Translation of annotated text

Abstract

A method for providing a vehicle with a command system that responds to natural-language commands in a target language includes receiving source training-data used for training in a source language, generating target training-data based on the source training-data, and using that target training-data to train the command system in the target language. Generating the target-training data includes generating rewritten source-sentences having concept tags derived from source annotations and using a translation model to translate these into corresponding target sentences, thus generating pairs of rewritten source-sentence and a target sentences. A first subset of these pairs is used to fine tune the translator model. The fine-tuned model then translates the pairs in a second subset of pairs to generate the target training-data based on the source training-data.

Description

TRANSLATION OF ANNOTATED TEXTCROSS-REFERENCE TO RELATED APPLICATIONS[00011 This application claims the benefit of U.S. provisional application Serial No. 63 / 676,542 filed July 29, 2024, the disclosure of which is hereby incorporated in its entirety by reference herein.TECHNICAL FIELD

[0002] Disclosed herein are systems and methods for translation of annotated text.BACKGROUND

[0003] Many vehicles feature a command system that accepts commands in the form of spoken utterances. Examples of such commands include navigation instructions (e.g., “Find me the nearest pharmacy.”), instructions to operate intra-vehicular machinery (e.g., “Make it two degrees cooler in here,”) and requests for information (e.g., “When will the airport tunnel re-open?”).

[0004] The process of causing such a command system to interpret spoken commands typically includes the use of a natural-language-understanding system. Such systems generally include machine-learning components that have to be trained. This requires considerable amounts of annotated training data. Annotated training data includes sentences that have been annotated in a way that helps a natural-language-understanding system learn to interpret language. The task of collecting and annotating such training data is burdensome.

[0005] Training data is often language specific. Training data in English is generally not all that useful for training a natural-language system to understand Urdu. Therefore, one who wishes to develop a command system for a different language would normally begin by creating new annotated training data for that language. This is quite burdensome. As a result, when a language has only a limited number of speakers, it is often not worth the effort to create a command systemthat responds to commands in that language. This creates an impediment to the widespread adoption of such systems.(0006] If training data for a first language already exists, it is tempting to simply use machine translation to translate that training data into a second language. After all, through the use of deep learning machine translation models, it is possible nowadays to create reasonably good translations more or less automatically.

[0007] Nevertheless, a problem remains: the annotations. Training a natural-language- understanding system requires more than just samples of sentences. It also requires the annotations. While machine translation is effective in translating a sentence from one language into another, it has no basis for also annotating the resulting translation.SUMMARY

[0008] Disclosed herein is a system and method including a way to leverage training data that was used to train a command system in a source language so that it can be used to train that command system to receive and execute commands in a target language. It does so by jointly translating annotated source sentences from the source language into corresponding annotated target sentences in the target language in such a way that the annotations in the source sentence are adjusted to correspond to the correct locations in the target sentence. These annotated target sentences can then be used to train the natural-language-understanding machinery for use with the target language.

[0009] The training data begins with a set of source sentences. These source sentences are in the source language. Each source sentence has a beginning and an end. A reader who reads this source sentence accumulates pieces of information along the way. By the end of the sentence, the reader is able to define the “semantic signature,” i.e., the “meaning” of the source sentence.

[0010] The information carried by a sentence is distributed along the sentence. Each piece of information pertains to a particular context. As an example, a sentence such as “Find me the nearest hotel” comprises a “concept sequence” having three concepts: an action (“Find me”), a categoryof destination (“hotel”), and a relative location (“nearest”). The concepts in a concept sequence are thus arranged in a particular concept order. That concept order depends in large part on the language.10011] A natural -language understanding machine that has been trained on a first language reads a sentence and thus collects information about the concepts within the sentence. It then uses stacked attention layers to relate each concept to a particular location within the sentence. Based on this information, the machine is able to define an overall “semantic signature,” i.e., the “meaning” of the source sentence.

[0012] A translated version of the source sentence, i.e., the target sentence, will comprise the same semantic signature. Its concept sequence will have the same concepts. However, the concept order may differ.

[0013] In the above example, a French translation of “Find me the nearest hotel” would be “Trouve-moi I’hotel le plus proche.” As is apparent, while the same concepts are present, they have been permuted. In particular, the concept related to “relative location” and that related to “category of destination” have been swapped.

[0014] A variety of metrics are available for determining an extent to which two concept orders differ. For each such metric, one can define a threshold. If the difference falls below that threshold, the source and translated sentences are said to have “translational equivalence.” Otherwise, they are said to lack “translational equivalence.”

[0015] As noted above, a sentence that is used as training data for a natural-language- understanding model includes annotations. These annotations identify where the sentence introduces particular concepts. Thus, in the sentence, “Find me the nearest hotel,” an annotation such as <CATEGORY> would point to the end of the English sentence and an annotation for <RELATIVE LOCATION> would point to a location just before the end of the English sentence. These locations would be swapped in the French translation.

[0016] The method contemplates training a translator model (hereafter referred to as a “trainee model”) using source sentences that have been rewritten to introduce tags into the correct placesin the source sentence. These tags define the locations of the concepts. For instance, the sentence “Find me the nearest hotel” might be rewritten as “<ACTION>Find<NULL> me<NULL>the<RELATIVE-LOCATION>nearest<CATEGORY>hotel.” In this example, words that do not assist in defining a concept are tagged as defining a null concept.[00171 As far as the trainee model is concerned, these tags are merely words that are part of the language. The trainee model will therefore observe, during its training period, that these tags often accompany particular words. As a result, the trainee model will learn to carry along those tags when translating the source sentence into a target sentence. For example, when trained with annotated data, the model is likely to believe that <CATEGORY> tends to be near “hotel.” Therefore, when translating “<CATEGORY>hotel,” the model is likely to return “<CATEGORY>h6tel .” As a result, the trainee model will inevitably develop the habit of including the concept tag close to the translated version of whatever was adjacent to the concept tag.

[0018] Another approach is to cluster the tags together in the source sentence and to have tags in the target sentence that point to the relevant tags in the source sentence. In this case, the sentence “Find me the nearest hotel” would be annotated as“Find me the nearest hotel<ACTION><NULL><RELATIVE- LOCATION><CATEGORY>” and the corresponding target sentence would be annotated as“Trouve-moi I’hotel le plus proche<l><2><4><3>.”

[0019] In effect, one can view the trainee model as one that has been trained to translate a peculiar English dialect, i.e., one in which the native speakers routinely tag words as they speak, into an equally peculiar French dialect having the same property.

[0020] Using this method, it becomes possible to develop a set of pairs. Each element in the set of pairs consists of two sentences with a common semantic signature: an annotated source sentence and an annotated translation thereof. A data miner then leverages multilingual sentence embeddings to identify those pairs that have properties that are most suitable for use as trainingdata for the target command system. In particular, the data miner separates those pairs that have translational equivalence from those pairs that do not. Those pairs that have both translational equivalence and a common semantic signature are then used as training data for training a model, which can then be used by the command system to operate in the target language.[00211 In one aspect, the disclosure herein features a method for providing a vehicle with a command system that responds to natural language commands in a target language by receiving source training-data that includes source sentences and corresponding source annotations. This source training-data is used for training a natural-language system in a source language. The method continues with generating target training-data based on the source training-data and using that target training-data to train the command system in the target language.[00221 The process of generating the target-training data comprises several steps. These include generating rewritten source-sentences that incorporate concept tags that are derived from the source annotations and using a translation model to translate the rewritten source-sentences into corresponding target sentences. This generates a set of pairs, each of which includes a rewritten source-sentence and a corresponding one of the target sentences.[0023| The resulting set of pairs can be divided into two subsets: a first subset of pairs and a second subset of pairs. The first subset is to be used for fine tuning the translator model. Having been thus fine-tuned, the translator model is let loose on the second subset. Accordingly, the finetuned translator model translates the rewritten source-sentences in the second subset. This results in the target training-data. Since this target-training data is essentially a translation of the sourcetraining data, the method described herein leverages the effort used to construct the source-training data to reduce the effort that would otherwise have been required to generate training data from scratch in a target language.[0024J The process then continues with using a first subset of the set of pairs to fine tune the translator model, thereby generating a fine-tuned translator model, and using the fine-tuned translator model to translate the rewritten source-sentences in a second subset of the pairs, thereby generating the target training-data based on the source training-data.

[0025] Different languages tend to have different volumes of training data available. Those with abundant training data are referred to as “high-resource languages.” Those that are poor in training data are referred to as “low-resource languages.” As a result of this inequality, models trained in a high-resource language tend to perform differently models trained in a low-resource language.

[0026] However, a useful byproduct of the method described herein is that the foregoing inequality will tend to diminish over time. This is because in the method described herein, the volume of target training-data will tend to be about the same as the volume of source training-data. As a result, if a particular language enjoys an abundance of training data, that abundance will spread to those languages that would otherwise have only a paucity of training data.

[0027] These and other features of the disclosure will be apparent from the following detailed description and the accompanying figures, in which:

[0028] In one aspect, the disclosure herein features a method for providing a command system in a vehicle. The command system in question is one that responds to natural language commands in a target language. Such a method includes receiving source training-data for training a naturallanguage system in a source language. This source training-data comprises source sentences and corresponding source annotations. The method continues with generating target training-data based on the source training-data and then using the target training-data to train the command system in the target language. The step of generating such target training-data includes generating rewritten source-sentences, the rewritten source-sentences incorporating concept tags that are derived from the source annotations; using a translation model, translating the rewritten sourcesentences into corresponding target sentences, thereby generating a set of pairs, each of the pairs including a rewritten source-sentence and a corresponding one of the target sentences; using a first subset of the set of pairs to fine tune the translator model, thereby generating a fine-tuned translator model; and using the fine-tuned translator model to translate the rewritten source-sentences in a second subset of the pairs, thereby generating the target training-data based on the source trainingdata.

[0029] Among the practices of the method are those that further include determining that a first pair’s constituent rewritten source-sentence and corresponding one of the target sentences havetranslational equivalence; determining that a second pair’s constituent rewritten source- sentence and corresponding one of the target sentences lack translational equivalence; placing the first pair in the first subset; and placing the second pair in the second subset.

[0030] Also among the practices of the method are those in which wherein generating the rewritten source-sentences comprises selecting a first source sentence and inserting the concept tags adjacent to each other in the first source sentence and those in which generating the rewritten source-sentences comprises selecting a first source sentence and inserting the concept tags between portions of the first source sentence.

[0031] In some practices, a pair from the set of pairs includes a rewritten source-sentence and a corresponding one of the target sentences. The rewritten source- sentence comprises a first set of concept tags and the corresponding one of the target sentences comprises a second set of concept tags. Each concept tag in the second set points to a corresponding one of the concept tags in the first set.BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The embodiments of the present disclosure are pointed out with particularity in the appended claims. However, other features of the various embodiments will become more apparent and will be best understood by referring to the following detailed description in conjunction with the accompany drawings in which:

[0033] FIG. 1 shows two vehicles having command systems;

[0034] FIG. 2 shows training sets for use in training the natural -language-understanders in the respective command systems shown in FIG. 1;

[0035] FIG. 3 shows an example of annotations applied to sentences in a source language and a target language;

[0036] FIGs. 4 and 5 show a process for constructing a training set for training a command system in a second language using training data for a first language;

[0037] FIG. 6 shows the source sentence example of FIG. 3 after having been transformed into a rewritten source-sentence; and

[0038] FIG. 7 shows a process for constructing a training set for training a command system using a fine-tuned translator model.DETAILED DESCRIPTION|0039| FIG. 1 shows a first vehicle 10 in which a first command-system 12 receives an utterance 14 from an occupant 16 via a microphone 18. In the course of operating the first vehicle 10, the occupant 16 issues many such utterances 14. Some of these utterances 14 include spoken commands. These utterance-borne commands are directed to the first command-system 12. Examples of such commands include instructions to navigate the vehicle 10 to a particular location, instructions to operate one of the many systems incorporated within the vehicle 10, examples of which include the climate control system or infotainment system, and instructions that require the command system 12 to interact with an external information source.

[0040] To assist it in understanding these spoken utterances 14, the first command-system 12 includes a first understander 20. This first understander 20 is a natural language understanding system that has been trained to understand an utterance 14 made using a first language. In order for it to understand such an utterance 14, the first understander 20 must be trained. Such training requires the use of source training-data 22, as shown in FIG. 2. This source training-data 22 consists of source sentences 24 and source annotations 26.

[0041] As shown in FIG. 3, a source sentence 24 comprises portions 28. Each portion 28 comprises one or more words. Because of their inherent meanings and their relationship to other words within the source sentence 24, these portions 28 communicate a particular “concept.” It is the purpose of the source annotations 26 to connect these portions 28 to individual concepts,

[0042] Each source annotation 26 identifies which portion 28 of the source sentence 24 communicates a particular concept. In the example shown in FIG. 3, the source annotations 26 identify the portions 28 “hotel,” “nearest,” and “find me” as corresponding to the concepts of“categories,” “relative location,” and “actions,” respectively. A common way to prepare these source annotations 26 is to do so manually. As a result of having to provide source annotations 26, the source training-data 22 is burdensome to prepare.

[0043] In addition to showing the first vehicle 10, FIG. 1 also shows a second vehicle 30. The second vehicle 30 is much like the first vehicle 10. It comprises a second command-system 32 that functions much like the first command-system 12 of the first vehicle 10. Just as the first commandsystem 12 relied on a first understander 20, the second command-system 32 relies on a second understander 34. However, the first understander 20 differs from the second understander 34. In particular, the second understander 34 has been trained to understand a second language, instead of the first language.

[0044] Because the first and second languages differ, one cannot train the second understander 34 using the source training-data 22. Instead, one must use target training-data 36, as shown in FIG. 2. Like the source training-data 22, the target training-data 36 comprises target sentences 38 and target annotations 40, also shown in FIG. 2.

[0045] As was the case with the source sentences 24, each target sentence 38 also has portions 28 that relate to distinct concepts. However, as is also apparent from FIG. 3, in some cases, the order of portions 28 in the target sentence 38 is a permutation of the order of portions 28 in the source sentence 24.(0046) Although the target sentences 38 and the source sentences 24 are in different languages, it is not unreasonable for them to mean the same thing. Thus, if one already has source trainingdata 22 with source sentences 24, one can, in principle, simply translate the source sentences 24 to create the target sentences 38. This is a task that can easily be carried out by a translator model 42. This makes it possible to leverage existing source sentences 24 in the course of developing the target training-data 36.

[0047] However, the target training-data 26 needs more than target sentences 38. It also needs target annotations 40. Unfortunately, although a translator model 42 excels at translating source sentences 24 into corresponding target sentences 38, it does not know how to annotate those targetsentences 38. The task of annotating the target sentences 38, i.e., producing the target annotations 40, continues to be one that is carried out manually. This poses a significant obstacle to the rapid and economical creation of target training-data 36.|0048] The method described herein provides a way to leverage not only the source sentences 24 but also the effort that went into creating the source training-data 22. The method proceeds by using the source training-data 22 to train the second understander 34 in a manner indicated in FIG. 4 and FIG. 5.

[0049] The vehicles 10, 30 may be configured to include various types of components, processors, and memory, and may communicate with a communication network. The communication network may be referred to as a “cloud” and may involve data transfer via wide area and / or local area networks, such as the Internet, Global Positioning System (GPS), cellular networks, Wi-Fi, Bluetooth, etc. The communication network may provide for communication between the vehicle and an external or remote server and / or database, as well as other external applications, systems, vehicles, etc. This communication network may provide navigation, music or other audio, program content, marketing content, internet access, speech recognition, cognitive computing, artificial intelligence, to the vehicle.

[0050] The remote server and the database may include one or more computer hardware processors coupled to one or more computer storage devices for performing steps of one or more methods as described herein and may enable the vehicle to communicate and exchange information and data with systems and subsystems external to the vehicle and local to or onboard the vehicle. The vehicle may include one or more processors configured to perform certain instructions, commands and other routines as described herein. Internal vehicle networks may also be included, such as a vehicle controller area network (CAN), an Ethernet network, and a media oriented system transfer (MOST), etc. The internal vehicle networks may allow the processor to communicate with other vehicle systems, such as a vehicle modem, a GPS module and / or Global System for Mobile Communication (GSM) module configured to provide current vehicle location and heading information, and various vehicle electronic control units (ECUs) configured to corporate with the processor.

[0051] The processor may execute instructions for certain vehicle applications, including navigation, infotainment, climate control, etc. Specifically the processor may execute instructions for the processes discussed herein, including those with respect to at least FIGs. 4„ 5 and 7. Instructions for the respective vehicle systems may be maintained in a non-volatile manner using a variety of types of computer-readable storage medium. The computer-readable storage medium (also referred to herein as memory, or storage) includes any non-transitory medium (e.g., a tangible medium) that participates in providing instructions or other data that may be read by the processor. Computer-executable instructions may be compiled or interpreted from computer programs created using a variety of programming languages and / or technologies, including, without limitation, and either alone or in combination, Java, C, C++, C#, Objective C, Fortran, Pascal, Java Script, Python, Perl, and PL / structured query language (SQL).

[0052] Referring to FIG. 4, the process begins with a rewriting step 44 in which the source annotations 26 for each source sentence 24 are integrated into the source sentences 24. This results in a set of rewritten source-sentences 46, as shown in FIG. 5 and in FIG. 6.|0053| FIG. 6 shows the source sentence 24 in FIG. 3 after having been transformed into a rewritten source-sentence 46. The rewritten source-sentence 46 comprises the original source sentence 24 in combination with concept tags 48. Each concept tag 48 corresponds to a source annotation 26. The order of concept tags 48 defines a concept sequence.

[0054] As is apparent from FIG. 6, a rewritten source-sentence 46, if read aloud, would sound peculiar. This peculiarity results from the concept tags 48, which, having been integrated into the original source sentence 24, would also have to be read aloud.

[0055] In effect, the original source sentence 24, which was in the source language, has been translated into a peculiar dialect of that source language. The native speakers of this peculiar dialect, were they to exist, would have developed the peculiar habit of uttering concept tags 48 associated with the words of their utterances during the course of the utterance. The phenomenon is not unlike the practice of those who, having acquired experience in working with a stenographer, have developed the skill of uttering punctuation marks during the course of dictation with that stenographer.

[0056] The incorporation of concept tags 48 turns out to be useful for two reasons. First, the rewritten source-sentences 46 can be used for fine tuning the translator model 42 so that it learns how to build target sentences 38 that have corresponding target annotations 40 integrated therein in a manner similar to that shown in FIG. 6. Second, because the existing source annotations 26 already point to relevant portions 28 of each source sentence 24, it is a simple matter to mechanize the process of integrating them as concept tags 48 into each source sentence 24.

[0057] In a first translation step 50 that follows, the translator model 42 translates the rewritten source-sentences 46 to produce target sentences 38. This results in a set of pairs 52, each of which consists of a rewritten source- sentence 46 and a corresponding target sentence 38, as shown in FIG. 6.

[0058] For each pair 52, the translator model 42 knows which portions 28 of the source sentence 24 correspond to which portions 28 of the target sentence 38. After all, it was the translator model 42 that carried out the translation. This means that the translator model 42 also knows which concept tags 48 correspond to which portions 28 of the source sentence. By combining these two pieces of information, the translator model 42 is able to determine the correct concept tags 48 for each portion of each target sentence 38. In effect, this amounts to automatically having annotated each target sentence 38.

[0059] Following the first translation step 50 is an alignment step 54. The alignment step 54 includes inspecting each pair 52 to determine whether it has the property of translational equivalence. A pair 52 that enjoys translation equivalence is included in a first set 56, or whether it lacks this property, in which case it is included in a second set 58. This first set 56 forms a portion of the target training-data 36.[0060| A filtering step 60 following this alignment step 54 filters out those pairs 52 in the first set 56 so that they can be used for fine-tuning the translator model 42.

[0061] The process continues with a fine-tuning step 62 in which the pairs 52 in the first set 56 are used to fine-tune the translator model 42. It was discovered that fine tuning such a translator model 42 has the effect of teaching that translator model 42 to translate a rewritten source-sentence46 in a way that creates a target sentence 38 that has been annotated with pointers 64 that point back to corresponding source annotations 32 in the source sentence 24, as shown in FIG. 6.(0062] The translator model 42, having been trained to translate source annotations 32 in the source sentences 30 into corresponding pointers 64 that point back to those source annotations 32, can now be considered a fine-tuned translator model 66, as shown in FIG. 7. As such, it is now in a position to undertake the task of reliably translating the rewritten source-sentences 46 in the second set 58 into additional annotated target sentences 38 for use as target training-data 36. This translation is carried out in a second translation step 68 shown in FIG. 4 to yield the complete target training-data 36.

[0063] The technical effect of the foregoing procedure is therefore to leverage existing source training-data 22 to create target training-data 36 by making a minor modification to the source training-data 22 to enable a translator model 42 to be fine-tuned in such a way that it can use the source annotations 26 to effectively create target annotations 40.

[0064] The descriptions of the various embodiments have been presented for purposes of illustration, but are not intended to be exhaustive or limited to the embodiments 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 described embodiments.

[0065] Aspects of the present embodiments may be embodied as a system, method or computer program product. Accordingly, aspects of the present disclosure may take the form of an entirely hardware embodiment, an entirely software embodiment (including firmware, resident software, micro-code, etc.) or an embodiment combining software and hardware aspects that may all generally be referred to herein as a “module” or “system.” Furthermore, aspects of the present disclosure may take the form of a computer program product embodied in one or more computer readable medium(s) having computer readable program code embodied thereon.

[0066] Any combination of one or more computer readable medium(s) may be utilized. The computer readable medium may be a computer readable signal medium or a computer readable storage medium. A computer readable storage medium may be, for example, but not limited to, anelectronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer readable storage medium include the following: an electrical connection having one or more wires, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (erasable programmable read-only memory (EPROM) or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing. In the context of this document, a computer readable storage medium may be any tangible medium that can contain, or store a program for use by or in connection with an instruction execution system, apparatus, or device.[0(167] Aspects of the present disclosure 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 disclosure. 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, enable the implementation of the functions / acts specified in the flowchart and / or block diagram block or blocks. Such processors may be, without limitation, general purpose processors, special-purpose processors, application-specific processors, or field- programmable.

[0068] The flowcharts 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 disclosure. 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.

[0069] While exemplary embodiments are described above, it is not intended that these embodiments describe all possible forms of the invention. Rather, the words used in the specification are words of description rather than limitation, and it is understood that various changes may be made without departing from the spirit and scope of the invention. Additionally, the features of various implementing embodiments may be combined to form further embodiments of the invention.

Claims

WHAT IS CLAIMED IS:

1. A method for providing a command system in a vehicle, wherein the command system responds to natural language commands in a target language, the method comprising: receiving source training-data for training a natural -language system in a source language, the source training-data comprising source sentences and corresponding source annotations; generating target training-data based on the source training-data; and using the target training-data to train the command system in the target language; wherein generating the target-training data comprises: generating rewritten source-sentences, the rewritten source-sentences incorporating concept tags that are derived from the source annotations; using a translation model, translating the rewritten source-sentences into corresponding target sentences, thereby generating a set of pairs, each of the pairs including a rewritten source-sentence and a corresponding one of the target sentences; using a first subset of the set of pairs to fine tune the translator model, thereby generating a fine-tuned translator model; and using the fine-tuned translator model to translate the rewritten sourcesentences in a second subset of the pairs, thereby generating the target training-data based on the source training-data.

2. The method of claim 1, further comprising: determining that a first pair’s constituent rewritten source-sentence and corresponding one of the target sentences have translational equivalence; determining that a second pair’s constituent rewritten source-sentence and corresponding one of the target sentences lack translational equivalence; placing the first pair in the first subset; and placing the second pair in the second subset.

3. The method of claim 1, wherein generating the rewritten source-sentences comprises selecting a first source sentence and inserting the concept tags adjacent to each other in the first source sentence.

4. The method of claim 1, wherein generating the rewritten source-sentences comprises selecting a first source sentence and inserting the concept tags between portions of the first source sentence.

5. The method of claim 1, wherein the pairs comprise a first pair, wherein the first pair comprises a rewritten source- sentence and corresponding one of the target sentences, wherein the rewritten source-sentence comprises a first set of concept tags, wherein the corresponding one of the target sentences comprises a second set of concept tags (), wherein each concept tag in the second set points to a corresponding one of the concept tags in the first set.

6. A system for providing a command system in a vehicle, wherein the command system responds to natural language commands in a target language, the system comprising: a microphone configured to receive a natural language command in a target language, a processor configured to: receive source training-data for training a natural -language system in a source language, the source training-data comprising source sentences and corresponding source annotations; generate target training-data based on the source training-data; and use the target training-data to train the command system in the target language; wherein generating the target-training data includes: generating rewritten source-sentences based on the natural language command, the rewritten source-sentences incorporating concept tags that are derived from the source annotations;using a translation model, translating the rewritten source-sentences into corresponding target sentences, thereby generating a set of pairs, each of the pairs including a rewritten source-sentence and a corresponding one of the target sentences; using a first subset of the set of pairs to fine tune the translator model, thereby generating a fine-tuned translator model; and using the fine-tuned translator model to translate the rewritten sourcesentences in a second subset of the pairs, thereby generating the target training-data based on the source training-data.

7. The system of claim 6, wherein the processor is further configured to: determine that a first pair’s constituent rewritten source-sentence and corresponding one of the target sentences have translational equivalence; determine that a second pair’s constituent rewritten source-sentence and corresponding one of the target sentences lack translational equivalence; place the first pair in the first subset; and place the second pair in the second subset.

8. The system of claim 6, wherein generating the rewritten source-sentences includes selecting a first source sentence and inserting the concept tags adjacent to each other in the first source sentence.

9. The system of claim 6, wherein generating the rewritten source-sentences includes inserting the concept tags between portions of the natural language command.

10. The system of claim 6, wherein the pairs comprise a first pair, wherein the first pair comprises a rewritten source- sentence and corresponding one of the target sentences, wherein the rewritten source-sentence comprises a first set of concept tags, wherein the corresponding one of the target sentences comprises a second set of concept tags, wherein each concept tag in the second set points to a corresponding one of the concept tags in the first set.

11. A method for providing a command system in a vehicle, wherein the command system responds to natural language commands in a target language, the method comprising: receiving source training-data for training a natural -language system in a source language, the source training-data comprising source sentences and corresponding source annotations; generating target training-data based on the source training-data; and using the target training-data to train the command system in the target language.

12. The method of claim 11, wherein generating the target-training data includes generating rewritten source-sentences, the rewritten source-sentences incorporating concept tags that are derived from the source annotations.

13. The method of claim 12, wherein generating the target-training data includes using a translation model and translating the rewritten source-sentences into corresponding target sentences, thereby generating a set of pairs, each of the pairs including a rewritten source-sentence and a corresponding one of the target sentences.

14. The method of claim 13, wherein generating the target-training data includes using a first subset of the set of pairs to fine tune the translator model, thereby generating a fine-tuned translator model.

15. The method of claim 14, wherein generating the target-training data includes using the fine-tuned translator model to translate the rewritten source-sentences in a second subset of the pairs, thereby generating the target training-data based on the source trainingdata.

16. The method of claim 14, further comprising: determining that a first pair’s constituent rewritten source-sentence and corresponding one of the target sentences have translational equivalence;determining that a second pair’s constituent rewritten source-sentence and corresponding one of the target sentences lack translational equivalence; placing the first pair in the first subset; and placing the second pair in the second subset.

17. The method of claim 14, wherein generating the rewritten source-sentences comprises selecting a first source sentence and inserting the concept tags adjacent to each other in the first source sentence.

18. The method of claim 16, wherein generating the rewritten source-sentences comprises selecting a first source sentence and inserting the concept tags between portions of the first source sentence.

19. The method of claim 16, wherein the pairs comprise a first pair, wherein the first pair comprises a rewritten source- sentence and corresponding one of the target sentences, wherein the rewritten source-sentence comprises a first set of concept tags, wherein the corresponding one of the target sentences comprises a second set of concept tags (), wherein each concept tag in the second set points to a corresponding one of the concept tags in the first set.

Images