JPH10247193A - Grammar check system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To display the grammatical errors with higher accuracy and to correct these errors with no overcorrection by adding tags to the sentences in regard to the parts of speech to check the sentences about inaccurate grammar and accessing the headwords of an English dictionary based on the context where the words appear. SOLUTION: The output of an analyzer 20 is connected to a part-of-speech sequence probability decision unit 22 to derive the sequence probability of an input sentence 10. The output of the unit 22 is used by various modules for analysis of the sentence 10. Then the hardness that is decided by the part-of- speech sequence is useful in a context sensitivity dictionary reference module 40. In general, the given words have many parts of speech corresponding to the semi-headwords of a dictionary other than the context. The module 40 accesses the dictionary and selects a proper definition based on the part of speech of the word that is obtained by the part-of-speech analyzer 20.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a grammar checking system, and more particularly, to displaying grammatical errors more accurately using various grammar checking modules and correcting such errors without overcorrection. It is about a system to do.

[0002]

BACKGROUND OF THE INVENTION As discussed in U.S. Pat. No. 4,868,750, issued to Henry Kucera et al. Receives the text and uses the stored word / analysis dictionary and analysis program to analyze the encoded text and perform automatic linguistic analysis by identifying errors. In particular, such systems are used within the Microsoft Word program for detecting grammatical errors.

[0003] One of the most troublesome problems associated with such systems is the extremely high error rate when the system suggests proper usage. The reason why the error rate is excessively high is that the system analyzes the sentence incorrectly.
Also, even assuming the sentence analysis is accurate, the Microsoft system often suggests incorrect words.

Some classes of systems attempt to analyze a sentence based on the probability that the entire sentence is accurate. The biggest problem with such a system is that it requires storage and processing power beyond the capabilities of current personal computers and associated memory.

[0005] Other systems attempt to detect incorrect grammar by analyzing the sentence based on a training corpus. However, due to system limitations, this type of system cannot be used in a personal computer environment due to the need for mass storage and high speed processing.

As an example, conventional grammar checking systems fail to routinely insert indefinite articles such as "a" and "an". This is, in fact, a major problem if a foreign language speaker attempts to translate into the natural language provided by the system.

Also, grammar rules or "the" or "a"
In the case of sentences composed by people who are not familiar with the proper colloquial usage of articles, the inability to insert such articles is very significant. In addition, common mistakes made by conventional grammar checking systems include not recognizing an incorrect verb sequence in which multiple verbs are used. Although multiple verbs are used properly in sentences, most foreign language speakers routinely use "He has rec.
ognize that somethingexis
ts ". Here, “has” is a verb, and “recognize” is a verb. As you can see, several verbs are obviously misused.

Most importantly, the problem of so-called determiners arises. For example, the sentence "I have ci
"garrate" is clearly missing the qualifier "a". Similarly, qualifiers such as "some" and "a new" are often dropped. Thus, the appropriate statement is "I
have a few cigarettes ". Note that the same sentence can be properly constructed by pluralizing the noun.
For example, "I have a new cigarat
tes "or" I have cigarette "
s ".

[0009] Other typical grammatical errors that are neither corrected by the spell checker nor by conventional grammar systems may not be able to correct improper inflections. For example, with respect to inappropriate verb variants, such a system may be referred to as "I driven to the
We rarely correct sentences such as "market".

[0010] The above problems are most acute when viewed from non-native speakers who are not familiar with the grammar and rules of the language. The rules are not as simple as one might think, especially in English, and the exact “grammar” is often determined by grammar or rules that the native speakers using the language are not familiar with.

[0011]

Therefore, it is important to provide a grammar checking system that takes into account the mistakes most often made by non-native speakers of a particular nationality. Thus, for example, a group of errors that a Japanese native speaker usually makes are translated into English in a predictable and thus correctable manner. Similarly, for example, for French or any Romance language, some characteristic errors made when translating to English can be detected and corrected.

[0012] Syntax recognition systems are generally limited to operations on text with a small, well-defined vocabulary or, more generally, operations on text that handles a limited range of syntax features. It had been. Increasing the vocabulary or syntax range requires more and more complex structures and an increased number of specialized recognition rules, and thus makes the system very large or generally available computer
It becomes troublesome for business operation on the system.

Another common system for detecting and correcting contextual errors in a text processing system is Frederick
No. 4,674,065 to Frederick B. Lang et al. In this patent application, a special dictionary of multiple sets of homonyms and easily confused words, and multiple sets of di-gram states and n-grams (n-grams) from which the proper usage of words can be determined statistically. Implement a system that proofreads the document for word usage confirmation and text processing by combining the n-gram) state. As discussed above, performing statistics on words rather than parts of speech requires a very large training corpus and fast computation, thus making the system somewhat cumbersome for personal computing applications.
In addition, the system detects words that are confused for words that have the same pronunciation, which is confused in common usage, but not enough to make corrections for words that are not the same.

Finally, US Pat. No. 4,830,521 describes:
This is a patent for an electronic typewriter having a spell check function and a proper noun recognition ability. It should be understood that the problem of noun recognition is centered on capitalization scenarios that may or may not be accurate in recognizing proper nouns. Most importantly, this patent
Although capitalization is obviously done for words anywhere in the sentence, testing a word and checking whether it is the first word in the sentence to determine the function of capitalization I can only do it.

As other prior arts, there are a number of patents which attempt to solve grammar problems by first using spelling correction. Such patents include US Patent No. 5,218,536.
No. 5,215,388, No. 5,203,705, No. 51
No. 61245, No. 5148367, No. 4995740
No. 4,980,855, No. 4,915,546, No. 49
Nos. 12671, 4903206, 4887920
No., 4887212, 4887334, 48th
No. 62408, No. 4852003, No. 4842428
No. 4829472, No. 4799191, No. 47
No. 99188, No. 4797855, No. 4689768
There is a number.

Also, US Pat.
No. 220503, No. 5200893, No. 516489
No. 9, No. 511389, No. 5029085, No. 5
No. 083268, No. 5068789, No. 500701
No. 9, No. 4994966, No. 4974195, No. 4
No. 958285, No. 4933896, No. 491459
0, 4816994, 477309, etc.
Many patents deal with text analysis. All of these patents relate specifically to systems that cannot be practiced to check grammar to the level required by non-native speakers who must provide written documents in a given natural language. Please understand that. It should also be understood that these patents relate to a general system that is not specifically intended to correct grammar and English idioms for non-native speakers.

Finally, there are several patents on how efficiently a dictionary can be encoded. These patents are disclosed in US Pat. Nos. 5,189,610, 506015.
No. 4, No. 4,959,785, and No. 4,782,264. It should be understood that encoding a dictionary is the only step in formalizing a system that can properly check the grammar.

[0018]

SUMMARY OF THE INVENTION To solve some of the major problems of conventional spell checking and grammar checking systems, the grammar checking system of the present invention comprises:
First, tag the sentence for part of speech, and then check the sentence for incorrect grammar. The system corrects word usage based on the probability that the part of speech has the correct sequence, rather than the probability that the entire sentence is correct. The system then checks for incorrect inflections, proper usage of indefinite articles, incorrect capitalization, incorrect auxiliary verb sequences, and inappropriate qualifiers. The system also accesses the English dictionary entry based on the context in which the word appears.

In particular, the system provides for inexact past tense and past participle forms of verbs ("d" instead of "dove").
driven, "hitten" instead of "hit"), plural forms of nouns (e.g., "childs" instead of "children"), inaccurate comparisons of adjectives or superlatives (e.g., "better" instead of "better") go
order, "best" instead of "best"
t ") in context and correct.

Further, the system uses the indefinite article "a"
And in context for appropriate usage of "an". The system also detects and corrects improper use of qualifiers by inserting or correcting words that are part of a sub-restrictive adjective that usually precedes the narrative adjective. These are the articles "the", "a",
"An", and "you" instead of them,
Includes correction of words such as "their", "some", "each". Corrections also generally involve the improper use of numbers that give the noun specificity, or extracting the referent from the class indicated by the noun.

In other features, the system first retrieves the underlying spelling and then uses a training sequence to determine the probability that the underlying word is or is not capitalized. To check for correct capitalization.

The system also checks for incorrect auxiliary verb sequences associated with the use of multiple verb sequences in a sentence and determines the correct shape based on a finite state machine or directed graph, especially a non-cyclic directed graph. Using this graph, the system first makes an error ("he
has accepted this fact), and then correct it to convert the multiple verb sequence into the appropriate form ("he has accepted the fact").
is fact ").

As another feature, the system accesses dictionary headwords based on the context in which the particular word is used in the sentence, as determined by the part of speech obtained from the part of speech tagger. The system can only display headwords in the dictionary that are relevant to the syntax context.

These and other features of the present invention will be better understood when read in detail with the drawings.

[0025]

BEST MODE FOR CARRYING OUT THE INVENTION

Embodiment 1 FIG. The configuration of the first embodiment of the present invention will be described with reference to FIG. FIG. 1 is a diagram showing a configuration of a grammar checking system according to Embodiment 1 of the present invention. In the drawings, the same reference numerals indicate the same or corresponding parts.

Referring now to FIG. 1, there is provided an instant grammar check of input sentences that is accurate, especially for those who spell foreign languages, and can be easily used, especially by those without computer knowledge. is important. In order to perform a grammar check, the input sentence 10 is
It is input to the CPU 14 of the processing system 16.

For reliable grammar confirmation, it is important to accurately determine each part of the pronunciation of the input sentence. Whereas conventional grammar checking systems used the input sentence directly, a feature of the present invention is to divide the input sentence into several parts of speech, thereby obtaining a part of speech sequence. This is "A
Stochastic Parts Program and Noun Phrase Parserfor
Unrestricted Text '' Proceedings of the Second Conf
erence on Applied Natural Language Processing, Aust
in, Texas, 1988, the part-of-speech analyzer 20 available as an embodiment of the Kenneth Church (Stochastic Parts Program). The result of deriving the part of speech is based on the input sentence "I hear
A part-of-speech sequence such as “pronoun, verb, qualifier, noun, verb” of “d this band play”.

Deriving the part of speech alone does not guarantee that the derived part of speech reflects an appropriate sentence.

In order to analyze and construct a proper sentence, it is important to ascertain the probability that the part of speech sequence corresponds to the correct word order. The output of the analyzer or tag 20 is coupled to a part-of-speech sequence probability determination unit 22 to derive the probability of the sequence of input sentences. The output of this unit is used by various modules in analyzing the input sentence 10.

The first module is a part of speech recognition unit 24 that selects a set of easily confused words or sentences based on the probability of the corresponding part of speech sequence. The selection of the correct word or sentence is determined in one embodiment by a probability higher than a predetermined threshold. The selection of the correct sentence is made by unit 26. The input to unit 26 is
Probability of various sentences as well as input sentences. As described below, unit 26 includes a list of easily confused words.

Although confusing sentences can be corrected in the manner described above, another module 28 is used to determine the spelling of the underlying word. Conventional spell checkers use look-up tables for spelling confirmation, but they do not consider capitalization, resulting in annoying suggestions for spelling errors. In addition, those grammar checking systems that use proper spelling are often confused by words, sentences, or words whose initials are capitalized.

To provide more reliable spell checking and grammar correction, the underlying spelling recovery unit 2
8 treats capitalized words as "confused" words. Then, using the techniques described above, derive, based on a training corpus, such as Brown's corpus, the probability that a capitalized word is in one category or in another category. I do.

Therefore, the conventional language processing system is
While restoring the underlying spelling of a word by imposing restrictions on whether the word is a common noun or proper noun, the recovery unit of the present invention uses context and probability to classify each word. This is done by analyzing sentences with words that are capitalized and words that are not capitalized to see which has a higher probability. The words analyzed for spelling then take the form of more probable words in the sentence. After restoring the most likely spelling, the output of the restoration unit 28 is passed to the inflection check / correction system 30.
To join. This spelling collector is a traditional spell check variant or tailored for the specific person who speaks a foreign language.

As an additional module, the auxiliary verb correction unit 32 also needs the exact part of speech derived from the part of speech sequence probability determination unit 22. An auxiliary verb correction problem arises when there are multiple verbs in a sentence that is partially inappropriate. This occurs when inaccurate tenses are used in complex auxiliary verb sequences. For example, the sentence "he world living" includes two verbs "would" and "living".
One precise form of the sentence is “he worldd live
e ”. Therefore, it is necessary to correct the tense of the verb "live".

To do this, the auxiliary verb correction unit 3
2 detects an incorrect auxiliary verb sequence and then suggests a correction. This is done first using a directed acyclic graph describing a finite set of verb sequences. It should be appreciated that before establishing the exact verb sequence, it is important to accurately identify the exact part of speech made by unit 22 as described above.

The output of the auxiliary verb correction unit 32 is coupled to a correct sentence selection unit 34 that indicates the appropriate alternative sentence.

An additional module that uses part of speech is the determiner correction unit 36. The purpose of this unit is to correct the words that determine the noun phrase referent. Examples of qualifiers are words such as "the", "a", "some". There are three types of errors detected and corrected by this unit. That is, there is a missing determiner, an unrelated determiner, and a mismatch between the determiner and the noun.

As an example of missing determiner, "John re
In “ad book”, “the” is missing. As an example of an unrelated quantifier, "John when to
In "the New York", "the" should be deleted. The sentence "Johnread many boo
For "k", the noun "book" must be pluralized to match the determiner "many", and the discrepancy is obvious. To detect inappropriate qualifiers, identify part of speech tags so that noun phrases can be identified. The system identifies noun phrases by maximizing matching rule expressions that define the sequence of part-of-speech tags that make up a valid noun phrase.

The system then tests each noun phrase to see if any qualifiers are missing. As part of this process, the leading noun is first detected, and then it is determined whether the leading noun is a collective noun, a collective noun, a idiom, or a missing qualifier. The system then tests each noun phrase to see if it has an extraneous determiner. Finally, the system tests whether the number of qualifiers and nouns in the noun phrase match. As a result, words are inserted, deleted or replaced as shown at 38.

Further, the module 42 corrects the usage of the indefinite articles “a” and “an” based on the input sentence 10.

Finally, the accuracy provided by the part-of-speech sequence becomes useful in the context-sensitive dictionary lookup module 40. In general, a given word has, out of context, a number of parts of speech, each corresponding to a sub-entry in the dictionary. The context sensitive dictionary reference module 40 accesses the dictionary and selects an appropriate definition based on the part of speech of the word obtained by the part of speech module 20. For example, the word "love"
A noun or verb, the noun "love" is a verb "l
For "ove", there are many different headwords in the dictionary. If the input sentence is "She was my first
Assuming that the word is "love", the word "love" becomes
Identified as a noun by the part of speech module, the context sensitive dictionary lookup module selects only dictionary entries for the noun "love" and dictionary entries for the verb "love".

The spelling of the word base is module 28
It is to be understood that, after recovery, this underlying spelling is used not only for inflection correction by module 30 but also within conventional spelling system 44. Thus, conventional spell-checking systems may overlook the initials during the spell-checking process, rather than providing inappropriate suggestions.

A) Grammar Correction Based on Part-of-Speech Probability In the past, some of the grammar checking systems described above
It attempts to correct English idioms by correcting improper usage of troublesome words, especially words with the same pronunciation but different spellings. For example, "too", "to" and "two", "their", "they're" and "there". Another common mistake is "maybe"
The main title is whether the word is one word or two words, such as and "may be". Also, "which"
And "whose" have the same pronunciation, but some words are often misused.

In the past, to confirm proper usage, the grammar of a sentence was calculated as the probability that the sentence would appear in English. Such statistical techniques assign high probabilities to grammatically correct sentences and low probabilities to non-grammatical sentences. Statistical techniques are obtained by training on sets of English sentences, ie, a training corpus. The corpus defines the exact usage. Therefore,
When a sentence is entered into such a grammar checking system,
Compute the probability that the entire sentence is associated with the corpus. It should be understood that to accept the entire English vocabulary, about 60,000 words, a corpus of hundreds of trillion words must be used. In addition, a corresponding number of probabilities must be stored in the computer. Thus, the task of analyzing an entire sentence is computationally and memory intensive.

To establish the correct usage in the system of the present invention, the probabilities of the part of speech sequence are derived. For this, depending on how sophisticated the system is,
It can be considered that there are 100 to 400 possible parts of speech. This translates to millions of word training corpora for hundreds of trillions. This type of analysis can be easily performed on standard computing platforms, including those used for word processing.

Therefore, in the system of the present invention, the sentence is first divided into several parts of speech. For example, the sentence "I he
"ard this band play" is analyzed as "pronoun, verb, determiner, noun, verb". The probability of this part of speech sequence is determined by comparing the sentence with the corpus. Also, this is not feasible without simply considering so-called trigrams. Trigrams are triple parts of speech that are adjacent in the input sentence. Usually, it is sufficient to analyze three adjacent parts of speech to determine accuracy. The probabilities of these trigrams are used to determine that a particular sentence has correct usage. Therefore, instead of examining the entire sentence, the probabilities of three adjacent parts of speech are calculated from the training corpus.

Assuming two sentences, one confused with the other, the techniques described above can be used to determine which is the correct usage. The system described above has two advantages because it can be determined with a low error rate. The first advantage is to clearly identify which of the two sentences is correct. A second advantage is that after determining the exact sentence, the part of speech can be used by other grammar checking modules for further processing.

Next, referring to FIG. 2, the input sentence S1 indicated by 30 is combined with the candidate sentence S2 indicated by the part of speech tags 32 and 34. The candidate sentence S2 is given an input that includes a list 36 of words to be confused. Tag 32 is 3
8, the sentence S1 is divided into the most likely part-of-speech sequence T1 and its probability P1. This is done by algorithms such as those described in Church. In this algorithm,
By calculating the most likely product of the probabilities of all possible trigrams of the part of speech, we obtain the most likely part of speech sequence.

The words in sentence S1 are part of a list 36 of easily confused words. In this case, all possible alternative sentences S2 for sentence S1 are generated according to list 36. The output of statement generator 34 is applied to tag 32,
Again, the algorithm yields the most likely part-of-speech sequence T2, shown at 40, and its probability P2.

At 38 and 40, the statements S1 and S2
After deriving the probabilities P1 and P2 of, it is then important to determine the most likely part-of-speech sequence. To determine the appropriate sentence to select, compare P2 and P1, as shown at 42, and if P2-P1 is greater than a certain threshold e, as shown at 44, the sentence S2
Is suggested. If P2-P1 ≤ e, no change is suggested, as shown at 46.

For example, if the input sentence is “I want to
"here this band" and "here"
Is misused in place of the exact word "ear", the two sentences S1 "I want to he
re this band "and S2" I want t
o hear this band ".

Assuming a statistical model of English sentences to compare these two sentences, one can try to compare the overall probabilities of the sentences. This method is called "Information Pr
Eric Mays, Fred Damer, published in Ocessing and Management, 27 (5): 517-422, 1991.
eau, Robert Mercer), "Context Bass"
Exploring in the "ed Spelling Correction," which is extremely expensive to compute, and therefore handles unrestricted text that requires a vocabulary of more than 40,000 words, is not feasible on a standard computer. To be able to calculate sentence probabilities directly, a huge amount of training data, for example at least 400,00
It requires 0000 training words and a huge amount of storage space.

In contrast, the system of the present invention shown in FIG. 2 compares the probability of the most likely part-of-speech sequence for a given input sentence with a sentence that is likely to be confused with it. For example, the sentence "I want to
Instead of calculating the probability of "here this band", the system derives the most likely part-of-speech sequence for the sentence, for example, "pronoun, verb, TO, adverb, qualifier, noun", and this part-of-speech for the input sentence. Compute the probability of the sequence. Similarly, the system
"I want to hear this ban
The most likely part-of-speech sequence for "d", for example, "pronoun, verb, TO, verb, qualifier, noun" is derived and its probability is calculated for the relevant sentence. Then
The system of the present invention compares "h
er "or" ear "is used.

Instead of calculating the above probabilities, in a preferred embodiment, the system of the present invention calculates the geometric mean of these probabilities by considering their word length,
For example, the logarithm of P1 divided by the number of words in S1;
It is calculated by comparing the logarithm of P2 with the number of words in S2. This means that the single word "ma
This is important when confused with a series of words such as "ybe" and "may be". When comparing the probabilities of part-of-speech sequences directly, statistical language models assign lower probabilities to longer sentences, so the results are not always accurate, but shorter sentences are more advantageous than longer sentences. is there. The above is shown in FIG.

The list of confused words 36 generally includes the following set: to, too, two; I, me; it's, it's; their, theth're, there; whose, whoch; then, tan; whose, who's; accept, exclude; adice, advise; lose, loose; write, right; youth, you're; affect, effect;

It should be noted that the system of the present invention is applicable to other confused words, and in particular to other languages, such as French, Italian, Spanish. Note that this method is universal as long as part-of-speech tagging can be performed using the method described in Church, a trigram model.

In summary, in addition to selecting sentences that are likely to be accurate, the systems of FIGS.
This is important when confirming other decisions about the grammatical nature of the sentence. The above provides a better and more reliable modality for breaking a sentence into several parts of speech.

In order to correct a sentence, it is important that the sentence can first be divided into several parts of speech. How accurate the grammar checker can work is
It is determined crucially by the accuracy of this division. With more reliable part-of-speech generation, the final result of a grammar check can be much more reliable.

B) Correction of "a" and "an" It should be understood that one of the most frequent errors made by foreign speakers is the usage of the indefinite articles "a" and "an". The English rules specify that the indefinite article "a" is used before the word pronounced in the first consonant, and "an" is used before the word pronounced in the first vowel. . The intrinsic and incorrect implementation of these English rules tests whether the first letter of the next word is a vowel or consonant. It is true that most words pronounced with the first consonant (vowel) are actually spelled with the first consonant (vowel), but this is not always the case, as in the following example. For example, the word "hour" has a first consonant (h), but a first consonant (h) corresponding to a vowel (eg, ow).
Pronounced with sound. Similarly, the word "European"
Starts with the first consonant (letter "E") but is pronounced with the first consonant (eg, "ye").

A previous solution to this problem consists in storing a dictionary of pronunciations of all English words. These solutions, while accurate, require a great deal of storage space for all English words.

Rather than using a dictionary lookup table for every word in English, the system of the present invention applies simple rules if no exception to the rule is found. Exceptions to the rules, each beginning with a vowel, are pronounced consonant first, words not processed by the rule, and beginning with a consonant, but pronounced first with a vowel,
It is stored in two small tables corresponding to the words processed by the rules. The lookup table for these words contains less than 300 words, unlike a general purpose dictionary-based system of 60000 words.

The following table is a list of words with English exceptions.

Table 1 Ewell Ewell's U U's U-boat U-boot's U-turn U-turn's UFO UFO's Uganda Uganda's Ugandan Utani's Utanian's Utanian's Uganda Uruguayan Utah Utah's Utopia Utopia's Utopian Utrecht Utrecht's ewe ewe's ewer ewer's once one one's one-armed one-eyed one-horse one-sided one-step one-time one-upmanship oneself ouija oija's ouij a-board ouija-board's ouijas u u's ubiquitous ubiquity ubiquity's ukase ukase's ukulele ukulele's ululate ululated ululating ululation ululation's unanimity unanimity's unanimous nuanimously unicorn unicorn's unification unification's unified uniform uniformed uniform uniform uniform ue uniquely uniqueness unisex unison unison's unit unit's unite united unitedly uniting unity unity's universal universality universality's universally universe university university's uranium uric urinal urinary urinate urinated urinating urine urine's urines usable usage usage's use use 's used useful usefulness useless uselessly uselessness user use 'S using usual usually usurer usurer's usurious usurp usurpation usurpation's usurper usurper's usury usury's utensil utnsil's uterine uterine's uterus uterus' utilitarian utilitarianism utilitarianism's utility utility's utilizable utilization utilization's utilize uvula uvula's uvular

Table 2 'em'un F F's H H's H-bomb L L's LSD M M's MP MP's N N's NB NHS R R's S S's SOS X X 's X-ray Xmas Yvonne f f's h h's hauteur heir heiress heirloom honest honestly honesty honorarium honorary honorific honor honorable honorably honour honourable honourably hour hourglass hourly l l's m m's n n's nb r r' s s s's xx's

From the above, it can be seen that part of the present invention focuses on the recognition that what is important in determining the exact usage of an indefinite article is the first sound emitted when a word is pronounced. I want to be understood.

After first establishing a limited list of exceptions, the following three rules apply: The first rule is that the word following the indefinite article "a" or "an" is the letter "eu"
Applies if it starts with. In this case, the indefinite article "a"
Must be used. The second rule is that the word following the indefinite article "a" or "an" is a vowel character "a",
It is applied when it starts with "e", "i", "o", "u". In this case, the indefinite article "an" must be used. The third rule is that the indefinite article "a" or "an"
Is applied if the word following the letter begins with a consonant letter. In this case, the indefinite article "a" must be used.

Referring to FIG. 4, it is determined by blocks 302, 304, 306 that each word w1 and subsequent word w2 of the input sentence 300 is determined by tracking the position of the current word i in the input sentence. You. If it is determined by 308 that the current word is not "a" or "an", the algorithm proceeds to block 30
Go to the next word via 4,306. The current word w1 is “a” or “an” and the next word w2 is
If found in block 310, it is specified by 312 that the current word w1 must be corrected to "a" if necessary. If the next word w2 is not found in Table 1 and is found by Table 314 to be found in Table 2, then if necessary, the current word w1 must be corrected to "an" by 316. It is specified. Otherwise,
Block 318 that next word starts with letter "eu"
The current word w1 if necessary
Must be corrected to "a" by 320. Otherwise, if it is determined by block 322 that the next word w2 begins with the letters “a”, “e”, “i”, “o”, “u”, the current word w1 is replaced, if necessary. 324 specifies that it must be corrected to "an". Otherwise, if necessary, 326 specifies that the current word w1 must be corrected to "a".

C) Correcting Incorrect Auxiliary Verb Sequences As mentioned above, non-native speakers often use incorrect intensions in complex auxiliary verb sequences when trying to write English sentences. One example is "he has co
nsider ". Here, the incorrect usage is the tense of the verb "consider". None of the current grammar checking systems check auxiliary verb sequences because it is clearly difficult to recognize auxiliary verb sequences and because part of speech tags are not usually calculated.

In the system of the present invention, referring now to FIG. 5, the sentence 410 is analyzed by the part of speech tag
The part of speech of the related sentence shown in 14 is derived.

To detect an error, the end and start of an incorrect auxiliary verb sequence must be detected.
For example, the sentence "he has been conside
In “rthis fact”, it is important to detect the end of the error, that is, the “consider”, which is the fourth word in the sentence. All words after "consider", ie, "this fact", do not affect the accuracy of the auxiliary verb sequence. Similarly, the starting point of the error, ie, "has", the second word in the sentence
It is important to detect All words before "has" are irrelevant for determining the correctness of the auxiliary verb sequence.

After generating the part of speech of the sentence, an end point detection step 424 is used to calculate the end position of the incorrect auxiliary verb sequence. To detect an incorrect verb sequence end, all correct part-of-speech sequences of all auxiliary verb sequences are
It is stored in the directed acyclic graph shown in FIG. 7 described below.

At block 422, from the directed acyclic graph of all correct auxiliary verb sequences, another directed acyclic graph corresponding to all possible incorrect auxiliary verb sequences is generated. After a graph corresponding to 422 has been obtained, the graph may be replaced with the incorrect auxiliary verb sequence "ha
ve-3rd-person verb-infini
active ". This corresponds to the incorrect auxiliary verb sequence "has consider". Following the graph from left to right to detect the end point of the error, the end state is reached while the input string is read from left to right. If the input sentence part-of-speech is read into the incorrect auxiliary verb sequence graph when the graph reaches its final state, since the part-of-speech corresponds to a word in the input sentence, this means that at the end of the auxiliary verb sequence, Uniquely identify. The identification of this word with respect to its position in the sentence is then indicated by the endpoint detector 424.

Similarly, the start point detector 426 detects a word corresponding to the start point of the auxiliary verb sequence. This is done by detecting the end point of the error and then traversing the graph in the reverse direction from right to left until reaching the start of the graph. For example, when going from left to right, the system:
"Has" to "have-3rd-singular"
As “consider” and “verb-i
nfinitive ". The system detects that there is an error in this regard and the word "cons"
ider "as the last word in the incorrect auxiliary verb sequence. Then, going backwards through the input sequence in the graph, it passes through "consider" and through "has". This reaches the beginning of this particular graph, thus identifying the word "has" as being the first word in the auxiliary verb sequence.

Referring again to FIG. 5, after determining the end of the auxiliary verb sequence, at block 428 the end of this incorrect sequence is determined as the last word in the incorrect sequence of the input sentence. At block 430, the start position of the incorrect sentence is determined as the position of the word that begins the incorrect sequence as a number reflecting that position in the input sentence. As shown in block 432, the other directed acyclic graph shown in FIG. 8 specifies a set of possible exact sequences for each incorrect auxiliary verb sequence.
Unit 432 then leads to the directed acyclic graph shown in FIG. 8 via the incorrect auxiliary verb sequence and outputs a set of possible correct auxiliary verb sequences for the user to view, as shown at block 434. I do.

Referring to FIG. 7, for every possible auxiliary verb sequence, a directed acyclic graph describing a set of possible exact auxiliary verb sequences is constructed. As can be seen from FIG. 7, on the left hand side of the graph, the starting point 44
From 0, "be", "were", "was", "i"
s "," am "," are "," been "," ha "
d "," have "," has "," could ",
"Should", "might", "may", "c"
an "," must "," would "," shal "
l "," will "," do "," does "," do
There is a box 442 containing all auxiliary verbs in English, such as "esn't". Words "be" to "been"
Is associated with node 444. In general, nodes specify that the verbs following their auxiliary verbs are the same. For example, the word "being" comes after "is". Similarly, in the case of the word "were", for example, it is "were being". Thus, node 444 indicates that there is a set of auxiliary verbs that are followed by the same verb. For example, a set of words "h
node 4 related to "ad", "have", "has"
After 46 comes the word "been". Similarly, in the case of the node 448, the word “have” comes after the words “could” to “will”. After these words, the word "do" comes. Finally, node 450
After the words "do", "does", and "doesn't", "have" is specified instead of "do".

This method of graph English idiom, in fact, absorbs all the rules into a compact graphical representation, thus obtaining the accuracy of the inaccuracy of the auxiliary verb sequence.

As can be seen from the figure, there is a box written “???” after the above-mentioned node. For example, box 452. Recall that the input to this graph is a sequence of words followed by a part of speech. This essentially tags the input using two variables. To keep the graph compact, the symbol "???" represents everything not described at this node. Referring to node 454, box 4
56 indicates that everything except "been" and "had" goes to node 458. Thus, it can be seen that the use of the box "???" indicates that symbols not described on the output of the state can be connected to the next node.

In addition to words, input sentences include parts of speech. For example, if the system reads the statement "haveconsider
When analyzing "ed", this graph is represented by the sentence "have h
aveconsidered vbn. "
“Vbn” represents a past participle. Starting from the left hand side of the graph, find the word "have" as shown at 460. From here, from the right past node 446, a box 462 that passes this word to node 454 as described above.
Proceed to. From node 454, possible words are 464
In “been”, “had” in 466, neither match the input sentence. Another alternative is node 45
Proceed to box 456, which allows you to pass to the right of 8,
Next, the process proceeds to a box 460 for designating “vbn” representing the past participle. This allows passage to node 470. The word under consideration is deemed acceptable as the analysis passes through box 456, which allows the sentence "have consulted" to proceed to the end point 472 of the graph. Between the intermediate node 470 and the end point 472, the symbol <E indicating an empty word
>. The use of a box denoted by <E> indicates that it is possible to go from one node to the next without considering such things as a following word or a following part of speech.

For words not found in box 442, they are placed in box 476 and node 48.
0 to the part-of-speech box 482, from which node 48
4 can be analyzed. Box 486 is an arc to the end point 472 (arc shape:
arc) or box 4 from node 484
Proceed to node 490 via 88, from which to part of speech box 492 or 494, and then reach endpoint 472. Finally, when node 484 combines words,
Proceed to node 496 and box 498 to go to node 50
Go to 0. Box 502 sets node 484 to node 5
04 to the part-of-speech box 506 and then to the end point 472 if appropriate. In 484, the word "havin
If it binds to "g" as well as "been", it sends it from node 508 to node 512 via box 510 and from there to node 516 via box 514. Thereafter, the process proceeds from the part of speech 518 to the end point 472 or to the box 520. Therefore, the input word “ha
wing "and" been "end point 4 via the above-mentioned route.
If it reaches 72, its proper usage is determined to be accurate. However, if the word "being" is added to this sequence, the output of node 512 will be
22 and box 524 to node 516.

In summary, a directed acyclic graph specifies all correct auxiliary verb usage. Thus, the same graph can be constructed for all incorrect auxiliary verb sequences. Thus, if a graph showing all the correct usages is constructed, a graph showing all the incorrect usages is immediately obtained. The compactness of this method is
It is very efficient when analyzing sentences.

Referring now to FIG. 8, using a finite state transformer in the form of a directed acyclic graph to suggest correction of an incorrect auxiliary verb sequence can be achieved by using the incorrectly generated incorrectly generated above. Determined by the non-ring graph of the verb sequence. Auxiliary verbs are paired to indicate proper correction, with the word on the left of each pair identified as incorrect and the word on the right being corrected. For example, after identifying that the auxiliary verb sequence "will had" is incorrect, the exact sequence is specified using the graph of FIG. Starting at input node 530, box 53
You can proceed to 2. The left side of this box is the same as the first word in the input. After passing through node 534 and box 536 and reaching node 538, the word under consideration is the word "had". Box 540
Indicates that "had" must be changed to "have". This is, in effect, output to node 542, from which it proceeds to end point 546 via box 544. After reaching end point 546 by this pass, the exact sequence suggested is "will have".

A more complicated case is a case where an incorrect sequence "world consid- ered" is considered. The corresponding part of speech tag is "world wo
oldconsidered vbn ". In this case, first, a box 5 indicating “would: would”
After passing through box 50 and proceeding from box 536 to box 538, node 538 is reached. Here, box 54
None of 0, 552, 554, 556, 558, 560 apply. This is because none of these boxes have the word "consider" in them.
Via box 562, appropriate and accurate suggestions via part-of-speech analysis box 564 are "worldconside
r ”. This is because the graph
"Considered" is the word "consider"
Reached for detecting that it was the past tense. This box suggests using the current tense, and thus suggesting the word “consider”. The analysis is "v
bd: / vbd / vb ". Note that vbd means past tense and vb means current tense. Node 538 that provides another exact change of the input
There are several alternative nodes from. For example, the suggested sequence may be "would have consid"
ered ". Here, box 566
Specifies that “have” is to be added. Box 5
Numeral 68 specifies that the part of speech hv of “have” is also added to the set in the case of a tag. After passing box 570, box 572 specifies that the past tense is to be changed to the current tense. In that case, the word "consid
"ered" is immutable because it is a past tense as well as a past participle. The input word is "considered"
Box 572 if not "new"
Specifies a change from the past tense “knew” to the present participle “know”.

The remainder of the graph of FIG. 8 is self-explanatory to make various suggested changes to incorrect verb sequences after they have been determined to be incorrect.

D) Inflection Correction for Non-Native Speakers As is usually the case, spell checking systems generally detect misspelled words via a dictionary lookup algorithm. While this is generally successful in detecting misspellings due to inadvertent keystrokes or character transpositions, these systems are ineffective against other types of misspellings. Most notably, non-native speakers' misspellings or, if not always, inadvertent transposition of letters in words, or inadvertent insertion or omission of letters, are primarily due to grammatical problems. For example, the sentence "He dri
Considering "ved his car yesterday", an error is not either carelessness or lack of knowledge of a particular spelling, in which case the verb "dri
uncertainty about the past tense of "ve".

In general, the spell checker suggests appropriate spelling based on the distance between the mistyped word and the word in the dictionary. This distance is typically the displacement, insertion,
Based on the number of characters that need to be transposed or deleted. The result is often strange. For example, in the example above, the exact suggestion is "drive"'s past tense, or "draw", but the current spell checker suggests "dried", especially "dripped". Note that the exact word "drive" is not suggested. This is because current spell checking systems do not analyze detected spelling errors for grammar.

Another example of a problem that current systems have when suggesting proper spelling is an inexact comparative adjective. For example, non-native speakers
When choosing a "good" comparison, one chooses a gooder based on the usual rules that often form comparison adjectives. As another example, if a non-native speaker wants to form the plural of the noun "child", then instead of "children" based on the usual rules for adding "s" to a singular noun, The word "childs" may be selected.

To show that the current spell checker cannot suggest a suitable word in the above example, typical spell checkers use the words “childs”, “child”, which are all incorrect in context. 's',
"Child's", "child", "childes"
Suggests. Even more inappropriate suggestions by current spell checkers are suggestions on how to spell "goodest" properly, namely "goodest" and "gooded".

These types of mistakes are annoying to native speakers, not only will they refuse to use the spell checking function, but also from within-context or out-of-context unfamiliar words. Non-native speakers have a higher level of frustration when forced to make a choice.

Referring now to FIG. 9, the present invention identifies typical examples of words that do not follow the usual rules for any of pluralization, past tense, past participle, comparative class formation, and superlative formation. It is important to. From this unique list of incorrect words generated based on the grammar, the system of the present invention suggests more appropriate replacement words. The spell checking system of the present invention typically operates to detect spelling errors through a dictionary lookup system. The correct word is then suggested based on a summary and morphology of representative incorrect words and prototyping, as described below.

In FIG. 9, the English word collector 600 includes:
A list of inaccurate English words 604 generated by comparing 604 the English word dictionary 602 and words from the English word dictionary 602 with words from the dictionary 606 generated by the usual rules 608 of English word formation. Contains. The result of the comparison is the above unique list of troublesome words based on incorrect grammar, not spelling errors.

Referring now to FIG. 10, in the process of actually correcting the detected inaccurate words, the English word dictionary 602 uses the previously generated inaccurate English words as discussed with respect to FIG. Used with the list 604 of FIG.
The detected incorrect word is obtained at 610 and is typically derived via a dictionary lookup. The list of inaccurate words detected, as well as inaccurate English words, is added to a unit 612 that determines the inaccurate word's original form and morphological characteristics such as tense, number, comparative vs. superlative. For example, for the incorrect word "driven", the original form of the word is "drive" and its morphological feature is "past tense or past participle". The archetypal and morphological features relate the archetypal and morphological features to the corresponding English words in the English word dictionary 602 to provide suggested corrected words, thereby also taking into account grammatical rules and exceptions. To the unit 614 that performs the processing.

In essence, a system that derives archetypal and morphological features based on typical incorrect usages will be able to suggest the appropriate words associated with these uncorrected usages. . The system performs a high level reference with the identified problem word that is a grammatical problem rather than a simple spelling error.

It should be understood that part-of-speech tags are useful in improving the accuracy of words suggested by the system. For example, a misspelled word is a past tense or a past participle. One example is "d", which can be "dove" or "driven" from the above.
lived ". Knowing the form in which the "inaccurate" word is used in the sentence allows the appropriate choice to be made based on part of speech.

E) Detection and Correction of Improper Use of Quantifiers One of the more difficult problems for non-native speakers
One is the problem of the usage of determiners. The qualifier is a word such as “the”, “a”, “some”, etc., which determines the referent of the noun phrase. There are three classes of errors related to determiners. The first is a missing determiner. For example, the sentence "John
"read book" is missing the qualifier of the noun phrase "book". A second class of determiner errors is the use of extraneous determiners. One example is "Joh
n sent to the New York ". Here, the qualifier "the" is inappropriate and should be deleted. A third class of determiner errors is that the noun associated with the determiner does not match. For example, "Joh
"n read many book" is "many"
And that the number of “book” does not match.

In order to detect improper usage of the determiner,
Use part-of-speech tags during analysis. For the part of speech tag,
This has been described with reference to FIGS. 2, 5, 15, 16, and 18. As an example of a tagged sentence, the sentence "J
"read read longlevel". Here, the tag for “John” is a proper noun, the tag for “read” is “past verb”, the tag for “long” is “adjective”,
The tag for "novel" is "singular noun".

As shown in FIG. 11, the system identifies the noun phrases shown in decision block 700. The decision block 700 identifies the noun phrases in the sentence by maximally matching the pattern defining the sequence of part-of-speech tags that make up the valid noun phrases. The pattern of the noun phrase is [DET] (MODS NOUN AND) * MO
Given by DS NOUNhead, the pattern of the MODS is given by (MOD + AND) * MOD. Here, DET, MODS, NOUN, and AND are defined as a set of part-of-speech tags of a determiner, a modifier, a noun, and a coordination conjunction, respectively. Notation [X]
Means zero or one occurrence of the enclosed expression X, and the notation (X) * means zero or more occurrences of the enclosed expression X. A + superscript, such as X +, (which is not actually a superscript) means one or more occurrences of the expression X.

The purpose of the above is to identify noun phrases. For example, in the above sentence, the noun phrase is “John” corresponding to the part of speech sequence “proper noun”, and “long” corresponding to the part of speech sequence “adjective singular noun”.
Novel ". The above uniquely identifies a noun phrase by identifying the beginning and end of the noun phrase. Noun phrase components are missing, irrelevant determiners,
Or it is very important to identify noun phrases to check for number discrepancies.

As shown at 702, when a noun phrase is found, the system tests whether the noun phrase's qualifier is missing. In this test, the entire noun phrase NP is checked, and the last word in the noun phrase, the first noun NOUN
Examine the head. head refers to the most important noun in the phrase, and has been found to be the last word in most cases. Also, in the test for quantifier omission, the qualifier DET of the noun phrase is examined. DET is the first word of a noun phrase or does not appear at all. FIG. 12 shows that the leading noun is a singular non-proper noun and DET does not exist.
704 and 706, the noun phrase is 7
At 08 it is tested to see if it is a name. The name is considered to be a capitalized phrase other than proper nouns. For example, "The Atlanta
"Policy Department" and "Gra
"dy Hospital" is a name. If the noun phrase is not known to be a name, the leading noun is tested at 710 to see if it is a collective noun. An aggregate noun is an unspecified amount of a substance, for example, "ri
ce "," fish ", and" carbon ". Note that collective nouns do not require a qualifier because they function effectively as plural nouns.

If the noun phrase is a name, an analysis is performed to determine whether the head noun is a collective noun, as shown at 712. A set noun is the same as a set noun, but appears in the name. For example, the sentence "She at
trended Harvard Universalsit
y ", the noun phrase" Harvard Universal "
"sity" is a name, and "University" is a set name noun. Therefore, "Universi
Note that "ty" appears in sentences without a determiner. Also note that a set noun is not the same as a set noun. For example, "University"
Is a set noun, but not a set noun. this is,
The sentence "She attached a fine uniti
version, the noun "universit"
It can be understood from the fact that the "y" has the qualifier "a". Therefore, it should be understood that no suggestion is made if it is determined that there is a set name noun.

However, there is a problem in the use of idioms.
As shown at 714, the noun phrase is analyzed to see if it is part of a idiom. This is done by looking up in the idiom dictionary. No suggestion is made if the noun phrase is part of a idiom. For example, the sentence "Th
In the event talk place, the noun phrase “place” is missing the qualifier, but is not suggested because it is part of the idiom “to take place”.

In the case of a singular non-proper noun without a qualifier, if the leading noun is neither a collective noun nor a collective noun and the noun phrase is not part of a idiom, the system will Suggest missing.

Next, referring to FIG. 13, the system comprises:
The noun phrase is checked for an unrelated qualifier (720 shown in FIG. 11). This is performed as follows. 724
As shown by, a noun phrase is introduced, and it is determined at 722 whether or not the leading noun is a proper noun by confirming whether or not a qualifier exists. If the above conditions are met, at 726 it is determined that there is an extraneous determiner. For example, "Johnwent to the Ne
"W York" is indicated as having an irrelevant determiner because the noun phrase "New York" includes the leading noun, which is a proper noun, and the noun phrase includes a qualifier, that is, "the". Proper nouns are identified by tags that determine the presence of the proper noun based on probability and context.

Referring again to FIG. 11, as shown at 730, the system of the present invention then checks the noun phrase for a number mismatch. FIG. 14 shows how this is done. The determination of a number match is, as shown at 732,
This is done by introducing the noun phrase into a detector that determines whether the leading noun in the noun phrase is a proper noun. This is because if the proper noun phrase contains a determiner, it has already been reported as an irrelevant determiner error. As shown at 734, assuming that the leading noun is not a proper noun, the system determines whether the noun phrase contains a determiner. If the noun phrase does not contain a qualifier, the problem of number mismatch does not arise.

As shown at 736, if a qualifier exists, the number of qualifiers is checked with the number of leading nouns to determine whether it is singular or plural. If they match, no error is signaled. On the other hand, if they do not match, a suggestion is made to change the number of leading nouns to match the number of qualifiers. Therefore, the sentence "John read onebo
In the case of "oks", it is suggested that the head noun "books" be changed to match the qualifier and singular. Similarly, for the sentence "John read many book", the system of the present invention suggests changing the head noun "book" to a plurality to match the qualifier. Alternatively, the system may change the qualifier to be different from the leading noun. However, this is the least likely strategy. The former gives better results because it is difficult to ascertain what the appropriate qualifier should be. Therefore, it is assumed that the correct quantifier for numbers has been properly entered.

In summary, the system of the present invention employs several techniques to detect noun phrases, head nouns, proper nouns, collective nouns, collective nouns, idioms using tagged sentences. It detects and corrects inappropriate use of the determiner. What is important for proper determination of abuse of the determiner is to detect noun phrases using pattern matching as described above with respect to FIG.

F) Recognition of proper nouns and other words that are essentially capitalized When analyzing a sentence, if the word is a proper noun, the word will be uniquely identifiable unlike all other nouns. It's pretty important to be able to recognize. In addition to proper nouns, "Harvard University
It can also recognize other essentially capitalized words, such as those appearing in names such as "y", so that the sentence can be parsed and understood, and thus the grammar analyzed. .

Words appear in capital letters in English sentences for two reasons. First, the words are proper nouns or other essentially capitalized words. Second
The word appears at the beginning of the sentence or after some punctuation, but otherwise is not capitalized. As an example, the sentence "Wells was an English"
Novelist, "Well
It should be understood that "s" is a proper noun and is therefore capitalized. The sentence "Wells were dug t
o provide drinking water "
, "Wells" is capitalized because it is the first word in the sentence.

Thus, in the first sentence, the grammar checking system recognizes that "Wells" is essentially capitalized and is therefore a proper noun. In the second sentence, the grammar checking system recognizes that "wells" is essentially not capitalized and is therefore a normal plural noun.

Prior to determining whether a noun was a proper noun, the system applied a relatively limited technique for recognizing words that were essentially capitalized. One approach has been to assume that the first word of the sentence is essentially not capitalized. This is useless for sentences that begin with proper nouns, as indicated by the first sentence.

Another approach is to use all words as proper nouns,
Or to distinguish them as ordinary words (but not both). From the above two sentences, "Wells"
Can be both proper nouns and ordinary words, so that this type of classification system is useless.

An obvious problem of not being able to properly identify whether a word is a proper noun is that when referring to a dictionary,
The wrong definition is searched. Simple grammar checks do not require a definition, but advanced word processing and grammar checking systems that provide instructional and informational data when determining appropriate usage will require proper nouns and other essentially uppercase letters. You need to identify exactly which words you have. Even when the dictionary lookup function is not part of the grammar checking system, the identification of proper nouns and other essentially capitalized words is important.

The importance of identifying whether a word is a proper noun affects the behavior of the part-of-speech tag that must use trigram probabilities to accurately determine the part of speech of each word in a sentence. Word capitalization and non-capitalization are
As having different trigram probabilities, it is important for Taga to know which form of a word is present in a sentence in order to apply accurate trigram probabilities. For example, the trigram probability of the proper noun “Wells” is the common noun “we”.
lls ”is different from the trigram probability. Therefore, Taga has the sentence "Wells was an Engl".
issue novelist), the word "Well"
s "is a proper noun and therefore" We
An uppercase trigram probability of "lls" must be applied.

To establish whether a word is an ordinary word rather than a proper noun or other essentially capitalized word, the system of the present invention employs one of two interpretations of each word. That is, whether the word is interpreted as a proper noun or a common noun is determined. The system does this by generating two forms of the sentence, one assuming the noun is a proper noun and the other as a common noun. The system then compares the trigrams of the two sentences. If a sentence that assumes that a word is a proper noun has a higher probability, the word is considered to be a proper noun. Otherwise, the word is considered a common noun.

Referring now to FIG. 15, there are two steps in the decision process to determine whether a noun is a proper noun. The first step, shown at 800, is a pre-processing step, starting at 802, with a tagged training corpus. This is related to a set of sentences where each sentence word is tagged with its part of speech. Next, the training corpus 802 contains 8
As shown at 04, the word is revised to capitalize words that are not proper nouns or words that are generally not essentially capitalized. If a word is tagged as a proper noun or name, or is an acronym, or is a pronoun "I", the word is considered to be essentially capitalized. In addition, if a word appears at the beginning of a sentence or after an open quote or colon, and if the word appears only at the beginning of the sentence or only after an open quote or colon, they are non- Capitalized.

More specifically, as shown in FIG.
At 10, a tagged training corpus 8 to get the next word / tag pair from the corpus, if any
Analyze 08. If found, at 812, analyze the word to see if it is capitalized. If the word is capitalized, check to see if the word is the first word in the sentence or if the word appears after an open quote or colon, as shown at 814. If so, as shown at 816, the word is tested to see if it is tagged as a proper noun or name, or is an initial or pronoun "I". Otherwise, 8
Capitalize the word in the revised training corpus shown at 18.

Next, referring again to FIG.
Analyze the revised training corpus to obtain a trigram probability model of the word. This allows
To eliminate the error associated with misrecognizing a word as a proper noun when it is in fact a common noun, or misrecognizing a word as a proper noun when it is in fact a proper noun A modified trigram model is obtained. After pre-processing the tagged training corpus to eliminate errors, at 822, a decision is made using the trigram model to determine whether the word in question is essentially capitalized. It requires a word in the sentence as input, and the output is the spelling of the base of the word.

As shown in FIG. 17, the decision making process described in 822 (FIG. 15) for determining whether a word is essentially capitalized, as shown in 850,
Start with a word in the sentence. Analyze this word to determine if its first letter is uppercase. Otherwise, as shown at 851, the interpretation of the word is given literally by the sentence. That is, if a word appears to be capitalized in the sentence, the word is interpreted as a proper noun. If the word does not appear to be capitalized in the sentence, the word is interpreted as a common noun. Therefore, no special action is taken if the word is not capitalized.

Next, assuming that the word is capitalized, as can be seen from 854, whether the word is the first word in a sentence, or whether the word comes after an open quote or colon Determine whether or not. Otherwise, no further action is taken. If so, 8
As shown at 56, the word is further processed to see if it is an initial. An initial is identified by all its alphabetic letters being capitalized or by being in the initials dictionary.
If it is determined that the word is an initial, no further processing is performed again.

If the word is not an initial, as shown at 858, the system treats the sentence as two forms, a proper noun with the word in question capitalized, according to the trigram model shown at 859. The other,
Calculate the probabilities Pc, Pu of the word in question being treated as a common noun without capitalization. The calculation is as described according to the part of speech tag described above.

As shown at 860, if the probability Pu of a sentence having a word that is not capitalized is higher than the probability Pc of a sentence that has a word that is capitalized, the system may consider the spelling of the word not being capitalized. Is returned as the most likely underlying spelling, so that this spelling can be used for later grammar checking. Otherwise, as shown at 864, the system returns the capitalized spelling of the word as the most likely underlying spelling.

It should be noted that restoring the spelling of the underlying word can make the grammar checking system more accurate and more useful. Restoring the underlying spelling requires the following two steps. The first step is to correct the part of speech in the training corpus for errors caused by misidentification of whether the word is a proper noun. The second step performs a series of analyzes to see if the capitalized or uncapitalized spelling of the word is more appropriate. This is done by the decision factor using the revised trigram probability model obtained in the preprocessing step to determine whether the word is essentially capitalized.

G) Context-based dictionary access When writing text, non-native speakers make use of the availability of a monolingual or bilingual dictionary. Dictionaries are one of the most useful sources of language for non-native speakers. It should be understood that the use of dictionaries is not limited to grammar checking problems, but is generally useful when writing text. Also, it should be understood that native speakers also make extensive use of dictionaries or thesaurus when constructing text.

Accessing dictionary entries is not as straightforward as one might expect, as words outside of context are very ambiguous in their syntax behavior as well as in their meaning. A given word in the dictionary is generally 20 words, 30 words,
Or have more headwords. Due to this large number of headwords, the use of dictionaries is very time consuming.

For example, outside the context, the word “left” is
It has a number of headwords in the English dictionary: As in the sentence "His left arm", the adjective "left
t ”, the sentence“ he moved ref
Tongue for the adverb “left”, such as “t on entering the room”, the sentence “Make
a left at the next cone
a headword for the noun “left” such as “r” and a headword for the past tense “left” of the verb “leave” such as the sentence “He left a minute ago”. However, when the word "left" appears in an English sentence, only one of the headwords is relevant to the context. Currently, there is no dictionary that can access the exact headword of a word based on context.

The system of the present invention selects dictionary headwords and ranks them based on the part of speech assigned to a given word in context. First, a headword corresponding to the word in the context is selected. Other headwords not relevant to the current context are still available at the request of the user. The part of speech of a given word in context is defined by the part of speech tag described above.

As an example, the sentence “He left a mi
Assuming the word “left” in “nute ago”, the part of speech tag assigns the tag “verb past tense” to the word “left” in the sentence. In this case, the system of the present invention uses the verb "l" corresponding to the usage of "left".
eave ", then a headword for" left "that is not used in that context,
In particular, the headword for "left" is selected as an adjective, an adverb and a noun.

The sentence "It has several ba"
Assuming the word "bases" in "ses", the part of speech tag assigns two tags "noun plural" to the word "bases" in the sentence. Out of context, the word "ba
ses ”is the plural of the noun“ basis ”and the noun“ ba
It should be understood that this can be the plural of the "se", as well as the third person of the verb "base". Context "It has s
In the case of "every bases", the system of the present invention selects headwords for the nouns "base" and "basis" corresponding to the word "bases" in that context, and then selects "bass" that are not used in that context.
Select a headword for "es", especially for the verb "base".

Next, referring to FIG. 18, in order to select a headword of a word appearing in the sentence 900 from the dictionary based on the context, a pair of the original form and the part of speech corresponding to the word irrelevant to the context is calculated. The word is analyzed by a morphology analyzer 910. As an example, for the word "left", the morphological analyzer outputs the following pairs of original and part of speech. ("Left", "adjective"),
("Left", "adverb"), ("left", "singular noun"), ("leave", "verb tense") The morphology analyzer 910 indexes by all inflections of all English words. It operates by examining the attached table and its headword, which is a set of original and part-of-speech pairs, and a part-of-speech tag 93 in the context to generate a unique part-of-speech tag T940 corresponding to the word in the context.
The word is analyzed by 0. This is Applied Natu
ral Language Processing, Austin Texas, Second in 1988
`` A Stochastic Parts Prog '' at Conference Proceedings
ram and Noun Phrase Parser for Unrestricted Text ''
Kenneth Church's Stochastic Parts Progra
m) by the part of speech tag available as an embodiment.

For example, if a word is in context "He left a
If it is “left” in “minute ago”, the part of speech tag outputs a part of speech tag “verb past tense”. To separate the morphological primitive corresponding to the context from the morphological primitive that does not correspond to the context, the unit 920 converts the pair 920 of the primitive and the part of speech into the set 950 corresponding to the part of speech tag 940 and the part of speech tag 940. No pair 960-2
Divide into two sets. In the previous example, the pair of the original form and part of speech pair corresponding to the context is (“leave”, “verb tense”). A pair of a primitive and a part of speech that does not correspond to a context is (“left”, “adjective”), (“left
t "," adverb "), (" left "," singular noun "). To display the headwords from the dictionary corresponding to the context, at 980 all the headwords in the dictionary 970 corresponding to the root forms found in the pair of the original and part of speech pairs corresponding to the context 950 are displayed. In the above example, the verb "l
All headwords for "eave" are displayed as context-related headwords. At 980, display all headwords in the dictionary 970 corresponding to the archetype found in the original and part-of-speech pair set corresponding to the context 960 to display all headwords from the non-contextual dictionary I do. In the above example, all headwords for the word “left” are displayed as headwords that are not relevant to the context.

It should be understood that for native or non-native speakers, the ability to select headwords from a context-based dictionary can be used for monolingual as well as bilingual dictionaries. The system of the present invention
Entry words that are relevant to the context can be selected, thus greatly reducing the number of entry words that the user must read.

While several embodiments of the system of the present invention have been shown, those skilled in the art will appreciate that modifications and variations can be made within the spirit of the present invention. Accordingly, the scope of the invention is limited only by the following claims.

[0132]

As described above, the grammatical inspection system according to the present invention includes means for analyzing a sentence for each word to determine whether the grammar is accurate, including means for tagging the words of the sentence with respect to part of speech, Module means for detecting improper usage; module means for determining whether the word is essentially capitalized; module means for detecting improper qualifiers, missing qualifiers, and irrelevant qualifiers; Module means for detecting improper inflections, module means for detecting improper auxiliary verb sequences, and module means for accessing the dictionary based on the context of the dictionary and the words in the sentence. Means for suggesting a word that has been corrected based on the output of one or more of the module means. It can more accurately display an effect that it is possible to correct such errors without excessive correction.

[Brief description of the drawings]

FIG. 1 is a block diagram of a complete grammar checking system showing various modules used for a full grammar check according to Embodiment 1 of the present invention;

FIG. 2 is a block diagram showing a method for correcting a word that is easily confused by using a part of the probabilities of a part of speech sequence according to Embodiment 1 of the present invention;

FIG. 3 is a block diagram illustrating a method for correcting words that are likely to be confused for the probability determining component of FIG. 2 using sentence length averaging to determine a probabilistic threshold according to Embodiment 1 of the present invention; FIG.

FIG. 4 uses a special table of English exceptions based on a method of pronouncing a word differently from a method of spelling a word to obtain an appropriate usage of an article according to the first embodiment of the present invention. FIG. 4 is a diagram showing a method of correcting indefinite articles “a” and “an”.

FIG. 5 is a block diagram illustrating a method for correcting an incorrect auxiliary verb sequence by obtaining a corrected sentence using a start point detector and an end point detector according to Embodiment 1 of the present invention; .

FIG. 6 is a block diagram of a configuration of a start point detector and an end point detector of FIG. 5 using a directed acyclic graph indicating an accurate verb sequence according to Embodiment 1 of the present invention;

FIG. 7 is a directed acyclic graph showing a set of English auxiliary verb sequences according to Embodiment 1 of the present invention;

FIG. 8 is a directed acyclic graph showing a finite state transducer that suggests correction for an incorrect auxiliary verb sequence according to Embodiment 1 of the present invention.

FIG. 9 calculates an incorrect word list corresponding to a non-genuine English dictionary based on the usual rules of English formation applied systematically to all English according to the first embodiment of the present invention. FIG. 3 is a block diagram illustrating an improved spell checking system for non-native speakers.

FIG. 10 uses original form and morphology in analyzing inaccurate words detected against a previously generated inaccurate English list, according to Embodiment 1 of the present invention; FIG. 3 is a block diagram illustrating a process of correcting an incorrect word detected using an incorrect word dictionary obtained from the system of FIG. 1 as well as an English dictionary.

FIG. 11 is a pattern for defining a sequence of part-of-speech tags constituting a valid noun phrase, including a check for confirming a missing determiner, an unrelated determiner, or a number mismatch according to the first embodiment of the present invention. Identifies noun phrases by maximizing matches, detects inaccurate use of qualifiers,
5 is a flowchart used when making a correction.

FIG. 12 shows whether a noun phrase is a name or a qualifier,
9 is a flowchart illustrating a method of determining whether a set noun, a set noun, or a part of a idiom, and reporting a missing determiner.

FIG. 13 determines whether the leading noun is a proper noun according to Embodiment 1 of the present invention, and then checks whether the noun phrase includes a qualifier, thereby reporting an irrelevant qualifier 4 is a flowchart illustrating a method of checking a noun phrase for an unrelated quantifier by performing the following.

FIG. 14 shows whether the leading noun is a proper noun or, if not, whether there is a qualifier and whether the number of qualifiers matches the leading noun according to Embodiment 1 of the present invention. 5 is a flowchart illustrating a method of checking a noun phrase for a number match, by making a decision by detecting a match and thereby reporting a mismatch.

FIG. 15 uses a revised training corpus for non-uppercase words that are not proper nouns or essentially not capitalized using the pre-processing module according to Embodiment 1 of the present invention. Directs the trigram model to the revised corpus to see if the noun is a proper noun, recognize proper nouns and other essentially capitalized words, and spell the base of the word FIG. 6 is a block diagram showing a method of recovering the spelling.

FIG. 16 is a flowchart illustrating a method for tagging the training corpus of FIG. 15 according to Embodiment 1 of the present invention; Use this to get the next word / tag pair and see if it is capitalized and if the word is the first word in a sentence or if it comes after an open quote or colon And whether the word is tagged as a proper noun or name, or whether the word is an initial or a pronoun "I", thereby checking whether the word is capitalized To test.

FIG. 17 according to Embodiment 1 of the present invention, as shown in FIG. 15, whether a word is capitalized, it is the first word of a sentence or comes after an open quote or colon Analyze whether the word is an initial, and if not, determine the probability of a sentence having a word that is not capitalized and calculate it from the probability of a sentence having a word that is capitalized 4 is a flow diagram illustrating a decision-making process for determining whether a word is essentially capitalized by checking if it is higher.

FIG. 18 shows a headword in a dictionary corresponding to a word used in a context and a word used in the context, using a part-of-speech tag and a morphology analyzer according to Embodiment 1 of the present invention. FIG. 9 is a block diagram illustrating a context-based dictionary access that determines a headword in the dictionary that does not correspond to the word at the time of entry.

[Explanation of symbols]

10 input sentences, 12 keyboards, 14 CPU, 16
Word processing system, 20 part-of-speech analyzer, 22 part-of-speech sequence probability determination unit, 24
Part-of-speech confirmation unit, 26 Accurate sentence selection unit, 28
Underground spelling recovery unit, 30 inflection check / correction system, 32 auxiliary verb correction unit, 34 exact sentence selection unit, 36 qualifier correction unit, 38
Word insertion / detection, 40 context sensitive dictionary reference module, 42 indefinite article correction module, 44 spelling system.

 ──────────────────────────────────────────────────続 き Continued on the front page (71) Applicant 597067574 201 BROADWAY, CAMBRIDGE, MASSACHUSETS 02139, U.S. S. A. (72) Inventor Yves Chaves United States, Massachusetts, Boston, Boylestone Street 770, Apartment 12 Bee (72) Inventor Emanuel Roche United States, Massachusetts, Boston, Commonwealth Avenue 37 (72) ) Inventor Andrew Earl Golding Five Frost Street, Cambridge, Massachusetts, United States, Apartment 1

Claims (1)

[Claims] 1. A means for analyzing a sentence for each sentence for correct grammar including a means for tagging a sentence word with respect to a part of speech, a module means for detecting inappropriate usage of an indefinite article, Module means for determining whether a word is capitalized, module means for detecting improper qualifiers, missing qualifiers, and irrelevant qualifiers; module means for detecting improper inflection of the word; Module means for detecting various auxiliary verb sequences,
And means for suggesting a corrected word based on the output of one or more of the module means coupled to the system, among the module means for accessing the dictionary based on the context of the word in the dictionary and the sentence. Grammar inspection system with.