Multilingual IP Document Translation

A method for improving multi-language and multi-domain machine translation, especially for low-resource and zero-resource languages, using pre-training, two-stage training, and back translation techniques. The method involves: 1) Pre-training a cross-language representation model XLM-R on a large multilingual corpus. 2) Adapting the pre-trained XLM-R for specific languages and domains using a new domain-specific transformation (CDSTX) to improve low-resource domain adaptation. 3) Fine-tuning the adapted XLM-R for translation tasks using back translation to further improve performance.

Source

A sentence translation method using bidirectional long short-term memory (LSTM) neural networks to improve accuracy and reduce difficulty compared to conventional machine translation. The method involves feeding a source sentence into a bidirectional LSTM to split it into two sequences, one in normal order and one reversed. The LSTM captures context between words using hidden states. It combines hidden states from both sequences at each step to generate a context vector. The context vectors and prior translations determine the translation at each step. This allows associating phrases with context and improves accuracy compared to just translating one sequence.

Source

Natural language processing using a synapper model unit that enables accurate translation and sentence parsing across multiple languages without requiring separate language-specific training. The method involves converting words into neural concept codes, feeding them into a language processing unit using a synapper model, parsing the codes, and converting back to words. This synapper model provides a unified structural representation of sentences across languages by integrating their different grammatical structures into multiple dimensions.

Source

Accurate machine translation method that improves translation of technical terms and proper nouns. It involves determining the domain-specific nouns in the source language text using forward maximum matching, translating those nouns separately, and replacing them in the translation. This improves accuracy for technical terms compared to using the standard translation model alone. The translation model is trained using filtered samples with optimal word ratios.

Source

Automated machine translation system that improves translation quality by preprocessing the source text before translation. The preprocessing involves steps like sentence splitting, simplification, named entity recognition, matching against existing translations, and applying transformational grammar. The goal is to generate semantically meaningful, less complex ordered tokens from the source text that are easier to translate accurately.

Source

Cross-lingual adaptation of language models for low-resource languages using disentangled syntax and shared conceptual latent space. The method involves converting multilingual input sentences into linearized constituency parse trees, masking leaf nodes to separate semantics, passing to a syntactic encoder, and determining if a new language is being learned. For similar scripts, transliteration aligns with syntax. For unique scripts, pseudo translation aligns semantics. This disentangles syntax from concepts, leverages relatedness for adaptation, and improves low-resource cross-lingual performance.

Source

Automated language translation using machine learning techniques to convert sentences from one natural language (like English) into another (like French). The translation is done by training a machine learning model on large datasets of translated sentences to learn the mapping between the source and target languages. The model can then accurately translate new, unseen sentences in the source language into the target language.

Source

A machine translation method using deep learning and attention models to improve the accuracy and efficiency of machine translation compared to existing methods like rule-based and statistical translation. The method involves training a neural network translation model using bilingual corpora, then deploying it to translate text. Attention mechanisms are used to focus the network on important parts of the input sequence during decoding. This allows the model to better capture and transfer meaning between languages. The attention weights can also be used to analyze the translation process. The method also includes steps like custom word segmentation to improve input processing.

Source

Chapter-level neural machine translation method that screens out irrelevant or weakly related words from chapter context to improve translation accuracy. It introduces a context filtering attention module in the Transformer encoder that gradually filters out strongly related words in the chapter context while encoding the current sentence vocabulary. This allows incorporating more relevant context words into the sentence encoding.

Source

Adaptive machine translation using reinforcement learning and domain classification to improve translation quality, especially for complex languages and specific domains. The method involves training a shared encoder-decoder transformer model on parallel corpora for multiple languages. Before feeding the source text to the encoder, it first passes through a multi-classification neural network to classify the domain. This allows distinguishing linguistic knowledge in specific fields. The classified domain is mapped to a specific encoder head for training. The cross-over loss function is used to learn domain-specific representations. Reinforcement learning techniques like rugged loss and feedback error are applied to encourage exploration and improve the model based on user feedback.

Source

Configuring multiple neural networks to translate documents with hierarchical structure by generating mappings between word locations, meanings, and grammatical rules in the source and destination languages. This provides a way to improve translation accuracy and efficiency by capturing the hierarchical structure of the languages and training neural networks to make corrections based on those mappings. The method involves collecting data in the destination language, generating mappings for word locations, meanings, and grammatical rules, training neural networks on corrections based on those mappings, and using the trained networks to translate documents.

Source

Configuring neural networks for document translation that improves accuracy by leveraging the hierarchical structure of documents. The method involves generating hierarchical mappings for the source and destination languages, including mappings between word locations, grammatical info, and grammatical rules. Corrections are generated based on these mappings, trained into neural networks, and used to translate documents. This automated hierarchy analysis and correction training allows the networks to identify and implement similar fixes across the document for more natural translation.

Source

Machine translation system using deep learning that aims to improve efficiency and accuracy over traditional rule-based or statistical machine translation methods. The system leverages deep learning techniques like recurrent neural networks (RNNs) and transformers to provide better contextual modeling of language. It uses a data preparation module, model selection unit, feature representation unit, model training module, evaluation and tuning module, and deployment and application module to train and optimize the translation models. The transformer architecture allows parallel computation and capturing long-distance dependencies compared to sequential RNNs. This provides better computational efficiency and modeling capabilities for translation.

Source

Machine translation system that customizes neural network translation models for specific text segments by dynamically retraining the model based on similar examples. When translating a new text, the system identifies matching phrases from a parallel corpus and ranks them based on applicability. It then retrains the neural network using the top-ranking matches to further customize the translation for those segments. This allows better translation accuracy for unique phrases that may not have been covered in the initial training.

Source

Multilingual, multidomain neural machine translation that leverages adapter layers to decouple language and domain feature spaces, allowing cross-lingual sharing of domain knowledge. Adapters are inserted inside the encoder-decoder model to mine language and domain knowledge separately. This decoupling enables transferring domain knowledge from a source language to a target language even if there's no parallel corpus in the target domain.

Source

Machine translation design method using deep learning that improves translation quality for small languages with limited data. The method involves preprocessing corpus data, pre-training a multilingual model, and fine-tuning the pre-trained model on specific language pairs. The preprocessing steps include segmenting text, word segmentation, language marking, and random replacement. This reduces corpus data dependence and allows better translation with less data.

Source

Neural machine translation system for accurately and efficiently translating documents using a cross-level attention mechanism that leverages the context of entire documents during translation. The system preprocesses the corpus to add document structure information. It then trains a basic translation model using this structured corpus. During translation, it calculates dependency weights between words and sentences to get global context. It combines sentence vectors with weighted context to accurately select the context for each word. This allows using full document structure instead of just sentence context.

Source

Building multilingual machine translation models that can accurately translate between languages while avoiding off-target translation issues. The method involves converting each language's semantics to a shared representation, translating from that space, then converting back. This allows leveraging shared parameters for better representation while avoiding issues with unique semantics. The method also uses techniques like pseudo-parallel data, pre-training, freezing parameters, and self-study to improve multilingual translation.

Source

Neural machine translation system for more accurate and coherent translation of longer texts like chapters by using a routing algorithm to selectively incorporate context information into the translation process. The system encodes the current sentence and surrounding sentences using self-attention and linear layers. A routing algorithm calculates a gate using the current sentence to screen the context information. This filtered context is fused with the current sentence encoding and passed through a decoder to generate the translation. The routing algorithm allows the current sentence to actively select and combine relevant context information for more coherent translations of longer texts.

Source

Natural language processing device and program that enables accurate machine translation by selectively using different types of bilingual data for training and data expansion. The device learns a translation model using balanced parallel data without omissions or excessive translations. It also learns a data augmentation model using only the balanced data. For translation, it uses the main model and expands with pseudo parallel data generated by translating monolingual text with the augmentation model. This avoids degradation from noisy bilingual sources.

Source

Thai-Chinese neural machine translation method that leverages unsupervised dependency syntax to improve translation quality when there is a lack of high-quality parallel corpora between Thai and Chinese. The method involves obtaining Thai dependency syntax information using unsupervised transfer techniques since Thai dependency parsing tools are limited. This unsupervised dependency syntax is then integrated into the Transformer neural network model using a modified attention mechanism that incorporates parent word positions. This allows the model to better conform to the Thai language structure during translation.

Source

A method and system for term translation that can quickly adapt to different translation requirements for specific words in different contexts. The method involves establishing a term base containing translations for key terms. This term base is stored in a database. During training, the term base is matched with sentence pairs in the training corpus and replaced with special markers. The modified corpus is then used to train the machine translation model. At translation time, the term base is retrieved and used with a forward maximum matching algorithm to identify terms in the input. The identified terms are then translated using the trained machine translation model.

Source

Hybrid translation system that combines the strengths of general-purpose neural machine translation with domain-specific translation dictionaries to improve accuracy and suitability for specific fields. The system receives a document, determines the field, translates using a general neural network, then converts some terms from the general translation into more specialized terms from a dictionary for the determined field.

Source

Machine translation method that extracts the key meaning of a source sentence using a knowledge base and then generates the translation based on that extracted meaning. It involves using a first machine learning module to map the source sentence to the semantic space defined by the knowledge base to extract the key information. A second machine learning module then generates the target sentence using that extracted meaning. This allows accurate translation by focusing on the essential meaning instead of trying to learn complex mappings between languages. The first module extracts the semantic tuples from the source sentence using the knowledge base, and the second module generates the target sentence using those extracted tuples.

Source

Low-resource neural machine translation using bidirectional dependency self-attention to improve translation quality in resource-constrained scenarios. The method involves integrating bidirectional dependency information into the Transformer encoder's multi-head attention mechanism. It fuses dependency knowledge from both parent-to-child and child-to-parent directions. This provides more comprehensive structural information for the model compared to just parent-to-child directions. The bidirectional dependency self-attention mechanism improves machine translation performance by effectively using dependency analysis for low-resource languages.

Source

Neural machine translation method for multilingual translation that improves accuracy and enables low-resource migration learning by using importance measurement and language-specific encoding. It addresses the issue of unused language-specific knowledge and low-resource transfer learning in multilingual translation. The method involves encoding and decoding sequences using attention and feedforward neural networks that have language-specific modules to capture unique features. This allows leveraging language-specific knowledge and semantics in translation. For low-resource migration learning, it involves training on a main language pair and then transferring to a new, related language pair with fewer data. The language-specific modules adapt to the new language.

Source

A multi-language universal translation method that improves the quality of machine translation between multiple languages using a technique called multi-granularity semantic alignment. The method leverages knowledge from pre-trained language models in multiple languages to improve the translation between those languages. It aligns semantic features at different granularities to better transfer and fuse knowledge between languages. This involves techniques like cross-language word alignment, cross-lingual semantic embedding, and multilingual semantic feature extraction. The method aims to improve the effectiveness of multilingual machine translation using pre-trained language models.

Source

A method for translating between multiple languages using a single neural network model. The method involves training a multilingual translation model on large bilingual datasets from various language pairs. The model uses a deep encoder with a segmented language analyzer to process the input text. This allows the model to simultaneously translate between multiple languages without requiring separate models for each language pair. The method aims to improve multilingual translation performance by leveraging knowledge migration between languages and reduce the cost and complexity of deploying multiple language-specific translation systems.

Source

Neural machine translation method that improves translation accuracy and reduces rare word issues by integrating scrutiny network and character encoding. The method involves using character-level encoding instead of word-level encoding to overcome problems with rare words and unseen words in translation. It maps input sequences to character embeddings, concatenates fragments, and uses Highway Networks and GRUs to encode the characters. This encoder is combined with a scrutiny network for translation that uses the encoded fragments instead of just the source sequence. The scrutiny network allows using previously translated fragments to aid in final translation, providing better context and accuracy compared to left-to-right decoding. The character encoding also allows dealing with spelling errors and special symbols without explicit segmentation.

Source

A neural machine translation model for multilingual text translation that improves translation quality compared to existing multilingual translation models. The key innovation is a specific design for the multilingual translation model that leverages the similarity and balanced data availability between the target languages. The model has multiple encoders and decoders, one for each source language, and shares the rest of the network architecture. This allows the model to translate between multiple source languages to multiple target languages using a single model. The shared parts of the network are trained on data from all the source languages. The unique encoders and decoders for each source language are trained separately on their respective parallel corpora. This leverages the language similarity between the source languages to improve translation accuracy.

Source

Chinese-Vietnamese neural machine translation method using dependency graph networks to improve translation performance in low-resource scenarios where parallel corpora are scarce. The method involves converting dependency syntax trees into dependency graphs, encoding the graph structure, fusing it with the source language sequence at the encoder, and using graph-sequence attention layers at both encoder and decoder to guide translation.

Source

A Uyghur-Chinese machine translation system that improves translation performance for the Uyghur language, which has complex morphology. The system uses a hybrid approach of words and morphemes as translation units. For high-frequency words, it uses whole words. For low-frequency words, it segments them into morphemes. This avoids long sentences from morpheme segmentation hurting attention mechanisms. A mixed granularity encoder processes words and morphemes. It improves translation by reducing vocabulary size and avoiding excessive sentence length. The hybrid strategy balances benefits of word-based vs morpheme-based translation.

Source

Customizable automatic document translation that allows users to specify translation attributes like adequacy or fluency. When translating a source document, the system uses a machine translator with the specified attribute. It then analyzes the resulting target document and provides a different translation attribute for feedback. This iterative process allows users to balance accuracy and readability in the translations.

Source

Automated machine translation of languages to convert sentences from one natural language into another language using computers. The goal is to have machines accurately translate sentences between different human languages without human intervention.

Source

Chinese-Vietnamese neural machine translation that incorporates syntactic parse trees to improve the quality of translation. The method involves combining lexical and syntactic parsing techniques with deep learning for machine translation. It uses a neural machine translation model trained on a Chinese-Vietnamese bilingual corpus, but additionally performs syntactic analysis on the source language to generate parse trees. These parse trees are converted into vectors and concatenated with the source word embeddings as input to the translation model. This fused input improves the translation accuracy and fluency compared to using just word embeddings.

Source

Natural language translation method that first processes sentences for translation into models with codes and directional operators, and then processes the models to generate translated sentences in the target language. The method involves separating sentences, parsing them with codes and operators, and translating by processing the parsed codes. It aims to mimic natural language processing like the human brain by using multi-dimensional parsing to improve accuracy and speed of machine translation.

Source

Neural machine translation (NMT) method that uses external information like dictionaries to improve translation quality without increasing model complexity. The method involves generating translations using both the source input and external information input. At each time step, it combines the original translation probability and the translation probability from the external information using a noise discrimination result. The noise discrimination is done by a neural network that determines if each word in the external input is helpful or noisy for the translation.

Source

A Chinese-Vietnamese neural machine translation method to improve translation quality when parallel corpus resources are limited. The method involves using coding induction and decoding scrutiny. It enhances decoding by incorporating a summary of the encoding stage. This provides additional semantic information to the decoder. The encoding stage is reversed and merged with the forward encoding to provide reverse enhancements. This supplements the forward sequence with semantic information from the reverse direction. This helps improve translation quality when parallel corpus resources are limited.

Source

Training a universal encoder component that can convert input documents in any natural language into a language-agnostic representation. This component can be used to train task-specific machine-trained components like sentiment analysis for multiple languages using a single universal encoder. The universal encoder is trained using a generative adversarial network (GAN) with a language model and discriminator. The GAN objective balances accuracy on language-specific tasks with coherence across languages. This allows the universal encoder to learn semantic meaning independent of syntax.

Source

Neural machine translation system that can accurately translate rare words by tracking their origin in the source sentence. The system uses a neural network trained to issue special "out-of-vocabulary" (OOV) marks for unknown words in the target sentence. These OOV marks can have different types like pointer marks that identify the corresponding source words, and empty marks that don't. The network learns to associate the OOV marks with the right source words through alignment data from a parallel corpus. This allows it to accurately translate rare words by keeping their connection to the source sentence.

Source

Text-level neural machine translation using a context memory network to improve the accuracy and context preservation of translating documents compared to sentence-level translation. The method involves encoding the entire input document instead of just sentence-by-sentence. A context memory network is used to efficiently represent the context before the current sentence being translated. This allows capturing longer context compared to encoding just previous sentences. The context memory is updated as the translation progresses.

Source

Neural machine translation system using external information like target language words, phrases, and sentences to improve translation quality without increasing model complexity. The system has separate encoders for source and external information, and a combined decoder. The external information encoder gets target words to generate hidden representations. A noise discriminator judges if hidden representations are noisy. The combined decoder uses source and external inputs along with noise discriminations to generate translations. The training corpus includes aligned source texts and external information sampled from target translations.

Source

An attentional neural machine translation system for translating a source sequence in a first language into a target sequence in a second language. The system uses a recurrent neural network grammar (RNNG) encoder to encode the source sequence and its phrase structures. An RNNG decoder with attention decodes the target sequence and its phrase structures. The attention mechanism compares the decoder state to encoder states to calculate phrase type vectors. This allows unsupervised induction of tree structures in the source and target sequences without explicit segmentation or parsing annotation. The system learns plausible segmentation and shallow parsing when trained on character-level datasets.

Source

A system for generating parallel sentences in multiple languages using a shared latent space. The system leverages a neural network trained for machine translation to build a shared latent space for the languages. A separate generative adversarial network (GAN) is then trained to generate codes in the shared space that can be decoded into sentences in either language. This allows concurrent generation of sentences in multiple languages using a single shared latent representation. The GAN is trained by feeding it inputs and comparing the generated codes to the actual translation codes from the shared space.

Source

Automatically translating patent claims from one language to another using a computer-based system that extracts key nouns and phrases from the claim definitions, translates those elements individually, and then combines them into a coherent translated claim. This approach improves accuracy compared to generic automatic translation tools which struggle with the complex grammar and syntax of patent claims.

Source

Neural network machine translation model for low-resource domains that leverages iterative two-way migration between source and target domains to improve translation performance. The model involves multiple rounds of feedback between domains to fully explore shared knowledge and avoid degradation. In each round, the source and target models optimize using each other's parameters as a teacher. This iterative migration allows learning from multiple source domains in a many-to-one setup.

Source

Chinese-Vietnamese neural machine translation method based on transfer learning to improve translation quality in low-resource scenarios like Chinese-Vietnamese where parallel corpus is scarce. The method involves using larger Chinese-English and English-Vietnamese parallel corpora to pre-train encoders and decoders separately. These pre-trained models are then used to initialize the Chinese-Vietnamese neural machine translation model instead of starting from scratch. This provides better semantic representation and translation accuracy for the Chinese-Vietnamese model.

Source

Artificial intelligence translation integration system that improves accuracy and flexibility of portable machine translation devices. The system uses AI techniques like feature extraction, classification, segmentation, and neural networks for translation. It converts data to standardized text, extracts features, classifies into language and format, segments into word modules, arranges format, and uses neural nets to translate. If translation fails, it falls back to manual translation. The system also enables offline language recognition and format classification using Adaboost.

Source

Automatically translating patent claims from one language to another in a way that makes them easier to understand. The translation is done by organizing the claim elements like nouns, verbs and objects into a structured sentence, then translating them in a specific sequence based on grammatical attributes, such as subordinate clauses, participles, and prepositions. This allows claims to be translated in a more accurate and reliable manner compared to standard machine translation methods.

Source

Machine translation system that combines syntax conversion and word translation models to improve naturalness and appropriateness of machine translations. The system learns syntax conversion and word translation knowledge from parallel corpora, applies them to source sentences in real time, and generates target sentences through decoding processes of a syntax converter and a word translator. This allows changing the word order in the target sentence like syntax-based translation while also translating consecutive words like phrase-based translation. The syntax converter extracts syntax features from the source sentence and the word translator generates target words constrained by the syntax. The syntax converter's output syntax and the word translator's output words are combined with weights learned from the parallel corpora to generate the final target sentence.

Source