JPH02284262A - Word correcting device - Google Patents

Abstract

PURPOSE:To correct an erroneous word to present another word as a candidate at the time of inputting the erroneous word by using micro features to express the meaning of a word. CONSTITUTION:A word dictionary 1 where words and their meanings are expressed with partial sets of micro features, a restriction holding means 3 where restriction items of micro features which the word to be substituted for an input word should have out of micro feature expressions stored in an erroneous input word storage means 2 are held, and a most approximate meaning selecting means 4 are provided, and this means 4 checks the word dictionary 1 to find the micro feature expression approximating the meaning expression of micro features in the erroneous input word storage means 2 and checks whether this found expression satisfies restriction items held in the restriction holding means 3, and this micro feature expression is outputted when it satisfies them. Thus, the correct word to be inputted is forecasted for the erroneously inputted word.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to an apparatus for modifying input words using micro-feature representations.

(Prior Art and its Problems) Conventional natural language analysis methods, including machine translation, have focused on processing based on syntactic information of sentences with good characteristics. References on natural language analysis include [Japanese Machine Translation System Using Semantic Structures] (Nikkei Electronics 1)
(April 17, 1984 issue) However, although the linguistic information that humans normally exchange with each other contains noise such as typos, omissions, grammatical deviations, and omissions, Regarding processing, the walking method shown in the reference literature is based on the premise that the input sentence given is a sentence of good quality. Therefore, with conventional analysis methods, the noise part in the input sentence makes it impossible to analyze it. As described above, conventional natural language analysis methods have problems that need to be solved regarding the analysis of incorrect input words. In order to deal with such noise, it is necessary to solve each of the problems listed above.The purpose of the present invention is to correct when an incorrect word is input, and to input another word. An object of the present invention is to provide a device that presents words as candidates.

(Means for Solving the Problems) The present invention is an apparatus for correcting words input in a natural language understanding system that has a set of micro features and the meaning of the word is expressed by a subset of the micro features. a word dictionary that stores pairs of words and meanings of the words expressed as "yes" or "absent" of micro features; erroneous input word storage means for storing the meaning of the input word; constraint holding means for retaining the nature of the meaning that the input word should originally have; The present invention is a word correction device comprising a nearest meaning selection means for inputting a word which is in a close relationship in meaning to the word in the present invention and which satisfies the property of the meaning of the constraint holding means.

(Operation and Principle) The present invention will be described in detail using a ninth example in which the problems remaining in the prior art are solved by the above-mentioned means.

A feature of the present invention is that the meaning of words is expressed using micro features. Micro features have a more detailed meaning for expressing various concepts. For example, the following collection of micro features can be considered.

(Blue, white, silver, soft, hard, pointy, killing, walking,...) Using these micro features to express ``knife'' and ``asparagus'', the following are respectively It can be expressed as a subset of micro features.

Knife = (silver, hard, sharp, kill,...
...) Asparagus = (green, pointy, plant, food, ...)
...) From this expression, we can see that "knife" and "asparagus" have something in common: they are "pointed."

As for the micro-features, it is sufficient to prepare a set that is convenient for the problem area of the specific application problem.For example, if the system deals with the field of political corruption, the micro-features
A set of features can be a set such as (recruitment, undersecretary, donation, transfer, etc.).

Once a set of micro-features is determined in this way, it can be expressed as described above using a dictionary in which the meanings of words are registered and a subset of micro-features for each word. Furthermore, the meanings of words that do not have specific meanings, such as ``desu'' and ``masu'', can be expressed as a complete set.

Next, assume that we are given a zero-order sentence that describes how to correct an incorrect input word and present another word candidate.

(1) Astronomer narried
5 tar.

(2) Astronolly married
5 tar.

(1) can be interpreted in two ways. One is, ``The astronomer married a movie star.'' The other is, ``The astronomer is obsessed with the starry sky.'' On the other hand,
(2) does not make sense, the subject is AStrOnOI
I mistyped 'ler as AStrOnOly. ^5 Start the association from tronoB, AStrOn
Finding a candidate word, OIIer or otherwise, will correct the incorrect input word. The procedure is as follows.

1) Find the semantic expression of ^S t ronoly using micro features.

2) AStrOnO expressed by micro features
If there are any meaning expressions for lly that contradict the properties that the word originally replacing Astrony should have, these are retained as constraints. For example, (2
)'s main verb is +tarry, so the subject should be human. However, in ^5tronony, the micro-feature "human" is written as "None." Therefore, when we bring in a word candidate with a semantic expression similar to AStrOnOIY, there is a constraint that "the micro-feature "human" must exist. Matters are created and maintained. Such constraints are given from higher levels such as syntactic analysis, semantic analysis, and context analysis.

3) Search the word dictionary and select the word with the micro-feature representation closest to Astronomy's micro-feature representation.

4) Check whether the selected micro-feature expression satisfies the micro-feature constraints determined in 2). If satisfied, it is output; if not, the word with the next closest micro-feature representation is selected. For example, Astrono
If m i l k-way is selected as my's nearest candidate, it is checked whether it has the "human" attribute.

Since the li l ky-way does not have the "human" attribute, it is rejected. As examples of micro-features that are represented by micro-feature expressions in Astronomy and have the attribute of human beings, candidates include "alien" and "astronaut." These words will then be output. However, the purpose of the present invention is to specify such candidates, and the extracted "alien" and "astronaut" are logically analyzed at a higher level - syntactic analysis, semantic analysis, and context analysis. It is outside the scope of the present invention to check whether the information is valid or not.

(Example) Hereinafter, details of the present invention will be explained according to examples.

FIG. 1 is a block diagram showing one embodiment of the present invention.

In Figure 1, 1 is a word dictionary in which words and their meanings are expressed as a subset of micro-features, and 2 is a micro-feature of a word that was input but was determined to be incorrectly input due to some problem. Erroneous input word storage means 3 stores the meaning by expression; erroneous input word storage means 2;
4 is a constraint holding means for holding constraints on micro features that a word to be replaced with an input word should have among the micro feature expressions stored in the erroneously input word storage means 2;・Find the micro-feature expression that is closest to the semantic expression by features, check whether it satisfies the constraints of the constraint holding means 3, and if it satisfies the constraints, output a micro-feature expression that is close in number to satisfy the requirement. If not, candidate words are taken out from the word dictionary 1 one after another in the order of closest micro-feature expressions, and when a word that satisfies the constraints of the constraint holding means 3 is found, it is output.

In this embodiment, a bit vector is used to express whether a micro feature is present or absent. That is, we give one bit for each micro-feature, and if a micro-feature is present, the corresponding bit is represented by a 1, and if it is absent, the corresponding bit is represented by a 0. Therefore, each word is rloololJ Its meaning is expressed by bit examples such as . Input word storage means 3
stores such a bit example.

The constraint holding means 3 sets bits corresponding to micro-features that correspond to the meaning they should originally have to 1, and corresponds to other micro-features, according to instructions from a higher-level analysis section such as a syntax analysis section or a semantic analysis section. The corresponding bit is set to 0, and a vector is generated and held.

FIG. 2 shows the detailed configuration of the nearest meaning selection means 4. In FIG. 2, 11 is exclusive fl! ! 12 is a bit adder, 13 is a comparator, and 14 is a constraint determiner.

In FIG. 2, the word dictionary 1 always outputs all its meanings to the exclusive OR circuit 11.
Only the words designated by the comparator 13 are output to the constraint determiner 14.

The exclusive OR unit 11 inputs the meaning of each word in the word dictionary 1 and the erroneous input word storage means 2, performs an exclusive OR on a bit by bit basis, and passes the result to the bit adder 12.

The bit adder 12 counts the number of 1 bits and outputs it. The comparator 13 compares the values output from the respective bit adders 12 and reads out from the word dictionary 1 the meaning of the word in the word dictionary corresponding to the bit adder 12 that takes the minimum value. The meaning of the read word is input to the constraint determiner 14 together with the value held by the constraint holding means 3. The constraint determiner 14 is
It is checked whether the meaning of the read word satisfies the constraints specified by the constraint holding means 3. Specifically, with respect to micro-features in which 1 is set in the constraint holding means 3, it is checked whether 1 is also set in the read word. When it is confirmed that this constraint is satisfied, the meaning of the word is output. if,
When the meaning of the read word does not satisfy the constraints of the constraint holding means 3, the constraint determiner 14 selects the comparator 1.
Step 3 is asked to find the meaning of the second closest word. This process is repeated until a word satisfying the constraints of the constraint holding means 3 is found.

The process described above will be explained using an example. Assume that the value of the erroneous input word storage means 2 is (1011).

Further, it is assumed that the following three words are stored in the word dictionary 1.

The bitwise exclusive OR of (0101), (0000), and (1010) is (1110), (1011), and (0001).The result of bitwise addition is (3), (3) , (1), and (1010) takes the minimum value, so (1011
), and (1010) is passed to the constraint determiner 14. Next, (1010) is compared with the value of the constraint holding means 3. If constraint holding means 3
If the value of is (0100), the second bit should be 1, but since it is 0, the constraint is not satisfied, and (1010) is rejected and the next candidate is selected.

If the value of constraint holding means 3 is (0010), (
1010) satisfies the constraint, so (1010) is output as the nearest word.

(Effects of the Invention) According to the present invention, in natural language processing, it is possible to predict the word that should originally be input from the input word, so it is possible to analyze a natural language sentence that contains a lot of noise. It can be an aid to the system.

The device provided by the present invention is suitable for constructing a natural language dialogue system as a system for extracting only the information required by a user from a large amount of natural language sentences, or as an interface for a database or an expert system.

[Brief explanation of drawings]

FIG. 1 is a configuration diagram showing one embodiment of the present invention, and FIG.
It is a block diagram of the embodiment which shows the nearest meaning selection means in the embodiment shown in detail. In the figure, 1 is a word dictionary, 2 is an input storage means, 3 is a constraint holding means, 4 is a nearest meaning selection means, and 11 is an exclusive O
12 is a bit adder, 13 is a comparator, and 14 is a constraint determiner. Figure 1

Claims (1)

[Scope of Claims] An apparatus for correcting words input in a system having a set of micro-features and aiming at natural language understanding in which the meaning of a word is expressed by a subset of the micro-features, comprising: and micro-features present or absent.
a word dictionary storing pairs of the meanings of the words expressed in a constraint maintenance means that maintains the nature of the meaning that a word should originally have; and a constraint maintenance means that searches the word dictionary to find words that are closely related in meaning to the words stored in the erroneously input word storage means and that are of the constraint maintenance means. A word correction device comprising nearest meaning selection means for outputting a word that satisfies the properties of its meaning.