Method and apparatus for normalizing and converting structured content6986104Abstract A method and apparatus are disclosed for transforming information from one semantic environment to another. In one implementation, a SOLx system includes a Normalization/Translation NorTran Workbench and a SOLx server. The NorTran Workbench is used to develop a knowledge base based on information from a source system, to normalize legacy content according to various rules, and to develop a database of translated content. During run time, the SOLx server receives transmissions from the source system, normalizes the transmitted content, accesses the database of translated content and otherwise translates the normalized content, and reconstructs the transmission to provide substantially real-time transformation of electronic messages. Claims What is claimed: Description FIELD OF THE INVENTION
The actual boundaries between phrases in data are known as phrase boundaries. Phrase boundaries are often covert—that is, not visibly marked. The SOLx tool of the present invention, as described in detail below, prepares data for translation by finding and marking phrase boundaries in the data. For example, it marks phrase boundaries in the string Acetic Acid Glass Bottle as follows:
This simple processing step—simple for a human, difficult for a language processing system—helps the machine translation system deduce the correct subgroupings of words within the input data, and allows it to produce the proper translation. The present invention is based, in part, on the recognition that some content, including business content, often is not easily searchable or analyzable unless a schema is constructed to represent the content. There are a number of issues that a computational system must address to do this correctly. These include: deducing the "core" item; finding the attributes of the item; and finding the values of those attributes. As noted above, conventional language processing systems have trouble telling which words in a string of words are more closely connected than other sets of words. They also have difficulty determining which word or words in the string represent the "core," or most central, concept in the string. For example, humans reading a string of words like Acetic Acid Glass Bottle in a catalogue of laboratory supplies may have no trouble telling that the item that is being sold is acetic acid, and that Glass Bottle just describes the container in which it is packaged. For conventional language processing systems, this is not a simple task. As noted above, a conventional language processing system may identify a number of possible word groupings, some of which are incorrect. Such a language processing system may deduce, for example, that the item that is being sold is a bottle, and that the bottle is made of "acetic acid glass." Obviously, this analysis leads to a faulty representation of bottles (and of acetic acid) in a schema and, therefore, is of little assistance in building an electronic catalogue system. In addition to finding the "core" of an item description, it is also useful to find the groups of words that describe that item. In the following description, the terms by which an item can be described are termed its attributes, and the contents or quantity of an attribute is termed its value. Finding attributes and their values is as difficult for a language processing system as is finding the "core" of an item description. For instance, in the string Acetic Acid Glass Bottle, one attribute of the core item is the package in which it is distributed. The value of this attribute is Glass Bottle. It may also be deemed that one attribute of the core item is the kind of container in which it is distributed. The value of this attribute would be Bottle. One can readily imagine other container types, such as Drum, Bucket, etc., in which acetic acid could be distributed. It happens that the kind of container attribute itself has an attribute that describes the material that the container is made of. The value of this attribute is Glass. Conventional natural language processing systems have trouble determining these sorts of relationships. Continuing with the example above, a conventional language processing system may analyze the string Acetic Acid Glass Bottle as follows:
This language processing system correctly deduced that Acetic and Acid go together. It incorrectly concluded that Acetic Acid go together to form the value of some attribute that describes a kind of Glass, and also incorrectly concluded that Acetic Acid Glass go together to give the value of some attribute that describes the bottle in question. The SOLx system of the present invention, as described in detail below, allows a user to provide guidance to its own natural language processing system in deducing which sets of words go together to describe values. It also adds one very important functionality that conventional natural language processing systems cannot perform without human guidance. The SOLx system allows you to guide it to match values with specific attribute types. The combination of (1) finding core items, and (2) finding attributes and their values, allows the SOLx system to build useful schemas. As discussed above, covert phrase boundaries interfere with good translation. Schema deduction contributes to preparation of data for machine translation in a very straightforward way: the labels that are inserted at the boundaries between attributes correspond directly to phrase boundaries. The next section describes a number of objectives of the SOLx system configuration process. All of these objectives relate to manipulating data from its native from to a form more amenable for translation or other localization, i.e., performing an initial transformation to an intermediate form. 2. Configuration Objectives Based on the foregoing, it will be appreciated that the SOLx configuration process has a number of objectives, including solving OOVs and solving covert phrase boundaries based on identification of core items and attribute/value pairs. Additional objectives, as discussed below, relate to taking advantage of reusable content chunks and resolving ambiguities. Many of these objectives are addressed automatically, or are partially automated, by the various SOLx tools described below. The following discussion will facilitate a more complete understanding of the internal functionality of these tools as described below. False OOV words and true OOV words can be discovered at two stages in the translation process: before translation, and after translation. Potential OOV words can be found before translation through use of a Candidate Search Engine as described in detail below. OOV words can be identified after translation through analysis of the translated output. If a word appears in data under analysis in more than one form, the Candidate Search Engine considers the possibility that only one of those forms exists in the machine translation system's dictionary. Specifically, the Candidate Search Engine offers two ways to find words that appear in more than one form prior to submitting data for translation: the full/abbreviated search option; and the case variant search option. Once words have been identified that appear in more than one form, a SOLx operator can force them to appear in just one form through the use of vocabulary adjustment rules. In this regard, the full/abbreviated search may output pairs of abbreviations and words. Each pair represents a potential false OOV term where it is likely that the unabbreviated form is in-vocabulary. Alternatively, the full/abbreviated search may output both pairs of words and unpaired abbreviations. In this case, abbreviations that are output paired with an unabbreviated word are potentially false OOV words, where the full form is likely in-vocabulary. Abbreviations that are output without a corresponding full form may be true OOV words. The machine translation dictionary may therefore be consulted to see if it includes such abbreviations. Similarly, some entries in a machine translation dictionary may be case sensitive. To address this issue, the SOLx system may implement a case variant search that outputs pairs, triplets, etc. of forms that are composed of the same letters, but appear with different variations of case. The documentation for a given machine translation system can then be consulted to learn which case variant is most likely to be in-vocabulary. To determine if a word is falsely OOV, words that are suspected to be OOV can be compared with the set of words in the machine translation dictionary. There are three steps to this procedure: 1) for each word that you suspect is falsely OOV, prepare a list of other forms that that word could take; 2) check the dictionary to see if it contains the suspected false OOV form; 3) check the dictionary to see if it contains one of the other forms of the word that you have identified. If the dictionary does not contain the suspected false OOV word and does contain one of the other forms of the word, then that word is falsely OOV and the SOLx operator can force it to appear in the "in-vocabulary" form in the input data as discussed below. Generally, this is accomplished through the use of a vocabulary adjustment rule. The vocabulary adjustment rule converts the false OOV form to the in-vocabulary form. The process for writing such rules is discussed in detail below. Problems related to covert phrase boundaries appear as problems of translation. Thus, a problem related to covert phrase boundaries may initially be recognized when a translator/ translation evaluator finds related errors in the translated text. A useful objective, then, is to identify these problems as problems related to covert phrase boundaries, rather than as problems with other sources. For example, a translation evaluator may describe problems related to covert phrase boundaries as problems related to some word or words modifying the wrong word or words. Problems related to potential covert phrase boundaries can also be identified via statistical analysis. As discussed below, the SOLx system includes a statistical tool called the N-gram analyzer (NGA) that analyzes databases to determine, among other things, what terms appear most commonly and which terms appear in proximity to one another. A mistranslated phrase, identified in the quality control analysis (described below in relation to the TQE module), which has a low NGA probability for the transition between two or more pairs of words suggests a covert phrase boundary. Problems related to covert phrase boundaries can also be addressed through modifying a schematic representation of the data under analysis. In this regard, if a covert phrase boundary problem is identified, it is often a result of attribute rules that failed to identify an attribute. This can be resolved by modifying the schema to include an appropriate attribute rule. If a schema has no yet been produced for the data, a schema can be constructed at this time. Once a categorization or attribute rule has been constructed for a phrase that the translator/translation evaluator has identified as poorly translated, then the original text can be re-translated. If the result is a well-translated phrase, the problem has been identified as one of a covert phrase boundary and the operator may consider constructing more labeling rules for the data under analysis. Covert phrase boundary problems can be addressed by building a schema, and then running the schematized data through a SOLx process that inserts a phrase boundary at the location of every labeling/tagging rule. The core item of a typical business content description is the item that is being sold/described. An item description often consists of its core item and some terms that describe its various attributes. For example, in the item description Black and Decker ⅜" drill with accessories, the item that is being described is a drill. The words or phrases Black and Decker, ⅜", and with accessories all give us additional information about the core item, but do not represent the core item itself. The core item in an item description can generally be found by answering the question, what is the item that is being sold or described here? For example, in the item description Black and Decker ⅜" drill with accessories, the item that is being described is a drill. The words or phrases Black and Decker, ⅜", and with accessories all indicate something about the core item, but do not represent the core item itself. A subject matter expert (SME) configuring SOLx for a particular application can leverage his domain-specific knowledge before beginning your work with SOLx by listing the attributes of core items before beginning work with SOLx, and by listing the values of attributes before beginning work with SOLx. Both classification rules and attribute rules can then be prepared before manipulating data with the SOLx system. Domain-specific knowledge can also be leveraged by recognizing core items and attributes and their values during configuration of the SOLx system and writing rules for them as they appear. As the SME works with the data within the SOLx system, he can write rules for the data as the need appears. The Candidate Search Engine can also be used to perform a collocation search that outputs pairs of words that form collocations. If one of those words represents a core item, then the other word may represent an attribute, a value, or (in some sense) both. Attribute-value pairs can also be identified based on a semantic category search implemented by the SOLx system. The semantic category search outputs groups of item descriptions that share words belonging to a specific semantic category. Words from a specific semantic category that appear in similar item descriptions may represent a value, an attribute, or (in some sense) both. Business content is generally characterized by a high degree of structure that facilitates writing phrasing rules and allows for efficient reuse of content "chunks." As discussed above, much content relating to product descriptions and other structured content is not free-flowing sentences, but is an abbreviated structure called a 'noun phrase'. Noun phrases are typically composed of mixtures of nouns (N), adjectives (A), and occasionally prepositions (P). The mixtures of nouns and adjectives may be nested. The following are some simple examples:
Adjective phrases also exist mixed with adverbs (Av). Table 2 lists some examples.
The noun phrase four-strand color-coded twisted-pair telephone wire has the pattern NNNAANNN. It is grouped as (fourN strandN)N (colorN codedA)A (twistedA pairN)N telephoneN wireN. Another way to look at this item is an object-attribute list. The primary word or object is wire; of use type telephone; strand type twisted-pair, color property color-coded, and strand number type is four-stranded. The structure is N1AN2N3N4. With this type of compound grouping, each group is essentially independent of any other group. Hence, the translation within each group is performed as an independent phrase and then linked by relatively simple linguistic rules. For example, regroup N1AN2N3N4 as NN3N4 where N=N1AN2. In Spanish this can be translated as NN3N4→N4'de' N3 'de' {N} where {N} means the translated version of N, and → means translated as. In Spanish, it would be N1AN2→N2A 'de' N1. The phrase then translates as N1AN2N3N4→N4 'de' N3 'de' N2A 'de' N1. In addition to defining simple rule sets for associating translated components of noun phrases, there is another factor that leads to the feasibility of automatically translating large component databases. This additional observation is that very few terms are used in creating these databases. For example, databases have been analyzed that have 70,000 part descriptions, yet are made up of only 4,000 words or tokens. Further, individual phrases are used hundreds of times. In other words, if the individual component pieces or "chunks" are translated, and there are simple rules for relating theses chunks, then the translation of large parts of the content, in principle, is straightforward. The SOLx system includes tools as discussed in more detail below for identifying reusable chunks, developing rules for translation and storing translated terms/chunks for facilitating substantially real-time transformation of electronic content. Another objective of the configuration process is enabling SOLx to resolve certain ambiguities. Ambiguity exists when a language processing system does not know which of two or more possible analyses of a text string is the correct one. There are two kinds of ambiguity in item descriptions: lexical ambiguity and structural ambiguity. When properly configured, the SOLx system can often resolve both kinds of ambiguity. Lexical ambiguity occurs when a language processing system does not know which of two or more meanings to assign to a word. For example, the abbreviation mil can have many meanings, including million, millimeter, military, and Milwaukee. In a million-item database of tools and construction materials, it may occur with all four meanings. In translation, lexical ambiguity leads to the problem of the wrong word being used to translate a word in your input. To translate your material, it is useful to expand the abbreviation to each of its different full forms in the appropriate contexts. The user can enable the SOLx system to do this by writing labeling rules that distinguish the different contexts from each other. For example, mil might appear with the meaning million in the context of a weight, with the meaning millimeter in the context of a length, with the meaning military in the context of a specification type (as in the phrase MIL SPEC), and with the meaning Milwaukee in the context of brand of a tool. You then write vocabulary adjustment rules to convert the string mil into the appropriate full form in each individual context. In schematization, resolving lexical ambiguity involves a number of issues, including identification of the core item in an item description; identification of values for attributes; and assignment of values to proper attributes. Structural ambiguity occurs when a language processing system does not know which of two or more labeling rules to use to group together sets of words within an item description. This most commonly affects attribute rules and may require further nesting of parent/child tag relationships for proper resolution. 3. Configuration Processes a. Normalization As the foregoing discussion suggests, the various configuration objectives (e.g., resolving false OOVs, identifying covert phrase boundaries, taking advantage of reusable chunks and resolving ambiguities) can be addressed in accordance with the present invention by transforming input data from its native form into an intermediate form that is more amenable to translation or other localization/transformation. The corresponding process, which is a primary purpose of SOLx system configuration, is termed "normalization." Once normalized, the data will include standardized terminology in place of idiosyncratic terms, will reflect various grammar and other rules that assist in further processing, and will include tags that provide context for resolving ambiguities and otherwise promoting proper transformation. The associated processes are executed using the Normalization Workbench of the SOLx system, as will be described below. There are two kinds of rules developed using the Normalization Workbench: grammatical rules, and normalization rules. The purpose of a grammatical rule is to group together and label a section of text. The purpose of a normalization rule is to cause a labeled section of text to undergo some change. Although these rules are discussed in detail below in order to provide a more complete understanding of the present invention, it will be appreciated that these rules are, to a large extent, developed and implemented internally by the various SOLx tools. Accordingly, SOLx operators need not have linguistics expertise to realize the associated advantages. i) Normalization Rules The Normalization Workbench offers a number different kinds of normalization rules relating to terminology including: replacement rules, joining rules, and ordering rules. Replacement rules allow the replacement of one kind of text with another kind of text. Different kinds of replacement rules allow the user to control the level of specificity of these replacements. Joining rules allow the user to specify how separated elements should be joined together in the final output. Ordering rules allow the user to specify how different parts of a description should be ordered relative to each other. With regard to replacement rules, data might contain instances of the word centimeter written four different ways—as cm, as cm., as c.m., and as centimeter—and the user might want to ensure that it always appears as centimeter. The Normalization Workbench implements two different kinds of replacement rules: unguided replacement, and guided replacement. The rule type that is most easily applicable to a particular environment can be selected. Unguided replacement rules allow the user to name a tag/category type, and specify a text string to be used to replace any text that is under that tag. Guided replacement rules allow the user to name a tag/category type, and specify specific text strings to be used to replace specific text strings that are under that tag. Within the Normalization Workbench logic, the format of unguided replacement rules may be, for example: For instance, the following rule says to find any [foot] category label, and replace the text that it tags with the word feet: If that rule was run against the following input, it would produce the following output: The second line is unchanged; in the first line, foot has been changed to feet. Guided replacement rules allow the user to name a tag/category type, and specify specific text strings to be used to replace specific text strings that are under that tag. This is done by listing a set of possible content strings in which the normalization engine should "look up" the appropriate replacement. The format of these rules is:
For instance, the following rule says to find any [length—metric] label. If you see mm, mm., m.m., or m. m. beneath it, then replace it with millimeter. If you see cm, cm., c.m., or c. m. beneath it, then replace it with centimeter:
it would produce the following output: From the user's perspective, such replacement rules may be implemented via a simple user interface such as shown in FIG. 1. FIG. 1 shows a user interface screen 100 including a left pane 102 and a right pane 104. The left pane 102 displays the grammar rules that are currently in use. The rules are shown graphically, including alternative expressions (in this case) as well as rule relationships and categories. Many alternative expressions or candidates therefor are automatically recognized by the workbench and presented to the user. The right pane 104 reflects the process to update or add a text replacement rule. In operation, a grammar rule is selected in the left pane 102. All text that can recognized by the rule appears in the left column of the table 106 in the right pane 104. The SME then has the option to unconditionally replace all text with the string from the right column of the table 106 or may conditionally enter a replacement string. Although not shown in each case below, similar interfaces allow for easy development and implementation of the various rules discussed herein. Joining rules allow the user to specify how separated elements should be joined together in the final output. Joining rules can be used to re-join elements that were separated during the process of assigning category labels. The user can also use joining rules to combine separate elements to form single delimited fields. Some elements that were originally adjacent in the input may have become separated in the process of assigning them category labels, and it may be desired to re-join them in the output. For example, the catheter tip configuration JL4 will appear as [catheter—tip—configuration] (JL4) after its category label is assigned. However, the customary way to write this configuration is with all three of its elements adjacent to each other. Joining rules allow the user to join them together again. The user may wish the members of a particular category to form a single, delimited field. For instance, you might want the contents of the category label [litter—box] (plastic hi-impact scratch-resistant) to appear as plastic,hi-impact, scratch-resistant in order to conserve space in your data description field. Joining rules allow the user to join these elements together and to specify that a comma be used as the delimiting symbol. The format of these rules is: The delimiter can be absent, in which case the elements are joined immediately adjacent to each other. For example, numbers emerge from the category labeler with spaces between them, so that the number twelve looks like this: A standard normalization rule file supplied with the Normalization Workbench contains the following joining rule: This rule causes the numbers to be joined to each other without an intervening space, producing the following output: The following rule states that any content that appears with the category label [litter—box] should be joined together with commas: If that rule was run against the following input, it would produce the following output: Ordering rules allow the user to specify how different parts of a description should be ordered relative to each other. For instance, input data might contain catheter descriptions that always contain a catheter size and a catheter type, but in varying orders—sometimes with the catheter size before the catheter type, and sometimes with the catheter type before the catheter size: The user might prefer that these always occur in a consistent order, with the catheter size coming first and the catheter type coming second. Ordering rules allow you to enforce this ordering consistently. The internal format of ordering rules is generally somewhat more complicated than that of the other types of rules. Ordering rules generally have three parts. Beginning with a simple example: The first part of the rule, shown in bold below, specifies that this rule should only be applied to the contents of a [catheter] category label: The second part of the rule, shown in bold below, specifies which labeled elements are to have their orders changed: Each of those elements is assigned a number, which is written in the format $number in the third part of the rule. The third part of the rule, shown in bold below, specifies the order in which those elements should appear in the output: The order $2 $1 indicates that the element which was originally second (i.e., $2) should be first (since it appears in the leftmost position in the third part of the rule), while the element which was originally first (i.e., $1) should be second (since it appears in the second position from the left in the third part of the rule). Ordering rules can appear with any number of elements. For example, this rule refers to a category label that contains four elements. The rule switches the position of the first and third elements of its input, while keeping its second and fourth elements in their original positions: FIG. 2 shows an example of a user interface screen 200 that may be used to develop and implement an ordering rule. The screen 200 includes a left pane 202 and a right pane 204. The left pane 202 displays the grammar rules that are currently in use—in this case, ordering rules for container size—as well as various structural productions under each rule. The right pane 204 reflects the process to update or add structural reorganization to the rule. In operation, a structural rule is selected using the left pane 202. The right pane 204 can then be used to develop or modify the rule. In this case, the elements or "nodes" can be reordered by simple drag-and-drop process. Nodes may also be added or deleted using simple mouse or keypad commands. Ordering rules are very powerful, and have other uses besides order-changing per se. Other uses for ordering rules include the deletion of unwanted material, and the addition of desired material. To use an ordering rule to delete material, the undesired material can be omitted from the third part of the rule. For example, the following rule causes the deletion of the second element from the product description: If that rule was run against the following input, it would produce the following output: To use an ordering rule to add desired material, the desired material can be added to the third part of the rule in the desired position relative to the other elements. For example, the following rule causes the string [real—cnx]'-' to be added to the product description: If that rule was run against the following input, it would produce the following output: After final processing, this converts the confusing 11/2 and 15/8 to 1½ ("one and a half") and 1⅝ ("one and five eighths"). In addition to the foregoing normalization rules relating to terminology, the SOLx system also involves normalization rules relating to context cues and phrasing. The rules that the SOLx system uses to identify contexts and determine the location and boundaries of attribute/value pairs fall into three categories: categorization rules, attribute rules, and analysis rules. Categorization rules and attribute rules together form a class of rules known as labeling/tagging rules. labeling/tagging rules cause the insertion of labels/tags in the output text when the user requests parsed or labeled/tagged texts. They form the structure of the schema in a schematization task, and they become phrase boundaries in a machine translation task. Analysis rules do not cause the insertion of labels/tags in the output. They are inserted temporarily by the SOLx system during the processing of input, and are deleted from the output before it is displayed. Although analysis tags are not displayed in the output (SOLx can allow the user to view them if the data is processed in a defined interactive mode), they are very important to the process of determining contexts for vocabulary adjustment rules and for determining where labels/tags should be inserted. The analysis process is discussed in more detail below. ii. Grammar Rules The various rules described above for establishing normalized content are based on grammar rules developed for a particular application. The process for developing grammar rules is set forth in the following discussion. Again, it will be appreciated that the SOLx tools guide an SME through the development of these rules and the SME need not have any expertise in this regard. There are generally two approaches to writing grammar rules, known as "bottom up" and "top down." Bottom-up approaches to writing grammar rules begin by looking for the smallest identifiable units in the text and proceed by building up to larger units made up of cohesive sets of the smaller units. Top-down approaches to writing grammar rules begin by identifying the largest units in the text, and proceed by identifying the smaller cohesive units of which they are made. Consider the following data for an example of building grammar rules from the bottom up. It consists of typical descriptions of various catheters used in invasive cardiology: Each of these descriptions includes some indication of the (diametric) size of the catheter, shown in bold text below: One can make two very broad generalizations about these indications of catheter size: all of them include a digit, and the digits all seem to be integers. One can further make two weaker generalizations about these indications of catheter size: all of them include either the letters Fr, or the word French; and if they include the letters Fr, those two letters may or may not be followed by a period. A subject matter expert (SME) operating the SOLx system will know that Fr, Fr., and French are all tokens of the same thing: some indicator of the unit of catheter size. Having noted these various forms in the data, a first rule can be written. It will take the form x can appear as w, y, or z, and this rule will describe the different ways that x can appear in the data under analysis. The basic fact that the rule is intended to capture is French can appear as Fr, as Fr, or as French. In the grammar rules formalism, that fact may be indicated like this: [French] is the name assigned to the category of "things that can be forms of the word that expresses the unit of size of catheters" and could just as well have been called [catheter—size—unit], or [Fr], or [french]. The important thing is to give the category a label that is meaningful to the user. (Fr), (Fr.), and (French) are the forms that a thing that belongs to the category [French] can take. Although the exact name for the category [French] is not important, it matters much more how these "rule contents" are written. For example, the forms may be case sensitive. That is, (Fr) and (fr) are different forms. If your rule contains the form (Fr), but not the form (fr), then if there is a description like this: The fr in the description will not be recognized as expressing a unit of catheter size. Similarly, if your rule contained the form (fr), but not the form (Fr), then Fr would not be recognized. "Upper-case" and "lower-case" distinctions may also matter in this part of a rule. Returning to the list of descriptions above, a third generalization can be made: all of the indications of catheter size include an integer followed by the unit of catheter size. This suggests another rule, of the form all x consist of the sequence a followed by b. The basic fact that the rule is intended to capture is: all indications of catheter size consist of a number followed by some form of the category [French]. In the grammar rules formalism, that fact may be indicated like this:
So far this example has involved a set of rules that allows description of the size of every catheter in a list of descriptions. The SME working with this data might then want to write a set of rules for describing the various catheter types in the list. Up to this point, this example has started with the smallest units of text that could be identified (the different forms of [French]) and worked up from there (to the [catheter—size] category). Now, the SME may have an idea of a higher-level description (i.e., catheter type), but no lower-level descriptions to build it up out of; in this case, the SME may start at the top, and think his way down through a set of rules. The SME can see that each of these descriptions includes some indication of the type of the catheter, shown in bold text below: He is aware that a catheter type can be described in one of two ways: by the tip configuration of the catheter, and by the purpose of the catheter. So, the SME may write a rule that captures the fact that catheter types can be identified by tip configuration or by catheter purpose. In the grammar rules formalism, that fact may be indicated like this:
Starting with tip configuration, the SME knows that catheter tip configurations can be described in two ways: 1) by a combination of the inventor's name, an indication of which blood vessel the catheter is meant to engage, and by an indication of the length of the curve at the catheter tip; or 2) by the inventor's name alone. The SME can write a rule that indicates these two possibilities in this way:
With these rules, the SME has a complete description of the [catheter—tip—configuration] category. Recall that the SME is writing a [catheter—tip—configuration] rule because there are two ways that a catheter type can be identified: by the configuration of the catheter's tip, and by the catheter's purpose. The SME has the [catheter—tip—configuration] rule written now and just needs a rule that captures descriptions of a catheter's purpose. The SME is aware that (at least in this limited data set) a catheter's purpose can be directly indicated, e.g. by the word angioplasty, or can be inferred from something else—in this case, the catheter's shape, as in pigtail. So, the SME writes a rule that captures the fact that catheter purpose can be identified by purpose indicators or by catheter shape. In the grammar rules formalism, that fact can be indicated like this:
"Top-down" and "bottom-up" approaches to writing grammar rules are both effective, and an SME should use whichever is most comfortable or efficient for a particular data set. The bottom-up approach is generally easier to troubleshoot; the top-down approach is more intuitive for some people. A grammar writer can use some combination of both approaches simultaneously. Grammar rules include a special type of rule called a wanker. Wankers are rules for category labels that should appear in the output of the token normalization process. In one implementation, wankers are written similarly to other rules, except that their category label starts with the symbol>. For example, in the preceding discussion, we wrote the following wanker rules:
Chunks of text that have been described by a wanker rule will be tagged in the output of the token normalization process. For example, with the rule set that we have defined so far, including the two wankers, we would see output like the following: Although the other rules are used in this example to define the wanker rules, and to recognize their various forms in the input text, since the other rules are not wankers, their category labels do not appear in the output. If at some point it is desired to make one or more of those other rules' category labels to appear in the output, the SME or other operator can cause them to do so by converting those rules to wankers. Besides category labels, the foregoing example included two kinds of things in rules. First, the example included rules that contained other category labels. These "other" category labels are identifiable in the example by the fact that they are always enclosed in square brackets, e.g.
The example also included rules that contained strings of text that had to be written exactly the way that they would appear in the input. These strings are identifiable by the fact that they are directly enclosed by parentheses, e.g.
There is a third kind of thing that can be used in a rule. These things, called regular expressions, allow the user to specify approximately what a description will look like. Regular expressions can be recognized by the facts that: unlike the other kinds of rule contents, they are not enclosed by parentheses, and they are immediately enclosed by "forward slashes." Regular expressions in rules look like this:
Although the foregoing example illustrated specific implementations of specific rules, it will be appreciated that a virtually endless variety of specialized rules may be provided in accordance with the present invention. The SOLx system of the present invention consists of many components, as will be described below. One of these components is the Natural Language Engine module, or NLE. The NLE module evaluates each item description in data under analysis by means of rules that describe the ways in which core items and their attributes can appear in the data. The exact (machine-readable) format that these rules take can vary depending upon the application involved and computing environment. For present purposes, it is sufficient to realize that these rules express relationships like the following (stated in relation to the drill example discussed above):
The NLE checks each line of the data individually to see if any of the rules seem to apply to that line. If a rule seems to apply, then the NLE inserts a label/tag and marks which string of words that rule seemed to apply to. For example, for the set of rules listed above, then in the item description Black and Decker ⅜" drill with accessories, the NLE module would notice that ⅜" might be a drill size, and would mark it as such. If the user is running the NLE in interactive mode, he may observe something like this in the output: In addition to the rules listed above, a complete set of rules for describing the ways that item descriptions for/of drills and their attributes would also include rules for manufacturers names, accessory lists, and whether or not the drill is battery powered. If the user writes such a set of rules, then in the item description Black and Decker ⅜" drill with accessories, the NLE module will notice and label/tag the following attributes of the description: The performance of the rules can be analyzed in two stages. First, determine whether or not the rules operate adequately. Second, if it is identified that rules that do not operate adequately, determine why they do not operate adequately. For translations, the performance of the rules can be determined by evaluating the adequacy of the translations in the output text. For schematization, the performance of the rules can be determined by evaluating the adequacy of the schema that is suggested by running the rule set. For any rule type, if a rule has been identified that does not perform adequately, it can be determined why it does not operate adequately by operating the NLE component in interactive mode with output to the screen. For tagging rules, test data set can be analyzed to determine if: every item that should be labeled/tagged has been labeled/tagged and any item that should not have been labeled/tagged has been labeled/tagged in error. In order to evaluate the rules in this way, the test data set must include both items that should be labeled/tagged, and items that should not be tagged. Vocabulary adjustment rules operate on data that has been processed by tagging/tagging rules, so troubleshooting the performance of vocabulary adjustment rules requires attention to the operation of tagging/tagging rules, as well as to the operation of the vocabulary adjustment rules themselves. In general, the data set selected to evaluate the performance of the rules should include: examples of different types of core items, and for each type of core item, examples with different sets of attributes and/or attribute values. b. Translation The SOLx paradigm is to use translators to translate repeatable complex terms and phrases, and translation rules to link these phrases together. It uses the best of both manual and machine translation. The SOLx system uses computer technology for repetitive or straightforward applications, and uses people for the complex or special-case situations. The NorTran (Normalization/Translation) server is designed to support this paradigm. FIG. 3 represents a high-level architecture of the NorTran platform 300. Each module is discussed below as it relates to the normalization/translation process. A more detailed description is provided below in connection with the overall SOLx schematic diagram description for configuration and run-time operation. The GUI 302 is the interface between the subject matter expert (SME) or human translator (HT) and the core modules of the NorTran server. Through this interface, SMEs and HTs define the filters for content chunking, access dictionaries, create the terms and phrases dictionaries, and monitor and edit the translated content. This N-Gram 304 filter for the N-gram analysis defines the parameters used in the N-gram program. The N-gram program is the key statistical tool for identifying the key reoccurring terms and phrases of the original content. The N-Gram and other statistical tools module 306 is a set of parsing and statistical tools that analyze the original content for significant terms and phrases. The tools parse for the importance of two or more words or tokens as defined by the filter settings. The output is a sorted list of terms with the estimated probabilities of the importance of the term in the totality of the content. The goal is to aggregate the largest re-usable chunks and have them directly translated. The chunking assembly and grammar rules set 308 relates the pieces from one language to another. For example, as discussed earlier, two noun phrases N1N2 are mapped in Spanish as N2 'de' N1. Rules may need to be added or existing ones modified by the translator. The rules are used by the translation engine with the dictionaries and the original content (or the normalized content) to reassemble the content in its translated form. The rules/grammar base language pairs and translation engine 310 constitute a somewhat specialized machine translation (MT) system. The translation engine portion of this system may utilize any of various commercially available translation tools with appropriate configuration of its dictionaries. Given that the translation process is not an exact science and that round trip processes (translations from A to B to A) rarely work, a statistical evaluation is likely the best automatic tool to assess the acceptability of the translations. The Translation Accuracy Analyzer 312 assesses words not translated, heuristics for similar content, baseline analysis from human translation and other criteria. The chunking and translation editor 314 functions much like a translator's workbench. This tool has access to the original content; helps the SME create normalized content if required; the normalized content and dictionaries help the translator create the translated terms and phase dictionary, and when that repository is created, helps the translator fill in any missing terms in the translation of the original content. A representation of the chunking functionality of this editor is shown in the example in Table 3.
The first column lists the original content from a parts list of cooking dishes. The term (A) etc. are dimensional measurements that are not relevant to the discussion. The second column lists the chunked terms from an N-gram
analysis; the third column lists the frequency of each term in the original content set. The fourth column is the number associated with the chunk terms in column 2. The fifth column is the representation of the first column in terms of the sequence of chunked content. If the translation of each chunk is stored in another column, and translation rules exist for reassembling the chunks, then the content is translated. It could be listed in another column that would have a direct match or link to the original content. Table 4 lists the normalized and translated normalized content. Finally, Table 5 shows the Original Content and the Translated Content that is created by assembling the Translated Normalized Terms in Table 4 according to the Chunked Original Content sequence in Table 3.
This example shows that when appropriately "chunked," machine translation grammar knowledge for noun phrases can be minimized. However, it cannot be eliminated entirely. Referring to FIG. 3, the Normalized Special Terms and Phrases repository 316 contains chunked content that is in a form that supports manual translation. It is free of unusual acronyms, misspellings, and strived for consistency. In Table 3 for example, Emile Henry was also listed as E. Henry. Terms usage is maximized. The Special Terms and Phrases Translation Dictionary repository 318 is the translated normalized terms and phrases content. It is the specialty dictionary for the client content. Other translation dictionaries 320 may be any of various commercially available dictionary tools and/or SOLx developed databases. They may be general terms dictionaries, industry specific, SOLx acquired content, or any other knowledge that helps automate the process. One of the tenets of the SOLx process is that the original content need not be altered. Certainly, there are advantages to make the content as internally consistent as possible, and to define some form of structure or syntax to make translations easier and more accurate. However, there are situations where a firm's IT department does not want the original content modified in any way. Taking advantage of the benefits of normalized content, but without actually modifying the original, SOLx uses a set of meta or non-persistent stores so that the translations are based on the normalized meta content 322. The above discussion suggests a number of processes that may be implemented for the automatic translation of large databases of structured content. One implementation of these processes is illustrated in the flow chart of FIG. 4 and is summarized below. It will be appreciated that these processes and the ordering thereof can be modified. First, the firm's IT organization extracts 400 the content from their IT systems—ideally with a part number or other unique key. As discussed above, one of the key SOLx features is that the client need not restructure or alter the original content in their IT databases. However, there are reasons to do so. In particular, restructuring benefits localization efforts by reducing the translation set up time and improving the translation accuracy. One of these modifications is to adopt a 'normalized' or fixed syntactic, semantic, and grammatical description of each content entry. Next, software tools identify (402) the most important terms and phrases. Nearest neighbor, filtered N-gram, and other analysis tools identify the most used and important phrases and terms in the content. The content is analyzed one description or item at a time and re-usable chunks are extracted. Subject matter experts then "internationalize" (404) the important terms and phrases. These experts "translate" the abbreviations and acronyms, correct misspellings and in general redefine and terms that would be ambiguous for translation. This is a list of normalized terms and phrases. It references the original list of important terms and phrases. Translators can then translate (406) the internationalized important terms and phrases. This translated content forms a dictionary of specialty terms and phrases. In essence, this translated content corresponds to the important and re-usable chunks. Depending on the translation engine used, the translator may need to specify the gender alternatives, plural forms, and other language specific information for the special terms and phrases dictionary. Referring again to an example discussed above, translators would probably supply the translation for (four-strand), (color-coded), (twisted-pair), telephone, and wire. This assumes that each term was used repeatedly. Any other entry that uses (color-coded) or wire would use the pre-translated term. Other dictionaries for general words and even industry specific nomenclature can then be consulted (408) available. This same approach could be used for the creation of general dictionaries. However, for purposes of this discussion it is assumed that they already exist. Next, language specific rules are used to define (410) the assembly of translated content pieces. The types of rules described above define the way the pre-translated chunks are reassembled. If, in any one description, the grammatical structure is believed to be more complicated than the pre-defined rule set, then the phrase is translated in its entirety. The original content (on a per item basis) is then mapped (412) against the dictionaries. Here, the line item content is parsed and the dictionaries are searched for the appropriate chunked and more general terms (content chunks to translated chunks). Ideally, all terms in the dictionaries map to a single-line item in the content database, i.e. a single product description. This is the first function of the translation engine. A software translation engine then assembles (414) the translated pieces against the language rules. Input into the translation engine includes the original content, the translation or assembly rules, and the translated pieces. A translation tool will enable a translator to monitor the process and directly intercede if required. This could include adding a new chunk to the specialty terms database, or overriding the standard terms dictionaries. A statistically based software tool assesses (416) the potential accuracy of the translated item. One of the difficulties of translation is that when something is translated from one language to another and then retranslated back to the first, the original content is rarely reproduced. Ideally, one hopes it is close, but rarely will it be exact. The reason for this is there is not a direct inverse in language translation. Each language pair has a circle of 'confusion' or acceptability. In other words, there is a propagation of error in the translation process. Short of looking at every translated phrase, the best than can be hoped for in an overall sense is a statistical evaluation. Translators may re-edit (418) the translated content as required. Since the content is stored in a database that is indexed to the original content on an entry-by-entry basis, any entry may be edited and restored if this process leads to an unsatisfactory translation. Although not explicitly described, there are terms such as proper nouns, trade names, special terms, etc., that are never translated. The identification of these invariant terms would be identified in the above process. Similarly, converted entries such as metrics would be handled through a metrics conversion process. The process thus discussed uses both human and machine translation in a different way than traditionally employed. This process, with the correct software systems in place should generate much of the accuracy associated with manual translation. Further, this process should function without manual intervention once sufficient content has been pre-translated. The various configuration processes are further illustrated by the screenshots of FIGS. 5-10. Although these figures depict screenshots, it will be appreciated that these figures would not be part of the user interface as seen by an SME or other operator. Rather, these screenshots are presented here for purposes of illustration and the associated functionality would, to a significant extent, be implemented transparently. These screenshots show the general processing of source content. The steps are importing the data, normalizing the data based on a set of grammars and rules produced by the SME using the NTW user interface, then analysis of the content to find phrases that need to be translated, building a translation dictionary containing the discovered phrases, translation of the normalized content, and finally, estimation of the quality of the translated content. The first step, as illustrated in FIG. 5 is to import the source structured content file. This will be a flat set file with the proper character encoding, e.g., UTF-8. There will generally be one item description per line. Some basic formatting of the input may be done at this point. FIG. 6 shows normalized form of the content on the right and the original content (as imported above) on the left. What is not shown here are the grammars and rules used to perform the normalization. The form of the grammars and rules and how to created them are described above. In this example, various forms of the word resistor that appear on the original content, for example "RES" or RESS" have been normalized to the form "resistor". The same is true for "W" being transformed to "watt" and "MW" to "milliwatt". Separation was added between text items, for example, "¼W" is now "¼ watt" or "750 OHM" is now "75 ohm". Punctuation can also be added or removed, for example, "RES,35.7" is now "resistor 35.7". Not shown in the screenshot: the order of the text can also be standardized by the normalization rules. For example, if the user always want a resistor description to of the form:
Another very valuable result of the normalization step can be to create a schematic representation of the content. In the phrase analysis step, as illustrated, the user is looking for the phrases in the now normalized content that still need to be translated to the target language. The purpose of Phrase Analysis, and in fact, the next several steps, is to create a translation dictionary that will be used by machine translation. The value in creating the translation dictionary is that only the phrases need translation not the complete body of text, thus providing a huge savings in time and cost to translate. The Phrase Analyzer only shows us here the phrases that it does not already have a translation for. Some of these phrases we do not want to translate, which leads us to the next step. In the filter phrases step as shown in FIG. 7, an SME reviews this phrase data and determines which phrases should be translated. Once the SME has determined which phrases to translate, then a professional translator and/or machine tool translates the phrases (FIGS. 8-9) from the source language, here English, to the target language, here Spanish. A SOLx user interface could be used to translate the phrases, or the phrases are sent out to a professional translator as a text file for translation. The translated text is returned as a text file and loaded into SOLx. The translated phrases become the translation dictionary that is then used by the machine translation system. The machine translation system uses the translation dictionary created above as the source for domain specific vocabulary. By providing the domain specific vocabulary in the form of the translation dictionary, The SOLx system greatly increases the quality of the output from the machine translation system. The SOLx system can also then provide an estimation of the quality of the translation result (FIG. 10). Good translations would then be loaded into the run-time localization system for use in the source system architecture. Bad translations would be used to improve the normalization grammars and rules, or the translation dictionary. The grammars, rules, and translation dictionary form a model of the content. Once the model of the content is complete, a very high level of translations are of good quality. Particular implementations of the above described configuration processes can be summarized by reference to the flow charts of FIGS. 11-12. Specifically, FIG. 11 summarizes the steps of an exemplary normalization configuration process and FIG. 12 summarizes an exemplary translation configuration process. Referring first to FIG. 11 , a new SOLx normalization process (1000) is initiated by importing (1102) the content of a source database or portion thereof to be normalized and selecting a quantify of text from a source database. For example, a sample of 100 item descriptions may be selected from the source content "denoted content.txt file." A text editor may be used to select the 100 lines. These 100 lines are then saved to a file named samplecontent.txt for purposes of this discussion. The core items in the samplecontent.txt file are then found (1104) using the Candidate Search Engine, for example, by running a words-in-common search. Next, attribute/value information is found (1106) in the samplecontent.txt file using the Candidate Search Engine by running collocation and semantic category searches as described above. Once the attributes/values have been identified, the SOLx system can be used to write (1108) attribute rules. The formalism for writing such rules has been discussed above. It is noted that the SOLx system performs much of this work for the user and simple user interfaces can be provided to enable "writing" of these rules without specialized linguistic or detailed code-writing skills. The SOLx system can also be used at this point to write (1110) categorization rules. As noted above, such categorization rules are useful in defining a context for avoiding or resolving ambiguities in the transformation process. Finally, the coverage of the data set can be analyzed (1112) to ensure satisfactory run time performance. It will be appreciated that the configuration process yields a tool that can not only translate those "chunks" that were processed during configuration, but can also successfully translate new items based on the knowledge base acquired and developed during configuration. The translation process is summarized below. Referring to FIG. 12, the translation process 1200 is initiated by acquiring (1202) the total set of item descriptions that you want to translate as a flat file, with a single item description per line. For purposes of the present discussion, it is assumed that the item descriptions are in a file with the name of content.txt. A text editor may be used to setup an associated project configuration file. Next, a sample of 100 item descriptions is selected (1204) from the content.txt file. A text editor may be used to select the 100 lines. These 100 lines to a file named samplecontent.txt. The translation process continues with finding (1206) candidates for vocabulary adjustment rules in the samplecontent.txt file using the Candidate Search Engine. The Candidate Search Engine may implement a case variant search and full/abbreviated variant search at this point in the process. The resulting information can be used to write vocabulary adjustment rules. Vocabulary adjustment rules may be written to convert abbreviated forms to their full forms. Next, candidates for labeling/tagging rules are found (1208) in the sample/content.txt file using the Candidate Search Engine. Labeling/tagging rules may be written to convert semantic category and collocation forms. Attribute rules can then be written (1210) following the steps set forth in the previous flow chart. Vocabulary adjustment rules are then run (1212) using the Natural LanguageEngine against the original content. Finally, the coverage of the data set can be analyzed (1214) evaluating performance of your vocabulary adjustment rules and evaluating performance of your attribute rules. At this point, if the proper coverage is being achieved by the vocabulary adjustment rules, then the process proceeds to building (1216) a domain-specific dictionary. Otherwise, a new set of 100 item descriptions can be selected for analysis and the intervening steps are repeated. To build a domain specific dictionary, the SME can run a translation dictionary creation utility. This runs using the rule files created above as input, and produces the initial translation dictionary file. This translation dictionary file contains the words and phrases that were found in the rules. The words and phrases found in the translation dictionary file can then be manually and/or machine translated (1218). This involves extracting a list of all word types using a text editor and then translating the normalized forms manually or through a machine tool such as SYSTRAN. The translated forms can then be inserted into the dictionary file that was previously output. Next, the SME can run (1220) the machine translation module, run the repair module, and run the TQE module. The file outputs from TQE are reviewed (1222) to determine whether the translation results are acceptable. The acceptable translated content can be loaded (1224) into the Localized Content Server (LCS), if desired. The remainder of the translated content can be analyzed (1226) to determine what changes to make to the normalization and translation knowledge bases in order to improve the quality of the translation. Words and phrases that should be deleted during the translation process can be deleted (1228) and part-of-speech labels can be added, if needed. The SME can then create (1230) a file containing the translated words in the source and target languages. Once all of the content is found to be acceptable, the systems is fully trained. The good translated content is then loaded into the LCS. It has been found that it is useful to provide graphical feedback during normalization to assist the SME in monitoring progress. Any appropriate user interface may be provided in this regard. FIG. 13 shows an example of such an interface. As shown, the graphical desktop 1300 is divided into multiple work spaces, in this case, including workspaces 1302, 1304 and 1306. One workspace 1302 presents the source file content that is in process, e.g., being normalized and translated. A second area 1304, in this example, functions as the normalization workbench interface and is used to perform the various configuration processes such as replacing various abbreviations and expressions with standardized terms or, in the illustrated example, defining a parse tree. Additional workspaces such as workspace 1306 may be provided for accessing other tools such as the Candidate Search Engine which can identify terms for normalization or, as shown, allow for selection of rules. In the illustrated example, normalized terms are highlighted relative to the displayed source file in workspace 1302 on a currently updated basis. In this manner, the SME can readily determine when all or enough of the source file has been normalized. In a traditional e-business environment, this translation process essentially is offline. It becomes real-time and online when new content is added to the system. In this case, assuming well-developed special-purpose dictionaries and linguistic information already exists, the process can proceed in an automatic fashion. Content, once translated is stored in a specially indexed look-up database. This database functions as a memory translation repository. With this type of storage environment, the translated content can be scaled to virtually any size and be directly accessed in the e-business process. The associated architecture for supporting both configuration and run-time operation is discussed below. B. SOLx Architecture 1. Network Architecture Options The SOLx system operates in two distinct modes. The "off-line" mode is used to capture knowledge from the SME/translator and knowledge about the intended transformation of the content. This collectively defines a knowledge base. The off-line mode includes implementation of the configuration and translation processes described above. Once the knowledge base has been constructed, the SOLx system can be used in a file in/file out manner to transform content. The SOLx system may be implemented in a variety of business-to-business (B2B) or other frameworks, including those shown in FIG. 14. Here the Source 1402, the firm that controls the original content 1404, can be interfaced with three types of content processors 1406. The SOLx system 1400 can interface at three levels: with a Local Platform 1408 (associated with the source 1402), with a Target Platform 1410 (associated with a target to whom the communication is addressed or is otherwise consumed by) and with a Global Platform 1412 (separate from the source 1402 and target 1408). A primary B2B model of the present invention focuses on a Source/Seller managing all transformation/localization. The Seller will communicate with other Integration Servers (such as WebMethods) and bare applications in a "Point to Point" fashion, therefore, all locales and data are registered and all localization is done on the seller side. However, all or some of the localization may be managed by the buyer or on a third party platform such as the global platform. Another model, which may be implemented using the global server, would allow two SOLx B2B-enabled servers to communicate in a neutral environment, e.g. English. Therefore, a Spanish and a Japanese system can communicate in English by configuring and registering the local communication in SOLx B2B. A third model would include a local seller communicating directly (via HTTP) with the SOLx B2B enabled Buyer. 2. Network Interface Previously, it was discussed how structured content is localized. The next requirement is to rapidly access this content. If there are ongoing requests to access a particular piece of localized content, it may be inefficient to continually translate the original entry. The issues, of course, are speed and potentially quality assurance. One solution is to store the translated content along with links to the original with a very fast retrieval mechanism for accessing the translated content. This is implemented by the SOLx Globalization Server. The SOLx Globalization server consists of two major components (1) the Document Processing Engine and (2) the Translated Content Server (TCS). The Document Processing Engine is a WebMethods plug-compatible application that objects. The TCS contains language-paired content that is accessed through a cached database. This architecture assures very high-speed access to translated content. This server uses a hash index on the translated content cross-indexed with the original part number or a hash index on the equivalent original content, if there is not a unique part number. A direct link between the original and translated content via the part number (or hash entry) assures retrieval of the correct entry. The indexing scheme also guarantees very fast retrieval times. The process of adding a new localized item to the repository consists of creating the hash index, link to the original item, and its inclusion into the repository. The TCS will store data in Unicode format. The TCS can be used in a standalone mode where content can be accessed by the SKU or part number of the original item, or through text searches of either the original content or its translated variant. If the hashed index of the translated content is known. It, of course, can be assessed that way. Additionally, the TCS will support SQL style queries through the standard Oracle SQL query tools. The Document Processing Engine is the software component of the Globalization Server that allows localized content in the TCS to be integrated into typical B2B Web environments and system-to-system transactions. XML is rapidly replacing EDI as the standard protocol for Web-based B2B system-to-system communication. There are a number of core technologies often call "adaptors" or "integration servers" that translate ERP content, structures, and formats, from one system environment to another. WebMethods is one such adaptor but any such technology may be employed. FIG. 15 shows a conventional web system 1500 where, the WebMethods integration server 1502 takes as input an SAP-formatted content called an IDOC 1504 from a source back office 1501 via API 1503 and converts it into an XML-formatted document 1506 for transmission over the Web 1508 via optional application server 1510 and HTTP servers 1512 to some other receiver such as a Target back office 1510 or other ERP system. The document 1506 may be transmitted to Target back office 1514 via HTTP servers 1516 and an integration server 1518. | |||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||