A machine translation model crowdsourcing incremental learning method

Abstract

The application particularly relates to a machine translation model crowdsourcing incremental learning method, which comprises the following steps: providing a translation model array composed of multiple machine translation models, introduction and details of each translation model; model owners can select whether to open crowdsourcing training, if the crowdsourcing training is opened, the system automatically preprocesses and cleans data uploaded by participants for participating in joint training; data quality scoring is performed, and a contribution value of the batch of data to model training is calculated according to the effective data amount and the quality score; after the participants confirm the contribution value, the model incremental training process is entered through a front-end interactive module; after the incremental training is completed, the model is updated to the machine translation array, the model related information is updated, and the opening is continued, and other participants can participate in the training. According to the method, data resources and the model after incremental training can be greatly shared through joint training and model sharing of multiple persons.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, specifically relating to a crowdsourced incremental learning method for machine translation models. Background Technology

[0002] Currently, there are two main technical solutions for incremental learning methods in machine translation based on neural networks. One is the manual approach, where the person responsible for model training mainly collects professional data and performs model training and quality tuning. However, this one-on-one processing by model training personnel results in high labor costs and the process is not automated.

[0003] Another approach is automatic data submission, also known as a machine learning platform. Users can upload data to train their own models and data, which are used exclusively by specific individuals. There is no shared training or model-sharing mechanism, and model training is handled centrally by designated personnel. However, due to the scarcity of training data for specialized fields, this method fails to address the data resource sharing issue. Furthermore, the repeated training and deployment of similar models leads to a waste of hardware and model resources.

[0004] Neural machine translation, after incremental training with domain-specific data, can significantly improve translation quality in specialized fields. Given the high training costs of professional-domain machine translation, this paper proposes a crowdsourced incremental learning method. By aggregating data from the same domain, multiple users can share an incrementally trained model, and the ownership of the shared model is resolved based on its contribution value. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this invention provides a crowdsourced incremental learning method for machine translation models. By enabling multiple users to jointly train and share models, data resources and incrementally trained models can be shared to a greater extent. On the one hand, this effectively alleviates the problem of data scarcity in professional fields; on the other hand, it improves the utilization efficiency of hardware resources and incremental learning models.

[0006] The technical problem to be solved by this invention is achieved through the following technical solution:

[0007] A crowdsourced incremental learning method for machine translation models includes the following steps:

[0008] Step 1: Provide a translation model array consisting of multiple machine translation models, and provide an introduction and details for each translation model, including language direction, training data size, whether crowdsourced incremental learning is enabled, and translation performance testing function;

[0009] Step 2: The model owner can choose whether to enable crowdsourced incremental learning. If the owner chooses not to enable it, the model training permission belongs only to the model owner. Crowdsourced incremental learning is not enabled, and the model owner can upload data for incremental training.

[0010] Step 3: If the model owner chooses open crowdsourced incremental learning, participants can see the initial model training data volume on the model details page. After participants upload the data for joint training, the system automatically preprocesses and cleans the data.

[0011] Step 4: After cleaning the data, compare it with the model's historical training data, perform secondary deduplication, count the remaining effective data volume, and score the data quality. Based on the effective data volume and quality score, calculate the contribution of this batch of data to the model training.

[0012] Step 5: After the participants confirm their agreement to the contribution value through the front-end interaction module, the incremental training process of the model begins.

[0013] Step Six: After incremental training is completed, update the model in the machine translation array, update the relevant model information, and keep it open for other participants to participate in the training.

[0014] Furthermore, the aforementioned crowdsourced incremental learning method for machine translation models also includes automatic quality assessment after incremental training is completed.

[0015] Furthermore, the aforementioned crowdsourced incremental learning method for machine translation models also includes manual quality assessment after incremental training is completed.

[0016] Furthermore, the manual quality assessment method described above is as follows: extract several sentences from the original text as a test set, calculate the BLEU value of the translated text in the test set, and perform a manual quality score on the test set.

[0017] Furthermore, 200 sentences from the original text were extracted as a test set.

[0018] Furthermore, the preprocessing and cleaning in step three above are as follows: calculate the original text similarity by editing distance to remove duplicate data, and at the same time calculate the semantic similarity of the original text and translation in the uploaded data to remove the original text and translation with poor semantic similarity.

[0019] A machine translation model crowdsourcing incremental learning system is provided, which uses the aforementioned machine translation model crowdsourcing incremental learning method.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] The crowdsourced incremental learning method for machine translation models of this invention, through collaborative training and model sharing by multiple participants, can significantly share data resources and incrementally trained models. On the one hand, this effectively alleviates the problem of data scarcity in specialized fields; on the other hand, it improves the utilization efficiency of hardware resources and incremental learning models. This method can solve the problems of crowdsourced training and model ownership in translation models, incentivizing data owners in the same domain to participate in training and contribute model training results. Attached Figure Description

[0022] Figure 1 This is a flowchart of the crowdsourced incremental learning method for the machine translation model of this invention.

[0023] Figure 2 This is a schematic diagram illustrating the use of the crowdsourced incremental learning method in the machine translation model of this invention. Detailed Implementation

[0024] The present invention will be further described in detail below with reference to specific embodiments, but the implementation of the present invention is not limited thereto.

[0025] This embodiment provides a crowdsourced incremental learning method for machine translation models, as detailed in the appendix. Figure 1 and attached Figure 2 The crowdsourced incremental learning method for this machine translation model includes the following steps:

[0026] Step 1: Provide a translation model array consisting of multiple machine translation models, and provide an introduction and details for each translation model, including language direction, training data size, whether crowdsourced incremental learning is enabled, and translation performance testing function;

[0027] Step 2: The model owner can choose whether to enable crowdsourced incremental learning. If the owner chooses not to enable it, the model training permission belongs only to the model owner. Crowdsourced incremental learning is not enabled, and the model owner can upload data for incremental training.

[0028] Step 3: If the model owner chooses open crowdsourcing incremental learning, participants can see the initial model training data volume on the model details page. After participants upload the data for joint training, the system automatically preprocesses and cleans the data, calculates the original text similarity through edit distance, removes duplicate data, and calculates the semantic similarity between the original and translated texts in the uploaded data, removing the original and translated texts with poor semantic similarity.

[0029] Step 4: After cleaning the data, compare it with the model's historical training data, perform secondary deduplication, count the remaining effective data volume, and score the data quality. Based on the effective data volume and quality score, calculate the contribution of this batch of data to the model training.

[0030] Step 5: After the participants confirm their agreement to the contribution value through the front-end interaction module, the incremental training process of the model begins.

[0031] Step Six: After incremental training is complete, perform automatic quality assessment;

[0032] Step 7: After incremental training is completed, perform manual quality assessment, extract several sentences of the original text as the test set, calculate the BLEu value of the translated text in the test set, and perform manual quality scoring on the test set. Preferably, extract 200 sentences of the original text as the test set.

[0033] Step 8: After incremental training is completed, update the model in the machine translation array, update the relevant model information, and keep it open for other participants to participate in the training.

[0034] In this embodiment, the order of automatic quality assessment and manual quality assessment is not required. That is, automatic quality assessment can be performed first, followed by manual quality assessment; or manual quality assessment can be performed first, followed by automatic quality assessment.

[0035] The crowdsourced incremental learning method for machine translation models in this embodiment can solve the problems of crowdsourced training and model ownership in translation models, and incentivize data owners in the same domain to participate in training and contribute model training results.

[0036] This embodiment also provides a machine translation model crowdsourcing incremental learning system, which uses the machine translation model crowdsourcing incremental learning method of this embodiment.

[0037] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A crowdsourced incremental learning method for machine translation models, characterized in that, Includes the following steps: Step 1: Provide a translation model array consisting of multiple machine translation models, and provide an introduction and details for each translation model, including language direction, training data size, whether crowdsourced incremental learning is enabled, and translation performance testing function; Step 2: The model owner can choose whether to enable crowdsourced incremental learning. If the owner chooses not to enable it, the model training permission belongs only to the model owner. Crowdsourced incremental learning is not enabled, and the model owner can upload data for incremental training. Step 3: If the model owner chooses open crowdsourcing incremental learning, participants can see the initial model training data volume on the model details page. After participants upload the data for joint training, the system automatically preprocesses and cleans the data, calculates the original text similarity through edit distance, removes duplicate data, and calculates the semantic similarity between the original and translated texts in the uploaded data, removing the original and translated texts with poor semantic similarity. Step 4: After cleaning the data, compare it with the model's historical training data, perform secondary deduplication, count the remaining effective data volume, and score the data quality. Based on the effective data volume and quality score, calculate the contribution of this batch of data to the model training. Step 5: After the participants confirm their agreement to the contribution value through the front-end interaction module, the incremental training process of the model begins. Step Six: After incremental training is complete, perform automatic quality assessment; Step 7: After incremental training is completed, perform manual quality assessment, extract 200 original sentences as the test set, calculate the BLEu value of the translated test set, and perform manual quality scoring on the test set. Step 8: After incremental training is completed, update the model in the machine translation array, update the relevant model information, and keep it open for other participants to participate in the training.

2. A crowdsourced incremental learning system for machine translation models, characterized in that, The machine translation model crowdsourcing incremental learning method described in claim 1 is used.

Images