JP2002197085A - Automatic natural language translation system - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide an automatic translation system from a certain natural language into another natural language, preferably from English into Japanese. SOLUTION: This automatic natural language translation system receives a text sentence in a source natural language (preferably English) and translate it into an objective natural language (preferably Japanese). In addition, an operator can automatically retranslate a selected part of a source text by using the system. The system includes some pieces of improvement to aim at use of values of probability in relation to direction words of a dictionary to make primary translation of a document including phrases with different translated words in different fields more exact, to specify some grammatical rules as a control means to exclude inappropriate translation and to remove the inappropriate translation.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to automatic translation from one natural language to another, preferably from English to Japanese.

[0002]

2. Description of the Related Art Various methods have been proposed for machine translation of a natural language. Typically, systems used for translation include a computer that receives input in one language, processes the received input, and outputs it in another language. This type of translation has heretofore not been accurate and requires a skilled operator to significantly modify the output. Translation work by conventional systems generally involves a structure conversion operation. The purpose of this structure conversion is to convert the parse tree of the source language sentence (that is, the syntax structure tree) into a tree corresponding to the target language. Up to now, two types of structure conversion methods have been tried. That is, conversion based on grammar and conversion between templates.

In the case of a grammar-based transformation, the domain of the structural transformation is the grammar rules used to obtain the source language parse tree (ie, the set of subnodes that are the direct children of a given node). Limited to the area. For example, when there is the following expression, VP = VT01 + NP (the verb phrase has one transitive verb and the noun phrase arranged in this order.) Then, for Japanese, 1 + 2 => 2 + 1 (V
The order of T01 and NP is reversed. )

[0004]

The parse tree of the source language, including the application of rules, is structurally transformed such that the order of verbs and objects is reversed. In Japanese, the verb comes after the object. This method is very efficient in that it is applied exactly where the rules were used to obtain the source language parse tree, and that it can quickly find where a particular transformation has taken place. On the other hand, as described above, the conversion mechanism is largely limited in that its area is greatly restricted, and that natural language may require conversion rules that span non-child nodes. Is weak.

In the conversion between templates, the structural conversion is
It is specified in the form of an input / output (I / O) template or subtree. If an input template matches a structure tree, the structure tree portion that matches that template is converted as specified in the corresponding output template. Although this is a very powerful transformation mechanism, it can take a considerable amount of time to find which part of an existing structure tree matches a given input template, which can be expensive.

[0006]

SUMMARY OF THE INVENTION The automatic natural language translation system of the present invention has many advantages over conventional machine translation devices. After the system automatically provides the possible and most appropriate translation of the input text information and supplies the user with its output (preferably a Japanese translation of the English input text), the user You can interact with this system to manipulate the translation results or to automatically get other translations. A person who operates the automatic natural language translation system of the present invention can automatically re-translate the remaining part while retaining the part of the translation result determined to be satisfactory. This selective retranslation saves time for the operator because only those parts that need to be retranslated are translated, and potentially many inaccurate parts are If there is, it can save the tedious task of examining the very high quality parts of the translation. In addition, since many translation adjustments can be made in this system, usually the final structure of the translation is often created by the system. Thus, by using this system, potential mistakes by humans (operators) are reduced, and the time required for reworking, such as changing sentence structure, personality or tense, can be saved. The system provides operators with extensive and accurate grammar and spelling knowledge.

According to the automatic natural language translation system of the present invention, the translated sentence becomes more accurate due to various ambiguity processing of sentence breaks included in the source language and a powerful semantic transmission function, and the translation of the operator's translation is made more accurate. The time required for rework is reduced. The quality of translation is further improved by the learning statistics that the system stores in accordance with the preferences of a particular user. The idiom processing method of the present system has the advantage that when there is a sentence containing a phrase that forms a idiom, it can be accurately translated without considering the meaning of the idiom itself. The system is not only efficient, but also has a variety of functions to match less relevant characteristics. The structural balance expert and the peer structural expert of the present system efficiently discriminate target analysis from non-target analysis. The capitalization expert efficiently interprets uppercase words in sentences efficiently and the capitalization procedure efficiently processes proper nouns in compound words without completely ignoring interpretation as common nouns.

[0008] In one aspect, the present invention relates to an improvement in an automatic natural language translation system, wherein the improvement includes an "automatic region identifier" to aid translation.
Regarding the use of The domain includes a set of terms and usage patterns belonging to a certain usage or a certain person.
For example, a business message, a marketing document, a computer-related document, a document in a technical field such as physics, or the like can be considered as an area. Some words included in the dictionary, when used in a certain domain, become a translation of the target natural language (for example, Japanese), but are used in other domains or regardless of the field. If used,
The meaning may be different. A list of domain keywords is also used. Keywords are domain-specific words or terms associated with each region, and are used to determine to which region a particular sentence of the source natural language (or source document) belongs. The "automatic region identifier" function includes a keyword (or document or part of a document) containing enough keywords to determine that the sentence (or document) is a sentence of a specific region. Judge whether it is. If the area can be identified, words that are listed in the dictionary and are also included in the sentence (or document) being translated are used with an increased probability value, and the sentence (or document) is used.
Translation is performed. The determination by "automatic area identifier"
It is based solely on source natural language and keywords. The "automatic region identifier" feature of the present invention does not include region mismatch analysis (analysis of not matching a particular region) in the resulting tree structure, thus speeding up the translation time of the system.

In another aspect, the present invention relates to another improvement of an automatic natural language translation system. The improvements involve parsing source natural language sentences that are marked as "almighty" or "part-of-speech," or use the grammar rule "unmarked." An "almighty" grammar rule is a rule by which all other rules that parse the same part of the sentence are excluded. The “part of speech first” grammar rule is a rule that excludes all other rules that thereby parse the same part of the sentence, but only if the other rules have the same part of speech. Done. The "unmarked" grammar rule has no precedence over any other rules. Marking grammar rules such as "almighty" or "part of speech priority" are effective ways to remove irrelevant grammatical interpretation trees from the tree structure created by the system's translation engine in the absence of these rules. It is. Such a marking grammar rule can be analyzed more effectively in terms of time and memory capacity. That is, the translation time of the system can be reduced. This is because the translation engine does not have to create large tree structures that might otherwise be created without marking. Also, the translation engine can free up memory (eg, RAM) space because it does not have to store large tree structures that must be stored if not marked.

[0010] In yet another aspect, the present invention relates to yet another improvement of an automatic natural language translation system. This is an improvement related to parsing using a dictionary with headwords that are very unlikely or never to be a component of the final tree structure. If the dictionary contains a compound headword and at least one of the compounds has a very low probability value, this low probability value indicates that the headword has any value in the source natural language. Indicates that it is unlikely that the correct parse tree of the sentence could function as a particular part of speech. If the parse tree structure of an input sentence includes a headword that is marked, the parse tree is included in the sentence because the probability value associated with the headword that is marked is low. On the other hand, the probability of being accurate is very small. By using this probability value, unnecessary parts of the tree structure created by the translation engine of the system can be eliminated.

The above and other objects, features, features and advantages of the present invention will become apparent upon reading the following description and the appended claims.

[0012]

DESCRIPTION OF THE PREFERRED EMBODIMENTS First, an outline of an automatic natural language translation system of the present invention will be described without referring to the drawings. After describing the outline, description will be made with reference to the drawings.

The automatic natural language translation system can translate a source natural language into a target natural language. In a preferred embodiment, the system translates English into Japanese.
An apparatus for receiving and storing a source natural language and a translation engine for producing a translation into a target natural language;
An apparatus for displaying a translation result to a user; and an apparatus for providing and displaying another translation result to a user. As a specific example of this system, the translation engine
Includes pre-parser, parser, graph creation device, evaluator, graph scorer, grammar structure extraction device and structure conversion device. The preparser examines the input text and analyzes the ambiguous parts of the sentence end recognition of the input sentence. The preparser then creates and displays the input text in an analysis chart that includes the dictionary headword. The parser performs an analysis of the chart to obtain a possible syntactic classification for the input text. The graphing device generates a graph of the possible parsing of the input text based on the analysis chart. The graph includes nodes and sub-nodes that are relevant to the possible interpretation of the input text. An evaluator containing a series of experts evaluates the interpretable graph and adds expert weights to the nodes and subnodes of the graph. The graph scorer uses expert weights to evaluate sub-nodes and then associates each of the N top scores with each node. The grammar structure extraction device
Assign the parse tree structure to the preferred interpretation determined by the graph scorer. The structure conversion device performs a structure conversion on the parse tree structure in order to obtain a translation in the target language.

In the following three paragraphs: (a) To calculate the final weighted score for each subnode,
How the scorer combines expert weights; (b) how the graph scorer combines subnode scores to arrive at the final node score; and (C) the linguistic information is a tree of nodes and subnodes. Is described.

To calculate the final weighted score for each subnode, the graph scorer associates a constant value with each subnode. Analysis of the linguistic information associated with each subnode determines the score of the subnode. For example, see FIG. 8 where a series of expert evaluators examine the linguistic information stored at each node and subnode. The graph scorer computes the sum of the individual weighted scores for each expert to obtain a final weighted average for a particular node or subnode. Combining multiple weighted scores into one weighted average score is a standard problem in computer science. One method that can be used is to multiply the result of each expert by a constant (weight) assigned to that expert. The weight assigned to each expert is a matter to be determined at the time of design.
The designer determines the priority (weight) assigned to each expert. The weighted average is the sum of each number multiplied by a constant and a series of numbers. For example, the following equation is obtained. Weighted average = (w ₁ ) * (x ₁ ) + (w ₂ ) * (x ₂ )
+. . . + (W _n ) * (x _n ) where the weights w ₁ , w ₂ ,. . . , W _n are non-negative numbers, and their sum is 1. See, for example, Spiegel, "Theory and Problems 76 of Probability and Statistics" (1975, McGraw-Hill, Inc.), which describes the use of weighted averages for statistical expectations.

[0016] To combine subnode scores to obtain the final node score, the graph scorer can communicate the subnode scores from the bottom to the top of the graph. In the case of a graph in which each node has a set of N scores, one or more transmission methods can be determined. One technique that can be used to communicate subnode scores is:
There are storage methods that are some kind of dynamic programming used to solve optimization problems. Solving the optimization problem can involve many possible numbers (results). The goal is to find the optimal number. The algorithm used for optimization solves each sub-sub-problem only once and stores the result, so that it is not necessary to recalculate the answer each time a sub-sub-problem is encountered. For a more detailed description that applies to optimization problems, see, for example, Komen et al., "Invitation to Algorithms", McGraw-Hill, 1990, pp. 301-314. Pages 301, 302, and 312 of this "Invitation to Algorithm" describe one method that can be used to convey subnode score information in the graph.

When transmitting linguistic information in a tree,
The semantic portion of the system operates to convey semantic information from smaller internal components to larger components. The transfer of semantics is based on the four classes of syntactic classification used in the parsing operation (SEMNP, SEMV
P, SEMADJ and VERB). Before meaning transmission can take place, the linguistic information stored in the node must be analyzed. Analysis of semantic information stored in the node checks which noun-like element and verb-like element of the grammar rule apply to which nominative object and which selective restriction slot of the verb-like element of the grammar rule Guided by a set of rules that guess. Gerald Gazda, in his book "Natural Language Processing of Prologues" (Addison Wesley Publishing Company, 1989), stores in nodes of non-ring-shaped graphs that have similar orientations to the graphs disclosed herein. Describes a set of rules that can be used to analyze the semantic information being performed. Gazda describes the use of property matching to match information about neighboring nodes. Gazda states that property matching involves the equations described below.

"Some properties that appear on one node must be the same as properties that appear on another node. Recent research shows the classification of parents and the morphemes associated with those properties. This also assumes the principle of equalizing the details of certain classes of properties that appear above, which is called the "head" of the phrase. Most phrases have only one head. Thus, for example, a verb phrase inherits the tense of that verb. Because the verb is the "head" of the verb phrase. Using the notational resources used so far, there is no easy way to specify this principle that applies to the entire grammar. However, if we assume that the relevant properties can all be found on one branch of the DAG, we can very easily state the effect of this principle on a rule-by-rule basis. 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 characteristic value of the head of the parent VP are the same. There must be. The rules discussed in Gazda can be readily applied to each class of syntax disclosed herein. The linguistic information assigned to each node using Gazda's rules can be propagated through the tree by storage techniques.

Here, to summarize the contents of the above three paragraphs, the weighted average is one method of determining the subnode score, and each subnode score is a well-known storage technique applied to the optimization problem. Can be conveyed through the graph, and the method described in Gazda's book is
The linguistic information stored at each node can be used to analyze, which means that the linguistic information can be conveyed in the grammar structure analysis chart using storage method techniques.

In the automatic natural language translation system, re-translation can be automatically performed after the first automatic translation is completed.
That is, the system automatically provides the possible and most appropriate translation of the input text information and outputs it to the user (preferably,
After providing the Japanese translation of the input English text), the user can interact with the system to modify the displayed translation or automatically obtain another translation.

An automatic natural language translation system uses a language model that breaks one sentence into substrings. A substring is one or more phrases that occur in the order specified as part of the sentence. For example, the substring "The man is happy" is "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 define substrings in different ways and at different levels of detail. For example, in the sentence "They wou1d 1ike an arrow.", "An arrow" is usually classified as a noun phrase (NP). Another model classifies “an arrow” by syntactic properties (eg, singular noun phrases) and literary properties (weapons). If the meaning of this phrase is ambiguous, there are several ways to classify it, for example, "an arrow"
It can also mean an arrow-shaped symbol. When language models provide a way to resolve ambiguities, they typically resolve ambiguities by combining smaller units into larger units. When evaluating larger units, these models consider only the information contained in the larger units.

As a specific example of this system, "an
The semantic characteristic of the arrow (symbol 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."
The semantic property of "an arrow" is not used in evaluating the verb phrase "shot it with an arrow".

For any substring of a sentence in which a particular language model (interpreted substring) is interpreted in one way, there is an exported attribute. The exported attribute is the interpreted substring. And all the attributes used to evaluate the combination with other units that form a larger substring. Export is an interpreted substring, interpreted along with the exported properties. Attributes included in the interpreted substring but not exported are called substructures.

The parser of the system includes a grammar database. The parser uses grammar rules to find all possible interpretations of the sentence. The grammar database is X = A1
A2. . . It consists of a series of context-free phrase structure rules in the form An. X represents A1, A2,. . . , An, or a higher-level node (sub-node) A1 to a lower-level node (sub-node) A
Called n.

The graphing system of the system graphically represents many possible interpretations of a sentence. Each node in the graph corresponds to the export of a substring.
As a specific example of the system, one export is represented by one node. The graph includes an arc emanating from a node associated with one export, where the arc represents a substructure of the export based on the application of a grammar rule. The graph represents at least two types of arcs, the export substructure. The graph shows at least two types of arcs: (1) a single arc pointing to one different export of the same substring;
(2) Two arcs containing a set of pointers to the two exports, ie, when concatenated, their substrings form a substring of the original export. Note that the formula in (2) assumes Chomsky's normal form grammar. The amended claim 35 applies, by paraphrasing type (2), to a grammar reflecting an arc with an N-fold pointer pointing to N exports, rather than a Chomsky normal form grammar.

The graph includes a single starting export point S from which all parts of the graph can be reached by following a series of arcs. The departure export corresponds to the entire sentence.

A plurality of arcs start from one node only if the same export can be formed from a plurality of exports. (A set of pointers within an arc of two arcs is not considered multiple arcs in this sense.)
Multiple arcs will point to one node only if the export is one element of multiple exports. Nodes with no arc correspond to dictionary entries assigned to the substring.

A plurality of language experts assign a numerical score to the set of exports. The language expert applies the score to each node of the graph. As a specific example of the system, a score matrix (each element of the matrix is
The weight for multiplying the score of a particular expert) is the fixed length "N" of the floating point number for any sentence.

The scores are evaluated by a scoring module incorporated into the graphing engine and / or parser. A score is calculated for all exports that make up the higher export. The score for the higher export is calculated as the sum of the scores of any expert applied to the combination of the export forming the higher level export and the score assigned by the structural adjustment expert.

The order of arriving at the node and examining the score is:
This is the standard depth first graph movement algorithm. In this algorithm, the scored nodes are marked and will not be scored again. During the scoring process, the scoring module will evaluate the higher unit of any node before evaluating it.
Evaluate the headword node of the dictionary. Each entry in the dictionary has one score.

When there are a plurality of methods for performing one export, a plurality of scores are obtained. That is, if there are k ways to export, there are k possible scores. The multiple scores are processed as follows.

(1) In a single element rule, each of the k lower export scores is added to the expert value applied to the single element rule, resulting in A vector of k scores is associated with the parent export.

(2) In the rule composed of two elements, it is considered that the left child has a g score and the right child has an h score. After that, the number obtained by multiplying the g score by the h score is calculated by adding each score of the right child to each score of the left child, and further adding the expert value applied to the rule of two elements. Is done. If the g score multiplied by the h score exceeds N, only the highest N score is kept with the parent node.

(3) If an export can be made in multiple ways, at most N scores are added to the score list for that node, and only the highest score is kept.

Once the score calculation has been completed, the method described above is such that each export includes its nodes and the attributes of all substructures not represented in the export, the g most likely to export. Associated with a set of g scores (g in the range from 1 to N) that represent the methods (with respect to the language model) that are high. In special cases, such as the root node S, this scoring method gives the g most likely ways to form the whole sentence.

Each score in the above score list has an associated pointer. The pointer, which provides information indicating which scores in the lower export score list have been combined to create a higher level score. By tracking each pointer, the g most likely interpretations of the sentence can be extracted as an unambiguous parse tree.

The automatic natural language translation system will be described in more detail with reference to FIGS. After that, FIG.
Various improvements of the present invention will be described with reference to FIGS.

Referring to FIGS. 1 and 2, the automatic natural language translation system 10 of the present invention includes an input interface 12, a translation engine 16, a storage device 18, a user input device 22, a display 20, and an output interface 14. The input interface is adapted to receive a series of texts written in a source language such as English. As the input interface,
It may include a keyboard, a voice interface or a digital electronic interface such as a modem or serial input. The translation engine translates the source language using the data in the storage device. The translation engine can be made entirely of hard-wired logic, or can include one or more processing units and associated storage instructions. The translation engine may include the following elements and parts thereof: pre-parser 24,
A user interface 42 including a parser 26, a graph creation device 28, a grammatical structure analysis / translation evaluator 30, a grammatical structure analysis extraction device 32, a structure conversion device 34, and another grammatical structure system 37. The structure conversion device includes a structure conversion device 36 by grammar rule control, a structure conversion device 38 by dictionary control, and a structure conversion device 4 by generation rule control.
0 may be included. The storage device 18 may include one or more disks (eg, hard disks, floppy disks, and / or optical disks).
And / or may include memory storage (eg, RAM) and the like. These storage devices can store all or some of the elements described below. That is, a basic dictionary 44, a technical term dictionary 46, a dictionary created by the user, a grammar rule 48, a generation rule 50, a semantic characteristic tree 52, a structural characteristic tree 54, and a graph 56. Storage 18 is used or useful for performing input text information written in source natural language, output text information written in target language, and translations including one or more dictionaries, domain keywords and grammar rules. Used to store all kinds of information. User input interface 22 includes a keyboard, mouse, touch screen, light pen, or other user input device and is available to the operator of the system. The display may be a computer display, a printer or other type of display, or may be another device for providing information to an operator. The output interface 14 exchanges the final translation of the source text in a target language such as Japanese. The interface may include an electronic interface such as a printer, display, audio interface, modem or serial line, or may include other devices for sending text to an end user.

As an operation of a specific example of the translation system of the present invention, as shown in FIGS. 1, 2 and 3, the preparser 24 first performs a preliminary analysis operation (step 102) on the source text 23. Do. This includes analyzing the ambiguity of the end-of-sentence recognition of the source text and creating a structural analysis chart that includes the dictionary headword 25. Thereafter, the parser 26 performs a structural analysis of the chart created by the preparser (step 104) in order to obtain a structural analysis chart in which the syntax possibility 27 is described. The graph creating device 28 creates a graph of the possible interpretation 29 based on the structural analysis chart obtained in the structural analysis step (step 106). The evaluator 30 accessing the series of experts 43 evaluates the stored interpretation graph (step 108) and adds expert weights to the graph 31. Graph scorer 33
Scores the nodes and associates the N (eg, 20) highest scores 35 with each. The grammatical structure extractor 32 assigns the structure analysis tree structure 39 to this preferred interpretation (step 110). Thereafter, the structure conversion device 34 accessing the conversion table 58 performs a structure conversion process (step 112) on the tree in order to obtain the translation 41 that has become the target language. The user may use another structural analysis system 3 to obtain another translation.
7 can be interacted with.

Referring to FIG. 4, the system of the present invention comprises:
Preliminary structural analysis is performed by dividing the input word string into tokens (step 114) containing the individual punctuation marks and the character groups forming the word. The appearance of spaces affects the interpretation of characters at this level. For example, "x
"-" In "-y" is a dash, but "-" in "xy"
Is a hyphen.

Thereafter, the preparser combines the tokens into words (step 116). At this level, the preparser recognizes special structures (eg, Internet addresses, telephone numbers and social security numbers) as a unit. The preparser also performs a dictionary lookup to separate groups. For example,
If "re-enact" is listed as "reenact" in the dictionary, it will be one word, otherwise it will be three separate words.

In the next preliminary structure analysis stage, sentence end recognition is performed to determine where the sentence ends (step 118). During this process, the preparser performs a series of steps to identify the possible end-of-sentence of each sentence (i.e., after each word of the source text), and provides a basic dictionary, a technical term dictionary, and an on-board dictionary. Access user-created dictionaries.
The preparser does not need to perform this step if a particular order is specified, and the steps can be performed as a series of ordered rules, or can be hard-coded into the hardware. .

Referring to FIG. 5, when there is a sequence of symbols that cannot be analyzed, such as a series of dashes "-----", the preparser does not translate each one, but interprets them as one "sentence" as a whole. And record (Step 1
20). The preparser requests two carriage returns at the end of the sentence (step 122). If the first letter of the next phrase is lowercase, the preparser does not consider it the end of a sentence (step 12).
4). If a sentence begins with a new line and is short (eg, a title), the preparser considers it a single sentence.

The preparser considers a terminator (.), Question mark (?), Or exclamation point (!) As the end of a sentence, unless it includes closing parentheses and closing quotes (step 128). In the case of a sentence ending with ".""Or"?"", The preparser uses virtual punctuation after the quotation marks in addition to punctuation before the quotation marks. The following example shows a method for punctuation that is virtually added to “?”.

Did he asks the question "What do you want?"? Are you concerned about "the other peop1e"? In English, the above sentences are replaced with "?" It is likely to be over. The presence of the virtual punctuation added by the preparser indicates that there is something like a question mark before the quotation mark, or nothing at all. Behind the quotes are something like a terminator or question mark. The remaining grammatical structure of this sentence allows the most appropriate choice to be made at a later processing stage.

The preparser also uses a number of additional methods in analyzing the terminator (steps 130, 13).
2, 134, 136 and 138). Some abbreviations included in the dictionary are marked at the beginning of the sentence as being never used, and some are marked at the end of the sentence as being never used (step 130). These rules are always respected. For example, "Ltd." is not used at the beginning of a sentence, and "Mr." is not used at the end of a sentence. The preparser also states that the following words are "the", "i
n "unless it is a frequent word (step 13
2) If a capital letter has a terminator, it is not considered to be the end of the sentence. If the word before the stop is in one of the dictionaries, the sentence ends at the stop (step 134). The word before the terminator is not listed in the dictionary, the term has an internal terminator (eg, IBM), and the next term is not listed in the dictionary as lowercase, or If the next phrase itself is uppercase, the sentence is considered not to end at the end (step 136). Otherwise, the terminator indicates the end of the sentence (step 138).

Referring again to FIGS. 2 and 3, when a sentence break is specified by the preparser, the parser places the phrase of the sentence into a syntactic classification and calculates a possible syntactic interpretation 25 of the sentence. Apply the grammar rules of the grammar database to those phrases (step 104). Grammar rules 48 may be implemented as a set of computer-capable rules that represent the grammatical restrictions of the language,
In English, there are hundreds of such rules, and these rules can apply to hundreds of syntactic classes. To reduce the extra time required to calculate this task, the different possible meanings of a phrase are ignored.

In the next step (step 106), the graph creation device fetches different meanings of the words and phrases,
Extend the parser results using a dictionary to create a non-ringed graph with a direction that represents all the semantic interpretations of the sentence. This graph is created with the help of a sequence of semantic communication described later. These procedures are based on a series of created grammar rules, and in some cases, access semantic property trees for semantic information. The semantic property tree is a tree structure including the semantic classification. This tree is roughly organized from abstract to concrete, with two terms meaningful both from how far apart in the tree and what each level in the tree is. Can determine how they are related. For example, "cat" and "do
"g" is more relevant than "cat" and "pudding", so two of "cat" and "dog" have a shorter distance in the tree of "animal", and "animal", "cat" Are examples stored at different levels of the tree. Because "animal" is a more abstract word than "cat".

Referring to FIG. 9, the graph of this figure includes a node 80 whose subnodes 82, 84, 86 are pointers 88, 89, 9 in a manner that indicates various types of associations.
It is linked at 0,91. The first type of association in the graph is that the node representing the phrase has pointers to the nodes of the words and subphrases that make up the phrase. For example, node 84 representing "the bank"
Are linked by pointers 92 and 93 to the words "the" 94 and "bank" 95 that make up them.
The second type of association in the graph is when the phrase interpretation has a pointer to another way to make the same higher level component from the lower level.
For example, a node 80 representing the phrase "by the bank" has two source interpretation locations 8 containing pointers 88 and 89 and 90 and 91 linking to each of the constituent words.
1,83. In this example, the different individual constituent words include different subnodes 84, 86, each representing a different meaning for the phrase "the bank". The structure of the graph is defined by the results of the parsing work and is restricted by the syntax of the source sentence. The nodes of this graph are associated with storage locations for the semantic information entered during the semantic transmission process.

The parts that convey the semantics of the system serve to convey semantic information from the smaller components that contain them to the larger components. The semantic information applies to the four classes of syntactic classification used in the initial parsing tasks. The four classes are SEMNP (including noun type objects and prepositional phrases), SEMVP (usually subject, object-like verb phrases), SEMADJ (adjective), and VEM
ERB (verb-type verb in a dictionary that often takes an object)
It is. Other syntax classifications are ignored by certain rules.
The grammar rule setter can override actions that do not appear on the surface by marking the rule with a specific mark. These special orders come first.

The method of transmitting semantic characteristics in the system has two aspects. The first aspect is to examine the noun-type and verb-type components of the grammar rules to find out which optional restriction slots of the verb-type component apply to the noun-type object. A set of rules that can be done. For example, the rule for the verb phrase of the sentence “I persuaded him to go.” Is roughly VP = VT11 + NP
+ VP (where VP is a verb phrase, VT11 is a transitive classification, and NP is a noun phrase). An exemplary default rule is that if the verb takes an object, the first NP to the right of the verb must apply a selection restriction. Another rule states that VP restrictions on the subject must apply to the first NP to the left of the VP. When these rules are put together, "pers
The evaluation is made so that the meanings of both "uade him" and "him go" are well understood. As already explained,
These rules reflect the complex rules of English, so their numbers are very limited.

Referring to FIG. 6, the semantic transfer operation includes copying the selection restrictions from the SEMVP to the statement (step 140). If SEMNP is used as a position expression, the validity of the semantic constant specifying a good position is evaluated (step 142). If a rule contains the union of two SEMNPs (detected by ANDing syntactic features), the graph creator will logically intersect the semantic properties and generate a semantic distance expert. (Step 144).

In examining the rules specified for the transmission of semantic properties, the grapher may use a higher level (eg, a SEMNP that contains more words).
If it finds the position of the "head" of the SEMNP to be transmitted to the (part of), the graph creation device also transmits the semantic property (step 146). However, if the “head” is a word for division (eg, “Portion”, “Part”), the “head” is transmitted from the SEMNP to the left or right.
SEMVP and SEMADJ are communicated in the same manner, except that they do not have a location for partitioning (step 148). Adjectives are part of SEMVP in this sense.

If the SEMVP was created from a rule that includes VERB, the graphing device propagates the subject constraint of VERB upwards, except when the VP is passive. If passive, the first object restriction of VERB is communicated (step 150). In the case of a rule including SEMVP, the graph creation device sets the selection restriction of SEMVP to S.
Attempt to apply to NPs encountered when moving left from EMVP (step 152). In the case of rules involving SEMADJ, the grapher will attempt to apply the SEMADJ selection restriction to any SEMNP encountered when first moving right from SEMADJ, and if that fails, the left direction (Step 154).

For any remaining unused object selection restrictions in VERB (which have not been communicated upwards since they are passive), the grapher will generate the SEMNP encountered on the right side of VERB. The above restrictions are applied in order (step 156). In all these rules, the verb selection restriction runs out as soon as it applies to something. For all previous rules, SEMNP
Are not exhausted if anything is applied to those SEMNPs. Starting from this rule, SEM
NP is exhausted. Finally, one rule is SEMV
When P is created, the graphing device determines whether a previously unused SEMVP or SEMADJ is included, and if so, communicates it upwards ( Step 158).

The system also performs language feature matching. Linguistic characteristics are the characteristics of words and other components. Syntax matching is used by the parser, and semantic matching is used by the graph creator. But,
The same scheme is used for both. For example, "the
"y" has a syntactic property of plural, while "he" has a syntactic property of singular. Feature matching marks grammar rules so that the features of the phrases that are applied only if the features of the phrases to which they apply meet certain conditions. For example, assume that there are the following rules.

S = NP {+ VP} where the symbol {} means that the characteristics of the numbers of NP and VP must match. Thus, this rule states that "they are" and "he is" are correct, but that "they are"
"is" and "he are" are not allowed.

The characteristic match restriction is divided into “local” and “wide area”. Extensive actions are calculated when the grammar is created, not when the sentence is actually processed. The sequence of the wide range of operations that must be performed is then coded as instruction bytes.

The calculation of "wide range" characteristic operation must start with a rule of n elements (ie, having more than one element to its right). afterwards,
The system assigns codes to the various two-part rules such that the set of properties ends up propagating between the rules in the correct way. By dividing a rule consisting of n elements into a rule consisting of two elements, the analysis task is greatly simplified. However, the system retains the power of "widespread" processing because the system tracks a set of properties across a two-element rule.

In the system of the present invention, the dictionary treats the words as even constituting phrases, but also includes "idioms" composed of a plurality of phrases. These two forms will eventually compete with each other for the most appropriate translation. For example, in the dictionary, the meaning of “b1ack sheep” is registered as “surplus”. However, in some cases, the phrase "b1ack sheep" may mean "black sheep." Since both of these forms are retained, this non-idiom translation is also selected as the correct translation.

This idiom also belongs to another classification. For example, the system can use three types of classification:

Almighty: United States of Americ
a Priority: long ago Normal: b1ack sheep Almighty idioms take precedence over possible interpretations of any of the words that make up the sequence. Preferred idioms take precedence over any possible interpretation when the words making up the sequence have the same general usage. Ordinary idioms compete with other headwords at the same level.

The resulting graph is evaluated by an expert (step 108, FIG. 3), which provides a score that indicates the likely accuracy of the interpretation of the graph. The system of the present invention includes a scoring method that applies to all parts of any length of the sentence, not just the entire sentence. An important factor in using a graph is that, even if it is used in so many statements,
That is, the subtree is completely scored and analyzed only once. For example, "Near the bank there
In the phrase "is a bank.", "Near the bank"
The phrase has at least two meanings, but only one decision is made to determine the most appropriate interpretation of the phrase. The phrase "there is a bank" can similarly be interpreted in two different ways, but the decision of which of these two interpretations is most appropriate is made only once. Thus, this sentence can be interpreted in four different meanings, but the subphrase is scored only once. Another feature of this graph is that each node has easily accessible information about the length of that part of the sentence. This feature allows the N most appropriate interpretations of any substring of the English sentence without having to re-analyze the English sentence.

In a single run, the most appropriate N analyzes of the sentence are obtained each time (N is some number of the order of 20), but by using a graph, the system allows more N that can incorporate the results of user selection for small components and respect user selection
Perform the most appropriate analysis. All of these analyzes can be done quickly, without re-analyzing the sentence or scoring any substrings again.

Referring to FIG. 8, expert evaluator 3
A zero action is based on various factors that characterize each translation and are processed by various experts. The probability rule expert 170 evaluates the average relative frequency of the grammar rules used to obtain the initial source language parse tree. Selection restriction expert 178 evaluates the degree of semantic matching of the resulting translation. The dictionary headword probability expert 172 evaluates the average relative frequency of a particular “part of speech” of some phrases in the sentence used to obtain the initial source language parse tree. The statistical expert evaluates the average relative frequency of a particular paraphrase selected for a translation.

The system automatically determines the English "part of speech" (POS) for individual English words, phrases, and word groups. When the system translates a sentence, it automatically determines the part of speech and usually makes the right decision. However, sometimes the sentence to be translated and its whiteness are ambiguous.
When words or phrases that can be interpreted as different parts of speech are included, a plurality of different and all "correct" interpretations are obtained. The system operator may ignore the part of speech that is automatically determined by the system and instead manually set any part of speech for a word, phrase or group of words. For example, "John saw a boy with
a telescope. ", the system operator considers" a boy with a telescope "to be a noun phrase,
The sentence would be interpreted to mean "the boy had a telescope" and not "John used a telescope to see the boy." If the operator overwrites the part-of-speech rules determined by the system by making several possible part-of-speech settings or manually setting a more restrictive part-of-speech setting, the translation result will be worse or at least not improved May happen. Noun phrases are less restrictive than nouns, and groups are the least restrictive parts of speech. 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 Preposition phrase Conjunction Group English
This is useful when the meaning depends on how the system interprets a particular prepositional phrase. For example,
The sentence "We need a book on the fourth of July" is interpreted as saying "on the fourth of July" has the meaning of an adjective. However, if "on the fourth of July" is interpreted as an adverb phrase, it means "On July 4, we want a book." The system automatically displays "on the fourth of Jul
If the operator thinks that he has assigned the wrong part of speech to "y", the operator will see "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. If the operator does not want the system to translate a particular word, phrase or group of words from English to Japanese, such words, phrases or groups of words can be set to the part of speech "English". The operator can remove one or more part-of-speech settings, whether the settings are made automatically by the system or manually by the operator.

The system tracks translation usage statistics at multiple levels for each user. For example, the system maintains statistics at the level of the surface form of a phrase (the frequency at which "leaving" was used as a transitive or intransitive verb), and at a semantic level (meaning "remaining after" or "~ Or "leave", "leaf", "lef"
The number of appearances is accumulated for each of the different variations of "t" and "leaving". The system may also keep a distinction between the usage statistics used in the last few sentences and the usage statistics at any time for the user. Further, the system can distinguish between cases where the user has intervened and instructed to use a particular meaning of a phrase, and cases where the system has used a particular meaning of the phrase without user intervention. .

The structure adjustment expert 182 is a feature related to the length of a certain sentence component, and is based on features common to English and many other European languages. With some (but not all) constructs, these languages do not welcome sentences in which light (short) elements have heavy (long) elements to the left of them. For example, Mary hit Bill with a broom. (Left is light, right is heavy)
(Applicable) Mary hit with a broom Bill. (Heavy left and light right)
(Non-conforming) Mary hit with a broom a dog that tried to bite he
r. (heavier on the left and heavier on the right) (conform) When there are two parsings of a sentence, one constructs a "heavier left and lighter right" sequence that includes structures that try to avoid such sequences If so, and the other parsing is not, the former is deemed not to represent the intended interpretation of the sentence. This expert is an effective way to distinguish between intended and unintended analyses.

In the equivalence structure of the pattern “A of B and C”, it is determined whether the intended interpretation is “A of {B and C}” or “A {of B} and C”. Can be difficult. Equivalent structure expert 180
Measures the semantic distance between BC and the semantic distance between ACs to determine which equivalence mode will combine the two elements that are closer together. This expert accesses the semantic property tree during processing. The expert is also an effective way to distinguish between intended and unintended analysis of a sentence.

Many English words contain potential ambiguity in their interpretation as common nouns and as proper nouns. The capitalization expert 176 uses the position of the capitalization in the sentence to determine whether capitalization is significant. For example, the following statement is Brown is my first choice. My first choice is Brown. The first sentence is inherently ambiguous, while the second is
"Brown" differs in that it is much more likely to be a person's name than a color name. The expert will determine whether words starting with capital letters are at the beginning of a sentence or not at the beginning of words (as in the example above), whether words in capital letters are included in the dictionary, and words that start with lowercase letters are registered in the dictionary. Factors are taken into account. This expert is an effective way to correctly interpret capitalized words in sentences.

When a sentence initially includes a sequence of uppercase words, the sequence is treated as a proper noun or a common noun. The system of the present invention uses a capitalized sequence procedure, giving preference to the former interpretation. If the sequence itself cannot be parsed according to normal grammatical rules, the sequence is processed without translation as a set of unparsed noun phrases. This step is
It has proven to be a very effective means of handling compound proper nouns without completely ignoring the interpretation of low-common-level common nouns.

Referring to FIG. 7, the machine translation system of the present invention has the efficiency of a structure conversion method based on a simple grammar rule, but has a grammar rule control structure conversion mechanism that is close to the capability of the inter-template structure conversion method. 162 is used. This method uses grammar rules 160 that can specify an uneven composite structure. The format of the rules used in other translation systems is shown below.

[0076] The system of the present invention uses the following grammar rule format.

[0077]

(Equation 1) In this syntax, a symbol preceded by a "#" is a symbol that is not visible for the purpose of sentence structure analysis, but is a virtual symbol used to construct substructures once the analysis is available. .

Given this type of grammar, it becomes possible to specify a plurality of structural transformations between any sequence of nodes in a child relationship of a substructure. As a result, the structure conversion mechanism based on the grammar rules is converted into a mechanism having some capabilities of the inter-template structure conversion mechanism. Although the system of the present invention is based on the second type of grammar rule above, it automatically creates a corresponding grammar rule in the first form. Thus, a first form of grammar rule can be used to parse a sentence, and a second form of grammar rule can be used to form a grammar parsing structure.

The structure conversion includes a dictionary control structure conversion operation 166 for accessing the dictionary 161 to operate on the parse tree after being operated by the grammar rule control structure conversion operation. Thereafter, a production rule structure conversion operation applies the production rules to the resulting parse tree to provide the target language text 41.

Referring again to FIGS. 1 and 2, after the system has produced the most desirable translation according to the above process, the translation is provided to the user through display 20. The user can then either adopt the translation or use another analysis system 37 via the user input device 22.
The operation can be performed by operating. During rework,
The user can request retranslation of the other parts while keeping the part of the translation result that has been correctly translated. This can be done quickly. Because the system keeps a graph containing the expert weights 31.

The automatic natural language translation system has been described in some detail with reference to FIGS. Hereinafter, 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, the translation engine 16 of the automatic natural language translation system receives the source text 23 and automatically translates it into the target natural language text 41. This translation is affected by the "automatic region identifier" function according to another aspect of the invention. With this feature,
When performing translation, the translation engine 16 lists one or a plurality of area keywords 200 (N is shown in the figure).
And access to the dictionary 202. The dictionary 202 includes headwords that receive different translations depending on the area where the headword is used. Each headword in dictionary 202 may be one or more phrases written in the source natural language.
Generally speaking, each separate area has a list of area keywords, and FIG. 10 shows N areas. Each entry in the list of keywords may also be one or more words written in the source natural language. The list of keywords 200 can be separated from the dictionary 202 as shown, or
As shown in FIG. 4, it can be built in the dictionary.

An area can include a set of terms and usage patterns that belong to a certain field of usage or a particular population. For example, one area may include documents in technical fields such as business communications, marketing documents, computer related documents, and physics.

The dictionary 202 has different target natural languages (for example, Japanese) when it is used in relation to a certain area, when the phrase is used in another area, and when it is not used in any specific area. Includes phrases with translations. The keyword 200 is associated with a specific sentence (or input document) written in the source natural language 23 in relation to each area.
Is a region-specific word or term used to determine if a region belongs to one of the possible regions. In each list of the area keywords 200, each keyword has a numerical value indicating the strength as a keyword for a specific area in association with the keyword. For example, the “computer” area list contains “open-architecture compu” with a value of 100%.
ter ”, the value of 100% means that the sentence containing this keyword is always a sentence in the“ computer ”area, and the keyword“ computer ”in the“ computer ”area is 25% The value of 25% means that a sentence having this keyword is less likely to exist in the “computer” area.

With the "automatic region identifier" function, the translation engine 16 preferably provides the strength of a keyword in a sentence, document, or part of a document so that the sentence or document can be determined to belong to a particular area. Determine the sum of the values of (or whether there are enough keywords). When a keyword exists in a certain area, the word or phrase is listed in the dictionary 202 as an area designation word, and the probability of the word or phrase included in the sentence or document to be translated is increased, and the sentence or document is translated. The above decisions made by translation engine 16 are based solely on the source natural language and keywords.

The translation engine 16 determines at least one area to which at least a part of the input text information 23 belongs by one of the following two methods, for example.
The first method is to address each sentence of the input text information 23 one at a time and count the number of occurrences of the keyword in each area. For each region having keywords equal to or greater than the predetermined number, the translation engine 16
By using a translation specific to the region, the probability of some phrases included in the sentence translated into the target natural language 41 is increased. The above-mentioned predetermined number can be fixed to the same value for all the sentences of the input text information 23, or can be changed for each sentence according to the length of the sentence (the number of phrases). The predetermined number can be set by the system 10 or can be set by a user of the system 10. The second method is a method of counting the number of appearances of keywords in each area appearing in a predetermined number of words of the input text information 23, such as the first N words. For each region having more than a predetermined number of keywords, the translation engine 16 uses a translation word specific to that region to convert several words contained in the sentence translated into the target natural language 41. Increase the probability. Similarly, the predetermined number can be set by the system 10 or by a user of the system 10.

In the preferred embodiment of the present invention, translation engine 16 does not count each keyword, as described in the previous paragraph, but rather, to obtain the sum of the strength values for each region. Input text information 2 by using the strength value associated with the keyword
3. Determine at least one region to which at least a part of 3 belongs. In this preferred embodiment, translation engine 16 otherwise operates as described in the previous paragraph. That is, in the first preferred method, each sentence of the input text information 23 is addressed one at a time, and for each region where the sum of the keyword strengths is greater than or equal to a predetermined number, The translation engine 16 increases the probability of some phrases included in a sentence translated into the target natural language 41 by using a translation specific to the region. The above-mentioned predetermined number can be fixed to the same value for all the sentences of the input text information 23, or can be changed for each sentence according to the length of the sentence (the number of phrases).
The predetermined number can be set by the system 10 or can be specified by a user of the system 10. In the second preferred method, the sum of the strength values of the keywords in each area is determined by the keywords that appear in a predetermined number of words in the input text information 23, such as the first N words of the source text 23. You. For each region where the sum of keyword strengths is greater than or equal to a given value,
The translation engine 16 increases the probability of some phrases included in a sentence translated into the target natural language 41 by using a translation specific to the region. As described above, the predetermined number can be set by the system 10 or can be set by a user of the system 10.

As an example of the “automatic area identifier” function,
"Markup language" and "help desk" are included in the list of domain keywords in the "computer" domain,
Assume that "endusers" has two different translations depending on whether "end users" is used in the "computer" domain. When the following sentence is part of source text 23, Markup language should be easily used by end users
without being supported by help desk. Translation engine 16
Are two keywords in the "computer" area, "mark
"up language" and "help desk". By finding these two keywords in the "computer" area, the translation engine 16 can determine, based on the threshold, that the sentence is a sentence in the "computer" area. Once the sentence is determined to belong to the "computer" domain, the translation engine 16 then translates the "end users" specific to the "computer" domain into target natural language text 41 provided to the user. To increase the chances of using the translation of "end users", the probability of translation of "end users" specific to the "computer" area is increased. In addition to the mode in which this function works for each sentence, the "automatic region identifier" function can also work on the entire input text. For example, the translation engine 16 determines that the determined number of keywords corresponding to a specific area is the source text 2
If it is within the first n words of n (where n is an integer), it can be determined whether the entire source text 23 belongs to the "computer" area. In the case of the full text mode, all the phrases of the source text 23 having the translation specific to the area included in the dictionary 202 are translated using the translation specific to the area.

Referring to FIG. 11, the translation engine 16 of the automatic natural language translation system 10
And automatically translates it into the target natural language text 41, which translation is affected by some settings and markings applied to various rules of grammar rules 206 according to other features of the present invention. receive. These settings are “almighty” and “part of speech priority”. The translation engine 16 of the system 10 uses a grammar rule 206 that can be marked as “almighty” or “part of speech priority” to parse the source natural language 23 sentence 2
Have 6. The parser 26 uses a grammar rule marked “almighty” as a rule that prevents any other rule from being applied to any part of the source natural language 23 to which the almighty rule applies. Use The parser 26 uses a grammar rule marked "part of speech priority" as a rule that blocks only other rules relating to the same part of speech as a particular part of speech. A grammar rule that is not marked has no priority over other rules. Marking the grammar rules as "almighty" or "part of speech priority" is a very effective way of removing unnecessary parsing branches from the tree structure otherwise created by the translation engine 16 of the system. is there. This marking of grammar rules also eliminates the need for the translation engine 16 or the large tree structure that would otherwise have to be created without the above marking,
In addition, if the above-mentioned marking is not performed, it is not necessary to store a tree structure which becomes large, so that analysis can be performed more efficiently in terms of time and storage capacity.

As an example of the grammar rule marking function, it is assumed that the grammar rule includes the following three rules.

(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) Further, it is assumed that the system 10 uses a dictionary including the following three headwords.

[0092] When there are these grammatical rules and dictionary headwords, the word string "one eighteen-hundredth" included in the source text
Can be analyzed by the following two methods.

[0093]

(Equation 2) If grammar rule (1) is marked as "almighty", all other parsing of the word string "one eighteen-hundredth" (as in the second parsing above)
Is excluded, so the first is used in these two analyses. The almighty rule (1) governing all other rules is used. When the grammar rule (1) is marked as “part of speech priority”, the first analysis is selected similarly to the second analysis. In this case, the rules (1) and (2) are The rule (1), which addresses the same part of speech and is provided with a control (ie, part of speech priority) mark, causes the grammar rule (1) to be used for word string analysis. "Almighty", as the following example shows:
And "part of speech priority" are different, and "almighty"
When the translation is performed according to the rule with the mark, the result is not always the same as when the “part of speech priority” mark is added.

In the following four grammatical rules, the third rule includes:
It is assumed that the mark of “part of speech priority” is attached.

(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) In addition, the system 10 uses a dictionary including the following headword.

[0096] Given these grammatical rules and dictionary headwords, the word string "I work four days" contained in the source text,
It can be analyzed at the following three methods and phrase levels.

[0097]

(Equation 3) In the case of grammar rule (3) marked as "part of speech priority", grammar rule (3) is the same word string ("four days")
Grammar rule (2) excludes the application of all other grammatical rules that parse
Since “u days” is analyzed as a noun phrase, the first of the above three analyzes is used in preference to the second analysis.
However, since the part of speech has an adverb phrase different from the noun phrase, the above-mentioned third analysis is not excluded. Thus, grammar rule (3) is changed to "almighty".
Would not apply any other grammar rule (3), but if grammar rule (3) is marked as “part of speech priority”, only other rules that address the same part of speech as grammar rule (3) Excluded.

Referring to FIG. 12, the translation engine 16 of the automatic natural language translation system 10
And automatically translates it into the target natural language text 41, which translation is affected by the probability value associated with the dictionary entry according to another feature of the present invention.
According to this probability value property, the parser 26 of the translation engine 16 creates a dictionary 208 containing headwords that are either very unlikely to be a component of the final tree structure or that are marked as not being a component. use. Dictionary 20
8 includes compound headwords, at least one of which is marked with a very low probability value, as a specified part of speech in the correct parse tree of any sentence in source natural language 23. Indicates little probability of functioning. If the headword with the mark is included in the parse tree structure of a certain input sentence, the parse tree is not included in the sentence because the probability value associated with the headword with the mark is low. Is very low. This use of probability values can have the effect of removing the tree structure created by the translation engine 16 of the system 10.

As an example of using the value of the probability, it is assumed that there is a dictionary containing the following compound word entry words.

[0100] assumption that the entry word part of speech grammar rules probability in today prepositional phrase part of speech priority P (n) in this case, P (n) is to find words in a given source text sentence "in today" is in fact a prepositional phrase Play the probability value. It is assumed that P (n) is set to an infinitely small numerical value as compared with a normal compound word entry in the dictionary. The input sentence is:

This sentence is syntactically ambiguous and can be interpreted as the following two sentences: He didn't {come in} today. And He didn't come {in today}. (In today) has a very infinitely small value of the associated probability of a matching dictionary entry word, P (n), so the probability that this sentence will be parsed as a second sentence is very high. Less. Since "in today" is marked with the part of speech priority, there is no analysis other than analyzing it as a prepositional phrase that can give a normal probability value. Thus, improper parsing of this sentence is undervalued. If P (n) for "in today" is set to zero, the parser will use the "intoday-preposition phrase" to parse the above sentence.
Is not used at all. However, in this particular example, it is not desirable to set P (n) to zero, as it may be used in statements such as:

[0102] I am interested in today, and not in t
omorrow. In this case, it is appropriate to analyze "in today" as a prepositional phrase. Using a low P (n) value instead of zero is an undesirable grammar structure analysis.

He didn't come {in day}. Can be excluded, while I am interested in today, and not in tommorow. Can be correctly interpreted as parsing "in today" as a prepositional phrase. Can be analyzed.

All of the above functions and processes can be implemented by various hard-wired logic design and / or programming techniques incorporated in a general-purpose computer. The steps shown in the flowchart typically do not need to be applied in order, and some steps may be combined. Also, the function of this system is
It can be divided into programs and data in various ways. In addition, it would be advantageous if the grammar and other operating rules were provided to the user in a compiled format, while being developed in one or more high-level languages.

Any of the above embodiments of the automatic natural language translation system, including all of the features disclosed herein, may be implemented on general purpose computers (eg, Apple Macintosh, IBM PC and compatibles, SUN workstations, etc.). And a computer readable medium such as an optical compact disc (CD).

Those skilled in the art will appreciate that various changes, modifications, and alternative implementations of the invention described herein can be made without departing from the spirit and scope of the invention as set forth in the appended claims. Will be able to. Accordingly, the invention is to be defined not by the preceding illustrative description but instead by the spirit and scope of the following claims.

[Brief description of the drawings]

The same reference numbers in different drawings generally identify 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 data flow chart showing the overall function of the system of FIG. 1;

FIG. 3 is a flowchart showing the operation of the system of FIG. 1;

FIG. 4 is a flowchart showing an operation of a sentence end recognition function of a preparser 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 an operation of meaning transmission of the system of FIG. 1;

FIG. 7 is a flowchart showing the structure conversion operation of the system of FIG. 1;

FIG. 8 is a flowchart of an expert evaluator of the system of FIG. 1;

FIG. 9 illustrates a sample graph used by the system of FIG. 1 for the example phrase “by the bank”.

FIG. 10 is a diagram of an “automatic region identifier” function performed by an automatic natural language translation system according to one aspect of the present invention.

FIG. 11 illustrates one feature according to another aspect of the invention performed by the automatic natural language translation system, which is a characteristic that affects which grammatical rule is applied to translate a source text into a target natural language. Figure

FIG. 12 is a feature performed by an automatic natural language translation system according to yet another aspect of the invention, including a probability value assigned to a dictionary entry to remove improper translations. Function diagram

Claims (1)

[Claims] 1. A storage device in which a dictionary is stored, and a translation engine that translates input text information written in a source natural language into output text information written in a target natural language using the dictionary. The dictionary includes a headword written in the source natural language, a mark associated with the headword, and a probability value assigned to the marked headword, wherein the probability value is the mark Indicates the probability that a headword with a mark is translated into a translation written in the target natural language as a specific part of speech, wherein the headword is composed of a plurality of phrases, and the translation engine comprises: An automatic natural language translation system, comprising: analysis means for analyzing the marked headword as the specific part of speech according to the probability value assigned to a word.