JPH06332940A - Example leading type natural language analyzing device - Google Patents

Abstract

PURPOSE:To select optimum structure with high accuracy according to any simple method concerning a vague sentence by selecting the maximum likelihood input structure according to calculated tolerance and outputting the selected maximum likelihood tolerant structure. CONSTITUTION:A sentence in a natural language is inputted from an input means 1, the input sentence is mapped to input structure by an analyzing means 2 and when the plural kinds of input structure are provided, the tolerance is calculated for each kind of structure. Since the maximum likelihood tolerant input structure is selected and outputted by a selecting means 4 according to the calculated tolerance, the optimum structure can be selected with high accuracy concerning the vague sentence. Namely, the maximum likelihood tolerant structure is outputted while referring to the example concerning the vague input sentence. Moreover, approximate collation can be provided by utilizing thesaurus, and the number of examples is reduced. Further, since the example is prepared for each field, the example can be easily adjusted for each field. On the other hand, this method can be applied to any arbitrary language and easily integrated into a lot of conventional analyzing methods (such as CFG and pattern matching).

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an example-driven natural language analysis device, and more particularly to a simple method for determining an optimum structure for an ambiguous sentence, which is used in a machine translation system, an information retrieval system, a question answering system, etc. The present invention relates to an example-driven natural language analysis device that can be selected with higher accuracy.

[0002]

2. Description of the Related Art Demand for computer language processing systems is increasing, and research and development are being actively conducted. Recently, machine translation systems have been commercialized in various fields such as information retrieval systems and question answering systems.

[0003]

The language analysis device, which is an essential element of a computer language processing system, has had a common problem in the past. That is, it is not possible to select the optimum structure with high accuracy for ambiguous sentences. For example, the preposition of a preposition is a typical example of causing structural ambiguity.

I present a paper at the conferenc
e. The prepositional phrase "at the conference" in this example sentence may modify both the verb "present" and the noun "a paper".
In general, it is difficult to uniquely determine a prepositional phrase to which a preposition is related by a conventional parsing device based on grammar rules. In computer processing, a person can usually determine the likelihood of a preposition even if there are a plurality of prepositions of a given preposition. In this example, the former is the more natural contact.

A method that is widely used at present is a method of designating priority information of an affiliated party by using a grammatical rule or a semantic marker in a dictionary. However, this method cannot handle the free case of a verb (for example, the above-mentioned prepositional phrase "at the conference") or the case where the prepositional phrase is associated with a noun, apart from the essential case of the verb.

On the other hand, in recent years, probabilistic grammar has been actively studied. Since this method optimizes the application of grammatical rules, it is necessary to give detailed grammatical rules close to words in order to resolve disambiguation of prepositional phrases. It is not easy to describe such fine grammar rules. Further, even if it can be described, the learning time for optimization becomes enormous (the calculation amount is O if the size of the grammar is G.
(G ³ )).

There is a method of using world knowledge and a discourse model to eliminate structural ambiguity. However, this method has some problems such as difficulty of knowledge creation and heavy processing,
There is no one that can process actual sentences.

Therefore, a main object of the present invention is to provide an example-driven natural language analysis device which can select an optimum structure for an ambiguous sentence by a simple method with high accuracy.

[0009]

The invention according to claim 1 is
Input means for inputting a sentence in natural language, analyzing means for mapping the input sentence input to an input structure, and likelihood calculating means for calculating a likelihood of each structure when a plurality of mapped input structures are obtained And selecting means for selecting the maximum likelihood input structure according to the calculated likelihood, and output means for outputting the selected maximum likelihood structure.

According to the second aspect of the present invention, the likelihood calculating section according to the first aspect stores an example storage section for storing a partial structure that frequently appears in the natural language sentence and a word tree based on the similarity of meanings. It is configured to have a thesaurus arranged in shape.

In the invention according to claim 3, the likelihood calculation section of claim 1 calculates the overall likelihood based on the partial likelihood calculated for each partial structure of the input structure.

According to the invention of claim 4, the partial likelihood of claim 3 searches the example storage unit for an example partial structure having a minimum meaning distance between the partial structures and the input partial structure, and the partial structure meaning It consists of the distance and the number of retrieved example substructures.

In the invention according to claim 5, the semantic distance between substructures according to claim 4 is determined based on the interword semantic distance calculated according to a thesaurus of words included in the input substructure and the example substructure.

In the invention according to claim 6, the example storage section of claim 2 is configured to be accessible using the partial structure as a key.

In the invention according to claim 7, the calculation of the likelihood and the partial likelihood of claim 1 is realized as a post-process of the analyzing means or directly incorporated in the analyzing means.

[0016]

In the example-driven natural language analysis apparatus according to the present invention, a natural language sentence is input from the input means, the input sentence is mapped to the input structure by the analysis means, and when a plurality of input structures are obtained, each structure is obtained. The maximum likelihood input structure is selected and output by the selecting means according to the calculated likelihood, and the optimum structure can be selected with high accuracy for an ambiguous sentence.

[0017]

The following description shows two embodiments in which the present invention is applied to the analysis of English. Here, in particular, the preposition dependency is dealt with, but the present invention is also effective for the dependency of other structural ambiguities such as to-infinitives, relative clauses, and subordinate clauses, if the examples are collected. is there.

In order to simplify the explanation, an example of the problem of determining whether the presumably predecessor of the prepositional phrase "at the conference" in the above example sentence is the verb "present" or the noun "a paper" is taken as an example. Treat as.

FIG. 1 is a schematic block diagram of an embodiment of the present invention, and FIG. 2 is a diagram showing a concrete example of the likelihood calculating section shown in FIG.

First, the input unit 1 is composed of a keyboard, a character recognition device, a voice recognition device, etc., inputs a sentence, and at the same time divides the sentence into words, adds information such as a part of speech to each word according to a dictionary, and analyzes it. Give to 2. The analysis unit 2 generates a possible structure (input structure) for an input by a method such as conventional syntax analysis or pattern matching. As shown in FIG. 2, the likelihood calculation unit 3 includes an example storage unit 6 and a thesaurus 7.
Compute the likelihood of each of the input structures, including and. The likelihood calculator 3 is a characteristic part of the present invention, and will be described in detail later. The selection unit 4 selects the maximum likelihood structure from the likelihood-added structures calculated by the likelihood calculation unit 3, and outputs the selected maximum likelihood structure to the output unit 5. The output unit 5 includes a display device, a printing device, and the like.

The above-mentioned likelihood calculation unit 3 calculates the overall likelihood based on the partial likelihood calculated for each partial structure of the input structure. As for the partial likelihood, the example storage unit 6 uses the example partial structure that minimizes the semantic distance (real value) between the partial structures and the input partial structure.
It is composed of the semantic distance between the substructures and the number (frequency) of the retrieved example substructures. The semantic distance between substructures is determined based on the semantic distance between words calculated according to the thesaurus 7 of words included in the input substructure and the example substructure. Here, the partial likelihood determination will be described in detail.

FIG. 3 corresponds to an ambiguous reference of an example sentence.
Two input structures are shown, and here, in order to simplify the description, they are represented by what is called a dependency structure expression. FIG. 3A corresponds to the case related to the verb “present”, and FIG. 3B corresponds to the case related to the “noun“ paper ”, and the dependency is indicated by a bold line. The likelihood calculator 3 uses these two possibilities, namely “present, at, the conf
erence ”,“ a paper, at, the conference ”. At that time, two of the example storage unit 6 and the thesaurus 7 shown in FIG. 2 are referred to. The example storage unit 6 is configured so that it can be accessed by using the partial structure as a key. In the case of the prepositions described here, the partial structure is a triple consisting of a verb (or noun), a preposition, and a noun.

FIG. 4 is a diagram showing an example of the example storage unit,
In particular, FIG. 4A shows a case related to a verb, and FIG.
Indicates a case involving a noun. FIG. 5 is a diagram for explaining the calculation of the semantic distance between a part of the thesaurus and words.

The thesaurus 7 is a dictionary in which words are systematized according to a common concept of meaning. A part of this is shown in FIG. The semantic distance between words is calculated using thesaurus 7. In this method, the semantic distance between words is
It is defined by the semantic distance between the above concepts. The meaning distance between concepts is set to a value from 0 to 1 according to the position of the smallest common superordinate concept in the thesaurus 7. A value of 0 means that the two concepts are the same, and a value of 1 means they are irrelevant. If the number of layers (n + 1) is 0, 1 / n, 2 /
Assign n, ..., 1 as the distance. Thesaurus 7 in FIG.
Assigns 0, 1/3, 2/3, 1 in the (3 + 1) layer. For example, the smallest common superordinate concept of the two concepts "human" and "monkey" is "animal", and the distance is 1/3.
Therefore, the distance between the word "she" having the concept "human" and the word "chimpanzee" having the concept "monkey" is 1/3. In FIG. 5, d (x, y) represents a semantic distance between words (or concepts) x and y.

The semantic distance between substructures is calculated as the sum of weighted distances of the words. Input substructure I
And the example substructure E is composed of a sequence of n words I _k and E _k , respectively, the intersubstructure semantic distance d (I, E) is calculated by the following equation.

[0026]

[Equation 1]

The weight w _k needs to be adjusted so that the correct answer rate is high, but here w _k = 1 for simplicity.

Next, the likelihood of the entire input structure will be described. Here, first, considering a simple case where there is only one prepositional phrase in the input sentence, the likelihood of the entire structure and the partial likelihood match. The selection unit 4 makes this comparison.

Next, the specific processing will be described. Let n be the number of candidates for the dependency, x _i (1 <= i <= n) be the candidates (verb or noun) for the dependency, and p be the preposition,
Let y be the object (noun) of p. The problem is the input "x ₁ ,
, X _n , p, y ”, the maximum likelihood destination“ x _k ”of“ p, y ”is selected. In the example sentence, the input is "pres
"ent, a paper, at, the conference", and the correct answer is "present". [Processing Procedure] (Step 1) The following processing is repeated (1 <= i <=
n).

The semantic distance between the partial examples with "x _i py" is calculated according to the above-mentioned equation (1), and the example storage unit 6 is searched for the example of the minimum distance. The frequency f _i of the example with the same distance as the minimum distance d _i is stored.

[0031] (Step 2) d _i is minimized x _i is 1
If there are only three, x _i is set as the dependent, and the process ends.

[0032] (Step 3) f _i among d _i is minimized x _i is if there is only one is x _i that maximizes, Exit to return the x _i as dependency destination.

(Step 4) Return all x _i for which d _i is the minimum as the related party, and end. Example 1 When ending in step 2.

I present a paper at the conferenc
e. The minimum distance for "present, at, the conference" is 0.0
The minimum distance for "a paper, at, the conference" is 0.33. Therefore, the former is decided. Example 2 When ending at step 3.

We have the next conference at the
hotel. In this case, "have, at, the hotel""the next confe
The shortest distance between both candidates "rence, at, the hotel" was 0.17. The former example of the shortest distance is "hold,
The frequency of "at, hotel" is 92, and the example of the shortest distance of the latter is "meeting, at, hotel", and the frequency is 15, so the reference of "at, the hotel" is , "Have".

The above procedure is a compound likelihood that can be calculated from the partial inter-example semantic distance d _i and the frequency f _i of examples with the same distance,
For example, it is almost equivalent to minimizing d _i −f _i / 10 ^m (m depends on the scale of the example storage unit 6, but empirically, m = 6 is sufficient).

The maximum likelihood is determined according to the total sum of the composite likelihoods of all the prepositions. Next, a second embodiment will be described. The second embodiment is basically the same as the first embodiment. The difference is that the calculation of the likelihood and the partial likelihood is incorporated in the analysis unit 2. When the grammar rule is applied, the related party candidates are extracted, and at the same time, the partial likelihood calculation is performed.

The following grammar rules are prepared. (1) VP PP → VP (2) NP PP → NP It is assumed that the input sentence is the following example sentence.

I present a paper at the conferenc
e. Substructure “present when grammar rule (1) is applied
, At, the conference ”can be extracted and the partial likelihood calculation of the partial structure is performed. When the grammar rule (2) is applied, the partial structure “a paper, at, the conference” can be extracted, and the partial likelihood calculation of the partial structure is performed. At the end of the analysis, the likelihood of the entire structure is obtained. The other points are the first
Same as the embodiment.

[0040]

As described above, according to the present invention, a natural language sentence is input and mapped to an input structure, and when a plurality of input structures are obtained, the likelihood of each structure is calculated and the likelihood is calculated. Since the maximum likelihood input structure is selected and output according to the degree, the maximum likelihood structure can be output with reference to an ambiguous input sentence. Moreover, approximate matching can be realized by using a thesaurus, and the number of examples can be reduced.
Furthermore, if an example is prepared for each field, it can be easily adjusted for each field. Further, this method can be applied to any language, and can be easily incorporated into many conventional analysis methods (CFG, pattern matching, etc.).

[Brief description of drawings]

FIG. 1 is a first example of an example-driven natural language analysis device according to the present invention.
It is a block diagram which shows the outline | summary of an Example.

FIG. 2 is a block diagram showing an outline of a likelihood calculation unit shown in FIG. ,

[Figure 3] Ambiguous input sentence "I present a paper at the con
is a diagram showing a possible input structure for "ference."

FIG. 4 is a diagram showing a part of an example storage unit, in particular,
(A) shows a case related to a verb, and (b) shows a case related to a noun.

FIG. 5 is a diagram for explaining a part of the thesaurus and calculation of a semantic distance between words.

[Explanation of symbols]

 1 input unit 2 analysis unit 3 likelihood calculation unit 4 selection unit 5 output unit 6 example storage unit 7 thesaurus

Front Page Continuation (72) Inventor Kura Furuse, Seika-cho, Soraku-gun, Kyoto Prefecture, Osamu Osamu Osamu Osamu, No. 5, Sanpeiya, ATR Co., Ltd. Automatic translation telephone laboratory (72) Ina, Hitoshi Iida Seiraku-cho, Soraku-gun, Kyoto Prefecture Daiji Intani, Shoji, Hiratani No. 5, Arai Co., Ltd. Automatic translation telephone laboratory

Claims (7)

[Claims] 1. Input means for inputting a sentence in natural language; analysis means for mapping the input sentence input by the input means onto an input structure; and a plurality of input structures obtained by the analysis means,
Likelihood calculation means for calculating the likelihood of each structure, selection means for selecting the maximum likelihood input structure according to the likelihood calculated by the likelihood calculation means, and maximum likelihood structure selected by the selection means An example-driven natural language analysis device, comprising: output means for outputting 2. The likelihood calculating means includes an example storage unit that stores an example substructure that frequently appears in the natural language sentence, and a thesaurus that organizes words into a tree shape based on similarity in meaning. An example-driven natural language analyzer according to claim 1, characterized in that it comprises: 3. The example-driven nature according to claim 1, wherein the likelihood calculating means calculates the overall likelihood based on a partial likelihood calculated for each partial structure of the input structure. Language analysis device. 4. The partial likelihood is searched from the example storage unit for an example partial structure having a minimum meaning distance between the partial structures and the input partial structure, and the meaning distance between the partial structures and the searched example part are searched. The example-driven natural language analysis device according to claim 3, characterized in that it comprises the number of structures. 5. The semantic distance between substructures is determined based on an interword semantic distance calculated according to the thesaurus of words included in the input substructure and the example substructure. 4. Example-driven natural language analyzer. 6. The example-initiated natural language analysis apparatus according to claim 2, wherein the example storage unit is accessible by using the partial structure as a key. 7. The calculation of the likelihood and the partial likelihood is
The example-initiated natural language analysis apparatus according to claim 1, characterized in that it can be directly incorporated into the analysis means other than a method realized as a post-processing of the analysis means.