Use of annotations in statistical machine translation

Description

BACKGROUND

1. Field of the Invention

The present invention relates generally to statistical machine translation of documents and more specifically to systems and methods for processing annotations associated with a translation memory.

2. Description of the Related Art

In the field of computer-generated translations, there are two approaches to translating a document from a source language into a target language. The first approach is statistical machine translation (SMT) which uses a set of statistical probabilities to match a word or phrase in one language to an equivalent word or phrase in another language. The set of probabilities is generated using a large quantity of documents that have been previously translated from the source language to the target language.

The second approach is translation memory (TM) which uses bilingual databases of parallel translations of sentence segments. A segment of a source document in a source language is matched to an entry in the TM. The corresponding entry, in the target language, is provided as a translation of the segment. However, TM techniques are limited to translating only the segments that have a corresponding entry in the TM database. For example, a sentence in French such as “Solaris—mise a jour (11-22 UC) serveur NET: 3 licence de utilization” can not be translated using a database with a similar entry such as:

	Solaris - mise à jour (33-64 UC)	Solaris - SPARC server
	serveur	(33-64 CPU)
	SPARC: 1 licence de utilisation	upgrade: 1 RTU license
	Voir page 122 pour le numéro de	See page 122 for the
	référence de la mise à niveau de	recommended warranty
	garantie recommandée.	upgrade part number.

because the database does not include an entry that matches the sentence exactly. To adapt the TM to identify tags that indicate occurrences such as words or values, the TM includes abstract tags in the entries. For example, the database above may be rewritten as:

	PRODUCTNAME - mise à jour	PRODUCTNAME -
	(NUM-NUM	PRODUCTNAME
	COMPONENTNAME) serveur	server (NUM-NUM
	PRODUCTNAME:	COMPONENTNAME)
	NUM licence de utilisation	upgrade: NUM RTU licence
	Voir page NUM pour le numéro de	See page NUM for the
	référence de la mise à niveau de	recommended warranty
	garantie recommandée.	upgrade part number.

where the tags (in capital letters) appear in place of the words indicating words such as the product name, the component name, and numerals. The rewritten database includes a match for the sentence above and can provide a perfect translation.

However, if the modified TM is used to train an SMT engine, the tags can interfere with the training process and result in less accurate translation probabilities when the system receives segments to be translated that have not been rewritten via the same process used to modify the TM. This may result in inaccurate translations.

SUMMARY OF THE INVENTION

A method, system, and computer readable medium for translating a document is provided. A statistical machine translation engine is trained using a translation memory comprising an annotation. A translation rule associated with the annotation is defined. A source document in a source language is received. The source document comprises an instance of the annotation and a string. The string is translated using the statistical machine translation engine. The instance of the annotation is processed according to the translation rule. A target document in a target language is generated based on the translated string and the processed annotation.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 depicts an illustration of an exemplary environment in which some embodiments may be practiced;

FIG. 2 depicts a block diagram illustrating an exemplary translation engine according to exemplary embodiments;

FIG. 3 depicts a flowchart illustrating an exemplary process for integrating a translation memory with a statistical machine translation engine according to exemplary embodiments; and

FIG. 4 depicts a flowchart illustrating an exemplary process for selecting a configuration set according to exemplary embodiments.

DETAILED DESCRIPTION

A system and method for translating a document having annotations is provided. The system and method may be used to integrate translation memory (TM) systems and statistical machine translation (SMT) systems. The system and method distinguishes between annotations to translate according to rules associated with a configuration set and word strings to translate using statistical machine translation to generate a translated document. The system and method may also be used by the translation engine 104 when no TM is available.

FIG. 1 depicts an illustration of an exemplary environment 100 in which some embodiments may be practiced. The environment 100 comprises a document library 102, a translation engine 104, and a network 106. The translation engine 104 is configured to communicate with the document library over the network 106. The network 106 may comprise a public network (e.g., the Internet) or a private network. The document library 102 comprises parallel translations of documents or segments of documents used in SMT. These parallel translations may be referred to as “parallel segments.” The document library may also include a TM database. The translation memory database may include entries having abstract tags. The abstract tags may include annotations indicating a value, a component name, a product name, or other lingo.

The translation engine 104 is configured to translate a document written in a source language into a target language using statistical machine translation (SMT). The source language is the language in which the source document appears. The target language is the language into which the source document is translated. The generated translation of the document is the target document. The translation engine 104 may retrieve documents from the document library 102 that are to be translated. In some embodiments, the translation engine 104 is trained based on the parallel segments stored in the document library 102. The translation engine 104 may also be trained using the TM database.

In operation, the translation engine 104 accesses the parallel segments and TM database stored in the document library 102 to generate sets of statistical probabilities used to translate a source document. The generated sets of statistical probabilities may be stored in the document library 102 and/or the translation engine 104. If the TM database comprises entries containing abstract tags, the document library 102 and/or the translation engine 104 may comprise a configuration set that defines the abstract tags such that the translation engine 104 is able to recognize the abstract tags as such during training and/or document translation. More specifically, the configuration set is a set of configuration values that define the processing behavior of annotations in a source document.

FIG. 2 depicts a block diagram illustrating a translation engine 104 according to exemplary embodiments. The translation engine 104 is configured to translate a source document into a target language to generate a target document. The translation engine 104 may comprise a computational device, such as a desktop computer, server, laptop computer, or the like. The translation engine 104 comprises a processor 202, a user interface 204, a communication interface 206, and a memory 208 that communicate via the bus 210. The processor 202 is configured to execute commands received via the communication interface 206 or the user interface 204. The commands may include an instruction to generate a set of statistical probabilities using a set of parallel segments, an instruction to identify annotations or abstract tags, and/or an instruction to translate a selected document. The processor 202 is further configured to retrieve and execute commands from the memory 208.

The user interface 204 is configured to display a graphical user interface (GUI) to a user and receive input from the user. The user interface 204 may comprise a monitor, speakers, a keyboard, a mouse, and so forth. The input received via the user interface 204 may include a selection of a document to be translated, an annotation definition, or corrections to the target document. The communication interface 206 is configured to receive and transmit data, such as source documents or target documents, over the network 106.

The memory 208 is configured to store instructions for translating a document using statistical machine translation. The memory 208 may further store a configuration set 212. The configuration set 212 defines annotations that occur in the TM database to indicate abstractions such as values, model numbers, and technical jargon. The configuration set 212 may further define one or more translating or formatting annotations associated with the TM database. The configuration set 212 is discussed further herein. Although translation engine 104 is illustrated as having the processor 202, the user interface 204, the communication interface 206, the memory 208, and the bus 210, fewer or more modules may comprise translation engine 104 and still fall within the scope of various embodiments.

The translation engine 104 is configured to integrate annotations appearing in a TM database into the SMT technique used by the translation engine 104. By integrating the annotations, a source document can be translated as a whole rather than broken into segments that are translated according to whether a match appears in the TM database.

To integrate the annotations, a user generates a configuration set that defines the annotations occurring in a TM database. For example, the annotation:

• • <NUM number=“33”/>
    may appear in the TM database within an entry such as:

Voir page <NUM number = “33”/> pour	See page <NUM number = “33”/>
le numéro de référence de la mise à	for the recommended warranty
niveau de garantie recommandée.	upgrade part number.

The configuration set 212, to instruct the translation engine 104 as to the functionality of the annotation, may include a translation rule such as:

• • <tag tag-name=“NUM” word-class=“number” />
    The configuration set 212, therefore, includes a translation rule for recognizing an annotation and distinguishing the annotation from other sentence fragments appearing in the source document.

By training the translation engine 104 to recognize the annotation “<NUM number=“33”/>”, the translation engine 104 can be further trained based on the entries in the TM database. Training the translation engine 104 using the parallel segments listed as entries in the TM database increases the accuracy of the set of statistical probabilities used in SMT. In the above example, the translation engine 104 is able to assign a higher probability that the French phrase, “Voir page <NUM number=“33”/> pour” is translated “See page <NUM number=“33”/> for.” Further, because of the annotation, the translation engine 104 is able to accurately translate the French phrase, “Voir page <NUM number=“56”/> pour” as “See page <NUM number=“56”/> for” without further training or input from the user.

When a user selects a source document to be translated, the annotations are added to the source document. The translation engine 104 translates the source document using SMT. Additionally, the translation engine 104 processes annotations according to the configuration set 212. For example, some annotations, as described below, may include a translation rule that instructs the translation engine 104 to process segments of the source document differently. The annotations, therefore, give a user greater control over how a source document is translated.

FIG. 3 depicts a flowchart illustrating an exemplary process 300 for integrating a translation memory with a statistical machine translation (SMT) engine according to exemplary embodiments. The statistical machine translation engine may comprise the translation engine 104. The exemplary process 300 allows the translation engine 104 to accurately translate source documents that include the annotations. The annotations may indicate abstract values and/or terms. According to exemplary embodiments, the annotations may comprise extensible markup language (XML) tags, hypertext markup markup language (HTML), other markups based on standard generalized markup language (SGML), a textual field, or other annotation mechanism. The annotations are defined in the configuration set 212 and may appear in the TM database.

In step 302, the translation engine 104 is trained using the TM database to identify and process the annotations in the source document. Training the translation engine 104 using the TM database also allows the translation engine 104 to learn general translation patterns within the database entries and modify the probability values associated with words or fragments accordingly. For example, the translation engine 104 may learn an output context for the annotations. An output context is the word or words that consistently appear in the target document around selected annotations, such as a defined semantic behavior for dates, numbers, product names, or the like.

In step 304, translation rules associated with the annotations are defined for a language pair. The translation rules may include an output context, a translation boundary, and/or a formatting instruction. The translation rules may be probabilistic or heuristic.

For example, a word class annotation indicates that a word or phrase belongs to an abstract word class. For example, the annotation:

• • <NUM numb er=“122”/>
    indicates that the text “122” is a number, such as a page number. Likewise, the annotation:
  • <VAL metric=“2.59 N*m”/>
    indicates that the text “2.59 N*m” is a metric measurement. The translation engine 104 may preserve the word class annotation in the target document. Other annotations may be used to indicate product names, component names, technical jargon, and so forth.

A translation boundary indicates that text occurring before the boundary is translated separately from text occurring after the boundary. By including a boundary, a user is able to prevent “translation swap” occurring across the boundary. For example, a user can indicate that text occurring between parentheses in a source document to be translated separately from the text occurring outside the parenthesis. For example, the annotation:

• • <segment x=“1”/>L=16<segment/>
    indicates that the fragment “L=16” is to be translated independently of the text surrounding the fragment.

In some embodiments, a second boundary annotation comprises a sentence boundary. A sentence boundary indicates a sentence break in instances where sentence punctuation does not normally appear. For example, a sentence boundary may be used to separate the following two sentences:

• • <sentence/>Pieza no reutilizable <sentence/>
  • MOLDURA EXTERIOR DEL PARABRISAS.
    The sentence boundary indicates that the phrases should be treated as separate sentences even though no punctuation appears in the source document.

Another annotation that may be included in the source document is a translation boundary. A translation boundary indicates a segment of the source document that is not to be translated into the target language. Translation boundaries can be nested such that a portion of a not-translated segment is translated in the target document. For example, the underlined text in the following sentence will be translated into the target language:

• • Here's an <off> comment <on> interesting </on> another comment </off> problem.
    where the word, “interesting” is translated because the “on” annotations defining the translation boundary are nested within the “off” annotations defining another translation boundary.

Formatting annotations may include space preservation around the annotations when the source document is translated into the target language. The translation engine 104 may insert extra spaces or delete space around words that may be undesirable in the target document. To preserve the spacing, the annotations are inserted with the desired spacing after the start tag and before the end tag. For example:

• • <textart id=“b111209e17-1-7”>from P/W RL Fuse </textart>
    indicates that there is no space before the word “from” and that there is a space after the word “fuse.” In some embodiments, word class annotations, described above, do not include space preservation.

Using annotations, the annotations themselves can be preserved within the target document. By preserving the annotations, the translation engine 104 does not translate data occurring within the annotation itself. For example, an annotation may include a part name or a component name that is not to be translated into the target language. Referring to the previous example, the fragment, “b111209e17-1-7”, will not be translated. The tag body, however, may be translated. In the above example, the phrase “from the P/W RL Fuse” will be translated into the target language.

In step 306, the source document comprising annotations and strings is received. The annotations include, for example, the XML tags described herein. The strings include the portions of the source document that are not annotated. The strings are processed by the translation engine 104 using statistical machine translation techniques.

In step 308, the language pair configuration set is selected. The language pair configuration set 212 is a group of configuration settings that defines the annotations associated with a source language and/or a target language. The configuration set 212 may additionally specify which annotations are to be removed from the target document after translation and which annotations are to be preserved in the target document. By editing the configuration sets, users of the translation engine 104 are able to define additional annotations and the behavior associated with the annotations. A configuration set may be accessed even if no TM database is available. An example of a configuration set 212 for documents in which the source language is English and the target language is empty is:

• • <config>
    • <filters>
      • <xml-filter src=“eng” tgt=“ ”>
        • <tags>
        •  <tag tag-name=“seg” break=“1” />
        •  <tag tag-name=“unit” word-class=“unit” />
        •  <tag tag-name=“val” word-class=“val” />
        •  <tag tag-name=“num” word-class=“num” />
        •  <tag tag-name=“tr-on” hands-on=“1” />
        •  <tag tag-name=“tr-off” hands-off=“1” />
        •  <tag tag-name=“segment” no-swap=“1” />
        •  <tag tag-name=“textart”/>
        •  <tag tag-name=“ptext” preserve-space=“1” />
        • </tags>
      • </xml-filter>
      • <xml-filter name=“Auto Parts Catalog”>
        • <tag tag-name=“unit” word-class=“unit” />
        • <tag tag-name=“val” word-class=“val” />
        • <tag tag-name=“num” word-class=“num” />
      • </xml-filter>
    • </filters>
  • </config>
    This exemplary configuration set can be used for any language pair where the source language is English. This configuration set also includes a second configuration set that is defined by a configuration set name, “Auto Parts Catalogue.” The user can identify specific configuration sets by using configuration set names. As such, in some embodiments, the configuration set 212 is selected and/or specified by the user. The configuration set 212 may be selected according to the exemplary process depicted in FIG. 4.

In step 310, the source document is translated according to the annotations. Using the annotations, a more accurate and more fluent translation can be generated by the translation engine 104. The annotations can be adapted to technical jargon occurring in the source document. The annotations may also be used to insert objects such as images, charts, tables, graphs, or the like appearing in the source document into the target document.

FIG. 4 depicts a flowchart illustrating an exemplary process 400 for selecting a configuration set, such as configuration set 212, according to exemplary embodiments. The configuration set may be stored in as a file. Further, a single file may contain more than one configuration set. In some embodiment, a configuration set may be identified by a configuration set name. The unique name allows a user to select a specific configuration set to be used for processing a document. Further, upon request, the system can supply the user with a list of all available configuration sets. The exemplary process 400 may be performed as step 306 by the translation engine 104, according to some embodiments.

In step 402, a determination is made as to whether a specific configuration set is selected by the user. In the above example, a user may select the configuration set for “Auto Parts Catalogue.” If a configuration set is selected by the user, a second determination is made as to whether the configuration set is found in step 404. If the configuration set is not found in step 404, an error is reported to the user in step 406. If the configuration set is found, the annotations in the source document are processed according to the configuration set selected by the user in step 408.

If no specific configuration set is selected by the user in step 402, a determination as to whether a configuration set for the exact language pair of the source language and the target language exists in step 410. In a first further step, if the exact language pair does not exist, another determination may be made as to whether a configuration set having a matching source language and no target language exists. In a second further step, if no match for the source language is found, a determination may be made as to whether a configuration set having no source language and a matching target language exists.

In step 410, and the first and second further steps discussed above, if a matching configuration set is found, the annotations in the source document may be processed according to the matching configuration set, as described in step 408. If no matching configuration set exists, the annotations in the source document may be processed according to an empty configuration set. The empty configuration set allows the translation engine 104 to translate the source document as though the annotations are not present.

The above-described functions and components can be comprised of instructions that are stored on a storage medium. The instructions can be retrieved and executed by a processor. Some examples of instructions are software, program code, and firmware. Some examples of storage medium are memory devices, tape, disks, integrated circuits, and servers. The instructions are operational when executed by the processor to direct the processor to operate in accord with various embodiments. Those skilled in the art are familiar with instructions, processor(s), and storage medium.

While various embodiments have been described above, it should be understood that they have been presented by way of example only, and not limitation. The scope of the present disclosure is in no way limited to the languages used to describe exemplary embodiments. Thus, the breadth and scope of a preferred embodiment should not be limited by any of the above-described exemplary embodiments.