus can extract a noun that well expresses the meaning as a meaning-characterizing word.

[3] One aspect of the present invention is the semantic analysis device described above, wherein the semantic characterization word extraction unit represents that a noun included in a final phrase of the definition sentence is one of a plurality. In the case of a specific word, a noun included in the phrase that modifies the final phrase is extracted as a meaning-characterizing word.
According to the present invention, the semantic analysis apparatus, when the final clause of the semantic definition sentence is a specific word representing one of a plurality of, for example, “one”, “one”, etc., Extract nouns that are semantically characterized words from the phrase that modifies the phrase.
Thereby, the meaning analysis apparatus can extract the noun used as a meaning-characterizing word accurately.

[4] One aspect of the present invention is the semantic analysis apparatus described above, wherein the set of text data is a set of text data to be searched based on the words.
According to this invention, the semantic analysis device extracts a related word of a word used as a keyword from a set of text data to be searched for keywords.
As a result, the semantic analysis apparatus can accurately rank the meaning of words having a plurality of meanings according to the search target.

[5] In one aspect of the present invention, a computer used as a semantic analysis apparatus is configured to use a semantically-characterized word that characterizes each of the semantic meanings from textual description text data in which a plurality of semantic meanings of words having a plurality of meanings are described. A semantically characterizing word extracting unit to extract, a related word extracting unit to extract a related word of the word from a set of text data based on a co-occurrence relationship with the word, and the semantic feature extracted by the semantically characterizing word extracting unit A similarity calculation unit for calculating the similarity between the attached word and the related word extracted by the related word extraction unit, and the meaning corresponding to the word meaning characterization word based on the similarity calculated by the similarity calculation unit Is a program for functioning as a ranking processing unit that determines the ranking in which the user is likely to use.

  According to the present invention, it is possible to rank the meaning of words having a plurality of meanings according to the search target.

It is a functional block diagram which shows the structure of the meaning analysis apparatus in one Embodiment of this invention. It is a figure which shows the example of the meaning description text in the same embodiment. It is a figure which shows the example of the base text set in the same embodiment. It is a flowchart which shows the process sequence of the meaning analysis apparatus in the embodiment. It is a figure which shows the example of the meaning characterization word in the same embodiment. It is a figure which shows the example of the mutual information amount of the related word with respect to the ranking preparation object word in the same embodiment. It is a figure which shows the example of the distribution similarity between the related word and the meaning characterization word in the embodiment. It is a figure which shows the example of the ranking result in the same embodiment.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
FIG. 1 is a functional block diagram showing a configuration of a meaning analysis apparatus 1 according to an embodiment of the present invention. The meaning analysis apparatus 1 is realized by one or a plurality of computer devices, and as shown in the figure, a storage section 10, a meaning characterization word extraction section 11, a related word extraction section 12, a similarity calculation section 13, and A ranking processing unit 14 is provided.

  The storage unit 10 stores various data used for processing of each unit. Specifically, the storage unit 10 stores a meaning-descriptive text and a base text set. The meaning description text is text data (meaning description text data) in which a plurality of meanings for ambiguous words having a plurality of meanings are described. The base text set is a set of text data serving as a base for extracting words related to the word for which the semantic ranking is to be created.

  The word meaning characterization word extraction unit 11 extracts a word characterizing the word meaning of the word for which the word meaning ranking is created from the word description text stored in the storage unit 10. In the following, a word that is a creation target of a meaning ranking is described as a “ranking creation target word”, and a word that characterizes the meaning is described as a “meaning characterization word”. The related word extraction unit 12 extracts words related to the ranking creation target word from the base text set stored in the storage unit 10. As the base text set, for example, a set of text data to be searched for using a ranking creation target word as a keyword is used. Hereinafter, a word related to the ranking creation target word is referred to as a “related word”. The similarity calculation unit 13 calculates the similarity between the word meaning characterization word extracted by the word meaning characterization word extraction unit 11 and the related word extracted by the related word extraction unit 12. Based on the similarity calculated by the similarity calculation unit 13, the ranking processing unit 14 determines an order in which the meaning corresponding to each meaning-characterizing word is likely to be used. Thereby, the ranking processing unit 14 generates a ranking of which meanings described in the meaning description text are likely to be used.

Next, data used in the meaning analysis apparatus 1 will be described.
FIG. 2 is a diagram illustrating an example of the meaning description text. As this meaning text, for example, an ambiguity avoidance page of Wikipedia (http://en.wikipedia.org/), which is an existing encyclopedia service provided on the Internet, can be used. For example, dictionary data such as a Japanese dictionary can be used as the meaning description text. In the example of the meaning description text shown in the figure, a sentence defining a plurality of meanings is described for the ranking creation target word “lightning”.

  FIG. 3 is a diagram illustrating an example of a base text set. In the figure, an example in which a program EPG (Electronic Program Guide) is used as a base text set is shown. The program EPG shown in the figure includes, for a plurality of programs, an identifier (Id) for identifying the program, a program name (Title), a short display of the program name (Short Title), a program description (Description), and a program content (Detail). ) Information is included.

Next, the operation of the meaning analysis apparatus 1 will be described.
FIG. 4 is a flowchart showing an operation procedure of the semantic analysis apparatus 1 shown in FIG.

[Step S1: Word Characterizing Characterizing Word Extraction Process]
The word meaning characterization word extraction unit 11 reads out the word meaning description text of the ranking creation target word from the word meaning description text published on the Internet, for example, and writes it in the storage unit 10. Alternatively, the word meaning characterization word extraction unit 11 may read the word meaning description text of the ranking creation target word from the word meaning description text such as the Japanese dictionary stored in the storage unit 10 in advance. The meaning-characterizing word extraction unit 11 extracts a meaning-characterizing word that characterizes each meaning from the meaning-description text in which a plurality of meanings for the ranking creation target word are described.

  Specifically, the word meaning characterization word extraction unit 11 reads and parses the first definition sentence defining each meaning from the meaning description text in which the meaning of the ranking creation target word is described, and analyzes the definition sentence. Let the last phrase be the target phrase from which the meaning-characterizing words are extracted. Hereinafter, the phrase from which the meaning-characterizing word is extracted is referred to as a “word extraction target phrase”. The word meaning characterization word extraction unit 11 extracts nouns in the word extraction target clause.

  However, in the case of a specific word indicating that the final clause is one of a plurality of “one”, “one type”, etc., the semantic characterization word extraction unit 11 modifies the final clause with the “case” ”And the phrase closest to the final phrase as a word extraction target phrase, nouns are extracted. The specific word is stored in the storage unit 10 in advance. For example, in the meaning description text of the ranking creation target word “lightning” shown in FIG. 2, the final sentence of the definition sentence “one of game terms” is “one”. Therefore, the word meaning characterization word extraction unit 11 extracts the noun “game term” using the phrase “game term” that modifies the last phrase “one” as the word extraction target clause.

  Further, if the noun extracted from the word extraction target clause includes an unnecessary suffix, the meaning-characterizing word extraction unit 11 deletes the suffix. For example, the word meaning characterization word extraction unit 11 deletes the suffix “version” from “cartoon version” and sets it as “cartoon”. Note that unnecessary suffixes are stored in the storage unit 10 in advance. In addition, the word meaning characterization word extraction unit 11 extracts a plurality of nouns by extracting a noun as a word extraction target clause from the parallel clause when a phrase having a parallel relationship with the word extraction target clause exists in the definition sentence. Forgive. In the case of the definition sentence “One of Natural Phenomena / Meteorology” shown in the example of the meaning-description text shown in FIG. 2, the phrase “Meteorological” that qualifies the last phrase “one” and is closest to the last phrase becomes the word extraction target phrase. In addition, the phrase “natural phenomenon” in parallel with this phrase is also a word extraction target phrase. Thereby, the word meaning characterization word extraction unit 11 extracts the noun “natural phenomenon” and the noun “weather” from each word extraction target clause.

When extracting a noun from a word extraction target clause, the word meaning characterization word extraction unit 11 deletes morphemes (minimum semantic units) constituting the extracted noun in order from the top in order to extract as many general nouns as possible. The process of determining whether or not it is a general noun is performed. Therefore, for example, the top 1 million nouns frequently appearing on the web are stored in the storage unit 10 in advance as frequent nouns. This may be, for example, acquiring and storing frequently used noun data published on the Internet, or storing frequently nouns selected based on the number of hits when each noun is searched on the Internet. But you can. The word meaning characterization word extraction unit 11 deletes the morphemes constituting the noun extracted from the word extraction target phrase in order from the top until the frequent noun stored in the storage unit 10 matches the frequent noun. For example, in the case of the fourth definition sentence “Japanese male idol group” in the meaning description text shown in FIG. 2, the semantic characterization word extraction unit 11 extracts the noun “male idol group” with the final phrase as a word extraction target phrase. To do. The meaning-characterizing word extraction unit 11 divides the extracted noun into “male / idol / group” by morphological analysis processing. First, the word meaning characterization word extraction unit 11 determines whether the “male idol group” is a general noun. The meaning-characterizing word extraction unit 11 determines that the “male idol group” is not a common noun, and therefore deletes the first morpheme “male” so that the “idol group” Judge whether it is a proper noun. The meaning-characterizing word extraction unit 11 determines that the “idle group” is a common noun because the “idle group” is included in the frequent nouns, and extracts the “idle group” from the fourth definition sentence.
The word meaning characterization word extraction unit 11 outputs the noun extracted by the above processing to the similarity calculation unit 13 as a word meaning characterization word.

  FIG. 5 is a diagram illustrating an example of the meaning-characterizing words extracted by the meaning-characterizing word extraction unit 11 by the above processing. In the same figure, the meaning characterization word extraction part 11 has shown the meaning characterization word extracted from each definition sentence of the meaning description text shown in FIG.

[Step S2: Related Word Extraction Process]
Next, the related word extraction unit 12 extracts a related noun of the ranking creation target word from the base text set (step S2). In this process, a set of text data to be searched using a ranking creation target word as a keyword can be used as a base text set. For example, when searching for a television program, text data such as a program EPG is used as a base text set. In the present embodiment, the program EPG shown in FIG. 3 is used as a base text set.

  The related word extraction unit 12 extracts a sentence describing the program contents (for example, a sentence described in Detail) from the program EPG stored in the storage unit 10, performs morphological analysis, and extracts a noun. At this time, the related word extraction unit 12 extracts only common nouns for each phrase. The determination of whether or not it is a general noun is made based on the match with the frequent noun stored in the storage unit 10 as in step S1.

  Next, the related word extraction unit 12 evaluates the relationship between the extracted noun and the ranking creation target word. For the evaluation of this relevance, for example, a conventional index of mutual information can be used. The mutual information MI (A, B) for the word A and the word B is defined by the following equation (1). However, the word A is a ranking creation target word and the word B is a related word. The related word is a noun that co-occurs with the ranking target word in the program content description sentence.

  In equation (1), P (A, B) is the probability value that word A and word B appear in the description text of the program content (Detail) of the same program, and P (A) is the description of the program content of all the programs. The probability value that the word A appears in the sentence, and P (B) indicates the probability value that the word B appears in the description sentences of the program contents of all programs. It can be said that the larger the value of the mutual information MI (A, B), the deeper the relationship between the word A and the word B. The related word extraction unit 12 calculates the mutual information amount of each related word (word B) with respect to the ranking creation target word (word A) using Expression (1).

  FIG. 6 is a diagram illustrating an example of the mutual information amount of each related word with respect to the ranking creation target word. This figure shows the mutual information about the related nouns that co-occur with the ranking creation target word “lightning” in the program content description sentence of the program EPG shown in FIG. The related word extraction unit 12 outputs each related word and the mutual information amount calculated for the related words to the similarity calculation unit 13.

[Step S3: Similarity Calculation Processing]
Next, the similarity calculation unit 13 obtains a similarity between the meaning-characterizing word extracted in step S1 and the related word extracted in step S2. In this embodiment, an index such as distribution similarity is used as the similarity. In the distribution similarity, each word is clustered based on the dependency relationship of words in the actual text, etc., and the distribution of the probability of belonging to the class of each word is obtained from the result of the clustering, and the distance between the probability distributions is calculated. Calculate the similarity between words. Regarding distribution similarity, for example, reference "Kasama, De Saeger, Torizawa, Murata," Creation of large-scale similar word list using dependency stochastic clustering, "Language Processing Society 15th Annual Conference Announcement" Papers, C1-6, pp.84-87. (2009)) ”. The similarity calculation unit 13 targets each related word up to the top 100 of the mutual information among the related words input from the related word extraction unit 12. The distribution similarity with the attached word is calculated using the description of the base text set.

  FIG. 7 shows an example of the calculation result of the distribution similarity between each related word and each meaning-characterizing word. In the same figure, the calculation result of the distribution similarity of each meaning-characterizing word shown in FIG. 5 and the related word whose mutual information amount is included in the top 100 among the related words shown in FIG. 6 is shown. The similarity calculation unit 13 outputs the calculation result of the distribution similarity between each related word and each meaning-characterizing word to the ranking processing unit 14.

  In the above description, the distribution similarity is used as the similarity, but another index value indicating the similarity between words as a quantitative value may be used. For example, the distance between words in the thesaurus can be used as the similarity.

[Step S4: Ranking process]
The ranking processing unit 14 ranks each meaning using the distribution similarity calculated in step S3. Accordingly, when the meaning of the word for which ranking is to be created is Sem, the ranking processing unit 14 calculates Weight (Sem), which is the weight of each meaning Sem, using the following equation (2).

In equation (2), “noun (Sem)” is the noun extracted in step S1 as the meaning characterization word of the meaning Sem. Further, D (noun (Sem)) indicates the number of nouns extracted as meaning-characterizing words from the meaning Sem. Dsim (t, es _i ) indicates the distribution similarity between the word t and the word es _i , the word t is a related word extracted in step S2, and the word es _i is a word-characterizing characterization word from the word meaning Sem The i-th noun noun (Sem) extracted as (i is an integer not less than 1 and not more than D (noun (Sem))). For example, as shown in FIG. 5, when the meaning Sem is “one of natural phenomena / weather. Lightning bolt”, noun (Sem) is “natural phenomenon” and “weather”, and D (noun (Sem)) is “2”, the word es ₁ is “natural phenomenon”, and the word es ₂ is “weather”.
The ranking processing unit 14 generates data indicating the result of ranking the meanings in descending order of the calculated Weight (Sem).

  FIG. 8 is a diagram illustrating a ranking result for the meaning of the ranking creation target word “lightning”. The ranking processing unit 14 obtains the ranking (ranking) of the meaning Sem as a ranking result, the noun noun (Sem) extracted as the meaning characterization word from the meaning Sem, the calculated weight Semit weight (Sem), and Is generated as a ranking result. The ranking processing unit 14 writes the generated ranking result data into the storage unit 10 or outputs the data to a display device or another computer device.

When searching for a program that the user is interested in on a website on the Internet that distributes the program on demand, for example, the program EPG is used as a search target. Therefore, as in the embodiment described above, the meaning ranking is created using the program EPG as the base text set. Then, it is possible to improve the accuracy of the program search by grasping the meaning of the keyword used by the user for searching the program from the ranking result and using the grasped meaning for the search of the program EPG. It becomes possible.
For example, when a set of news texts is used as the base text set, it is expected that the meaning of “natural phenomenon, weather” as the word characterizing meaning will be higher in the ranking.
In this way, by changing the base text set for extracting related words, it is possible to obtain the meaning ranking result depending on the search target.

  As described above, according to the semantic analysis device 1 of the present embodiment, how is an ambiguous word having a plurality of meanings using a large text set and without manual intervention? It can be estimated whether it is easy to use in a certain sense. Furthermore, according to the semantic analysis device 1 of the present embodiment, the large-scale text set used as the base text set is set as a search target document set or a document set of the same or similar category as the search target document set. Thus, it is possible to obtain a ranking of ease of use of meaning for each search target.

  The above semantic analysis device 1 has a computer system therein. The operation process of the meaning analysis apparatus 1 is stored in a computer-readable recording medium in the form of a program, and the above processing is performed by the computer system reading and executing this program. The computer system here includes a CPU, various memories, an OS, and hardware such as peripheral devices.

Further, the “computer system” includes a homepage providing environment (or display environment) if a WWW system is used.
The “computer-readable recording medium” refers to a storage device such as a flexible medium, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Furthermore, the “computer-readable recording medium” dynamically holds a program for a short time like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, a volatile memory in a computer system serving as a server or a client in that case, and a program that holds a program for a certain period of time are also included. The program may be a program for realizing a part of the functions described above, and may be a program capable of realizing the functions described above in combination with a program already recorded in a computer system.

DESCRIPTION OF SYMBOLS 1 Word meaning analysis apparatus 10 Memory | storage part 11 Word meaning characterization word extraction part 12 Related word extraction part 13 Similarity calculation part 14 Ranking processing part

Claims (5)

A meaning-characterizing word extraction unit that extracts a meaning-characterizing word that characterizes each of the meanings from the meaning-description text data in which a plurality of meanings of words having a plurality of meanings are described;
A related word extraction unit that extracts a related word of the word from a set of text data based on a co-occurrence relationship with the word;
A similarity calculation unit that calculates a similarity between the meaning characterization word extracted by the meaning characterization word extraction unit and the related word extracted by the related word extraction unit;
A ranking processing unit for determining a ranking in which the meaning corresponding to the meaning-characterizing word is likely to be used based on the similarity calculated by the similarity calculating unit;
A semantic analysis device comprising: The meaning characterization word extraction unit extracts a noun included in a final sentence of a definition sentence of each meaning described in the meaning description text data as a meaning characterization word;
The semantic analysis apparatus according to claim 1, wherein: The meaning-characterizing word extraction unit, when the noun included in the final clause of the definition sentence is a specific word indicating that the noun is one of a plurality, the noun included in the clause that modifies the final phrase Are extracted as meaning-characterizing words,
The semantic analysis apparatus according to claim 2, wherein: The set of text data is a set of text data to be searched based on the word.
The semantic analysis device according to any one of claims 1 to 3, wherein the semantic analysis device is provided. A computer used as a semantic analysis device
A meaning-characterizing word extraction unit that extracts a meaning-characterizing word that characterizes each of the meanings from the meaning-description text data in which a plurality of meanings of words having a plurality of meanings are described;
A related word extraction unit that extracts a related word of the word from a set of text data based on a co-occurrence relationship with the word;
A similarity calculation unit for calculating a similarity between the meaning characterization word extracted by the meaning characterization word extraction unit and the related word extracted by the related word extraction unit;
A ranking processing unit that determines a ranking in which the meaning corresponding to the meaning-characterizing word is likely to be used based on the similarity calculated by the similarity calculation unit;
Program to function as.

Images