JPH0785041A - Relational concept extraction device - Google Patents

Abstract

(57) [Summary] [Purpose] To automatically collect and create the semantic descriptions of a dictionary used for natural language processing, to extract two related words from the input sentence together with the relationship. [Structure] A morpheme analysis means 1 for performing morpheme analysis that divides an input sentence into morphemes, and a pair of two words that can be related by a relationship established between words from the morpheme analysis result obtained by this morpheme analysis means 1 are extracted. And a related concept extracting means 2 for

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a word extracting device for creating a dictionary, and in particular, in order to automatically collect and create the semantic description of a dictionary used for natural language processing, it is related from input sentences. The present invention relates to a relational concept extraction device that extracts two words together with their relations.

[0002]

2. Description of the Related Art When performing natural language processing for describing the meaning of a natural language, in order to perform highly accurate processing for a wide range of inputs, a semantic description corresponding to the wide range of inputs is required. It is necessary to have a large and accurate dictionary.

Particularly in the case of technical documents and the like, in order to register new words that are born every day in a dictionary, it is difficult for humans to select words and describe their meanings in the conventional manner, in order to cope with a large amount of vocabulary and its increase. It is impossible to follow. Therefore, there is a strong demand for the development of a device that automatically performs such work.

[0004]

Conventionally, there has been proposed a device for automatically selecting and extracting a keyword, a technical term, a proper noun or the like from an input sentence for such a purpose (Japanese Patent Laid-Open No. 10-102638). Japanese Patent Laid-Open No. 1219/1989
28, JP-A-3-286372).

However, in these devices, the emphasis is placed on the recognition of the target word and only the words can be extracted one by one, so that only the portion corresponding to the entry word in the dictionary can be created. . For example, when there is a sentence "auxiliary storage medium such as floppy disk ...", the word "floppy disk" can be extracted, but "floppy disk" and "auxiliary storage medium" can be extracted in association with each other. I didn't have it, so I could extract the part corresponding to the entry word "floppy disk",
The semantic information that "floppy disk" is "auxiliary storage medium" could not be extracted. Therefore, in such a device, it is impossible to automatically create a dictionary including semantic information.

An object of the present invention is to automatically recognize a relationship not only with a word but with another word in order to solve the problem of automatically creating a dictionary containing such semantic information, and An object of the present invention is to provide a relational concept extraction device that can obtain semantic information to be registered in a dictionary by extracting it as a pair of attached words.

[0007]

To achieve the above object, in the present invention, as shown in FIG. 1 (A), a morphological analysis unit 1, a syntax analysis unit 2, a context analysis unit 3, and a relational concept extraction unit 4 are provided. Set up.

For example, when a sentence "using an auxiliary storage medium such as a floppy disk ..." Is input as an input sentence, the morphological analysis unit 1 uses it as a noun, case particle, adverb, verb, auxiliary verb. , Etc., the words are divided into morphemes, and morpheme analysis is performed.

The relational concept extraction unit 4 holds a relational concept extraction rule as shown in FIG. 4A, and detects whether or not the morphological analysis result transmitted from the morphological analysis unit 1 corresponds to these patterns. To do. In the above example, it is recognized as corresponding to the pattern of "Y such as X" because it is "auxiliary storage medium such as floppy disk", and as a result,
"Y s" has the meaning "X is a subordinate concept of Y"
In the output result "ubc X", the words in the sentence are respectively assigned to X and Y, and "auxiliary storage medium subc
The related concept of "floppy disk" is extracted.

There are types of such relational concepts, an example of which is shown in FIG. For example, "X comp Y" means "device X".
One of the elements (modules) that make up the device is the device Y. "
Means "X t-elem Y" means "technology Y is one of the elemental technologies for realizing technology X", and "X func Y" means "when device X is moved, The process Y is performed as one of the functions ".

By the way, there are cases where the relational concept cannot be accurately extracted only by morphological analysis. For example, instead of looking at "the floppy disk magnetically stores the data transmitted from the computer body", judge that "the floppy disk is transmitted from the computer body" and extract the wrong relational concept. .

In order to prevent such an error, the parsing unit 2
The syntactic analysis identifies the correct dependency relationship "data is magnetically stored on the floppy disk." And sends it to the relation concept extraction unit 4 to accurately determine the pattern. Extract.

The present invention also relates to an auxiliary storage medium.
Especially, it relates to a floppy disk. It is necessary to extract the relational concept that "the floppy disk is an auxiliary storage medium" from two sentences such as ". Is this necessary for the context analysis unit 3 to analyze a plurality of adjacent sentences? Judge whether or not.

[0014]

According to the present invention, instead of a heading such as "floppy disk", "auxiliary storage medium such as floppy disk", "function of floppy disk stores data", "floppy disk is driven by floppy disk drive" It is possible to extract a relationship that holds between two words such as "...". Therefore, by collecting these for a certain period of time, it is possible to collect the definition information of what the floppy disk is like, and thus it is possible to facilitate the creation of the dictionary.

[0015]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described with reference to FIGS. 2 is a block diagram of an embodiment of the present invention, FIG. 3 is a sentence sequence,
4 is an explanatory diagram of a pattern rule example, FIG. 5 is an explanatory diagram of a relation concept extraction state, and FIG. 6 is another example of a rule applied to morphological analysis / syntactic analysis.

In FIG. 2, the same symbols as in FIG. 1 indicate the same parts, 1 is a morpheme analysis part, 2 is a syntactic analysis part, 3 is a context analysis part, 4 is a relational concept extraction part, 5 is a word dictionary, and 6 is an identification. It is a dictionary.

The morphological analysis unit 1 divides an input sentence consisting of a plurality of input sentences into sentences, and collates each input sentence with a dictionary 5 in which words are stored, as shown in FIG. 3 (A). It is divided into morphemes using a morpheme rule of dividing into parts of speech and dividing into words.

The syntax analysis unit 2 constructs a syntax based on the morpheme analysis result for each input sentence sent from the morpheme analysis unit 1, that is, the analysis result of the part of speech which is divided into word units and constitutes each word. The parsing process is performed to analyze the structure of the input sentence using the parsing rules.

For example, as shown in FIG. 3A, "the floppy disk stores data transmitted from the computer main body."
3 is input, the morpheme analysis unit 1 refers to the word dictionary 5 and performs a morpheme analysis to analyze the sentence into words and parts of speech as shown in FIG. Then, these are transmitted to the syntax analysis unit 2.

The syntactic analysis unit 2 applies the result to the syntactic analysis rule to perform syntactic analysis, and "transmits data from computer""storesdata""storesdata""stores in floppy disk". The dependency concept extracting unit 4 recognizes the dependency relationship
Output to.

The context analysis unit 3 receives the morpheme analysis result from the morphological element analysis unit 1, the syntactic analysis result and the input sentence from the syntactic analysis unit 2, and performs context analysis of the input sentence. For example, as shown in FIG. 3B, "the present invention relates to a floppy disk" is input as the first sentence (preamble) and "particularly for word processing" is input as the second sentence (postscript). When this is done, it recognizes that there is a relational concept that spans these two sentences, and notifies the relational concept extraction unit 4 that these two sentences should be processed collectively.

Of course, the target pattern of the context analysis is not limited to this. For example, as shown in FIG. 4B, "... refers to X." is input as the first sentence (preamble) and the second sentence ( 2) in the case of “If you do P ...” as the latter sentence)
It is determined that the two sentences are connected, and the relational concept extraction unit 4 is notified that the two sentence sets should be processed together.

In a typical input sentence, it is efficient to select and switch the pattern rule for extracting the relationship according to the type of the input sentence. For example, in the case of patent publications, "industrial application field" and "problems of conventional technology"
Since the formal characteristics of the sentence forming the item differ depending on the item such as, the execution efficiency and the extraction accuracy can be improved by switching the pattern rule group for each item.

The relational concept extraction unit 4 includes the morpheme analysis unit 1
The relational expression pattern prepared in advance is applied to the information of the morpheme analysis result in, the syntactic analysis result in the syntactic analysis unit 2, and the context analysis result in the context analysis unit 3, and the designation is made in advance by referring to the identification dictionary 6. And a pair of two words related by the relationship.

Extraction of relational expression pattern according to FIG.
The action will be described. (1) Enter "M₁Part improved M₂The putter
When the existence of a user is recognized by referring to the identification dictionary 6,
Looking at the restrictions entered in the identification dictionary 6, M ₁, M₂
Is a device type, and if these are device types,
M₁, M₂And cut out (M₂comp M₁)
It is stored and stored in a file (not shown). Device type here
Is, for example, in "PC with improved input section"
Indicates something like "input section" or "personal computer".
The “upper” and “thing” in the “improved top”
Do not distinguish from device type.

(2) The input sentence contains "M ₁ {includes | has |
When the existence of the pattern "equipped | equipped | equipped | built-in | held | having} M ₂ " is recognized, it is determined whether or not M ₁ and M ₂ are device types, and when these are device types, (M ₂ comp
Extract M ₁ ) and retain. Here, {A | B | ...
.} Represents that any one of the elements A, B ... Is included.

(3) In the input sentence, "M ₂ built in M ₁ "
When the existence of such a pattern is recognized, it is determined whether or not M ₁ and M ₂ are device types, and when these are device types, (M ₁ co
mpM ₂ ) is extracted and retained.

(4) When the existence of the pattern "M ₁ and its M ₂ " is recognized in the input sentence and it is determined that M ₁ and M ₂ are device types, (M ₁ comp M ₂ ) is extracted. To do. (5) Input text "Currency changer with automatic bill recognition function"
Such as "F {function | means} {has / has}
The existence of the pattern "M" is recognized, and F is a functional type, M
Is a device type, (M func
F) is extracted.

(6) In the input sentence, "F {perform |
When the existence of the pattern [{to allow | for}} M] is recognized and F is determined to be a functional type and M is a device type, (M func F) is extracted. Here, [A] indicates that the element A may or may not be present.

(7) When the existence of the pattern "F [ga] {able | possible} M" is recognized in the input sentence and it is determined that F is a functional type and M is a device type (M func
F) is extracted.

(8) The existence of a pattern "P ₁ {to | corresponds to P} P ₂ " is recognized in the input sentence, and P ₁ and P
_When it is determined that ₂ is a processing type that indicates input / output or calculation (P ₁ t-elem P ₂ ), it is extracted.

(9) In the input sentence, "for X {to | to}}
When the existence of the pattern "P" is recognized and P is a processing type, (P obj X) is extracted regardless of X.

(10) "Y of X {etc. Etc." in the input sentence
If the existence of the pattern is recognized, and X and Y belong to the same type, (Y subc X) is extracted. The relational concept extraction unit 4 holds the result of such extraction in a file (not shown).

The word dictionary 5 is filled with words, parts of speech, etc. necessary for the morphological analysis unit 1 to perform morphological analysis on the input sentence. The identification dictionary 6 stores relational expression patterns, constraints, etc., such as those shown in FIG. 4A, which the relational concept extraction unit 4 needs for actions. In addition, examples of device type, functional type, processing type, etc. are also stored. Note that the relation expression pattern further includes that shown in FIG. It should be noted that the display of constraints is omitted in FIG. Of course, the relational expression patterns are not limited to these.

The operation of the present invention shown in FIG. 2 will be described. For the morphological analysis unit 1, the input sentence “using an auxiliary storage medium such as a floppy disk as shown in FIG.
When "..." is input, the morpheme analysis unit 1 refers to the word dictionary 5, analyzes each word, adds a part of speech to this, and sends it to the syntax analysis unit 2 and the related concept extraction unit 4.
The syntax analysis unit 2 analyzes the dependency relationship as described above and sends it to the context analysis unit 3 and the relation concept extraction unit 4.

As a result of the analysis in the syntactic analysis unit 2, it can be understood that the relational expression pattern can be extracted based on the morphological analysis processing. Therefore, the relational concept extraction unit 4 refers to the identification dictionary 6,
It is identified that they correspond to the relational expression pattern "Y such as X", that X is a floppy disk, Y is an auxiliary storage medium, and both are device type. As a result, the related concept of "floppy disk Subc auxiliary storage medium" is extracted and held in a file (not shown).

In this way, related concepts are extracted from a large number of input sentences during a period of, for example, half a year or one year. By grouping these with a specific word, for example, using "floppy disk" as a key, floppy disk"
As an attribute of, a meaning definition such as one of auxiliary storage media, used in a computer, used in a word processor, etc. can be obtained, and thus a dictionary, for example, a technical terminology can be extremely easily edited.

In the case of a simple sentence such as a title of a paper or a headline of a newspaper as an input sentence,
Since the predetermined information can be extracted only by performing the morpheme analysis process, in such a case, the syntactic analysis unit 2 and the context analysis unit 3 are stopped and can be efficiently extracted based on the processing from the morpheme analysis unit 1.

Further, depending on the type of the input sentence and the input sentence, one of the syntactic analysis unit and the context analysis unit can be omitted. For example, if it is decided not to extract a sentence that spans a plurality of sentences, the context analysis unit may not be used.

[0040]

As described above, according to the present invention, the relation concept of a pair of two words can be automatically collected. Therefore, as compared with the conventional term dictionary editing, when the relation is manually collected, a comparison is made. Therefore, the editing effect can be greatly improved.

[Brief description of drawings]

FIG. 1 shows a principle diagram of the present invention.

FIG. 2 shows a configuration diagram of an embodiment of the present invention.

FIG. 3 shows example sentences.

FIG. 4 shows an example of pattern rules.

FIG. 5 shows an explanatory diagram of an extraction state.

FIG. 6 shows another example of pattern rules applied to morphological analysis / syntactic analysis.

[Explanation of symbols]

 1 Morphological analysis unit 2 Syntax analysis unit 3 Context analysis unit 4 Relational concept extraction unit 5 Word dictionary 6 Discrimination dictionary

Claims (4)

[Claims] 1. A morpheme analysis means (1) for performing a morpheme analysis for dividing an input sentence into morphemes, and a morpheme analysis result obtained by this morpheme analysis means (1) can be related by a relationship established between words. A relational concept extraction device comprising: a relational concept extraction means (4) for extracting a pair of two words. 2. A syntactic analysis unit (2) for analyzing the structure of the input sentence is provided, and the sentence parsing result obtained by the syntactic analysis unit (2) enables relations to be established by a relationship established between words. 2. The relational concept extraction device according to claim 1, wherein a pair of words is extracted. 3. A context analysis means (3) for analyzing a relation established between a plurality of sentences of the input sentence is provided, and a word space is established from a context analysis result obtained by the context analysis means (3). 2. The relationship concept extracting device according to claim 1, wherein a pair of two words that can be related by the relationship is extracted. 4. A context analysis means (3) for analyzing a relationship established between a plurality of input sentences is provided, and a context analysis result obtained by the context analysis means (3) is used to establish a relationship between words. 3. The relational concept extraction device according to claim 2, wherein pairs of two words that can be related are extracted.