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

Direct speech-to-speech translation through machine learning models that can translate speech from one language to another without intermediate text conversion. The translation is done end-to-end using sequence-to-sequence models with encoder, attention, and decoder components. The encoder converts input acoustic features to hidden states, the attention module processes the hidden states, and the decoder generates output acoustic features representing translated speech in the target language. This allows fluent direct translation without compound errors or loss of paralinguistic information.

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

Generating natural language models with improved accuracy by training on masked sentence pairs. The method involves masking out words in a sentence pair, predicting the masked words, and predicting the relationship between the masked sentences. This provides a masked sentence pair group representing the original pair. The group is fed through an initial neural network to learn word vectors and relationships. The initial network is adjusted based on the original and predicted pairs to improve accuracy. The trained network can then be used for tasks like language processing. Masking helps the model learn global context and semantics beyond isolated words.

Source

Training natural language processing models to better convert natural language to logical forms like SQL queries, by augmenting the training data with contrastive and alternative examples to address catastrophic forgetting and overgeneralization. The augmentation involves revising examples with forgotten syntax to cause regression during training, and modifying operators to mitigate bias. This expanded training data is then used to train the model. This allows it to better handle unseen natural language queries and avoid forgetting previously learned mappings.

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

Non-autoregressive neural machine translation method that improves the translation quality of non-autoregressive neural machine translation models by incorporating context feature extraction. The method involves using hybrid grouping linear transformation modules at both the encoder and decoder ends of the CMLM model to enhance local feature capabilities. This allows context information to be effectively utilized during decoding by grouping and integrating source sentence features.

Source

Modular semantic machine translation method that improves stability, interpretability, and flexibility of machine translation compared to end-to-end and generative models. The method involves decoupling understanding and generation modules. It encodes source language into a semantic representation, translates it into an intermediate semantic representation, then generates target language from that. This allows separate training and interpretation of understanding vs generation. It also uses perturbed semantic representations to improve translation consistency.

Source

Semi-supervised machine translation using multi-task learning and feature self-distillation to improve translation accuracy when labeled data is limited. The method involves jointly training translation models on a mix of bilingual and monolingual data. For the monolingual data, the models learn to self-distill features by comparing outputs. This involves a teacher model built from previous rounds' student models, with soft labels used to guide training. By directly mining high-level features from monolingual data without pseudo parallel corpus, it avoids issues of existing methods.

Source

Chinese-Vietnamese neural machine translation that accurately translates entity words in low-resource scenarios. The method involves fusing block and entity category information at the encoding stage and using a pointer network at decoding to ensure correct entity output. It also introduces constraint prompts based on bilingual dictionaries to guide entity translation. This improves entity accuracy compared to baseline models without damaging overall translation quality.

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

Text generation model for improving translation quality in resource-scarce scenarios by reducing differences in semantic representations of different languages. The model enhances representation capabilities of a text generation model in latent space to improve translation quality when resources are scarce. It samples semantic information from hidden representations using gated units that capture semantic details from specific positions. By selectively extracting semantic information from each position and mapping sampled semantics to a common space, the model learns to reduce differences in semantic structures between languages. This helps overcome issues like redundant or missing words due to cultural and regional factors.

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

Intelligent translation system using machine learning to improve accuracy and efficiency compared to traditional rule-based methods. The system has a visual interface to select translation modes. It uses a neural network architecture with components like speech recognition, feature extraction, and semantic translation to process input speech. It generates a primary recognition result from the extracted features. It then compares this against a large database to measure difference values for vocabulary, syntax, and semantics. If the differences exceed a threshold, it biases the extracted features towards the original meaning. This iterative biasing improves accuracy by adjusting features based on comparison with known meanings.

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

Reducing cross-modal and cross-language barriers in end-to-end speech translation to improve the quality of translation in scenarios with limited training data. The method involves a speech encoder and text decoder that share a semantic space for better cross-modal alignment. The speech encoder generates a specific state sequence for each decoder layer without introducing extra modules. A hybrid attention sub-module at the decoder uses softmax to generate word embeddings by attending to both speech states and other words in the sentence. This allows the speech states and target language words to share a semantic space, reducing barriers between modalities and languages.

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