JP2003196282A - Automatic natural language translating system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an automatic natural language translating system. <P>SOLUTION: This automatic natural language translating system for receiving a text sentence in source natural language (preferentially, English), and for translating it into target natural language (preferentially, Japanese) allows an operator to automatically re-translate the selected part of the source text. This system is improved to make more accurate the first translation of a document including words and phrases having a different translation in a different field, to designate several grammatical rules as a control means to remove any improper translation, and to use the value of probability associated with the entry word of a dictionary to remove any improper translation. <P>COPYRIGHT: (C)2003,JPO

Description

DETAILED DESCRIPTION OF THE INVENTION [0001] BACKGROUND OF THE INVENTION The present invention relates to a natural language
Automatic translation from other natural languages, preferably from English to Japanese
About translation. [0002] 2. Description of the Related Art Various types of machines are available for machine translation of natural languages.
Formulas have been proposed. The system usually used for translation
Receives input in a language and processes the received input.
Including computers that output in other languages. This
Traditionally, translations of Ip have been inaccurate, and
If the trained operator does not significantly modify the output
No. Translation work using conventional systems is generally
Includes the conversion operation. The purpose of this structural transformation is the source language
A parse tree for the sentence (ie, a syntax tree)
It is to convert to a tree corresponding to the language. to date
Two types of structural transformation schemes have been tried. Sand
That is, conversion based on grammar and conversion between templates
You. In the case of conversion based on grammar, the structure conversion
The domain is given in the source language parse tree (ie, given
Set of subnodes that are direct children of
Is limited to the area of grammar rules used for example
If we have VP = VT01 + NP
Lyrics and noun phrases are arranged in this order. ) And for Japanese, 1 + 2 => 2 + 1 (VT01 and N
The order of P is reversed. ) [0004] A saw including the application of rules
The linguistic tree of language is structurally the order of verbs and objects.
Is converted to be the opposite. Verbs are purpose in Japanese
Because it comes after the word. This method uses the source language
Exactly where the rules were used to get the parse tree
Applies and allows you to quickly find where certain transformations have taken place.
It is very efficient in that Meanwhile, on
As mentioned, the territory is said to be severely restricted
Points, and natural language spans non-child nodes
In that it may require such conversion rules.
The exchange mechanism is weak. In the conversion between templates, the structural conversion is
In the form of input / output (I / O) templates or subtrees
It is specified. A structure tree with an input template
If they match, the structure tool that matches the template
Directory part is specified in the corresponding output template.
Converted to cages. This is a very powerful conversion mechanism
Is the structure tree where the given input template exists.
Time to find which part of the match
And the processing cost may increase. [0006] SUMMARY OF THE INVENTION The automatic natural language translation of the present invention
Translation systems have many advantages over traditional machine translation devices.
With benefits. The system automatically enters the text information
Provide the most appropriate translation possible and provide the user with the output (good
Suitably after supplying a Japanese translation of the English input text)
In order for the user to modify the displayed translation result
To obtain other translations automatically or automatically.
Can interact with the stem. Automatic natural language of the present invention
The person who operates the translation system decides that this is good.
Keep the rest of the translation and keep the rest
It can be re-translated automatically. This selective retranslation
By performing the translation, the parts that need retranslation
Only the translation takes place, so the operator saves time
As there are potentially many inaccuracies
Also considers very high quality translations
You can do tedious tasks quickly. In addition,
The system can make various translation adjustments.
And usually the final structure of the translation is often
Created. Therefore, using this system
Reduces potential errors by humans (operators)
Modification of combs, sentence structure, personality and tense matching, etc.
Saves time. This system is an operator
Can provide extensive and accurate grammar and spelling knowledge
It has become so. The automatic natural language translation system of the present invention
Various ambiguities in sentence breaks in source languages
Translations are more accurate thanks to the power of meaning and powerful semantic communication
And the time required for operator translation rework
Will be shorter. Tailored to specific user preferences
The learning statistics that the system remembers
Translation quality is further improved. Idiom processing method of this system
Is used when there is a sentence that contains a phrase
It is possible to translate correctly without considering the meaning of the word itself
It has the advantage of being able to. If the system is efficient
Diversified to match less relevant characteristics
Has functions. The structural balance expert of this system
And the peer structure expert,
Efficiently distinguish between no analysis. Capitalized Expert
Efficiently interprets uppercase words in sentences accurately and
String procedures completely ignore interpretations as common nouns.
Instead, it processes compound proper nouns efficiently. [0008] From a certain point of view, the present invention provides an automatic
This is related to the improvement of a natural language translation system.
In this case, the improvement is "automatic region identifier" to help translation
About the use of Areas may contain certain uses or certain
Includes a collection of terms and usage patterns belonging to people.
For example, business domains, marketing
Technical documents such as computer documents, computer documents, and physics.
Documents in the field can be considered. In words included in the dictionary
Is the target natural language when used in a specific area
(E.g. Japanese), but used in other areas
If used or used regardless of discipline,
The meaning may be different. Squirrel of domain keywords
Is also used. Keywords are the areas associated with each area
A unique word or term that is a source natural language (or source
To which area a particular sentence belongs to
Used to determine "Automatic area identifier" function
Means that the sentence (or document) is a sentence in a specific area
Sufficient keywords are used in the sentence (or its
Document or part of a document)
to decide. If the area can be identified, it is listed in the dictionary
The words that are also included in the sentence (or document) being translated
Use a higher probability value and use the sentence (or document)
Is translated. The determination by "automatic area identifier"
Based solely on source natural language and keywords
You. The "automatic region identifier" function of the present invention
Area mismatch analysis (not fit to specific area)
Analysis), so the translation time of the system
Be faster. In another aspect, the present invention is directed to an automatic nature
Another improvement of the language translation system. Its improvement
This is because "almighty" or "part of speech priority"
Or use the grammar rule "unmarked".
Parsing of source natural language sentences used. "Oh
The "Lumighty" grammar rule allows the same part of the sentence
Rule that excludes all other rules that parse
You. The "part of speech priority" grammar rule is
Rules that exclude all other rules such as parsing parts
But this is the case when other rules have the same part of speech.
It is done only in case. No other grammar rule
It does not have priority over the discrepancy rules. "All my
Grammar rules such as
If there is no rule, it is created by the system's translation engine
Irrelevant grammar interpretation tree from the tree structure
This is an effective way to remove. With such a mark
Grammar rules are more effective in terms of time and memory
Analysis can be performed In other words, the translation of the system
Time can be reduced. Because the translation engine
Is a big thing that could be made without marking
This is because there is no need to create a tree structure. In addition, translation
The engine must remember if not marked
Note that you do not have to remember large tree structures
(For example, RAM) area can be released. From a further point of view, the present invention provides
Still another improvement of the dynamic natural language translation system. This
Very likely to be a component of the final tree structure
Headwords marked low or absolutely absent
An improvement related to parsing using dictionaries with. dictionary
Contains compound entry words, and at least
Even if one compound has a very low probability value
This low probability value indicates that the headword is
Even a correct parse tree for any sentence in a word as a particular part of speech
Indicates that it is unlikely that it will work. An input
In the parse tree structure of the statement,
If the term is included, the heading marked
Because the probability value associated with a word is low, its parse tree is
Is very small for a sentence of this
By using the probability values, the system's translation engine
Unnecessary parts of the tree structure created by
Wear. The above and other objects, features and characteristics of the present invention.
And advantages will be apparent upon reading the following description and claims.
become. [0012] BRIEF DESCRIPTION OF THE DRAWINGS FIG.
The outline of the automatic natural language translation system of the invention will be described. This
Will be described with reference to the drawings. [0013] The automatic natural language translation system uses
The language can be translated into the target natural language. Suitable tool
As an example, the above system translates English into Japanese.
The above system is designed to convert source natural language to target natural language.
A device for receiving and storing a translation engine for producing translations;
A device for displaying translation results to a user;
And a device for providing and displaying another translation result.
No. As a specific example of this system, the translation engine
Pre-parser, parser, graph creation device, evaluator, group
Rough scorer, grammar structure extraction device and structure conversion device
including. The preparser examines the input text and
Analyze ambiguous parts of sentence end recognition. After that,
Is an analysis chart that includes dictionary headwords.
Create and display text. Parser input text
To obtain a possible syntactic classification for
Analyze the report. The graph creation device converts the analysis chart
Create a graph of the possible parsing of the input text based on
You. This chart shows the possible interpretations of the input text
Nodes and sub-nodes. A series of experts
The evaluator containing the list evaluates the interpretable graph and
Add expert weights to nodes and subnodes
You. Graph scorer is used to evaluate subnodes
Use the expert weights and then the N top
Associate the core with each node. The grammar structure extraction device
A suitable solution whose parse tree structure is determined by the graph scorer
Assign to The structure conversion device translates in the target language.
Perform a structural transformation on the parse tree structure to obtain
U. In the following three paragraphs, (a) each sub node
Graph to calculate the final weighted score for the
How Scorer Combines Expert Weights
(B) To reach the final node score,
How Rough Scorer builds subnode scores
(C) the language information is the number of nodes and subnodes
Describe how to communicate Lee. Calculate the final weight score for each subnode
For calculation, a graph scorer is defined for each subnode.
Associate numerical values. Language information related to each subnode
The analysis determines the score of the subnode. An example
For example, a set of expert evaluators is
Checking the linguistic information stored in the code See FIG.
No. A graph scorer can be used for a particular node or subnode.
To obtain the final weighted average for
Calculate the sum of the individual weighted scores for the report. Multiple
The weighted scores of the two into one weighted average score
Is a standard problem in computer science
You. One method that can be used is
The result contains the constants assigned to the expert
There is a method of multiplying them. Assign to each expert
The weight to be assigned is a matter to be determined at the time of design.
The designer assigns priority (weight) to each expert
decide. A weighted average is a series of numbers multiplied by a constant
It is the sum of the characters. For example, the following equation is obtained. Weighted average = (w₁) * (X₁) + (W_Two) * (X_Two)
+. . . + (W_n) * (X_n) Where weight w₁, W_Two,. . . , W_nAre not negative
The total number is 1. For example, regarding the expected value of statistics
Spiegel discusses the use of
Rate and Statistics Theory and Problems 76 ”(1975, Tuna
-Hill Inc.). In order to obtain the final node score,
Graph scorer to connect
Is a subnode from the bottom to the top of the graph
・ Score can be transmitted. Each node has N score
In the case of a graph with a set of
A method can be determined. Propagate subnode scores
One technique that can be used to reach
Certain dynamic programs used to solve optimization problems
There is a storage method that is ramming. For solving optimization problems
Can include many possible numbers (results). Eye
The goal is to find the optimal number. Used for optimization
Algorithm solves each sub-sub problem only once,
The result is memorized, so every time you encounter a sub-sub-problem,
There is no need to calculate again. Applied to optimization problems
For a more detailed description, see, for example,
"Invitation to Algorithms" (McGraw Hill, 1990
Pp. 301-314. This "A
301, 302 and 312 of "Invitation to Lugo Rhythm"
The page conveys subnode score information in the graph.
One method that can be used to accomplish this is described. When transmitting linguistic information in a tree,
The part that conveys the meaning of the system is
Communicate from smaller components to larger ones
Works. Semantic communication is used in parsing operations
Four classes of syntactic classification (SEMNP, SEMV
P, SEMADJ and VERB). meaning
Before the communication is performed, the linguistic information stored in the node is
Information must be analyzed. Stored in the node
Analysis of semantic information is based on the noun-like elements and verbs
By checking new elements, you can determine which noun-like object
The optional restriction slot of the verb-like element of the grammar rule
Is guided by a set of rules that guess
You. Gerald Gazda writes in his book, The Prologue
Natural Language Processing ”(Addison Wesley, 1989)
Company), a direction similar to the graph disclosed in this specification
Is stored in the nodes of the non-ring shaped graph with
Rules that can be used to analyze semantic information
Is explained. Gazda has an adjacent
Use of property matching to match information about
Is explained. Gazda has the following characteristics matching
It includes the expression described in. "Some properties appearing on a node
Is the same property that appears on another node
Must be. Recent research on parental classification and their characteristics
The features of certain classes that also show related morphemes
The principle of equality of gender is assumed. this
The child is called the "head" of the phrase. Most
The phrase has only one head. So, for example, a verb phrase
Inherits the verb's tense. Because verbs are verbs
Because it is the "head" of the phrase. Table used so far
This source, which uses notational resources but is applicable to the entire grammar,
There is no easy way to specify rules. But if relevant
The properties can all be found on one branch of the DAG
If we assume that this principle is very simple for each rule
Can be described. So we have the usual V
The P rule can be written as: VP-> V NP PP <Head of V> = <Head of VP> In this case, the characteristic value of the “head” of V and the value on the parent VP
The characteristic values of "head" must be the same. " The rules discussed in Gazda have the structure disclosed herein.
It can be easily applied to each sentence classification. Gazda's rule
The linguistic information assigned to each node using the rules
With the technique of storage, you can propagate through the tree
You. Here, the contents of the above three paragraphs are summarized.
And the weighted average is one of the sub-node scores
Method, where each subnode score is appropriate for the optimization problem.
Using a well-known storage technique
The method described in Gazda's book
Used to analyze linguistic information stored in each node
This linguistic information can use storage method technology
Can be transmitted in the grammar structure analysis chart
It will be said. In an automatic natural language translation system, the first
After the dynamic translation, the translation can be automatically re-translated.
That is, the system can automatically input text information
To provide the most appropriate translation and output it to the user (preferably,
After providing the Japanese translation of the input English text),
The user can modify the displayed translation or
Interact with the system to obtain another translation dynamically
Can be. An automatic natural language translation system converts one sentence
Use a language model that breaks down into substrings. sub
The strings appear in the order specified as part of the statement.
One or more phrases that appear. For example, Subst
The rings “The man is happy” are “The”, “The ma”
n "," man is happy. "," is "and" The man is
happy "it contains white but" is man "," man m
"an" and "The is" are not included. Different language models can be implemented in various ways and
Define substrings at various levels of detail. example
For example, in the sentence "They wou1d 1ike an arrow."
Means that "an arrow" is usually categorized as a noun phrase (NP)
You. In another model, "an arrow" is a syntactic property (eg,
For example, singular noun phrases) and literary characteristics (weapons)
Classify. If the meaning of this phrase is ambiguous,
There are several ways to categorize, for example, "an arrow"
It can also mean an arrow-shaped symbol. Language model
Provide a way to resolve ambiguities,
By combining small units with larger units,
Resolve ambiguity. When evaluating larger units, this
These models only use information contained in larger units.
Be considered. As a specific example of this system, "an
 The meaning of the arrow (sign or weapon) is changed to "They
 "1ike an arrow" in the sentence wou1d 1ike an arrow.
Used to evaluate the verb phrase On the other hand, "an a
The syntax of the phrase "rrow" is "Heshot it with an arrow."
In the sentence, the meaning of "an arrow"
The characteristic is to comment on the verb phrase "shot it with an arrow".
Not used for value. A specific language model (interpreted substring)
Any substring of a sentence that is interpreted in one way
For which the exported attribute exists,
The ported attributes are the interpreted substring and the
And other units that form large substrings
All attributes used to evaluate the match. D
Kissport interprets along with the exported properties
It is an interpreted substring. Interpreted sub
Included in the string but not exported
Attributes are called substructures. The system parser is a grammar database
including. The parser uses grammar rules to
Find possible interpretations. The grammar database is X = A1
A2. . . From a series of context-free phrase construction rules in the form An
Become. X is A1, A2,. . . , An
Or formed, high-level nodes (subnodes
D) A node (subnode) A whose level is lower than A1
Called n. [0027] The graph creation device of the system is a single sentence.
Many possible interpretations of this are represented graphically. Each no in the graph
The code corresponds to the export of a substring.
As an example of a system, one export is one
Represented by a node. The graph is related to one export.
Includes arcs from connected nodes, where the arcs are
Represents an export substructure based on the application of a rule. The
Roughs have at least two types of arcs:
Represents the Sport substructure. At least two graphs
Type of arc, ie, (1) the same substring
A single arc, pointing to one different export of
(2) A set of pointers pointing to two exports
Two arcs, that is, when connected,
Substring is the substring of the original export
Form a tag. The formula of (2) is the normal form of Chomsky
Note that grammar is assumed. Amended billing
Item 35 is, by paraphrasing type (2),
N exports, not Jomsky normal form grammar
Syntax that reflects an arc with an N-fold pointer pointing to
Applied. The graph shows one starting export point S
Include, and from there, a series of arcs
You can reach all parts. Departure export
Corresponds to the entire sentence. Is the same export a plurality of exports?
Multiple arcs from one node only if they can be formed from
Starts. (A set of two arcs
Pointers are not considered multiple arcs in this sense. )
The export is one element of multiple exports
Only in some cases can multiple arcs point to a node.
And Nodes without arcs are substrings
Corresponds to the headword of the dictionary assigned to. Multiple language experts can export
Assign a numerical score to the set of. This language expert
The score applies the score to each node of the graph. Shi
As a specific example of the stem, a score matrix (each element of the matrix is
Weight to multiply the score of a particular expert
Is a fixed-length floating-point number for any statement
It is "N". The score is calculated by a graph creation engine and
Scoring module built into parser
Is evaluated by The higher the score, the higher the
Calculated for all exports forming
It is. The score for a higher export is
With the export forming a high bell export
Of the score assigned by the structural adjustment expert
Of any expert score applied to the combination
Calculated as the sum. The order of arriving at the node and examining the score is:
This is the standard depth first graph movement algorithm. This
In the algorithm, the scored nodes are marked
They will be scored and will not be scored again. Score
Scoring module during the process of being marked
Before evaluating any higher-order node,
Evaluate the headword node of the dictionary. Each entry in the dictionary is
Has one score. Multiple ways to perform one export
If there is, there will be multiple scores. That is,
Yes, if there are k ways to do the sport
K scores. Multiple scores are processed as follows:
Is managed. (1) In the rule consisting of a single element,
And each of the k low export scores are single components
Is added to the expert value applied to the rule
And the resulting vector of k scores is the parent's
Related to Kissport. (2) In the rule composed of two elements, the left
The child has a g score and the right child has an h score
Will be considered. Then g score multiplied by h score
The numerical value is the score of the left child and the score of the right child
And apply to the two-element rule
Calculated by adding the expert values
You. When g score multiplied by h score exceeds N
Only the highest N scores are kept with the parent node.
Be held. (3) One export can be performed by a plurality of methods
If you can make it, at most N scores
Is added to the score list of the node with the highest score
Only be preserved. When the score calculation is completed, the above method
The export is displayed in that node and in the export
Export, including attributes of all substructures not
G most likely ways to do
G scores (ranging from 1 to N)
Confirm that it is related to the set of g). Roux
In special cases such as node S
The method determines the g most likely ways to form the whole sentence
give. Each score in the above score list is associated with
Has a pointer. Pointer, lower export
Which score in the core list indicates a higher level score
Provide information indicating whether they were combined to create
You. By tracking each pointer, g
The most likely interpretation as an unambiguous parse tree
Can be extracted. Referring to FIGS. 1 to 9, an automatic natural language
The translation system will be described in further detail. After that, FIG.
0, FIG. 11 and FIG.
The improvement of will be described. Referring to FIG. 1 and FIG.
The dynamic natural language translation system 10 has an input interface
12, translation engine 16, storage device 18, user input device
22, display 20 and output interface
14 inclusive. The input interface is
Can receive a series of texts written in English language
It has become so. As the input interface,
Keyboard, voice interface or modem or
Including digital electronic interface such as serial input
be able to. The translation engine uses data from the storage device.
To translate the source language. Translation engine, the whole
Can be made with hard-wired logic circuits.
Or contain multiple processing units and associated storage instructions
You can also. The translation engine is responsible for the elements and parts
Minutes, ie, the preparser 24,
Parser 26, graph creation device 28, grammatical structure analysis / translation
Translation evaluator 30, grammar structure analysis and extraction device 32, structure conversion device
User, including user 34 and another grammar structure system 37
An interface 42; Grammar
Structure conversion device 36 by rule control, structure by dictionary control
Conversion device 38 and structure conversion device 4 based on generation rule control
0 may be included. The storage device 18 has one or more storage devices.
Number of disks (eg hard disk, floppy
(Registered trademark) disc and / or optical disc)
And / or memory storage (eg, RAM), etc.
Can be included. These storage devices are required for
All or part of the element can be stored. Sand
Basic dictionary 44, technical term dictionary 46, created by the user
Dictionary, grammar rules 48, generation rules 50, semantic property tree
52, a structural property tree 54 and a graph 56. Record
The storage device 18 receives input text written in the source natural language.
Information, output text information written in the target language, and
Contains one or more dictionaries, domain keywords and grammar rules
All types of information used or useful for performing translations
Used to store User input interface
Keyboard 22, mouse, touch screen,
System, including a stylus or other user input device.
Can be used by operators. As a display,
Computer display, printer or other type of
The display can be used, and the operator can
Other devices for communicating information can also be used. Out
Interface 14 provides the final translation of the source text.
Exchange the translation in a target language such as Japanese. Above Inn
Interfaces include printers, displays, and audio interfaces.
Interface, modem or electronic interface such as a serial line.
Interface and can be used by end users
Other devices for sending strikes can also be included. The operation of a specific example of the translation system of the present invention will be described.
As shown in FIG. 1, FIG. 2 and FIG.
24 is a preliminary solution to the source text 23
An analysis operation (step 102) is performed. This involves a saw
Includes analysis of ambiguity in end-of-sentence recognition of texts and dictionaries
A structural analysis chart including the headword 25 is created. So
After the parser 26, the syntax possibility 27 is described.
Preparser to obtain the structural analysis chart
(Step 104) Structural analysis of the above chart
I do. The graph creation device 28 obtains the data in the structural analysis step.
29 possible interpretations based on the structural analysis chart
A graph is created (step 106). A series of experts
The evaluator 30, which accesses the port 43, uses the stored interpretation
Is evaluated (step 108), and
Add the weight of the kiss part. Graph scorer 33
Gives the scores of the nodes and N (eg, 20
The highest score 35). grammar
The structure extraction unit 32 uses the structure analysis tree for this preferred interpretation.
The structure 39 is allocated (step 110). Then change
The structure conversion device 34 accessing the conversion table 58
Tree to obtain the translation 41 that has become the target language
Above, the structure conversion processing (step 112) is performed. A user
Is another structural analysis system 3 to obtain other translations
7 can be exchanged. Referring to FIG. 4, the system of the present invention comprises:
Character strings that form the input word string into individual punctuation marks and words
Into tokens containing loops (step 114)
Performs a preliminary structural analysis. The appearance of the space
Affects the interpretation of characters at the level of. For example, "x
"-" In "-y" is a dash, but "-" in "xy"
Is a hyphen. After that, the preparser replaces the above token with
Combine with a phrase (step 116). This level
The preparser uses a special structure (for example, Internet
Such as credit addresses, telephone numbers and social security numbers
Is recognized as one unit. The preparser is
Dictionary to separate groups. For example,
"Re-enact" is listed as "reenact" in the dictionary
Is one word, but if not listed, three separate words
It becomes. In the next preliminary structural analysis stage, where the sentence ends
The end-of-sentence recognition is performed (step 118). this
During processing, the preparser sends the possible end of each sentence
That is, after each word in the source text)
When performing a series of steps, the basic dictionary, technical term dictionary and
And access the built-in user-created dictionary.
The preparser will use this if a particular order is specified.
Need not be performed, these steps are a series of
It can be run as an ordered rule, or it can be hard
It can also be coded by incorporating it into software. Referring to FIG. 5, the preparser
For symbols that cannot be analyzed, such as
When there is a sequence, do not translate each one,
Is interpreted and recorded as one "sentence" (step 1
20). The preparser uses two characters at the end of the sentence.
Request a ridge return (step 122). next
If the first letter of the phrase is lowercase, the preparser
Is not considered the end of a sentence (step 12
4). If the sentence starts on a new line and is short (for example,
If, for example, the title), the preparser puts it in one sentence
Is considered. The preparser uses closing parentheses and closing quotes.
Terminator (.), Suspect
Consider a question mark (?) Or exclamation point (!) As the end of a sentence
(Step 128). Ends with "." Or "?"
In the case of a sentence, the preparser
Use virtual punctuation after the quotes in addition to the point.
You. For punctuation added virtually to "?"
The method is illustrated in the following example. [0047] The question was "What do you want?". Did he ask the question "What do you want?"? Are you concerned about "the other peop1e"? In English, each sentence may end with a "?"
high. There is virtual punctuation added by the preparser
The thing is that there is something like a question mark before the quotes
Or nothing at all. After the quotes
There is something like a stop or question mark
You. The remaining grammatical structure of this sentence will
The most appropriate choice can be made at the management stage. The preparser is also capable of analyzing the
Use several methods (steps 130, 13)
2, 134, 136 and 138). Short in dictionary
Marks that can never be used at the beginning of contracted forms
It cannot be used at the end of a sentence
Some are marked (step 130). This
These rules are always respected. For example, "Ltd."
Is not used in the text, and "Mr."
Not. The preparser also states that the following words are "the", "i
n "unless it is a frequent word (step 13
2) When there is a terminator in one capital letter, the sentence ends
I don't think there is. The word before the end of the word
If in the book, the sentence ends at the end
(Step 134). The word before the stop is in the dictionary
The phrase has a terminator inside it (e.g.,
For example, IBM), the following words are in the dictionary as lowercase
If it is not listed, or if the next phrase is in uppercase
Sentence does not end at its end
(Step 136). Otherwise, end
A stop indicates the end of the sentence (step 138). Referring again to FIG. 2 and FIG.
When a break is specified by the preparser, the parser
Put the words of the sentence into the syntactic classification and solve the possible syntax of the sentence
A grammar database must be added to those phrases to calculate
The source grammar rules are applied (step 104). Grammar rules
Rule 48 is a computer that represents the grammatical restrictions of the language.
Can be executed as a series of rules that can be processed,
There are hundreds of such rules in English, and these rules
Can be applied to hundreds of syntactic categories. Calculation of this work
In order to reduce the extra time required for
Are ignored and ignored. In the next step (step 106)
Thus, the graphing device captures different meanings of the words and phrases,
A ring with a direction that represents all the semantic interpretations of the sentence
Use a dictionary to create non-graphs
Extend the result of This graph is a series of
Created with the help of semantic communication procedures. These hands
The order is based on a set of established grammar rules.
Access the semantic property tree for semantic information.
To A semantic property tree is a tree structure that contains semantic classifications
It is. This tree is roughly from abstract to concrete
And how far apart in the tree
Each level in the tree
From the side, how are the two terms related in meaning?
Can be determined. For example, "cat" and "do
g "is more relevant than" cat "and" pudding "
Therefore, "cat" and "dog" are two "animal" trees.
The distance in the tree is short, "animal" and "cat" are trees
Are stored at different levels. Because
"Animal" is a more abstract word than "cat"
Because. Referring to FIG. 9, the graph of FIG.
Node 80, whose subnodes 82, 84, 86 are seeds.
Pointers 88, 89, 9 in a way to indicate various types of associations
It is linked at 0,91. Graph first related ties
The node that represents a phrase is the node of the phrase that composes it.
Has a pointer to the node of the node or subphrase.
Things. For example, node 84 representing "the bank"
Is the word that composes it by pointers 92 and 93
It links to the phrases "the" 94 and "bank" 95.
The second type of association in the graph is that the phrase interpretation is lower
From the level one, make the same higher level component
You have a pointer to another method.
For example, node 80 representing the phrase "by the bank"
A pointer 88 linked to each of the words constituting each
89, and two source interpretation positions 8 including 90 and 91
1,83. In this case, different
Each constituent word is different for the phrase "the bank"
Sub-nodes 84 and 86 representing different meanings. Gra
The structure of the file is defined by the results of
Is restricted by the syntax The nodes in this graph are
Memory for semantic information entered during the taste transmission process
Associated with location. The parts that convey the meaning of the system are
From smaller components to larger components containing
It acts to convey information. Semantic information is analyzed at an early stage
Applied to the four classes of syntactic classification used in the work
You. The four classes are SEMNP (noun type object and
SEMVP (including subject, object)
Verb phrases such as), SEMADJ (adjective), and V
ERB (Verb-type verb of a dictionary that often takes an object)
It is. Other syntax classifications are ignored by certain rules.
The grammar rule setter should mark the rules with a specific mark
Can override actions that do not appear on the surface
You. These special orders come first. There are two methods of transmitting semantic characteristics in the system.
With two sides. The first aspect is the noun forms and actions of grammar rules.
By examining the components of the noun type, the purpose of the noun type
Which optional restriction slots of verb-type components
This is a set of rules that can be applied. example
For example, the verb phrase "I persuaded him to go."
The rule to do is, roughly speaking, VP = VT11 + NP
+ VP (where VP is the verb phrase and VT11 is the transitive verb
NP is a noun phrase). Default rules as examples
The rule is to the right of the verb if it takes the object
The first NP must apply selection restrictions
Is Umono. Another rule states that the VP for that subject
Restrictions must apply to the first NP to the left of the VP
There is no provision. When these rules are put together, "pers
"Uade him" and "him go" are both meaningful
I try to evaluate it. As already explained,
These rules reflect the complex rules of English,
Because of that the number is very limited. Referring to FIG. 6, the semantic transmission work is optional.
Including copying the limit from SEMVP to the statement
Step 140). SEMNP is used as a representation of location
If the semantic constant specifies a good position,
The validity is evaluated (step 142). A rule
The rule is the union of two SEMNPs (logical AND of syntactic features
If it is detected by
The rough creation device calculates the semantic characteristics and calculates the semantic distance.
Apply to expert (step 144). The rules specified for the transmission of semantic characteristics are as follows:
During the inspection, the graphing device
(For example, SEMNP it contains more words
The position of the “head” of the SEMNP to be transmitted to
If found, the graphing device also conveys semantic properties.
Is reached (step 146). However, "head" is for classification
(Eg "Portion", "Part")
The "head" propagates left or right from the SEMNP.
SEMVP and SEMADJ also have positions for sorting
In the same way, except that it is not
Is transmitted. Adjectives are part of SEMVP in this sense
is there. SEMVP made from rules containing VERB
If the VP is the passive voice,
Except for the case, the verb subject restriction is transmitted in the upward direction.
You. In case of passive voice, VERB's first object restriction is communicated
(Step 150). Rule space including SEMVP
In this case, the graph creation device sets the SEMVP selection restriction to S
Apply to NPs encountered when moving left from EMVP
(Step 152). Regulations including SEMADJ
In the case of the rule, the graph creation device selects SEMADJ
Restrictions encountered when first moving right from SEMADJ
Try to apply to any SEMNP that
If not, try to move to the left
Step 154). Any remaining unused object selection in VERB
Option restriction (to be transmitted in the upward direction to be passive
Not shown), the graph creation device is on the right of VERB.
Apply the above restrictions in order to the SEMNP encountered by the
(Step 156). In all these rules, the verb selection
Choice limits run out as soon as they apply to something
Would. For all previous rules, SEMNP
Is used if anything applies to those SEMNPs.
It is not done. Starting from this rule, SEM
NP is exhausted. Eventually, one rule is SEMV
When P is created, the graph creation device is not used until then.
SEMVP or SEMADJ not included
Decide whether or not it is included
(Step 158). The system also provides language feature matching.
Do. Linguistic characteristics are the characteristics of words and other components.
You. Syntax parsing is used by parsers,
Gender matching is used by the graph creation device. But,
The same scheme is used for both. For example, "the
"y" has the syntactic property of plural, while "he" is singular.
Has the syntactical characteristics Characteristic matching means that the characteristics of a phrase
If the characteristics of the words to which they apply meet certain conditions
Mark grammar rules so that they apply only when
U. For example, assume that there are the following rules. S = NP {+ VP} Here, the symbol ＠ indicates that the characteristics of the numbers of NP and VP do not match.
Means you have to. So this rule
Says "they are" and "he is" are correct, but "they
"is" and "he are" are not allowed. The characteristic match restrictions are “local” and “broad
Enclosure ". Extensive behavior is actually handled by the statement
Calculated when the grammar is created, not when
It is. And a wide range of actions that must be performed
The sequence is coded as instruction bytes. The calculation of the "wide-range" characteristic operation consists of n elements
Rule (ie, two or more elements to its right
Have). afterwards,
The system communicates between rules in the correct set of properties.
Code into various two-element rules to end
assign. a rule consisting of n elements from two elements
By dividing the rules into different rules, the analysis work is very simple.
Purified. However, the system is a two-part rule.
Tracking the set of properties across rules,
System retains the power of "widespread" processing. In the system of the present invention, a dictionary is not configured.
Although it is treated as an even phrase,
It also includes the constituent "idioms". These two shapes are
Ultimately, compete with each other for the most appropriate translation.
And For example, in a dictionary, it means "b1ack sheep"
Are registered as "surplus". However,
The phrase "b1ack sheep"
Sometimes it means "sheep." Both of these shapes are retained
Translations that are not idioms
Selected as a translation. This idiom also belongs to another classification. An example
For example, the system may use three types of classification:
Can be. [0064] Almighty: United States of America Preferred: long ago Normal: b1ack sheep Almighty idioms make up that sequence
Takes precedence over possible interpretation of any phrase. The preferred idiom is
Words that make up the sequence
Precedence over any possible interpretation
You. Normal idioms compete with other headwords at the same level.
You. The resulting graph is
(Step 108, FIG. 3)
The spurt is a score that indicates the likelihood of the graph being interpreted correctly.
Supply. The system of the present invention is not limited to whole sentences.
Scoring applied to all parts of the sentence of any length
Including the law. Important factors for using graphs
Says that even if it is used in so many statements,
The subtree is fully scored and analyzed only once
That is. For example, "Near the bank there
In the phrase "is a bank.", "Near the bank"
Has at least two meanings,
The best interpretation of a phrase is determined only once
Not. The phrase "there is a bank" is likewise
Can be interpreted, but these two solutions
You only need to determine which one is the most appropriate
Not done. So this sentence has four different meanings
Can be interpreted, but subclauses are scored only once
No. Another feature of this graph is that each node
Have easily accessible information about the length of the part
That is. Due to this feature, English text is analyzed again
The N sub-strings of any of the English sentences
The most appropriate interpretation can be made. In one execution, the most appropriate
N analyzes are obtained each time (N is about 20
System by using a graph with a number of degrees)
Is the result of the user's choice for smaller components.
N that respects user choices
Make the most appropriate analysis. All these analyzes are sentence
Without re-analyzing the
It is done without having to repeat the points, so it can be done quickly
Wear. Referring to FIG. 8, expert evaluator 3
A behavior of 0 characterizes each translation and gives various experts
This is done based on various factors that will be further processed. probability
The rules expert 170 provides the first source language parsing tree.
Average relative frequency of the grammar rules used to obtain
evaluate. Selection Restriction Expert 178 reports the resulting translation.
Evaluate the degree of semantic matching of the translation. Dictionary headword probability
Expert 172 creates an initial source language parse tree.
The specifics of some phrases in the sentence used to obtain
Evaluate the average relative frequency of "part of speech". Statistical expert
Is the flatness of the particular paraphrase selected for a translation.
Evaluate the average relative frequency. [0068] The system uses individual English words, phrases and
POS for English word group
Determine dynamically. When the system translates a sentence,
Automatically determine and usually make the right decision. But with
Sometimes, the sentence to be translated and the whiteness are ambiguous.
Contains words that can be interpreted as different parts of speech
When you have multiple different interpretations, all of which are "correct"
Is obtained. The system operator
Ignore the part of speech that Tem automatically determines, instead of words or
Manually add any part of speech for a phrase or group of words
Can also be set. For example, "John saw a boy with
a telescope. "
Considers "a boy with a telescope" a noun phrase,
The sentence means that the boy had a telescope
"John used a telescope to see the boy.
Used ". Operet
Data may have multiple possible part-of-speech settings or may be more restrictive.
The system can be set up manually by setting the
Translation result deteriorates when the part of speech rule determined by
Or at least not improve
There is. Noun phrases are more restrictive than nouns
Group is the least restrictive part of speech setting
You. The table below shows the various possible parts of speech settings. Part of speech (POS) noun Noun phrase Verb (transitive or intransitive) Verb phrase adjective Adjective phrase adverb Adverb phrase preposition Prepositional phrase conjunction group English The part-of-speech settings for "adjective phrases" and "adverb phrases"
Depends on how the system interprets certain prepositional phrases.
This is effective when the meaning is different. For example,
The sentence "We need a book on the fourth of July"
Means "on the fourth of July" as an adjective
If we interpret that we have
I want a book about German Day. Means
Interprets "on the fourth of July" as an adverb phrase
In this case, "We want a book on July 4th."
To taste. The system automatically displays "on the fourth of Jul
The operator thinks that an incorrect part of speech has been assigned to "y".
Operator, "Weneed a book on the f
`` on the fourth of July '' in the sentence `` ourth of July ''
Can be manually set to another part of speech. operator
The system will enter a specific word, phrase or group of words.
If you do not want to translate words into Japanese,
Set the part of speech "English" for a word, phrase, or word group
can do. The operator must make settings in the system
Manual, even if performed automatically by the operator.
Remove one or more part of speech settings, even if done
Can be The system has multiple levels for each user.
Keep track of translation usage statistics. For example, the system
Is the level of the surface form of the phrase ("leaving" is a transitive
Or as a frequency used as an intransitive verb)
And the semantic level (meaning "remains")
Or "meaning" started from) "
And the latter types are "leave", "leaves", "lef
t "," leaving "
Can be raised. The system is also used in the last few statements.
Usage statistics used and user usage statistics at any time
It can be stored separately from the value. In addition, the system
System intervenes so that the user uses certain meanings of the phrase
And the system is subject to user intervention.
Without using the specific meaning of the phrase
Can be. The structure adjustment expert 182 is able to
Features related to the length of the component, such as English and many other
-It is based on features common to European languages. A few
(But not all) structures, in these languages,
Sentences with heavy (long) elements on the left side of (short) elements
Is not welcome. For example,       Mary hit Bill with a broom. (Light on left, heavy on right) (Applicable)       Mary hit with a broom Bill. (Heavy on the left and light on the right) (nonconforming)       Mary hit with a broom a dog that tried to bite her.                         (Heavy on the left, heavier on the right) When there are two parses of a sentence, one is
Includes a structure that tries to avoid cans.
'' Sequence, and other analysis is not
Considers that the former does not represent the intended interpretation of the sentence.
Is done. This expert is not
This is an effective way to distinguish between different analyzes. An equivalent structure of the pattern "A of B and C"
In construction, the intended interpretation is “Aof {B and C}”.
Or "A {of B} and C"
Can be difficult. Equivalent structure expert 180
Which combines two elements with which the peer mode is closer
To determine the semantic distance between BCs.
Measure the separation and the semantic distance between ACs. This extract
The part accesses the semantic property tree during processing. this
Experts should be able to analyze a sentence
It is also an effective way to distinguish between analyzes. Many English words are interpreted as common nouns
And potential interpretations as proper nouns
You. Capital expert 176 states that capitalization is significant.
Capital letters in the sentence to determine if they have
Use the position. For example, the statement below is Brown is my
first choice. My first choice is Brown. The first sentence is inherently ambiguous, while the second is
"Brown" may not be a color name, but a person name
It differs in a very high point. This expert is in uppercase
The beginning of the phrase is at the beginning or not
Is there (as in the example above),
Words that start with lowercase letters in the dictionary
Factors such as whether they are registered
You. This expert uses words in capital letters
Is an effective way to interpret correctly. If the sentence contains a sequence of uppercase words and phrases,
, The sequence is treated as a proper noun or common noun.
Is managed. The system of the present invention uses a capitalized sequence
Procedures are used and the interpretation of the former takes precedence. The above sea
Kens can't parse itself according to normal grammar rules
Sometimes the sequence is not analyzed
Is processed without being translated as a noun phrase. This step is
Don't completely ignore interpretations of low-level common nouns
Is a very effective means of handling compound proper nouns
It has been proven. Referring to FIG. 7, the machine translation system of the present invention is shown.
System can improve the efficiency of structural transformation methods based on simple grammar rules.
Yes, but also the ability of the structure conversion method between templates
A close grammar rule control structure conversion mechanism 162 is used. this
The method is a statement that allows you to specify an uneven composite structure
The legal rule 160 is used. Used by other translation systems
The format of the rule is shown below. [0076] The system of the present invention uses the following grammar rule format:
Use. [0077] (Equation 1)In this syntax, the symbol preceded by "#" is the sentence structure
A symbol that is invisible for analysis purposes, but once analyzed
Virtual, used to build substructures if available
Symbol. Given this type of grammar,
Between any sequence of nodes in the child structure
Allows you to specify multiple structural transformations.
You. As a result, the structure conversion mechanism based on the grammar rules is a template.
Converts the rate conversion mechanism into a mechanism with some capabilities.
Is replaced. The system of the present invention is the second type described above.
Grammar rules, but the first form
Automatically create corresponding grammar rules. Therefore, the statement
Use grammar rules of the first form to parse
Use second form of grammar rules to form parsing structures
Can be The structure conversion is also performed by a grammar rule control structure conversion operation.
Work on the parse tree after receiving operations
To convert the dictionary control structure to access the dictionary 161
166. After that, the production rule structure conversion work
To provide the language text 41, the resulting
Apply the production rules to the parse tree. Referring again to FIG. 1 and FIG.
System uses the above process to determine which translation is most desirable.
After making it, the translation is sent to the user through the display 20.
Provided to Then the user adopts the translation
Or another analysis system 37 through the user input device 22.
The operation can be done by operating. During rework,
Users can leave the translated part as it is
Request retranslation of other parts while keeping
Can be. This can be done quickly. Because,
The stem holds a graph containing the expert weights 31
Because it is. Up to now, referring to FIGS.
Natural language translation systems have been described in some detail. Less than
Then, various improvements of the present invention will be described with reference to FIGS.
This will be described with reference to FIG. Referring to FIG. 10, an automatic natural language translation system is provided.
Stem's translation engine 16 translates source text 23
Automatically translates it into the target natural language text 41
translate. This translation is referred to as an "automatic
Affected by the "area identifier" function. With this feature,
When performing translation, the translation engine 16 uses the area keyword
One or more lists of 200 (N shown)
And access to the dictionary 202. The dictionary 202
Depending on the area in which the word is used,
Including words. Each headword in the dictionary 202 is a source natural
It may be one or more phrases written in a language.
Generally speaking, each separate area is a list of area keywords.
FIG. 10 shows N areas. Key
Each headword in the list of words is the source natural language
One or more words written in can also be used
You. The list of keywords 200, as shown in the figure,
Letter 202 or arrow 20
As shown in FIG. 4, it can be built in the dictionary. One area is the usage or identification of a certain field.
Include a set of patterns of terms and usage belonging to
Can be. For example, one area is business communication
Letter, marketing document, computer document, physics
It can include documents in technical fields such as science. The dictionary 202 is used for a certain area.
When the phrase is used in other areas, which particular
Different eyes when not used in the area
Include words that have a translation in a natural language (for example, Japanese)
It is. Keywords 200 are associated with each area
Specific sentence (or input document) written in natural language 23
Is used to determine if an item belongs to one of the possible areas.
Is a word or term specifying the area to be Area keyword
In each list of C200, each keyword is
And the strength as a keyword for a specific area
Has a numerical value indicating For example, the "computer" area squirrel
The open-architecture compu with 100% value
ter "and this 100%
The sentence containing this keyword is always
Data in the "computer" area.
The keyword "computer" in the region has a value of 25%
And the value of 25% is a sentence with this keyword
Are more likely to be in the "computer" area
Means that A translation engine is provided by the "automatic area identifier" function.
16 preferably the sentence or document belongs to a specific area.
Of a sentence, document or document.
Sum of the keyword's strength values (or
The word exists). keyword
If there is a certain area, the dictionary 202
Listed as area-specific words and trying to translate
The probability of words contained in a sentence or document is increased,
Or translation of the document. On the translation engine 16
The decision is based solely on the source natural language and keywords.
Good. The translation engine 16 has, for example, the following two
Either of the two methods, at least the input text information 23
Also determines at least one area to which a part thereof belongs.
In the first method, one sentence is input to each sentence of the input text information 23 at a time.
Address, and count the number of occurrences of keywords in each area.
It is to do. At least as many keys as the given number
For each region with words, the translation engine 16
Using a translation specific to the area, the target natural language 41
Increase the probability of some words in the translated sentence
You. The above-mentioned predetermined number is all of the input text information 23.
The same value can be fixed to the statement of
In addition, it can be changed according to the length of the sentence (number of phrases)
Wear. The predetermined number is set by the system 10.
Can be set by the user of the system 10.
You can also. The second way is to say the first N words
In each of the input text information 23,
This is a method of counting the number of appearances of keywords in an area. Place
For each region with more than a certain number of keywords
Will use the translation engine 16
To be included in the sentence translated into the target natural language 41
Increase the probability of some phrases. Similarly, the predetermined number is
It can be set by the system 10,
It can be set by 10 users. In a preferred embodiment of the present invention, a translation engine
16 counts keywords as described in the previous paragraph.
Strength values for each area, not by
Related to each keyword to get the sum of
Input text information 2 by using the strength value
At least one area to which at least a part of 3 belongs
To determine. In this preferred embodiment, the translation engine
If not, we've explained in the previous paragraph
Works like that. That is, in the first preferred method, the input
Each sentence in text information 23 is addressed one at a time.
And the sum of the keyword strengths is greater than
For each region with
Translated into the target natural language 41 using translations specific to the domain
To increase the probability of some words in the sentence. Place
The above-mentioned number is used for all sentences in the input text information 23.
The same value can be fixed, and for each statement,
Can be changed according to the length (the number of phrases).
The predetermined number may be set by the system 10.
Yes, or it can be specified by the user of the system 10.
You. In a second preferred method, the keywords in each area
Is the sum of the first N units of source text 23.
Like a word, it appears in a predetermined number of words in the input text information 23.
Determined by the keywords that appear. Keyword strength
For each region whose sum is greater than or equal to a given value,
The translation engine 16 uses translations specific to the area.
The number of sentences included in the sentence translated into the target natural language 41
Increase the probability of that phrase. As already explained,
The fixed number can also be set by the system 10.
However, it can also be set by the user of the system 10. As an example of the “automatic area identifier” function,
"Markup language" and "help desk"
Data area is included in the list of area keywords,
"Endusers" means that "end users"
Has two different translations depending on whether it is used in the domain
And The following sentence is part of source text 23
Come Markup language should be easily used by end users
 without being supported by help desk. According to the "automatic region identifier" function, the translation engine 16
Are two keywords in the "computer" area, "mark
"up language" and "help desk". "Con
Find these two keywords in the "Pewter" area
Thus, the translation engine 16 uses the threshold to
Determine that a sentence is a "computer" domain sentence
Can be. If the sentence belongs to the "computer" area
If the translation engine 16 determines that the
To the target natural language text 41 provided to the user.
"End users" specific to the "computer" area to translate
To increase the chances that the translation will be used,
The probability of translation of "end users" specific to the "data" area
I'm sorry. In addition to the mode in which this function works for each sentence,
The "area identifier" function can work on the entire input text.
Wear. For example, the translation engine 16 responds to a specific area
Source text 2
3 if it is in the first n words (n is an integer)
The entire source text 23 belongs to the “computer” area
You can decide whether or not. This all text mode
In the case of C, the translation specific to the area included in the dictionary 202 is
All words in source text 23 that have
It is translated using the original translation. Referring to FIG. 11, an automatic natural language translation system is provided.
The translation engine 16 of the stem 10
And automatically convert it to the target natural language text 41
Translated, this translation is a grammar rule according to other features of the present invention
Some settings and rules that apply to the various rules in Rule 206
Affected by marking. These settings are
"Lumighty" and "part of speech priority". Translation of system 10
The translation engine 16 outputs “almighty” or “speech
Use grammar rule 206, which can be marked
Parser 2 that parses sentences in source natural language 23 using
Have 6. Parser 26 is subject to the almighty rules
Every part of the source natural language 23
Rules that prevent the application of all other rules
Use grammatical rules marked "
You. Parser 26 relates to the same part of speech as a particular part of speech
The rule that blocks only the other rules is “part of speech priority”.
Use the grammar rules for which the markings are made. Mark
Unattended grammar rules take precedence over other rules.
Do not have. "Almighty" or "Good"
Marking "lyrics first"
Is a tree created by the translation engine 16 of the system.
-A very effective way to remove unnecessary analysis branches from the structure
It is. This marking of the grammar rules will also
Gin 16 or must be made if not marked above
You no longer need to create a large tree structure
Also, if the above mark is not made, the tree will grow
No need to memorize the structure
Analysis can also be performed more efficiently in terms of capacity.
You. As an example of the marking function of this grammar rule,
And the grammar rules include the following three rules. [0091] (1) Noun = number + hyphen (-) + ordinal hundred (for example,
three-hundredth) (2) Noun = noun + hyphen (-) + noun (for example, sc
hool-building) (3) Noun phrase = number + noun (for example, one book) The system 10 also creates a dictionary containing the following three headwords.
Suppose you use it. [0092] When you have these grammar rules and dictionary entries, the source
The word string "one eighteen-hundredth" contained in the text
Can be analyzed by the following two methods. [0093] (Equation 2) Marking grammar rule (1) as "almighty"
Has been done, a simple phrase "one eighteen-hundredth"
All other parsing of the word sequence (as in the second parsing above)
Is excluded, so the first of these two analyses is
Is used. All My Dominating All Other Rules
Tee rule (1) is used. If the grammar rule (1) is "
If it is marked Priority,
First analysis is selected, but in this case the rule
(1) and rule (2) address the same part of speech,
Rules with a Go (ie, part of speech) mark
According to (1), grammar rule (1) is used to analyze word strings.
It is. As the following example shows, "almighty"
And "part of speech priority" are different, and "almighty"
If you translate with the marked rules,
The result is not necessarily the same as if the
Produces no fruit. In the following four grammatical rules, the third rule includes:
It is assumed that the mark of “part of speech priority” is attached. [0095] (1) Noun = noun + noun (2) Noun phrase = noun (3) Noun phrase = number + noun {part of speech priority} (for example, in F
OUR DAYS) (4) Adverb phrase = number + adverbial noun (for example, I work FOUR
 DAYS a week) The system 10 also uses a dictionary containing the following headwords:
I do. [0096] When there are such grammar rules and dictionary entry words,
The word string "I work four days" contained in the text is
It can be analyzed at the following three methods and phrase levels. [0097] [Equation 3] For grammar rule (3) marked as "part of speech priority"
Has the same grammar rule (3) as the same word string ("four days")
Other parsing the same part of speech as a noun phrase
All grammatical rules have been excluded, and grammar rule (2) has been changed to "fo
u days ”as a noun phrase,
The first analysis is used in preference to the second analysis.
However, it has another adverb phrase as a part of speech
Does not exclude the third analysis above.
No. Thus, grammar rule (3) is changed to "almighty".
Mark any other grammar rule (3) applies
However, the grammar rule (3) is changed to "part of speech priority".
Address the same part of speech as in grammar rule (3)
Only other rules are excluded. Referring to FIG. 12, an automatic natural language translation system is provided.
The translation engine 16 of the stem 10
And automatically convert it to the target natural language text 41
Translate, this translation is a dictionary of other features of the present invention.
Affected by the probability value associated with the headword.
According to the probability value characteristic, the parser of the translation engine 16
26 is very likely to be a component of the final tree structure
Marked low or absolutely non-constituent
The dictionary 208 containing the headword is used. Dictionary 20
8 includes headwords of compound words,
At least one is marked with a very low probability value.
To the correct parse tree of any sentence in source natural language 23
Has little probability of functioning as a specified part of speech
It indicates that. An entry with a headword marked
If it is included in the statement's parse tree structure,
Low probability associated with headwords with
The parse tree is correct for the statement
The probability is very low. Use probability values in this way
This allows the translation engine 16 of the system 10 to
It can have the effect of removing the tree structure created by. As an example of the use of the probability value, look at the following compound word:
Suppose there is a dictionary that contains the words. [0100]Headword Part of speech Grammar rules probability       in today preposition phrase part of speech P (n) In this case, P (n) is a given source text sentence
The idea that the headword "in today" is actually a prepositional phrase
The specified probability value is played. P (n) is an ordinary copy of the dictionary.
Set to an infinitely small number compared to the headword of the ligature
Suppose you have The input sentence is: He didn't come in today. This sentence is syntactically ambiguous and can be interpreted like the following two sentences:
Wear He didn't {come in} today. When He didn't come {in today}. (In today) is the associated confirmation of the matching dictionary entry.
Whether the rate has a very infinitely small value, P (n)
The probability that this sentence will be parsed as a second sentence is very low.
Gone. "In today" has a part of speech priority mark
Prepositions that can give ordinary probability values
There is no parsing other than parsing as a phrase. Thus, this
Improper parsing of a sentence is undervalued. Against "in today"
If P (n) is set to zero, the parser
ー is “intoday-preposition phrase” to analyze the above sentence.
Is not used at all. However, in this particular example,
, So P (n) is set to zero
Is not desirable. [0102] I am interested in today, and not in t
omorrow. In this case, analyzing "in today" as a prepositional phrase is
Reasonable. Use low P (n) values instead of zero
If not, this is an unfavorable grammar structure analysis. [0103] He didn't come {in day}. Can be excluded, while I am interested in today, and not in tommorow. Parses "in today" as a prepositional phrase.
It can be analyzed properly. All of the above functions and processes are performed by a general-purpose computer.
Various hard wiring logic designs and
And / or can be implemented by programming technology
You. The steps shown in the flowchart are usually performed in order.
Do not need to be applied to the
Can be matched. Also, the function of this system is
Can be divided into programs and data in various forms
You. In addition, grammar and other rules of operation are
One or more while providing it to the user in a format
It is advantageous to develop in a high-level language. Including all functions disclosed in this specification
Any of the above examples of automatic natural language translation systems
Also, general-purpose computers (for example, Apple Macintosh
, IBM PC and compatible, SUN Workstation
Disk or optical compact disk
Computer readable media such as disk (CD)
Computer software.
Wear. A person skilled in the art will be described in the claims.
Without departing from the spirit and scope of the claimed invention.
Various changes or modifications to the invention described in the specification
Could be implemented separately. Therefore, the present invention
Rather than by way of illustrative example,
Defined by spirit and scope.

BRIEF DESCRIPTION OF THE DRAWINGS In the drawings, like reference numbers generally refer to the same parts, even in different drawings. The drawings are not necessarily to scale, emphasis instead generally being placed upon illustrating the principles of the invention. FIG. 1 is a block diagram of a system for automatically translating a natural language according to the present invention; FIG. 2 is a flowchart of data showing overall functions of the system of FIG. 1; FIG. 3 is a flowchart showing operations of the system of FIG. 1; 4 is a flowchart showing the operation of the sentence end recognition function of the pre-parser of the system of FIG. 1. FIG. 5 is a flowchart showing the operation of the parser of the system of FIG. 1. FIG. 6 is a flowchart showing the operation of meaning transmission of the system of FIG. FIG. 7 is a flow chart showing the operation of the structure conversion of the system of FIG. 1; FIG. 8 is a flow chart of an expert evaluator of the system of FIG. 1; FIG. 9 is a flow chart of the system of FIG. FIG. 10 shows a sample graph used by the automatic natural language translation system according to one aspect of the present invention. FIG. 11 is a feature performed by an automatic natural language translation system according to another aspect of the present invention, which is a characteristic that affects which grammatical rule is applied to translate a source text into a target natural language. FIG. 12 is a feature according to yet another aspect of the present invention, which is executed by the automatic natural language translation system, in which the value of the probability assigned to the headword of the dictionary in order to remove inappropriate translation is calculated. Diagram of features included

Continuation of front page    (72) Inventor Kuno ▲ Success ▼             United States Massachusetts             02178, Belmont, Glendale             Road 47 F-term (reference) 5B091 CA22 CC12 DA14

Claims (1)

Claims: 1. A storage device storing a dictionary, and using the dictionary, input text information written in a source natural language is translated into output text information written in a target natural language. A translation engine, wherein the dictionary includes headwords of compound words written in the source natural language, and at least one of the headwords is specified in any correct parse tree of the input text information. A mark indicating that there is almost no probability of functioning as a part of speech, the translation engine is configured to analyze the headword with the mark as the designated part of speech according to the probability indicated by the mark. An automatic natural language translation system, including