A5038537171- Natural Language Processing Techniques
- Topic Modeling
- Multimodal Machine Learning Applications
- Text Readability and Simplification
- Speech Recognition and Synthesis
- Speech and dialogue systems
- Adversarial Robustness in Machine Learning
- Music and Audio Processing
- Biomedical Text Mining and Ontologies
- Laser and Thermal Forming Techniques
- Criminal Justice and Corrections Analysis
- Hepatitis B Virus Studies
- Metallurgy and Material Forming
- Algorithms and Data Compression
- Industrial Automation and Control Systems
- Industrial Vision Systems and Defect Detection
- Handwritten Text Recognition Techniques
- Text and Document Classification Technologies
- Artificial Intelligence in Law
- Optical measurement and interference techniques
- Pharmacological Effects and Assays
- Metal Forming Simulation Techniques
- Domain Adaptation and Few-Shot Learning
- Vibration and Dynamic Analysis
- Evaluation Methods in Various Fields
Huzhou University
2024
Changzhou Academy of Intelli-Ag Equipment (China)
2022-2024
Huawei Technologies (China)
2019-2024
Huawei Technologies (Sweden)
2021-2023
Huaibei Normal University
2022
Monash University
2021
University of Macau
2021
Dublin City University
2014-2017
Large Language Models (LLMs) trained on extensive textual corpora have emerged as leading solutions for a broad array of Natural Processing (NLP) tasks. Despite their notable performance, these models are prone to certain limitations such misunderstanding human instructions, generating potentially biased content, or factually incorrect (hallucinated) information. Hence, aligning LLMs with expectations has become an active area interest within the research community. This survey presents...
LLM-as-a-Judge, which generates chain-of-thought (CoT) judgments, has become a widely adopted auto-evaluation method. However, its reliability is compromised by the CoT reasoning's inability to capture comprehensive and deeper details, often leading incomplete outcomes. Existing methods mainly rely on majority voting or criteria expansion, insufficient address limitation in CoT. We propose Crowd-based Comparative Evaluation, introduces additional crowd responses compare with candidate...
Simultaneous translation (ST) starts translations synchronously while reading source sentences, and is used in many online scenarios. The previous wait-k policy concise achieved good results ST. However, faces two weaknesses: low training speed caused by the recalculation of hidden states lack future information to guide training. For speed, we propose an incremental Transformer with average embedding layer (AEL) accelerate calculation during future-guided training, a conventional as teacher...
End-to-end simultaneous speech translation (SST), which directly translates in one language into text another real-time, is useful many scenarios but has not been fully investigated. In this work, we propose RealTranS, an end-to-end model for SST. To bridge the modality gap between and text, RealTranS gradually downsamples input with interleaved convolution unidirectional Transformer layers acoustic modeling, then maps features space a weighted-shrinking operation semantic encoder. Besides,...
Previous works have shown that contextual information can improve the performance of neural machine translation (NMT). However, most existing document-level NMT methods failed to leverage contexts beyond a few set previous sentences. How make use whole document as global is still challenge. To address this issue, we hypothesize be represented graph connects relevant regardless their distances. We employ several types relations, including adjacency, syntactic dependency, lexical consistency,...
Previous work on document-level NMT usually focuses limited contexts because of degraded performance larger contexts. In this paper, we investigate using large with three main contributions: (1) Different from previous which pertrained models large-scale sentence-level parallel corpora, use pretrained language models, specifically BERT, are trained monolingual documents; (2) We propose context manipulation methods to control the influence contexts, lead comparable results systems small and...
We argue that the vulnerability of model parameters is crucial value to study robustness and generalization but little research has been devoted understanding this matter. In work, we propose an indicator measure neural network by exploiting their via parameter corruption. The proposed describes maximum loss variation in non-trivial worst-case scenario under For practical purposes, give a gradient-based estimation, which far more effective than random corruption trials can hardly induce...
To alleviate the data scarcity problem in End-to-end speech translation (ST), pre-training on for recognition and machine is considered as an important technique. However, modality gap between text prevents ST model from efficiently inheriting knowledge pre-trained models. In this work, we propose AdaTranS end-to-end ST. It adapts features with a new shrinking mechanism to mitigate length mismatch by predicting word boundaries. Experiments MUST-C dataset demonstrate that achieves better...
This paper describes Huawei’s neural machine translation systems for the WMT 2019 biomedical shared task. We trained and fine-tuned our on a combination of out-of-domain in-domain parallel corpora six directions covering English–Chinese, English–French English–German language pairs. Our submitted achieve best BLEU scores pairs according to official evaluation results. In English–Chinese task, are in second place. The enhanced performance is attributed more training sophisticated models...
Learning multilingual and multi-domain translation model is challenging as the heterogeneous imbalanced data make converge inconsistently over different corpora in real world. One common practice to adjust share of each corpus training, so that learning process balanced low-resource cases can benefit from high resource ones. However, automatic balancing methods usually depend on intra- inter-dataset characteristics, which agnostic or requires human priors. In this work, we propose an...
This paper describes the DCU submission to WMT 2014 on German-English translation task. Our system uses phrasebased model with several popular techniques, including Lexicalized Reordering Model, Operation Sequence Model and Language interpolation. final is result of combination systems which have different pre-processing alignments.
We argue that the vulnerability of model parameters is crucial value to study robustness and generalization but little research has been devoted understanding this matter. In work, we propose an indicator measure neural network by exploiting their via parameter corruption. The proposed describes maximum loss variation in non-trivial worst-case scenario under For practical purposes, give a gradient-based estimation, which far more effective than random corruption trials can hardly induce...
Dependency structure provides grammatical relations between words, which have shown to be effective in Statistical Machine Translation (SMT).In this paper, we present an open source module Moses implements a dependency-to-string model.We propose method transform the input dependency tree into corresponding constituent for reusing tree-based decoder Moses.In our experiments, achieves comparable results with standard model.Furthermore, enrich model via decomposition of structure, including...
This paper describes our work in the WAT 2020 Indic Multilingual Translation Task. We participated all 7 language pairs (En Bn/Hi/Gu/Ml/Mr/Ta/Te) both directions under constrained condition—using only officially provided data. Using transformer as a baseline, Multi->En and En->Multi translation systems achieve best performances. Detailed data filtering domain selection are keys to performance enhancement experiment, with an average improvement of 2.6 BLEU scores for each pair system 4.6...
This paper describes our work in participation of the IWSLT-2021 offline speech translation task. Our system was built a cascade form, including speaker diarization module, an Automatic Speech Recognition (ASR) module and Machine Translation (MT) module. We directly use LIUM SpkDiarization tool as The ASR is trained with three datasets from different sources, by multi-source training, using modified Transformer encoder. MT pretrained on large-scale WMT news dataset fine-tuned TED corpus....