Efficient online domain adaptation

Abstract

Systems and methods for efficient online domain adaptation are provided herein. Methods may include receiving a post-edited machine translated sentence pair, updating a machine translation model by adjusting translation weights for a translation memory and a language model while generating test machine translations of the machine translated sentence pair until one of the test machine translations approximately matches the post-edits for the machine translated sentence pair, and retranslating the remaining machine translation sentence pairs that have yet to be post-edited using the updated machine translation model.

Description

CROSS REFERENCE TO RELATED APPLICATIONS

This Non-Provisional patent application is related to U.S. patent application Ser. No. 13/685,372, titled “Personalized Machine Translation via Online Adaptation”, which was filed on Nov. 26, 2012, which is hereby incorporated by reference herein in its entirety including all references cited therein.

FIELD OF THE TECHNOLOGY

Embodiments of the disclosure relate to machine translation systems and methods. More specifically, but not by way of limitation, the present technology includes systems and methods that provide efficient online domain adaptation where updates to a machine translation system occur as soon as post-edits to machine translations are received by the machine translation system.

BACKGROUND OF THE DISCLOSURE

Post edit data, such as human translator feedback, may be created by human translators in order to correct a machine translation sentence pair. For example, a machine translation sentence pair may include a source sentence unit, such as a word or phrase, as well as a machine translation generated target sentence unit. If the target sentence unit generated by the machine translation system is incorrect, a human translator may generate post-edits that correct the error. While these post-edits are a valuable resource for customizing and adapting statistical machine translation models, updating the machine translation system with these post-edits remains a difficult endeavor.

SUMMARY OF THE DISCLOSURE

According to some embodiments, the present technology may be directed to a method of immediately updating a machine translation system with post-edits during translation of a document, using a machine translation system that comprises a processor and a memory for storing logic that is executed by the processor to perform the method, comprising: (a) receiving a post-edited machine translated sentence pair, wherein the post-edited machine translated sentence pair comprises a source sentence unit and a post-edited target sentence unit; (b) updating a machine translation model by: (i) performing an alignment of the post-edits of the machine translated sentence pair to generate phrases; and (ii) adding the phrases to the machine translation model; (c) adapting a language model from the target sentence unit of the post-edits; (d) calculating translation statistics for the post-edits; (e) adjusting translation weights using the translation statistics while generating test machine translations of the machine translated sentence pair until one of the test machine translations approximately matches the post-edits for the machine translated sentence pair; and (f) retranslating the remaining machine translation sentence pairs that have yet to be post-edited using the updated machine translation model and the adjusted translation weights.

According to other embodiments, the present technology may be directed to a machine translation system that immediately incorporates post-edits into a machine translation model during translation of a document, the machine translation system comprising: (1) a processor; and (2) a memory for storing logic that is executed by the processor to: (a) receiving a post-edited machine translated sentence pair, wherein the post-edited machine translated sentence pair comprises a source sentence unit and a post-edited target sentence unit; (b) updating a machine translation model by: (i) performing an alignment of the post-edits of the machine translated sentence pair to generate phrases; and (ii) adding the phrases to the machine translation model; (c) adapting a language model from the target sentence unit of the post-edits; (d) calculating translation statistics for the post-edits; (e) adjusting translation weights using the translation statistics while generating test machine translations of the machine translated sentence pair until one of the test machine translations approximately matches the post-edits for the machine translated sentence pair; and (f) retranslating the remaining machine translation sentence pairs that have yet to be post-edited using the updated machine translation model.

BRIEF DESCRIPTION OF THE DRAWINGS

The accompanying drawings, where like reference numerals refer to identical or functionally similar elements throughout the separate views, together with the detailed description below, are incorporated in and form part of the specification, and serve to further illustrate embodiments of concepts that include the claimed disclosure, and explain various principles and advantages of those embodiments.

The methods and systems disclosed herein have been represented where appropriate by conventional symbols in the drawings, showing only those specific details that are pertinent to understanding the embodiments of the present disclosure so as not to obscure the disclosure with details that will be readily apparent to those of ordinary skill in the art having the benefit of the description herein.

FIG. 1 is an exemplary computing architecture that may be used to practice aspects of the present technology;

FIG. 2 is an example of an algorithm utilized by a machine translation system to immediately update a machine translation system with post-edits during translation of a document;

FIG. 3 is an example of an algorithm utilized by a machine translation system to update a probability table used by the machine translation system.

FIG. 4 is a flowchart of a method for immediately updating a machine translation system with post-edits during translation of a document

FIG. 5 illustrates an exemplary computing system that may be used to implement embodiments according to the present technology.

DETAILED DESCRIPTION

In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of the disclosure. It will be apparent, however, to one skilled in the art, that the disclosure may be practiced without these specific details. In other instances, structures and devices are shown at block diagram form only in order to avoid obscuring the disclosure.

Reference throughout this specification to “one embodiment” or “an embodiment” means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the present invention. Thus, the appearances of the phrases “in one embodiment” or “in an embodiment” or “according to one embodiment” (or other phrases having similar import) at various places throughout this specification are not necessarily all referring to the same embodiment. Furthermore, the particular features, structures, or characteristics may be combined in any suitable manner in one or more embodiments. Furthermore, depending on the context of discussion herein, a singular term may include its plural forms and a plural term may include its singular form. Similarly, a hyphenated term (e.g., “on-demand”) may be occasionally interchangeably used with its non-hyphenated version (e.g., “on demand”), a capitalized entry (e.g., “Software”) may be interchangeably used with its non-capitalized version (e.g., “software”), a plural term may be indicated with or without an apostrophe (e.g., PE's or PEs), and an italicized term (e.g., “N+1”) may be interchangeably used with its non-italicized version (e.g., “N+1”). Such occasional interchangeable uses shall not be considered inconsistent with each other.

The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. As used herein, the singular forms “a”, “an” and “the” are intended to include the plural forms as well, unless the context clearly indicates otherwise. It will be further understood that the terms “comprises” and/or “comprising,” when used in this specification, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof.

It is noted at the outset that the terms “coupled,” “connected”, “connecting,” “electrically connected,” etc., are used interchangeably herein to generally refer to the condition of being electrically/electronically connected. Similarly, a first entity is considered to be in “communication” with a second entity (or entities) when the first entity electrically sends and/or receives (whether through wireline or wireless means) information signals (whether containing data information or non-data/control information) to the second entity regardless of the type (analog or digital) of those signals. It is further noted that various figures (including component diagrams) shown and discussed herein are for illustrative purpose only, and are not drawn to scale.

The present technology provides for fast online updates to a machine translation system that uses language pairs, immediately after the machine translation system receives a new sentence pair. For example, the machine translation system may receive a new sentence pair that includes a source sentence unit and a post-edited target sentence unit. The post-edited target sentence unit includes changes to an original target sentence unit that was provided to a translator.

For context, the phrase “sentence unit” may be understood to include any sub-sentential unit that is obtained from a sentence in a document that is to be translated. In general, a machine translation system receives a document for translation and uses machine translation techniques to translate the document in a source language into a target language. This machine translation process includes breaking the source document into sentences and further dividing the sentences into sub-sentential units. Using methods that would be known to one of ordinary skill in the art, the machine translation system outputs sentence pairs for the sentence that include source sentence units and target sentence units that are machine translations of the source sentence units.

These sentence pairs are evaluated by human translators who may edit or change the target sentence unit of a sentence pair to correct errors in the machine translation. These changes are referred to as post-edits of target sentence units.

New information obtained from the post-edited target sentence unit is incorporated into the statistical models of the language pairs and may be passed to the decoder. With a fast online tuning method, the translation model of the present technology may utilize this new information in subsequent translations. This method yields significant bilingual evaluation understudy (BLEU) improvements over both small-scale and strong baselines when translating coherent in-domain sentences.

As opposed to adding new data to a static translation memory, the present technology may employ a translation model update procedure to modify relevant parts of probabilistic models of the translation model in order to update machine translations of subsequent sentences of a document.

FIG. 1 illustrates an exemplary architecture, which implements a methodology for rapidly adapting a machine translation system such that it can better anticipate the behavior of a human post-editor who is tasked with converting automatic translations of a source language text into high-quality translations in a desired target language. The architecture may include a plurality of clients 105A-C that are coupled with both a generic machine translation (MT) engine 115 and a machine translation system 100. Generally, the generic MT engine 115 initially generates machine translation pairs for a document received from one of the clients. The generic MT engine 115 utilizes a translation model (TM) 115A, a language model (LM) 115B and generic weights 115C that affect how the generic MT engine 115 translates the document received from the client, by affecting application of the translation model 115A and the language model 115B.

The machine translation system 100 is configured to utilize post-edits received from the clients to generate in-domain adapted and retuned machine translation models. For example, the machine translation system 100 generates both an in-domain translation model (IDTM) 100A, an in-domain language model (IDLM) 100B, and in-domain weighting (IDW) 100C. In some embodiments, each client may have their own IDTM, IDLM, and IDW due to the machine translation system's ability to re-translate remaining sentence pairs for a translator as post-edits from the translator are received by the machine translation system 100.

Thus, in some embodiments, the IDTM, IDLM, and IDW are invoked and utilized by the machine translation system 100 during a single translation session for a translator. In other embodiments, the IDTM, IDLM, and IDW may persist and can be utilized for multiple translation projects.

The following description provides an example of the use of the machine translation system 100 by a client, such as client 105A. Initially the post-editor associated with client 105A receives a source language text and automatic translations in the target language generated by a baseline machine translation engine 115. The source language text and automatic translations are also referred to as a sentence pair. The sentence pair includes a source sentence unit and a target sentence unit that is a machine translation of the source sentence unit.

Changes to the target sentence unit by the client 105A are referred to as a post-edited target sentence unit. As the post-editor corrects a target sentence unit, the post-edits are transmitted to the machine translation system 100. The machine translation system 100 then automatically learns new sub-sentential translation correspondences and enhances the generic MT engine 115 with these correspondences. The machine translation system 100 also adjusts the parameters of the generic MT engine 115 to optimize translation performance on the corrected translations. The result is a generic MT engine 115 that is better equipped to handle the vocabulary and phrasing choices desired by the post-editor on his/her current workflow.

FIG. 1 illustrates a server-side instantiation of the machine translation system 100 that continuously updates translations and personalizes on a per-user, per-document basis. It will be understood that many other variants of this approach are possible using the same technology. For instance, the updating may be done client side, the updating may be periodic instead of continuous, and the scope of personalization may be wider or narrower than per-user, per-document.

The translation model 115A and language model 115B are typically very large databases, where the weights are a vector of numbers. The machine translation system 100 has the ability to instantiate and reset the per-user, and in-domain translation models IDTM 100A, in-domain language models IDLM 100B, and in-domain weighting IDW 100C. It will be understood that IDTM 100A, IDLM 100B, and IDW 100C may belong to client 105A and are illustrated as a set of personalized machine translation tools 100D. Indeed, each client is provided with their own IDTM, IDLM, and IDW. Upon instantiation and reset the IDTM 100A and IDLM 100B are empty databases and the IDW 100C is equal to the generic weights 115C used by the generic MT engine 115. The IDTM 100A and IDLM 100B are typically very small databases compared to the generic TM 115A and LM 115B.

In some embodiments, the clients 105A-C communicate with the machine translation system 100 via a REST API over a network 110 such as the Internet. The API responds to requests for translation keyed to a user, requests to update the IDTM 100A, the IDLM 100B, and the IDW 100C, and requests to reset a per-user IDTM 100A, the IDLM 100B, and the IDW 100C.

When a client 105A begins to post-edit a document, the client 105A requests that the machine translation system 100 instantiate or reset the per-user IDTM 100A, IDLM 100B, and IDW 100C, and requests translations for the document segments, which are carried out initially by the generic MT engine 115. The generic MT engine 115 responds to the client request with a set of machine translated sentence pairs.

When a user post-edits a translation, and specifically a target sentence unit, the client 105A requests that the machine translation system 100 update the IDTM 100A, IDLM 100B, and IDW 100C for the user with new translational correspondences, target language phrases, and parameter weights to remember the post-edit corrections made by the user. The client 105A then requests re-translation of the remaining sentence pairs by the generic MT engine 115. The generic MT engine 115 may then re-translate the documents consulting the user's IDTM 100A, IDLM 100B, and IDW 100C via calls to the machine translation system 100, and also using generic models where appropriate. In some embodiments, once translation of a document is completed and a new document translation process begins, the client 105A may request a reset of their IDTM 100A, IDLM 100B, and IDW 100C.

In some embodiments, the machine translation system may execute test machine translations of the original sentence pairs and adjust the IDW 100C until the test machine translations generated by the generic MT engine 115 approximately match the post edit sentence pair received from the client 105A, as determined by BLEU. That is, once the machine translation system 100 populates the translation model with words and phrases from forced alignment of the post-edit sentence pair and creates the IDLM 100B, the machine translation system 100 may iteratively adjust the components of the IDW 100C until the translations generated by the generic MT engine 115 approximate the translation quality of the human translator.

While the above embodiments contemplate a separate generic MT engine 115 and machine translation system 100, it will be understood that the functionalities of these systems may be combined into a single machine translation system. Further, the functionalities of these systems may be executed on the client device 105A, rather than the client 105A interacting with the generic MT engine 115 and the machine translation system 100 over a network 110.

The following description provides details regarding the processes used by the machine translation system 100 to create and utilize the IDTM 100A, IDLM 100B, and IDW 100C for a client. The machine translation system 100 provides both a machine translation model adaptation process and a retuning of parameter weights. FIG. 2 illustrates an exemplary algorithm that includes both the machine translation model adaptation and weight retuning processes. Generally, new information obtained from the post-edits of machine translated sentence pairs are added to the model by updating vocabularies, databases, and phrase tables. Next, the algorithm adjusts existing translation model weights to encourage usage of new in-domain phrases using an online discriminative ridge regression technique.

The following variables are defined for purposes of clarity: Vs is a vocabulary of source words; Vt is a vocabulary of target words encountered by the machine translation system; (ŝ,{circumflex over (t)}) a post-edited machine translated sentence pair that is used to adapt a translation model M (e.g., IDTM 100A); and (s⁻, t⁻) is a phrase pair generated from the post-edited machine translated sentence pair.

The following steps are performed in sequence with parallelism where possible. First, (ŝ,{circumflex over (t)}) is tokenized and modified into lowercase. Next, Vt and Vs are updated with unknown words from (ŝ,{circumflex over (t)}). Subsequently, the machine translation system 100 uses use existing, static alignment models, which are pre-trained on original training data in both directions, to “force align” (ŝ,{circumflex over (t)}) and run regular alignment refinement heuristics to produce a word alignment for (ŝ,{circumflex over (t)}).

Next, the machine translation system 100 builds a small in-domain language model (IDLM 100B) during adaptation of the machine translation model. For each (ŝ,{circumflex over (t)}), the machine translation system performs an ngram-count to update an existing count file, also the machine translation system 100 recompiles the in-domain language model (IDLM 100B) using a smoothing algorithm, such as Witten-Bell smoothing. Since the amount of in-domain data is fairly small, the machine translation system 100 re-builds the IDLM 100B quickly and efficiently.

Next, the machine translation system 100 updates a fractional count table by extracting fractional counts from (ŝ,{circumflex over (t)}) and adds these new counts to the fractional count table. In some embodiments the machine translation system 100 extracts and filters lexicon entries from (ŝ,{circumflex over (t)}) and its alignments. An exemplary algorithm for building the fractional count table includes, but is not limited to, the IBM Model 4 Table, M4={cf(tj|si)}|Vs|×|Vt|.

After updating the fraction count table, the machine translation system 100 then updates a probability distribution table to change the distributions for each source term s_iε(ŝ,{circumflex over (t)}). In some embodiments, this includes the determination of maximum likelihood word alignments from (ŝ,{circumflex over (t)}) and its alignments.

To avoid dumping M1 databases to disk in text format or storing counts and re-normalizing, the machine translation system 100 may perform a heuristic update. Assuming that Vs and Vt were already updated, the machine translation system 100 uses the following equation M1_{ŝ,{circumflex over (t)}}={{circumflex over (p)}(tj|si)}Vs×Vt to determine Viterbi alignments extracted from (ŝ,{circumflex over (t)}). In some embodiments, the machine translation system 100 may utilize IBM Model 1 Tables algorithm (M1={p(tj|si)}|Vs|×|Vt|)

Also, FIG. 3 illustrates an example of a heuristic algorithm that may be utilized to alter the probability distributions for the translation model.

Next, the machine translation system 100 may be configured to adapt a phrase table used to generate phrases from sentence pairs. To obtain a set of new phrases P={(e⁻,f⁻)}₁, the machine translation system 100 executes a phrase extraction pipeline, which includes forward/inverse phrase extraction and sorting.

The process of adapting the phrase table used by the machine translation system 100 may include the pre-computing of various phrase features for the phrases (e⁻, f⁻) that were extracted from the post-edited machine translation sentence unit. Various features that are computed by the machine translation system include “missing word,” for which the machine translation system 100 uses the previously updated IBM Model 4 table as described above. Also, the machine translation system 100 may use an “inverse IBM Model 1” feature which employs the previously (heuristically) updated IBM Model 1 tables. Another exemplary feature includes “inverse phrase probability”, which is computed using counts of inversely extracted phrases.

The machine translation system 100 is also configured to execute a process of retuning of model weights. The machine translation system 100 is configured to adjust model weights to encourage the inclusion of new phrases and adaptation to the target domain. In some instances, the machine translation system 100 may adjust translation model weights using a tuning method such as discriminative ridge regression (DRR), although other translation model tuning methods known to one of ordinary skill in the art may also likewise be utilized.

The discriminative ridge regression method includes the determination of an nbest(ŝ) list for the re-decode of a source sentence ŝ, ordered by decreasing derivation scores. The machine translation system 100 builds an n×m matrix R_ŝ that contains the difference vectors between each feature vector h_ŝⁱ and h_ŝ*, the feature vector for the best hypothesis in terms of BLEU+1 to reference {circumflex over (t)}.

The goal of this process is to allow the machine translation system 100 to find a vector w such that R_ŝ·w∝I_ŝ, where I_ŝ is a column vector of n rows containing the difference in BLEU+1 scores for each h_ŝⁱ from h_ŝ*. This may be expressed by the equation

$w=\underset{w}{\arg \min }{R_{\hat{s}}·w-I_{\hat{s}}}^2,$ which is a regression problem with an exact solution w=(R′_ŝ·R_ŝ+βI)⁻¹R′_ŝ·I, where β=0.01 is the regularization parameter that stabilizes R′_ŝ·R_ŝ.

The solution w^t for sentence pair (ŝ,{circumflex over (t)})^t at update time t is interpolated with the previous weight vector w^t−1 and used for re-translating of sentence s^t+1 in the next iteration (e.g., remaining sentence pairs that have yet to be post-edited by the client 105A. In accordance with the present technology, the machine translation system 100 is configured to initially generate a set of machine translation pairs. After receiving a post edit of one of these machine translation pairs in the set, the machine translation system 100 executes the adaptation and tuning methods described herein to update the translation model/methodologies used by the machine translation system 100. Once updated, the machine translation system 100 retranslates any remaining machine translations that have yet to be translated from the set. Each time post-edits are received the machine translation system 100 updates and retranslates any remaining machine translations. Thus, the machine translation system 100 may, for each post-edit received, iteratively update the translation model and retranslate any previous machine translations, assuming the machine translations have not been post-edited by a human translator.

In some instances the interpolation weight is set to 0.5. This allows the machine translation system 100 to utilize a combination between the generic weights 115C of the generic MT engine 115 and the IDW 100C. The machine translation system 100 uses the interpolation weight of 0.5 and multiplies both the generic weights 115C and the IDW 100C by 0.5. The machine translation system 100 then utilizes an average of these values. The use of interpolation ensures moderation of large discrepancies between the generic weights 115C and the IDW 100C, which may lead to poor translations for the client.

FIG. 4 is a flowchart of an exemplary method for immediately updating a machine translation system with post-edits during translation of a document. The method is executed by a machine translation system 100, which is configured to execute the method. In some instances, the method includes receiving 405 a post-edited machine translated sentence pair. As mentioned above, the post-edited machine translated sentence pair comprises a source sentence unit and a post-edited target sentence unit. Again, this post-edited machine translated sentence pair includes a machine translation sentence pair that has been post-edited by a human translator in order to alter or modify the target sentence unit that was generated by a generic MT engine.

Next, the method includes updating 410 a machine translation model by performing an alignment of the post-edits of the machine translated sentence pair to generate phrases, and adding the phrases to the machine translation model.

In some embodiments, the method includes adapting 415 a language model from the target sentence unit of the post-edits, as well as calculating 420 translation statistics for the post-edits.

Next, the method includes adjusting 425 translation weights using the translation statistics while generating test machine translations of the machine translated sentence pair until one of the test machine translations approximately matches the post-edits for the machine translated sentence pair. Finally, the method includes retranslating 430 the remaining machine translation sentence pairs that have yet to be post-edited using the updated machine translation model. In some embodiments, retranslated machine translation sentence pair may be post-edited by a human translator. Thus, the method may return to step 405 when a poste-edit to a retranslated machine translation sentence pair, which would result in incremental retranslation of the machine translation sentence pairs, even those pairs that have been retranslated one or more times. Advantageously, each time remaining sentence pairs are retranslated it is assumed that the retranslations will cause the sentence pairs to require a lesser amount of post-editing, or no post-editing at all.

FIG. 5 illustrates an exemplary computing device (also referred to as computing system or system) 1 that may be used to implement an embodiment of the present systems and methods. The system 1 of FIG. 5 may be implemented in the contexts of the likes of computing devices, radios, terminals, networks, servers, or combinations thereof. The computing device 1 of FIG. 5 includes a processor 10 and main memory 20. Main memory 20 stores, in part, instructions and data for execution by processor 10. Main memory 20 may store the executable code when in operation. The system 1 of FIG. 5 further includes a mass storage device 30, portable storage device 40, output devices 50, user input devices 60, a display system 70, and peripherals 80.

The components shown in FIG. 5 are depicted as being connected via a single bus 90. The components may be connected through one or more data transport means. Processor 10 and main memory 20 may be connected via a local microprocessor bus, and the mass storage device 30, peripherals 80, portable storage device 40, and display system 70 may be connected via one or more input/output (I/O) buses.

Mass storage device 30, 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 10. Mass storage device 30 can store the system software for implementing embodiments of the present technology for purposes of loading that software into main memory 20.

Portable storage device 40 operates in conjunction with a portable non-volatile storage medium, such as a floppy disk, compact disk or digital video disc, to input and output data and code to and from the computing system 1 of FIG. 5. The system software for implementing embodiments of the present technology may be stored on such a portable medium and input to the computing system 1 via the portable storage device 40.

Input devices 60 provide a portion of a user interface. Input devices 60 may include an alphanumeric keypad, such as a keyboard, for inputting alphanumeric and other information, or a pointing device, such as a mouse, a trackball, stylus, or cursor direction keys. Additionally, the system 1 as shown in FIG. 5 includes output devices 50. Suitable output devices include speakers, printers, network interfaces, and monitors.

Display system 70 may include a liquid crystal display (LCD) or other suitable display device. Display system 70 receives textual and graphical information, and processes the information for output to the display device.

Peripherals 80 may include any type of computer support device to add additional functionality to the computing system. Peripherals 80 may include a modem or a router.

The components contained in the computing system 1 of FIG. 5 are those typically found in computing 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 computing system 1 can be a personal computer, hand held computing system, telephone, mobile computing system, workstation, server, minicomputer, mainframe computer, or any other computing system. The computer can also include different bus configurations, networked platforms, multi-processor platforms, etc. Various operating systems can be used including UNIX, Linux, Windows, Macintosh OS, Chrome OS, and other suitable operating systems.

Some of the above-described functions may be composed of instructions that are stored on storage media (e.g., computer-readable medium). The instructions may be retrieved and executed by the processor. Some examples of storage media are memory devices, tapes, disks, and the like. The instructions are operational when executed by the processor to direct the processor to operate in accord with the technology. Those skilled in the art are familiar with instructions, processor(s), and storage media.

It is noteworthy that any hardware platform suitable for performing the processing described herein is suitable for use with the technology. The terms “computer-readable storage medium” and “computer-readable storage media” as used herein refer to any medium or media that participate in providing instructions to a CPU for execution. Such media can take many forms, including, but not limited to, non-volatile media, volatile media and transmission media. Non-volatile media include, for example, optical or magnetic disks, such as a fixed disk. Volatile media include dynamic memory, such as system RAM. Transmission media include coaxial cables, copper wire and fiber optics, among others, including the wires that comprise one embodiment of a bus. Transmission media can also take the form of acoustic or light waves, such as those generated during radio frequency (RF) and infrared (IR) data communications. Common forms of computer-readable media include, for example, a floppy disk, a flexible disk, a hard disk, magnetic tape, any other magnetic medium, a CD-ROM disk, digital video disk (DVD), any other optical medium, any other physical medium with patterns of marks or holes, a RAM, a PROM, an EPROM, an EEPROM, a FLASHEPROM, any other memory chip or data exchange adapter, a carrier wave, or any other medium from which a computer can read.

Various forms of computer-readable media may be involved in carrying one or more sequences of one or more instructions to a CPU for execution. A bus carries the data to system RAM, from which a CPU retrieves and executes the instructions. The instructions received by system RAM can optionally be stored on a fixed disk either before or after execution by a CPU.

Computer program code for carrying out operations for aspects of the present technology may be written in any combination of one or more programming languages, including an object oriented programming language such as Java, Smalltalk, C++ or the like and conventional procedural programming languages, such as the “C” programming language or similar programming languages. The program code may execute entirely on the user's computer, partly on the user's computer, as a stand-alone software package, partly on the user's computer and partly on a remote computer or entirely on the remote computer or server. In the latter scenario, the remote computer may be connected to the user's computer through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection may be made to an external computer (for example, through the Internet using an Internet Service Provider).

The corresponding structures, materials, acts, and equivalents of all means or step plus function elements in the claims below are intended to include any structure, material, or act for performing the function in combination with other claimed elements as specifically claimed. The description of the present technology has been presented for purposes of illustration and description, but is not intended to be exhaustive or limited to the invention in the form disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the invention. Exemplary embodiments were chosen and described in order to best explain the principles of the present technology and its practical application, and to enable others of ordinary skill in the art to understand the invention for various embodiments with various modifications as are suited to the particular use contemplated.

Aspects of the present technology are described above with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and/or block diagrams, and combinations of blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer readable medium that can direct a computer, other programmable data processing apparatus, or other devices to function in a particular manner, such that the instructions stored in the computer readable medium produce an article of manufacture including instructions which implement the function/act specified in the flowchart and/or block diagram block or blocks.

The computer program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other devices to cause a series of operational steps to be performed on the computer, other programmable apparatus or other devices to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide processes for implementing the functions/acts specified in the flowchart and/or block diagram block or blocks.

The flowchart and block diagrams in the Figures illustrate the architecture, functionality, and operation of possible implementations of systems, methods and computer program products according to various embodiments of the present technology. In this regard, each block in the flowchart or block diagrams may represent a module, segment, or portion of code, which comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that, in some alternative implementations, the functions noted in the block may occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently, or the blocks may sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and/or flowchart illustration, and combinations of blocks in the block diagrams and/or flowchart illustration, can be implemented by special purpose hardware-based systems that perform the specified functions or acts, or combinations of special purpose hardware and computer instructions.

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 of immediately updating a machine translation system with post-edits during translation of a document, using a machine translation system that comprises a processor and a memory for storing logic that is executed by the processor to perform the method, comprising: receiving a post-edited machine translated sentence pair, the post-edited machine translated sentence pair comprising a source sentence unit and a post-edited target sentence unit;updating a machine translation model by: performing an alignment of the post-edits of the machine translated sentence pair to generate phrases; andadding the phrases to the machine translation model;adapting a language model from the post-edited target sentence unit;calculating translation statistics for the post-edits;adjusting translation weights using the translation statistics while generating test machine translations of the machine translated sentence pair until one of the test machine translations approximately matches the post-edits for the machine translated sentence pair; andretranslating remaining machine translation sentence pairs that have yet to be post-edited using the updated machine translation model and the adjusted translation weights.

2. The method according to claim 1, further comprising: receiving a document for translation from a source language into a target language; andperforming a machine translation of the document to generate a set of machine translated sentence pairs.

3. The method according to claim 1, wherein the updating of the machine translation model further comprises tokenizing the post-edits of the machine translated sentence pair.

4. The method according to claim 3, wherein the updating of the machine translation model further comprises updating a vocabulary of source sentence units with unknown source sentence units included in the post-edits of the machine translated sentence pair and updating a vocabulary of target sentence units with unknown target sentence units included in the post-edits of the machine translated sentence pair.

5. The method according to claim 4, further comprising extracting fractional counts from the post-edits of the machine translated sentence pair and adding the extracted fractional counts to a fractional count table.

6. The method according to claim 5, further comprising adjusting probability distributions for the source sentence unit of the post-edits of the machine translated sentence pair.

7. The method according to claim 1, further comprising updating a phrase table with counts that define a number of occurrences of phrases in the phrase table; and reordering feature values for the phrases in the phrase table based upon the counts.

8. The method according to claim 1, wherein the translation weights for the machine translation system are adjusted using discriminative ridge regression.

9. The method according to claim 1, wherein the adapting of the language model includes executing an ngram-count of the post-edited machine translated sentence pair to update a count file that comprises counts for input sentence pairs; and recompiling the language model using a smoothing algorithm.

10. The method according to claim 1, wherein the alignment includes both forward and reverse alignments of the post-edited machine translated sentence pair.

11. A machine translation system that immediately incorporates post-edits into a machine translation model during translation of a document, the machine translation system comprising: a processor; anda memory for storing logic that is executed by the processor to: receive a post-edit of a target sentence unit of a machine translated sentence pair of a set of machine translated sentence pairs, wherein a machine translated sentence pair comprises a source sentence unit and the target sentence unit;receive a post-edited machine translated sentence pair, the post-edited machine translated sentence pair comprising a post-edited source sentence unit and a post-edited target sentence unit;update the machine translation model by: performing an alignment of the post-edits of the machine translated sentence pair to generate phrases; andadding the phrases to the machine translation model;adapt a language model from the post-edited target sentence unit;calculate translation statistics for the post-edits;adjust translation weights using the translation statistics while generating test machine translations of the machine translated sentence pair until one of the test machine translations approximately matches the post-edits for the machine translated sentence pair; andretranslate remaining machine translation sentence pairs that have yet to be post-edited using the updated machine translation model.

12. The machine translation system according to claim 11, wherein the processor further executes the logic to: receive the document for translation from a source language into a target language; andperform a machine translation of the document to generate the set of machine translated sentence pairs.

13. The machine translation system according to claim 11, wherein the machine translation system updates the machine translation model by further tokenizing the post-edits of the machine translated sentence pair.

14. The machine translation system according to claim 13, wherein the machine translation system updates the machine translation model by updating a vocabulary of source sentence units with unknown source sentence units included in the post-edits of the machine translated sentence pair and updating a vocabulary of target sentence units with unknown target sentence units included in the post-edits of the machine translated sentence pair.

15. The machine translation system according to claim 14, wherein the processor further executes the logic to extract fractional counts from the post-edits of the machine translated sentence pair and adding the extracted fractional counts to a fractional count table.

16. The machine translation system according to claim 15, wherein the processor further executes the logic to adjust probability distributions for the source sentence unit of the post-edits of the machine translated sentence pair.

17. The machine translation system according to claim 11, wherein the processor further executes the logic to update a phrase table with counts that define a number of occurrences of phrases in the phrase table; and reorder feature values for the phrases in the phrase table based upon the counts.

18. The machine translation system according to claim 11, wherein the translation weights for the machine translation system are adjusted using discriminative ridge regression.

19. The machine translation system according to claim 11, wherein the machine translation system is configured to adapt the language model by executing an ngram-count of the post-edited machine translated sentence pair to update a count file that comprises counts for input sentence pairs; and recompiling the language model using a smoothing algorithm.

20. The machine translation system according to claim 11, wherein the machine translation system is configured to: receive post-edits for a retranslated machine translation sentence pair;re-updat the machine translation model;calculate translation statistics for the post-edits of the retranslated machine translation sentence pair;adjust translation weights using the translation statistics while generating test machine translations of the retranslated machine translated sentence pair until one of the test machine translations approximately matches the post-edits for the retranslated machine translation sentence pair; andretranslate any remaining retranslated machine translation sentence pairs that have yet to be post-edited using the updated machine translation model and the translation weights.