A Long Short-Term Memory(LSTM) Recurrent Neural Network(RNN) has driven tremendous improvements on an acoustic model based on Gaussian Mixture Model(GMM). However, these models based on a hybrid method require a force...A Long Short-Term Memory(LSTM) Recurrent Neural Network(RNN) has driven tremendous improvements on an acoustic model based on Gaussian Mixture Model(GMM). However, these models based on a hybrid method require a forced aligned Hidden Markov Model(HMM) state sequence obtained from the GMM-based acoustic model. Therefore, it requires a long computation time for training both the GMM-based acoustic model and a deep learning-based acoustic model. In order to solve this problem, an acoustic model using CTC algorithm is proposed. CTC algorithm does not require the GMM-based acoustic model because it does not use the forced aligned HMM state sequence. However, previous works on a LSTM RNN-based acoustic model using CTC used a small-scale training corpus. In this paper, the LSTM RNN-based acoustic model using CTC is trained on a large-scale training corpus and its performance is evaluated. The implemented acoustic model has a performance of 6.18% and 15.01% in terms of Word Error Rate(WER) for clean speech and noisy speech, respectively. This is similar to a performance of the acoustic model based on the hybrid method.展开更多
针对畜禽疫病文本语料匮乏、文本内包含大量疫病名称及短语等未登录词问题,提出了一种结合词典匹配的BERT-BiLSTM-CRF畜禽疫病文本分词模型。以羊疫病为研究对象,构建了常见疫病文本数据集,将其与通用语料PKU结合,利用BERT(Bidirectiona...针对畜禽疫病文本语料匮乏、文本内包含大量疫病名称及短语等未登录词问题,提出了一种结合词典匹配的BERT-BiLSTM-CRF畜禽疫病文本分词模型。以羊疫病为研究对象,构建了常见疫病文本数据集,将其与通用语料PKU结合,利用BERT(Bidirectional encoder representation from transformers)预训练语言模型进行文本向量化表示;通过双向长短时记忆网络(Bidirectional long short-term memory network,BiLSTM)获取上下文语义特征;由条件随机场(Conditional random field,CRF)输出全局最优标签序列。基于此,在CRF层后加入畜禽疫病领域词典进行分词匹配修正,减少在分词过程中出现的疫病名称及短语等造成的歧义切分,进一步提高了分词准确率。实验结果表明,结合词典匹配的BERT-BiLSTM-CRF模型在羊常见疫病文本数据集上的F1值为96.38%,与jieba分词器、BiLSTM-Softmax模型、BiLSTM-CRF模型、未结合词典匹配的本文模型相比,分别提升11.01、10.62、8.3、0.72个百分点,验证了方法的有效性。与单一语料相比,通用语料PKU和羊常见疫病文本数据集结合的混合语料,能够同时对畜禽疫病专业术语及疫病文本中常用词进行准确切分,在通用语料及疫病文本数据集上F1值都达到95%以上,具有较好的模型泛化能力。该方法可用于畜禽疫病文本分词。展开更多
为研究基于长短期记忆(Long Short-TermMemory,LSTM)网络的语音转文字系统的优化方法,首先说明LSTM在语音转文字任务中的基本原理和架构,其次分析自适应矩估计(Adaptive Moment Estimation,Adam)优化算法的核心机制及其在LSTM网络中的应...为研究基于长短期记忆(Long Short-TermMemory,LSTM)网络的语音转文字系统的优化方法,首先说明LSTM在语音转文字任务中的基本原理和架构,其次分析自适应矩估计(Adaptive Moment Estimation,Adam)优化算法的核心机制及其在LSTM网络中的应用,最后在Mozilla DeepSpeech框架中嵌入基于Adam优化的LSTM模型,并使用THCHS-30数据集进行实验。实验结果表明,基于Adam优化的LSTM模型在词错率和F1分数上均表现出显著的优越性。展开更多
Background: In the context of the fight against HIV, a lack of skills in monitoring and evaluating the personnel in charge of activities has been identified at the national level. It was the subject of a priority axis...Background: In the context of the fight against HIV, a lack of skills in monitoring and evaluating the personnel in charge of activities has been identified at the national level. It was the subject of a priority axis of the national plan for monitoring and evaluating the fight against HIV (2006-2010) that was aimed at strengthening the capacities of actors in this area. To increase the critical mass of competent human resources in the short term, the National Institute of Public Health (NIPH) of Côte d’Ivoire organized monitoring and evaluation training sessions for healthcare professionals from 2011 to 2016. Methods: A single case study with multiple levels of analysis was carried out, combining a qualitative survey and a literature review. An evaluation was carried out six months after each training session. In addition, the results of the pre- and post-tests and of the daily and final evaluations that accompanied the various training sessions were used to provide further information. The qualitative data collected were analyzed using INVIVO 15 software. Results: Some 89 health professionals (69% men and 31% women) working at the national level (51% at the central level, including 58% in health programs) and in development partner agencies (37%) participated in this capacity building program. Most participants were senior health managers (56%), data managers (23%), and statisticians and computer scientists (10%). Almost all the trainings were financed by 16 technical and financial partners (85%), mainly the MEASURE Evaluation project (27%). Conclusion: M&E training, despite all its imperfections, has made it possible to identify M&E training needs at the national level and to increase the critical mass of national skills and to have some culture in M&E.展开更多
随着数字农业的快速发展,农作物命名实体识别作为农业领域知识图谱构建的基础,成为一种高效率的农作物研究领域识别方法。由于农作物实体识别呈现结构复杂、实体指称不一致、干扰因素多等特征,严重制约了农作物领域实体识别的性能,提出...随着数字农业的快速发展,农作物命名实体识别作为农业领域知识图谱构建的基础,成为一种高效率的农作物研究领域识别方法。由于农作物实体识别呈现结构复杂、实体指称不一致、干扰因素多等特征,严重制约了农作物领域实体识别的性能,提出一种基于预训练语言模型的实体识别模型,使用BERT为文本中词进行编码、采用双向LSTM(Long-Short Term Memory)获取句子中关键词的上下文,采用CRFs(Conditional Random Fields)捕获词之间的依赖关系,并结合所构建的农作物命名实体识别数据集进行验证。实验证明该模型能够有效对农作物实体进行识别,且性能优于当前已有的实体识别模型。展开更多
基金supported by the Ministry of Trade,Industry & Energy(MOTIE,Korea) under Industrial Technology Innovation Program (No.10063424,'development of distant speech recognition and multi-task dialog processing technologies for in-door conversational robots')
文摘A Long Short-Term Memory(LSTM) Recurrent Neural Network(RNN) has driven tremendous improvements on an acoustic model based on Gaussian Mixture Model(GMM). However, these models based on a hybrid method require a forced aligned Hidden Markov Model(HMM) state sequence obtained from the GMM-based acoustic model. Therefore, it requires a long computation time for training both the GMM-based acoustic model and a deep learning-based acoustic model. In order to solve this problem, an acoustic model using CTC algorithm is proposed. CTC algorithm does not require the GMM-based acoustic model because it does not use the forced aligned HMM state sequence. However, previous works on a LSTM RNN-based acoustic model using CTC used a small-scale training corpus. In this paper, the LSTM RNN-based acoustic model using CTC is trained on a large-scale training corpus and its performance is evaluated. The implemented acoustic model has a performance of 6.18% and 15.01% in terms of Word Error Rate(WER) for clean speech and noisy speech, respectively. This is similar to a performance of the acoustic model based on the hybrid method.
文摘针对畜禽疫病文本语料匮乏、文本内包含大量疫病名称及短语等未登录词问题,提出了一种结合词典匹配的BERT-BiLSTM-CRF畜禽疫病文本分词模型。以羊疫病为研究对象,构建了常见疫病文本数据集,将其与通用语料PKU结合,利用BERT(Bidirectional encoder representation from transformers)预训练语言模型进行文本向量化表示;通过双向长短时记忆网络(Bidirectional long short-term memory network,BiLSTM)获取上下文语义特征;由条件随机场(Conditional random field,CRF)输出全局最优标签序列。基于此,在CRF层后加入畜禽疫病领域词典进行分词匹配修正,减少在分词过程中出现的疫病名称及短语等造成的歧义切分,进一步提高了分词准确率。实验结果表明,结合词典匹配的BERT-BiLSTM-CRF模型在羊常见疫病文本数据集上的F1值为96.38%,与jieba分词器、BiLSTM-Softmax模型、BiLSTM-CRF模型、未结合词典匹配的本文模型相比,分别提升11.01、10.62、8.3、0.72个百分点,验证了方法的有效性。与单一语料相比,通用语料PKU和羊常见疫病文本数据集结合的混合语料,能够同时对畜禽疫病专业术语及疫病文本中常用词进行准确切分,在通用语料及疫病文本数据集上F1值都达到95%以上,具有较好的模型泛化能力。该方法可用于畜禽疫病文本分词。
文摘为研究基于长短期记忆(Long Short-TermMemory,LSTM)网络的语音转文字系统的优化方法,首先说明LSTM在语音转文字任务中的基本原理和架构,其次分析自适应矩估计(Adaptive Moment Estimation,Adam)优化算法的核心机制及其在LSTM网络中的应用,最后在Mozilla DeepSpeech框架中嵌入基于Adam优化的LSTM模型,并使用THCHS-30数据集进行实验。实验结果表明,基于Adam优化的LSTM模型在词错率和F1分数上均表现出显著的优越性。
文摘Background: In the context of the fight against HIV, a lack of skills in monitoring and evaluating the personnel in charge of activities has been identified at the national level. It was the subject of a priority axis of the national plan for monitoring and evaluating the fight against HIV (2006-2010) that was aimed at strengthening the capacities of actors in this area. To increase the critical mass of competent human resources in the short term, the National Institute of Public Health (NIPH) of Côte d’Ivoire organized monitoring and evaluation training sessions for healthcare professionals from 2011 to 2016. Methods: A single case study with multiple levels of analysis was carried out, combining a qualitative survey and a literature review. An evaluation was carried out six months after each training session. In addition, the results of the pre- and post-tests and of the daily and final evaluations that accompanied the various training sessions were used to provide further information. The qualitative data collected were analyzed using INVIVO 15 software. Results: Some 89 health professionals (69% men and 31% women) working at the national level (51% at the central level, including 58% in health programs) and in development partner agencies (37%) participated in this capacity building program. Most participants were senior health managers (56%), data managers (23%), and statisticians and computer scientists (10%). Almost all the trainings were financed by 16 technical and financial partners (85%), mainly the MEASURE Evaluation project (27%). Conclusion: M&E training, despite all its imperfections, has made it possible to identify M&E training needs at the national level and to increase the critical mass of national skills and to have some culture in M&E.
文摘随着数字农业的快速发展,农作物命名实体识别作为农业领域知识图谱构建的基础,成为一种高效率的农作物研究领域识别方法。由于农作物实体识别呈现结构复杂、实体指称不一致、干扰因素多等特征,严重制约了农作物领域实体识别的性能,提出一种基于预训练语言模型的实体识别模型,使用BERT为文本中词进行编码、采用双向LSTM(Long-Short Term Memory)获取句子中关键词的上下文,采用CRFs(Conditional Random Fields)捕获词之间的依赖关系,并结合所构建的农作物命名实体识别数据集进行验证。实验证明该模型能够有效对农作物实体进行识别,且性能优于当前已有的实体识别模型。
