Personalized machine translation via online adaptation

Abstract

Personalizing machine translation via online adaptation is described herein. According to some embodiments, methods for providing personalized machine translations may include receiving translator feedback regarding machine translations generated by a machine translation system for a translator, determining translator feedback that improves translations generated by the machine translation system, and incorporating the determined translator feedback into the translation methodology of the machine translation system to personalize the translation methodology.

Description

FIELD OF THE PRESENT TECHNOLOGY

The present technology relates generally to machine translation methodologies and systems, and more specifically, but not by way of limitation, to personalized machine translation via online adaptation, where translator feedback regarding machine translations may be intelligently evaluated and incorporated back into the translation methodology utilized by a machine translation system to improve and/or personalize the translations produced by the machine translation system.

BACKGROUND

The quality of translations produced by machine translation systems often depends upon the quality of the translation methodologies utilized by the machine translation systems. Commonly utilized translation methodologies may comprise language models, translation memories, parallel corpora, translation dictionaries, phrase tables, and so forth. Additionally, translations produced by machine translation systems may be improved by adding new information to the translation methodologies utilized by the machine translation systems. The present technology contemplates utilizing translator feedback regarding machine translations to improve and/or personalize the translations produced by these machine translation systems.

SUMMARY OF THE PRESENT TECHNOLOGY

According to some embodiments, the present technology may be directed to methods for managing a personalized machine translation system. These methods may comprise: (a) receiving translator feedback regarding machine translations generated by a machine translation system for a translator, the machine translation system using a translation methodology; (b) determining translator feedback that improves translations generated by the machine translation system; and (c) incorporating the determined translator feedback into the translation methodology of the machine translation system to personalize the translation methodology.

According to other embodiments, the present technology may be directed to personalized machine translation systems. These systems may comprise: (a) a feedback processor that receives translator feedback regarding machine translations generated by a machine translation system for a translator, the machine translation system using a translation methodology; (b) a feedback classifier that determines translator feedback that improves translations generated by the machine translation system; and (c) a dictionary pusher that incorporates the translator feedback into the translation methodology of the machine translation system to personalize the translation methodology.

According to additional embodiments, the present technology may be directed to methods for managing a personalized machine translation system. These methods may comprise: (a) receiving translator feedback regarding machine translations generated by a machine translation system for a translator, the machine translation system using a translation methodology; (b) determining translator feedback that improves translations generated by the machine translation system; (c) determining one or more translation preferences for the translator; and (d) incorporating the determined translator feedback and the translation preferences into the translation methodology of the machine translation system to personalize the translation methodology.

According to some embodiments, a method may comprise: (a) generating a personalized translation of a target sentence using a language model, the language model used to determine a likelihood that a target sentence is grammatically well-formed, independent of translation quality; (b) receiving a feedback-adjusted translation for the language model translation; and (c) incorporating the feedback-adjusted translation into a training corpus.

BRIEF DESCRIPTION OF THE DRAWINGS

Certain embodiments of the present technology are illustrated by the accompanying figures. It will be understood that the figures are not necessarily to scale and that details not necessary for an understanding of the technology or that render other details difficult to perceive may be omitted. It will be understood that the technology is not necessarily limited to the particular embodiments illustrated herein.

FIG. 1 illustrates an exemplary architecture for practicing aspects of the present technology, the exemplary architecture including an exemplary machine translation personalization system.

FIG. 2A is a table of exemplary translation feedback sets.

FIG. 2B is a table of exemplary annotated feedback entries categorized for usefulness.

FIG. 3A is a table of best scoring precision and accuracy values calculated for useful translation feedback.

FIG. 3B is a table that illustrates a variety of metrics for filtered versus unfiltered feedback.

FIG. 3C is a table that includes a variety of metrics for determining dictionary entries for a baseline machine translation versus dictionary entries as affected by clean feedback.

FIG. 4 is a flowchart of an exemplary method for processing translation feedback, extracting dictionary entries, and updating a translation methodology of a machine translation system;

FIG. 5 is a flowchart of another exemplary method for managing a personalized machine translation system;

FIG. 6 is a block diagram of an exemplary computing system for implementing embodiments of the present technology.

DETAILED DESCRIPTION OF EXEMPLARY EMBODIMENTS

While this technology is susceptible of embodiment in many different forms, there is shown in the drawings and will herein be described in detail several specific embodiments with the understanding that the present disclosure is to be considered as an exemplification of the principles of the technology and is not intended to limit the technology to the embodiments illustrated.

It will be understood that like or analogous elements and/or components, referred to herein, may be identified throughout the drawings with like reference characters. It will be further understood that several of the figures are merely schematic representations of the present technology. As such, some of the components may have been distorted from their actual scale for pictorial clarity.

Generally speaking, the present technology may be directed, in some embodiments, to personalized machine translation via online adaptation. More specifically, but not by way of limitation, the present technology may comprise machine translation systems that receive and process translator feedback regarding machine translations. It will be understood that the machine translation system may utilize one or more types of translation methodologies to translate a source text in a source language into one or more target languages.

These and other advantages of the present technology will be described in greater detail below with reference to the collective drawings (e.g., FIGS. 1-6).

FIG. 1 illustrates an exemplary architecture 100 for practicing aspects of the present technology. According to some embodiments, the exemplary architecture 100, hereinafter “architecture 100,” may generally include a personalized machine translation system, hereinafter “system 105.” Translators or other end users may interact with the system 105 using a client device 110. In other embodiments, translators may interact with the system 105 via a web based interface, or an application resident on the client device 110.

The client device 110 and the system 105 may be communicatively coupled via a network 115. It is noteworthy to mention that the network 115 may include any one (or combination) of private or public communications networks such as the Internet. Additionally, various components of the system 105 may be communicatively coupled together via the network 115, where network connections of system 105 are shown in FIG. 1 as flow arrows between components.

According to some embodiments, individual components of the system 105 such as feedback classifier 145, dictionary extractor 155, and dictionary pusher 170 (or the entire system 105) may be implemented within a cloud-based computing environment. In general, a cloud-based computing environment is a resource that typically combines the computational power of a large grouping of processors and/or that combines the storage capacity of a large grouping of computer memories or storage devices. For example, systems that provide a cloud resource may be utilized exclusively by their owners, such as Google™ or Yahoo!™; or such systems may be accessible to outside users who deploy applications within the computing infrastructure to obtain the benefit of large computational or storage resources.

The cloud may be formed, for example, by a network of web servers, with each web server (or at least a plurality thereof) providing processor and/or storage resources. These servers may manage workloads provided by multiple users (e.g., cloud resource consumers or other users). Typically, each user places workload demands upon the cloud that vary in real-time, sometimes dramatically. The nature and extent of these variations typically depend on the type of business associated with the user.

Generally speaking, a source text 120 in a source language may be received from a translator using the client device 110. In some instances, the client device 110 may utilize a translation API for providing source texts to the system 105 and receiving machine translations therefrom. Translations of the source text 120 may be generated by a machine translation system, also referred to as a translation server farm 120A.

As would be known to one of ordinary skill in the art, the translation server farm 120A may comprise any, or a number of, machine translation systems that employ any number of translation methodologies. Broadly speaking, a translation methodology may be described as any method, process, algorithm, or combination that allows the translation server farm 120A to translate a source text in a source language to a target text in one or more target languages. Exemplary translation methodologies may comprise corpus-based machine translation, rule-based machine translation, transfer-based machine translation, hybrid-based machine translation, statistic-based machine translation, a translation memory, a language model, a translation dictionary, a phrase table, or any combinations or permutations thereof.

According to some embodiments, the translation server farm 120A may utilize a baseline translation methodology that comprises a translation model that utilizes source phrases to generate target translations that are contextually and syntactically relevant. Additionally, the translation methodology may comprise a language model that predicts how likely a particular target translation is to be correct. The translation methodology may utilize translation dictionaries (e.g., phrase tables) and weight vectors that weight each of the translation information generated by the language and translation models, as well as the translation dictionaries. Translator feedback may be utilized to improve the language model, translation model, and/or the translation dictionaries of the translation methodology. The personalization and/or updating of translation dictionaries will be discussed in greater detail below.

Once the translation for the source text 120 has been generated, a source and translation text pair 125 may be provided back to the client device 110 via the translation API. In other instances, the source and translation text pair 125 may be provided to a review environment 130. The review environment 130 may comprise a web-based feedback input system that receives translator feedback 135 from the client device 110. According to some embodiments, translation preferences may also be received using the review environment 130. In some instances, translators may input various translation preferences, such as translator domain. A translator domain may include a reference that may be utilized to guide or affect translations. For example, if the translator works in the financial services industry, the domain for the translator may be “financial services.” Thus, when the machine translation system encounters the word “bank” in a source text, the machine translation system may default to translating “bank” into a corresponding term in the target language, rather than the “bank” of a river, for example. The domain for the translator may be specified by the translator or may be inferred from an analysis of the subject matter or topics included in the source texts provided by the translator.

The translator feedback 135 may comprise any number of different types of information that reflect the translator's response to a translation generated by the translation server farm 120A. The translator's response may comprise, for example, an accuracy rating, corrections for the translation, and/or open-ended commentary, as well as other translation related feedback that would be known to one of ordinary skill in the art with the present disclosure before them. In some instances, the translator feedback 135 may comprise linguistic and qualitative annotations that support translation modeling and feedback research.

According to some embodiments, translator feedback 135 may comprise a triple that comprises a source unit, a translation unit, and a feedback unit. These various units may, in turn, be comprised of segments such as words, phrases, sentences, paragraphs, and so forth. In practice, the source unit may comprise a source word, the target translation unit may comprise a target translation word, and the feedback unit may comprise an accuracy rating for the translated word.

In other instances, the source unit may comprise multiple words or a phrase. The translation unit may comprise a translation of these words or phrases into a target language. Correspondingly, the feedback unit may comprise corrections for the target unit. For example, corrections may include the number of transformations, changes, edits, or other modifications that had to be performed on the translation unit to make it acceptable to the translator.

Exemplary translator feedback is shown in FIG. 2A in table format. The table 200 includes a plurality of translator feedback groups, such as group 205, which are arranged onto the table 200. The table 200 may comprise a series of columns that correspond to units of the feedback groups. For example, the table 200 may comprise a source unit column 210, a translation unit column 215, a correction unit column 220, and an optional commentary unit column 225.

The review environment 130 may then provide translator feedback 135 to a feedback data store 140. At this stage, the translator feedback 135 may comprise raw or noisy translation feedback, meaning that the translation feedback has not been processed to determine if the translation feedback is understandable, useable, actionable, or otherwise available for processing and/or incorporation by the system 105.

Using a process feedback API, the feedback classifier 145 may receive the translator feedback 135 from the feedback data store 140. Generally speaking, the feedback classifier 145 may be utilized to separate useful feedback from useless feedback. Again, a “useful” feedback entry is defined as one where the correction may be utilized to produce a better translation of the source text than translation generated by the machine translation system. Any other feedback content, such as a correction that makes the translation worse, commentary, junk, or badly formed data, may be classified by the feedback classifier as “not useful” and may be discarded or ignored by the feedback classifier 145.

FIG. 2B is a table 230 that includes exemplary annotated feedback entries categorized for usefulness. In some instances, translator feedback may be evaluated by human linguistic experts who annotate the translator feedback as either useful or non-useful.

Rather than using linguistic experts, the feedback classifier 145 may analyze the translator feedback using one or more surface classes of information to determine the “usefulness” of the translator feedback 135. Exemplary surface classes may comprise evaluations of: (1) a (word|character) Levenshtein distance between correction and translation, divided by translation length; (2) the correction/translation word overlap, divided by translation length; (3) translation words not included in a correction, divided by translation length; (4) correction words not included in the translation, divided by correction length; (5) (translation|correction)/source word overlap, divided by source length; (6) the length of (source|translation|correction); (7) the length of (translation|correction) divided by length of source; (8) (average|maximum) length of (source|translation|correction) word; (9) (average|maximum) length of (translation|correction) word, divided by (average|maximum) source word length; (10) (average|maximum) length of translation unit, divided by (average|maximum) correction word length; (11) determining if (correction|source) are included in the same string.

According to additional embodiments, the feedback classifier 145 may utilize back-translation (BT) features to determine the “usefulness” of the translator feedback 135. Generally speaking, the back-translation of translator feedback may be calculated by first generating translation-bt and correction-bt, the result of translating the translation or, respectively, the correction, back into the source language. It will be understood that extraction of BT features by the feedback classifier 145 may rely on the creation of back-translations of translated and corrected sentences.

Exemplary BT features may comprise: (1) (word|character) Levenshtein distance between (translation-bt|correction-bt) and the source segment; (2) (word|character) Levenshtein distance between translation-bt and correction-bt; (3) (translation-bt|correction-bt)/source word overlap, divided by source segment length; and (4) (translation-bt|correction-bt) words not in source, divided by translation-bt length.

FIG. 3A includes a table 300 that includes precision (Prec) values, which includes percentages of sentences classified that were classified as useful that actually generated an improved translation. The table 300 also includes accuracy (Acc) values, which represents percentages of translator feedback samples that were classified correctly, and recall (Rec) values, which represents percentages of useful sentences that have been categorized. As can be seen, Prec, Acc, and Rec values are calculated for both surface features and BT features. These values may also be calculated for the baseline machine translation system (e.g., machine translation system that does not utilize translator feedback).

In some instances, useful feedback may be stored in a second feedback data store 150 that is communicatively coupled with the feedback classifier and a dictionary extractor 155.

After translator feedback has been determined to be useful by the feedback classifier 145, the dictionary extractor 155 may be utilized to determine dictionary entries from the feedback. In some embodiments, the dictionary entries may include sub-sentential segments. The dictionary extractor 155 may locate phrase pairs that are implied by the feedback entries to be missing from or insufficiently weighted in a phrase table utilized by the machine translation system, such as the translation server farm 120A.

The dictionary extractor 155 may extract dictionary entries from a corpus of translator feedback triples (e.g., source unit, translation unit, correction unit). According to some embodiments, the dictionary extractor 155 may construct a first translation table and a second translation table. Generally speaking, the first translation table may comprise bitext pairs of source units and correction units. The dictionary extractor 155 may extract dictionary entries from the translator feedback by preparing the bitext pairs for alignment, aligning the words included in the bitext pairs, and extracting first phrase pairs based upon word alignment.

Next, the dictionary extractor 155 may construct a second translation table that comprises first pairs of source units and translation units. The dictionary extractor 155 may extract dictionary entries from the translator feedback by again preparing the bitext pairs for alignment, aligning the words included in the bitext pairs, and extracting second phrase pairs based upon word alignment.

According to some embodiments, the dictionary extractor 155 may prepare bitext pairs for alignment by segmenting words, tokenizing the segments, decapitalizing both sides of the bitext pairs, or any combinations thereof. Next, the dictionary extractor 155 may align the words in the bitext pairs using, for example, GIZA++ commercially available from International Business Machines, Inc.

In some embodiments, the dictionary extractor 155 may extract phrase pairs from the bitext pairs that are consistent with the alignments and subject to typical restrictions such as phrase length, unaligned word restrictions, and so forth, which will be described in greater detail below.

The dictionary extractor 155 may identify phrase pairs from the first and second translation tables that are likely good dictionary corrections, subject to the following exemplary restrictions. One exemplary restriction specifies that only phrase pairs with three or more words in either phrase will be considered by the dictionary extractor 155. Another exemplary restriction specifies that only phrase pairs with terminal words aligned will be considered by the dictionary extractor 155. Exemplary restrictions specify that for considered phrase pairs, for a given source side, if the (source, translation) and (source, correction) phrase tables do not share a target side, and the (source, correction) table has exactly one target side for that source, the phrase from the (source, correction) table is taken as an entry.

In an exemplary use case, the aforementioned methods were utilized by the dictionary extractor 155 to extract dictionary entries from the filtered noisy translator feedback acquired by the feedback classifier 145. Following the methodology described above, the feedback classifier 145 was trained on 814 English->Spanish and English->French annotated feedback entries, using baseline and back-translation features. The feedback classifier 145 was utilized to process 2,382 English->Spanish entries containing correction feedback, which captured by the feedback API and/or the review environment 130.

The feedback classifier 145 filtered 1,749 of the entries as “Useful.” From those entries a dictionary was constructed by the dictionary extractor 155, as described above. The dictionary contained approximately 10,286 entries. By way of contrast a dictionary having 10,677 entries was also constructed from the entire unfiltered 2,382-entry set, to determine the effects of filtering on dictionary quality.

FIG. 3B includes a table 305 that comprises a comparison the translation accuracy produced by a machine translation system using a dictionary-free baseline, as well as exemplary machine translation system that utilize various dictionaries. An original-source test set was used. Also included is an average mixed-case BLEU on two references. The table also includes percentages of phrases used in the dictionary-enhanced translations that come from the dictionary, and the percent of the dictionary entries that are good-quality, which were judged by human linguistic experts 160 who annotated randomly sampled entries.

FIG. 3C is a table 310 that shows the effects of using these entries on single-reference, mixed-case bilingual evaluation understudy (BLEU) values for various machine translations (using baseline and dictionary-enhanced machine translation). The basis for table 310 assumes that good-quality dictionary entries and translations are available to the machine translation system. Extractions for a corpus of 3,000 feedback entries in the English->Spanish technical manual domain that were professionally corrected were analyzed. The extractions were generated by repeating the dictionary extraction methods described above. Using the above-described methodology, 43,481 entries were extracted from the feedback. An exemplary held-out test corpus of 2,839 sentences was processed using both baseline and machine translation methodologies that were enhanced with the dictionary.

According to some embodiments, the dictionary extractor 155 may be configured to store selected first and second phrase pairs in a translation phrase table that resides on a data repository, such as the dictionary data store 165.

Once the dictionary entries have been extracted from the translator feedback 135, the system 105 may optionally provide the dictionary entries to human linguistic experts 160 for validation. Validated dictionary entries may be stored in a dictionary data store 165. In other embodiments, dictionary entries may be stored in the dictionary data store 165 without validation.

Next, a dictionary pusher 170 may provide the validated and/or non-validated dictionary entries to the translation server farm 120A. In some instances, the dictionary pusher may incorporate, integrate or otherwise combine the dictionary entries into the translation methodology utilized by the translation server farm 120A.

In addition to updating the translation methodology with the information extracted by the dictionary extractor 155, the dictionary pusher 170 may also provide the translation server farm 120A with translator preferences that may be utilized to further personalize the machine translations generated by the translation server farm 120A.

While the process illustrated in FIG. 1 contemplates the processing of useful translator feedback by extracting dictionary entries that may be incorporated into a phrase table (e.g., translation methodology), one of ordinary skill in the art will appreciate that translator feedback may be processed into other types of useable data that may be incorporated into translation methodologies other than phrase tables, such as parallel corpora, translation memories, as well as other translation methodologies.

The machine translation systems of the system 105 may process translator feedback regarding these machine translations and filter the translator feedback for usefulness. In accordance with the present disclosure, the “usefulness” of translator feedback may be determined by the system 105 evaluating whether the feedback, when incorporated into the translation methodology utilized by the machine translation system, produces superior machine translations as compared to machine translations systems that do not incorporate such feedback. The system 105 may produce a translation that is more accurate (e.g., a translation that required fewer manipulations or corrections). Additionally, the system 105 may personalize machine translations by gathering knowledge about the translation preferences of translators, such as common corrections, translator domains, and so forth and applying the translation preferences to machine translations generated for a particular translator.

According to some embodiments, the system 105 may utilize translation model (TM)-based personalization. The goal of TM-based personalization is similar to that of dictionary-based personalization (e.g., using a parallel corpus), to increase coverage of the translation model in response to correction feedback. While a dictionary-based approach may be executed efficiently, these approaches may not take into account a context of the translations it handles. That is, dictionary-based personalization may find source phrases that it has entries for and replace them with memorized target phrases.

Statistical MT in general may be seen as a context-aware methodology that combines learned phrase pairs together to form a translation. TM-based personalization uses feedback to learn new phrase pairs and then includes them with previously learned phrase pairs in an extant statistical MT system.

Similarly to dictionary-based personalization, a TM-based personalization method used by the system 105 may extract phrases from source and correction sentence pairs by inducing a word alignment between the elements of the pairs, for example, by using an unsupervised technique, such as GIZA++, and a sufficiently large corpus of pairs.

A standard phrase pair extraction methodology may be utilized to obtain sub-sentential phrase pairs that are consistent with the alignment and various externally imposed restrictions such as phrase length and treatment of unaligned words. Since the purpose of TM-based personalization is to expand coverage, TM-based personalization may be concerned with those phrase pairs that are not already in the phrase table.

TM-based personalization methods described above are similar to those used for dictionary-based personalization. Additionally, feedback-based phrase pairs may be incorporated into an extant phrase pair database by establishing features for the phrase pairs. A feature is a property of a phrase pair that provides some indication of the pair's utility. Some examples of features are “observed likelihood of target words given source” and “length of target phrase”. Each phrase pair has numeric values for each feature. The feature values for a phrase pair are combined together to give a total score for the phrase pair, which indicates how “good” the pair according to the current translation model.

Features whose values can be determined from the phrase pair text itself (e.g., “length of target phrase”) are easy to calculate for these new phrase pairs, but features based on statistics databases (e.g., “observed likelihood of target words given source”) may be more difficult to calculate due to the lack of appropriate statistics (e.g., the source and target words may not have been observed together sufficiently frequently to obtain an accurate likelihood). For such features the system 105 may simulate appropriate statistics by using either accumulated statistics for low-frequency words or established default values.

Having determined the new phrase pairs, the new phrase pairs may be incorporated into the extant phrase table and a statistical MT may be performed by the system 105 as described above.

In some instances, the system 105 may utilize tuning-based personalization. As mentioned above the score of a phrase pair may be determined by combining its feature values together. According to some embodiments, the “combining” may be performed as follows: For each feature, a numeric weight (which may be positive, negative, or zero) may be determined that indicates the usefulness of the feature relative to determining a phrase pair's overall quality. A positive weight indicates a high feature value (feature values are always zero or positive), which corresponds to high quality. A negative weight means a high feature value, which corresponds to low quality. A zero weight means a feature has no effect on quality. The score of a phrase pair is determined by summing the product of each feature value with its corresponding weight. As an example, let there be three features: A, B, and C. Let the weights for each feature be, respectively, W_A=2, W_B=1, and W_C=−3. For a given phrase pair x, let the feature values be, respectively, FxA=5, FxB=3, and FxC=2. Then the score of x, sX, is W_AF^XA+W_BF^XB+W_CF^XC=10+3+(−6)=7.

According to some embodiments, weights within an exemplary MT system may be determined by iteratively generating a large number of possible translations of a corpus of sentences (called the tuning corpus), determining the weights that enable the highest scoring translations to correspond to the highest possible extrinsic evaluations of those translations (e.g., via the BLEU metric). Additionally, the system 105 may use the calculated weights to generate more translations, until no further gains can be made. This iterative procedure is called tuning. Another commonly used approach to tuning, which differs only in the optimization algorithm used, is the Minimum Error Rate Training (“MERT”) approach.

In some embodiments, the system 105 may utilize feedback as a tuning corpus. Careful selection of a tuning corpus may be utilized to obtain good MT performance. Any tuning corpus should have broad coverage of words and grammar and thus come from a “general” domain. However, if it is known that translation may be skewed toward a particular other domain, performance can be improved by including in-domain sentences in the tuning corpus, in addition to the established general domain set. Since personalization may be viewed as a domain adaptation problem, the present technology may consider the (source, correction) pairs as a domain-specific set and append them to our tuning corpus. The system 105 may then use an established tuning methodology, such as those described above. These methods may be performed iteratively, by incorporating feedback pairs into the tuning set, choosing new weights, using those new weights to generate responses to more translation requests, and then collecting yet more feedback pairs, until it is judged that no additional benefit is gained.

The system 105 may also utilize feedback as tunable feature. In some instances, statistics may be collected based on feedback. Tuning features may be derived from those statistics. For example, the system 105 may be used to calculate a statistic that represents the likelihood that a phrase pair is “licensed” by the feedback. A phrase pair may be “licensed” if it is used both in the original machine translation of a sentence and in the correction of that sentence. This statistic is calculated from a (source, translation, correction) corpus and the new feature may then be used when tuning in order to determine a weight for it. Intuitively, licensed phrase pairs may be favored over unlicensed phrase pairs.

In some instances, the system 105 may employ language model (LM)-based personalization. It is noteworthy that an information source usable in MT is a language model (LM) of the target language. An LM provides a likelihood that a target sentence is well-formed, independent of its translation quality. Some MT systems use very large language models, constructed from general-domain language. These MT systems produce good quality translation results. However, a system that incorporates domain specific language models can improve performance if it is known that translation will be in a specific domain. Thus, the system 105 may consider personalization as a form of domain adaptation, and use given corrections as sentences in a target domain. From these sentences a small, domain-specific LM may be generated and used as an additional information source in translation.

Additionally, the system 105 may modify individual probabilities in the existing LM to correspond to differences between the automatically translated and feedback-adjusted corpora. Probabilities in a LM are generally expressed as the conditional likelihood of a word given a fixed number of words observed before it. If, for example, the phrase “fired the big guns” is observed to be frequently corrected to “fired the big cannons” then we adjust the likelihood of “guns” given “fired the big” lower and that of “cannons” given “fired the big” higher.

FIG. 4 illustrates another exemplary method for processing translation feedback, extracting dictionary entries, and updating a translation methodology of a machine translation system.

FIG. 5 is a flowchart of an exemplary method 500 for managing a personalized machine translation system. The method may comprise a step 505 of receiving translator feedback regarding machine translations generated by a machine translation system for a translator. Again the machine translation system may utilize a translation methodology (or multiple methodologies) for translating source language texts into target language texts. The method may also comprise a step 510 of determining translator feedback that improves translations generated by the machine translation system, as well as a step 515 of incorporating the determined translator feedback into the translation methodology of the machine translation system to personalize the translation methodology.

FIG. 6 illustrates an exemplary computer system 600 that may be used to implement an embodiment of the present technology. The computer system 600 of FIG. 6 may be implemented in the contexts of the likes of computing systems, networks, exchanges, servers, or combinations thereof disclosed herein. The computer system 600 of FIG. 6 includes one or more processors 610 and main memory 620. Main memory 620 stores, in part, instructions and data for execution by processor 610. Main memory 620 may store the executable code when in operation. The computer system 600 of FIG. 6 further includes a mass storage device 630, portable storage medium drive(s) 640, output devices 650, user input devices 660, a graphics display 670, and peripheral devices 680.

The components shown in FIG. 6 are depicted as being connected via a single bus 690. The components may be connected through one or more data transport means. Processor unit 610 and main memory 620 may be connected via a local microprocessor bus, and the mass storage device 630, peripheral device(s) 680, portable storage device 640, and graphics display 670 may be connected via one or more input/output (I/O) buses.

Mass storage device 630, which may be implemented with a magnetic disk drive or an optical disk drive, is a non-volatile storage device for storing data and instructions for use by processor unit 610. Mass storage device 630 may store the system software for implementing embodiments of the present technology for purposes of loading that software into main memory 620.

Portable storage device 640 operates in conjunction with a portable non-volatile storage medium, such as a floppy disk, compact disk, digital video disc, or USB storage device, to input and output data and code to and from the computer system 600 of FIG. 6. The system software for implementing embodiments of the present technology may be stored on such a portable medium and input to the computer system 600 via the portable storage device 640.

Input devices 660 provide a portion of a user interface. Input devices 660 may include an alphanumeric keypad, such as a keyboard, for inputting alpha-numeric and other information, or a pointing device, such as a mouse, a trackball, stylus, or cursor direction keys. Additionally, the computer system 600 as shown in FIG. 6 includes output devices 650. Suitable output devices include speakers, printers, network interfaces, and monitors.

Graphics display 670 may include a liquid crystal display (LCD) or other suitable display device. Graphics display 670 receives textual and graphical information, and processes the information for output to the display device.

Peripherals devices 680 may include any type of computer support device to add additional functionality to the computer system. Peripheral device(s) 680 may include a modem or a router.

The components provided in the computer system 600 of FIG. 6 are those typically found in computer systems that may be suitable for use with embodiments of the present technology and are intended to represent a broad category of such computer components that are well known in the art. Thus, the computer system 600 of FIG. 6 may be a personal computer, hand held computing system, telephone, mobile computing system, workstation, server, minicomputer, mainframe computer, or any other computing system. The computer may also include different bus configurations, networked platforms, multi-processor platforms, etc. Various operating systems may be used including Unix, Linux, Windows, Macintosh OS, Palm OS, Android, iPhone OS and other suitable operating systems.

It is noteworthy that any hardware platform suitable for performing the processing described herein is suitable for use with the technology. Computer-readable storage media refer to any medium or media that participate in providing instructions to a central processing unit (CPU), a processor, a microcontroller, or the like. Such media may take forms including, but not limited to, non-volatile and volatile media such as optical or magnetic disks and dynamic memory, respectively. Common forms of computer-readable storage media include a floppy disk, a flexible disk, a hard disk, magnetic tape, any other magnetic storage medium, a CD-ROM disk, digital video disk (DVD), any other optical storage medium, RAM, PROM, EPROM, a FLASHEPROM, any other memory chip or cartridge.

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 descriptions are not intended to limit the scope of the technology to the particular forms set forth herein. Thus, the breadth and scope of a preferred embodiment should not be limited by any of the above-described exemplary embodiments. It should be understood that the above description is illustrative and not restrictive. To the contrary, the present descriptions are intended to cover such alternatives, modifications, and equivalents as may be included within the spirit and scope of the technology as defined by the appended claims and otherwise appreciated by one of ordinary skill in the art. The scope of the technology should, therefore, be determined not with reference to the above description, but instead should be determined with reference to the appended claims along with their full scope of equivalents.

Claims

1. A method for managing a personalized machine translation system, the method comprising: receiving translator feedback regarding machine translations generated by a machine translation system for a translator, the machine translation system using a translation methodology, the translator feedback comprising a triple having a source unit, a translation unit of the source unit, and a correction unit;determining translator feedback that improves translations generated by the machine translation system;incorporating the determined translator feedback into the translation methodology of the machine translation system to personalize the translation methodology, the incorporation of the translator feedback comprising: constructing a first translation table that comprises pairs of source units and correction units;extracting dictionary entries from the translator feedback by: preparing the pairs for alignment,aligning words included in the pairs, andextracting first phrase pairs based upon word alignment;constructing a second translation table that comprises pairs of source units and translation units; andextracting dictionary entries from the translator feedback by: preparing the pairs for alignment,aligning words included in the pairs, andextracting second phrase pairs based upon word alignment; andfiltering the first and second phrase pairs by: selecting first and second phrase pairs having a threshold amount of words included in the phrase;selecting first and second phrase pairs that have terminal words which are aligned; andincluding the selected first and second phrase pairs in the translation methodology of the machine translation system.

2. The method according to claim 1, wherein the translation methodology comprises any of corpus-based machine translation, rule-based machine translation, transfer-based machine translation, hybrid-based machine translation, statistic-based machine translation, a translation memory, a language model, a translation dictionary, a phrase table, or any combination thereof.

3. The method according to claim 1, further comprising evaluating the translator feedback prior to the step of incorporating the translator feedback into the translation methodology to determine if the translator feedback comprises an improved translation relative to a translation generated by the machine translation system without the translator feedback.

4. The method according to claim 1, wherein the translator feedback includes corrections that are received from the translator relative to a previous translation generated by the machine translation system.

5. The method according to claim 4, further comprising providing the corrections to a human translator and receiving verified corrections from the human translator, wherein the verified corrections are incorporated into the translation methodology.

6. The method according to claim 1, further comprising evaluating translator feedback using a parallel corpus to validate the translator feedback.

7. The method according to claim 6, wherein the parallel corpus is selected by determining a domain associated with the translator and selecting a parallel corpus having a domain that is substantially similar to the domain of the translator.

8. The method according to claim 1, further comprising providing an accuracy rating for any of the translation unit, the correction unit, or combinations thereof.

9. A personalized machine translation system, the system comprising: a feedback processor that receives translator feedback regarding machine translations generated by a machine translation system for a translator, the machine translation system using a translation methodology;a feedback classifier that determines translator feedback that improves translations generated by the machine translation system, the translator feedback comprising a triple having a source unit, a translation unit of the source unit, and a correction unit;a dictionary pusher that incorporates the translator feedback into the translation methodology of the machine translation system to personalize the translation methodology; anda dictionary extractor that: constructs a first translation table that comprises pairs of source units and correction units;extracts dictionary entries from the translator feedback by: preparing the pairs for alignment,aligning words included in the pairs, andextracting first phrase pairs based upon word alignment;constructs a second translation table that comprises pairs of source units and translation units; andextracts dictionary entries from the translator feedback by: preparing the pairs for alignment,aligning words included in the pairs, andextracting second phrase pairs based upon word alignment; andfilters the first and second phrase pairs by: selecting first and second phrase pairs having a threshold amount of words included in the phrase,selecting first and second phrase pairs that have terminal words which are aligned, andstoring the selected first and second phrase pairs in a translation phrase table that resides on a data repository.

10. The system according to claim 9, wherein the translation methodology comprises any of corpus-based machine translation, rule-based machine translation, transfer-based machine translation, hybrid-based machine translation, statistic-based machine translation, a translation memory, a language model, a translation dictionary, a phrase table, or any combination thereof.

11. The system according to claim 9, wherein the feedback classifier further evaluates the translator feedback prior to the dictionary pusher incorporating the translator feedback into the translation methodology to determine if the translator feedback comprises an improved translation relative to a translation generated by the machine translation system without the translator feedback.

12. The system according to claim 9, wherein the feedback classifier further evaluates translator feedback using a parallel corpus to validate the translator feedback, wherein the parallel corpus is selected by determining a domain associated with the translator and selecting a parallel corpus having a domain that is substantially similar to the domain of the translator.

13. The system according to claim 9, wherein the translator feedback includes corrections that are received from the translator relative to a previous translation generated by the machine translation system.

14. The system according to claim 13, wherein the feedback processor further provides the corrections to a human translator and receives verified corrections from the human translator, wherein the verified corrections are incorporated into the translation methodology.

15. A method for managing a personalized machine translation system, the method comprising: receiving translator feedback regarding machine translations generated by a machine translation system for a translator, the machine translation system using a translation methodology;determining translator feedback that improves translations generated by the machine translation system;determining one or more translation preferences for the translator;incorporating the determined translator feedback and the translation preferences into the translation methodology of the machine translation system to personalize the translation methodology; andgenerating a tuning corpus by: calculating a weighted value for each feature of a feedback-based phrase pair,adding the weighted values together into an overall quality value, andincorporating the feedback pair into the tuning corpus if the overall quality value meets or exceeds a threshold value.

16. The method according to claim 15, wherein the one or more translation preferences comprises a domain for the translator.

17. The method according to claim 15, wherein the translation methodology further comprises feedback-based phrase pairs that include word alignments between source and correction sentence pairs.

18. The method according to claim 17, further comprising incorporating the feedback-based phrase pairs into a phrase pair database by analyzing features of the feedback-based phrase pairs, the features comprising properties that provide information indicative of a utility of the feedback-based phrase pairs.

19. The method according to claim 15, further comprising: selecting new weights for features values to improve overall quality of future translations;generating translations for translation requests; andcollecting new feedback-based translation pairs.

20. The method according to claim 15, further comprising determining if a feedback set is licensed, the feedback set comprising source text, a translation of the source text, and at least one correction for the translation, wherein the feedback set is licensed if the feedback set is included in an original machine translation of a sentence and in the at least one correction of the translation.

21. A method for managing a personalized machine translation system, the method comprising: receiving translator feedback regarding machine translations generated by a machine translation system for a translator, the machine translation system using a translation methodology comprising feedback-based phrase pairs that include word alignments between source and correction sentence pairs;determining translator feedback that improves translations generated by the machine translation system;determining one or more translation preferences for the translator;incorporating the determined translator feedback and the translation preferences into the translation methodology of the machine translation system to personalize the translation methodology;incorporating the feedback-based phrase pairs into a phrase pair database by analyzing features of the feedback-based phrase pairs, the features comprising properties that provide information indicative of a utility of the feedback-based phrase pairs; anddetermining features of a feedback-based phrase pair by simulating phrase pair statistics using any of accumulated statistics for low-frequency words or established default values.