CHIP2020评测任务1概述:中文医学文本命名实体识别

第六届中国健康信息处理会议(China Conference on Health Information Processing,CHIP2020)组织了中文医疗信息处理方面的6个评测任务,其中任务1为中文医学文本命名实体识别任务,该任务的主要目标是自动识别医学文本中的医学命名实体。共有253支队伍报名参加评测,最终37支队伍提交了80组结果,该评测以微平均F₁值作为最终评估标准,提交结果中最高值达68.35%。     

面向对话的融入交互信息的实体关系抽取

实体关系抽取旨在从文本中抽取出实体之间的语义关系,是自然语言处理的一项基本任务。在新闻报道、维基百科等规范文本上,该任务的研究相对丰富且已取得了一定的效果,但面向对话文本的相关研究还处于起始阶段。相较于规范文本,对话是一个交互的过程,大量信息隐藏在交互中,这使得面向对话文本的实体关系抽取更具挑战性。依据对话的特点,该文提出了融入对话交互信息的实体关系抽取方法,通过交叉注意力机制获取对话交互信息,提升性能,并结合多任务学习来解决语料库数据分布不均衡的问题。在DialogRE公开数据集上实验得到,F₁值为54.1%,F_1c值为50.7%,证明了该方法的有效性。     

CHIP2019评测任务1概述:临床术语标准化任务

第五届中国健康信息处理会议(China Conference on Health Information Processing, CHIP2019)组织了中文临床医疗信息处理方面的三个评测任务,其中任务1为临床术语标准化任务。该任务的主要目标是对中文电子病历中挖掘出的真实手术实体进行语义标准化。评测数据集中所有手术原词均来自于真实医疗数据,并以《ICD9-2017协和临床版》手术词表为标准进行了标注。共有56支队伍报名参加了评测,最终有20支队伍提交了47组结果。该评测以准确率作为最终评估标准,提交结果中最高准确率达到94.83％。     

面向司法案件的案情知识图谱自动构建

以法学知识为中心的认知智能是当前司法人工智能发展的重要方向。该文提出了以自然语言处理(NLP)为核心技术的司法案件案情知识图谱自动构建技术。以预训练模型为基础,对涉及的实体识别和关系抽取这两个NLP基本任务进行了模型研究与设计。针对实体识别任务,对比研究了两种基于预训练的实体识别模型;针对关系抽取任务,该文提出融合平移嵌入的多任务联合的语义关系抽取模型,同时获得了结合上下文的案情知识表示学习。在“机动车交通事故责任纠纷”案由下,和基准模型相比,实体识别的F₁值可提升0.36,关系抽取的F₁值提升高达2.37。以此为基础,该文设计了司法案件的案情知识图谱自动构建流程,实现了对数十万份判决书案情知识图谱的自动构建,为类案精准推送等司法人工智能应用提供语义支撑。     

基于BiLSTM-CRF的古汉语自动断句与词法分析一体化研究

古汉语信息处理的基础任务包括自动断句、自动分词、词性标注、专名识别等。大量的古汉语文本未经标点断句,所以词法分析等任务首先需要建立在断句基础之上。然而,分步处理容易造成错误的多级扩散,该文设计实现了古汉语断句与词法分析一体化的标注方法,基于BiLSTM-CRF神经网络模型在四种跨时代的测试集上验证了不同标注层次下模型对断句、词法分析的效果以及对不同时代文本标注的泛化能力。研究表明,一体化的标注方法对古汉语的断句、分词及词性标注任务的F₁值均有提升。综合各测试集的实验结果,断句任务F₁值达到78.95%,平均提升了3.5%;分词任务F₁值达到85.73%,平均提升了0.18%;词性标注任务F₁值达到72.65%,平均提升了0.35%。     

基于位置降噪和丰富语义的电子病历实体关系抽取

当前的电子病历实体关系抽取方法存在两个问题: 忽视了位置向量噪声和语义表示匮乏。该文提出一种基于位置降噪和丰富语义的实体关系抽取模型。模型首先利用位置信息和专业领域语料训练的词向量信息获取每个词的注意力权重,然后将此权重与通用领域语料训练的词向量结合,实现位置向量降噪和丰富语义引入,最后根据加权后的词向量判断实体关系类型。该方法在2010年i2B2/VA语料上进行实验评估,F₁值为76.47%,取得了基于该语料的最优结果。     

基于混合神经网络的实体关系抽取方法研究

实体关系抽取是信息抽取领域的重要研究内容,对知识库的自动构建起着至关重要的作用。针对非结构化文本实体关系抽取存在上下文环境信息难以准确表征,致使现有抽取模型准确率不能满足实际应用需求的问题,该文提出了一种新型的实体关系抽取模型BiGRU-Att-PCNN。该模型是基于混合神经网络,首先,构建双向门控循环单元(BiGRU)以更好地获取文本序列中的上下文语序的相关信息;然后,采用注意力(Attention)机制来达到自动关注对关系影响力高的序列特征的目的;最后,通过采用分段卷积神经网络(PCNN),从调整后的序列中较好地学习到了相关的环境特征信息来进行关系抽取。该模型在公开的英文数据集SemEval 2010 Task 8上取得了86.71%的F₁值,实验表明,该方法表现出了较好的性能,为信息抽取领域实体关系的自动获取提供了新的方法支持。     

基于图注意力卷积神经网络的文档级关系抽取

关系抽取作为信息抽取的子任务,旨在从非结构化文本中抽取出便于处理的结构化知识,对于自动问答、知识图谱构建等下游任务至关重要。该文在文档级的关系抽取语料上开展工作,包括但不局限于传统的句子级关系抽取。为了解决文档级关系抽取中长距离依赖问题,并且对特征贡献度加以区分,该文将图卷积模型和多头注意力机制相融合构建了图注意力卷积模型。该模型通过多头注意力机制为同指、句法等信息构建的拓扑图构建动态拓扑图,然后使用图卷积模型和动态图捕获实体间的全局和局部依赖信息。该文分别在DocRED语料和自主扩展的ACE 2005语料上进行实验,与基准模型相比,基准模型上融入图注意力卷积的模型在两个数据集上的F₁值分别提升了2.03％和3.93％,实验结果表明了该方法的有效性。     

基于BERT的心血管医疗指南实体关系抽取方法

实体关系抽取是医疗领域知识问答、知识图谱构建及信息抽取的重要基础环节之一。针对在心血管专病知识图谱构建的过程中尚无公开数据集可用的情况,收集了心血管疾病领域的医疗指南并进行相应的实体和关系类别的专业标注,构建了心血管专病知识图谱实体关系抽取的专业数据集。基于该数据集,首先提出双向变形编码器卷积神经网络（BERT-CNN）模型以实现中文语料中的关系抽取,然后根据中文语义中主要以词而不是字为基本单位的特性,提出了改进的基于全词掩模的双向变形编码器卷积神经网络（BERT（wwm）-CNN）模型用于提升在中文语料中关系抽取的性能。实验结果表明,改进的BERT（wwm）-CNN在所构建的关系抽取数据集上准确率达到0.85,召回率达到0.80,F₁值达到0.83,优于对比的基于双向变形编码器长短期记忆网络（BERT-LSTM）模型和BERT-CNN模型,验证了改进网络模型的优势。     

基于笔画中文字向量模型设计与研究

中文汉字在横向、纵向展开具有二维的复杂结构。现有的中文词向量研究大都止步于汉字字符,没有利用中文笔画序列生成字向量,且受限于统计模型本质,无法为低频、未登录字词生成高质量向量表示。为此,该文提出了一种依靠中文笔画序列生成字向量的模型Stroke2Vec,扩展Word2Vec模型CBOW结构,使用卷积神经网络替换上下文信息矩阵、词向量矩阵,引入注意力机制,旨在模拟笔画构造汉字的规律,通过笔画直接生成字向量。将Stroke2Vec模型与Word2Vec、GloVe模型在命名实体识别任务上进行评测对比。实验结果显示,Stroke2Vec模型F₁值达到81.49%,优于Word2Vec 1.21%,略优于GloVe模型0.21%,而Stroke2Vec产生的字向量结合Word2Vec模型结果,在NER上F₁值为81.55%。     

SemEval-2010 task 18: disambiguating sentiment ambiguous adjectives

Sentiment ambiguous adjectives, which have been neglected by most previous researches, pose a challenging task in sentiment analysis. We present an evaluation task at SemEval-2010, designed to provide a framework for comparing different approaches on this problem. The task focuses on 14 Chinese sentiment ambiguous adjectives, and provides manually labeled test data. There are 8 teams submitting 16 systems in this task. In this paper, we define the task, describe the data creation, list the participating systems, and discuss different approaches.     

Language independent search in MediaEval's Spoken Web Search task

In this paper, we describe several approaches to language-independent spoken term detection and compare their performance on a common task, namely “Spoken Web Search”. The goal of this part of the MediaEval initiative is to perform low-resource language-independent audio search using audio as input. The data was taken from “spoken web” material collected over mobile phone connections by IBM India as well as from the LWAZI corpus of African languages. As part of the 2011 and 2012 MediaEval benchmark campaigns, a number of diverse systems were implemented by independent teams, and submitted to the “Spoken Web Search” task. This paper presents the 2011 and 2012 results, and compares the relative merits and weaknesses of approaches developed by participants, providing analysis and directions for future research, in order to improve voice access to spoken information in low resource settings.     

Comparing the task effectiveness of instant messaging and electronic mail for geographically dispersed teams in Taiwan

To improve task effectiveness of geographically dispersed teams, media selection approaches a complex issue involving media features and user characteristics. This study compares user perceptions about task effectiveness between instant messengers (IM) and e-mail (EM). Using a field experiment, the design contrasted teams using EM with teams using IM. Thirty-four teams of two graduate students were enrolled as participants from five universities in Taiwan. Two people were separated geographically and had to communicate accordingly. Results indicate that IM-supported teams generated more ideas and task success. However, no significant difference was found in task difficulty between the IM and EM users.     

Guanxi Networks and Creativity in Taiwanese Project Teams

Personal connections, also called guanxi in Chinese culture, are embedded within a network of interdependent social exchanges, providing access to resources, information and supports, therefore facilitating performance. Building guanxi networks means building social networks. Project teams are commonly used as fundamental units in organizations, partly in order to generate more creative ideas, and so this paper evaluates how project teams use their guanxi networks within and outside of their teams and across boundaries to stimulate their creative performance. Fifty‐four product development project teams with 293 members from high‐tech Taiwanese firms participated in this study. Findings suggested that, in a guanxi culture such as Taiwan's, in order to facilitate a project team's creative performance, the most important task is to develop a more cohesive social unit before the team extends its external networks, since Chinese culture regards people first as members of groups, and secondly as individuals.     

基于核函数中文关系自动抽取系统的实现

实体关系抽取是信息抽取的重要组成部分.基于核函数的中文实体关系自动抽取系统应用改进的语义序列核函数,结合KNN机器学习算法构造分类器来分类并标注关系的类型.通过对ACE评测定义的三大类6子类实体关系的抽取,关系抽取的平均精度可以达到88%,明显高于基于特征向量和传统的序列核函数方法,该方法适合小训练集,易于学习新的实体关系.系统由8个独立的模块构成,便于维护和升级.系统既可以独立运行,也可以嵌入在开放的文本处理平台GATE环境.为了更好地利用关系抽取的结果,系统扩展传统的二元关系,抽取关系的同时,抽取该关系的描述,形成完整的中文实体关系抽取系统.     

汉语“比”字句关键要素的常规序列模式探索

表达“差比”义的“比”字句,是比较句的主要句型,也是比较句关键要素抽取研究中不可回避的主要课题。该句型的关键要素(SUB、BI、OBJ、ITM、DIM、RES、EXT)在语义上互相交织,在表层句法上可以实现为多种多样的序列模式。该文面向中文“比”字句关键要素抽取这个目标,对于表示“差比”义的460多个“比”字句文本进行了七种关键要素的标注。在此基础上,利用Apriori和PrefixSpan算法找出这些要素的关联规则及其序列模式,并归纳出六种“比”字句关键要素的分布规律。此外,该文还进一步说明了产生这六种模式规则的动因,为“比”字句特征选取和处理提供了重要的语言学理论依据。     

Group development in virtual teams: An experimental reexamination

Virtual teams are thought to be experienced differently and to have poor outcomes because there is little or no face-to-face interaction and a tendency for virtual team members to use different communication techniques for forming relationships. However, the expanding use of virtual teams in organizations suggests that virtual teams in real world contexts are able to overcome these barriers and be experienced in much the same way as face-to-face teams. This paper reports the result of an experiment in which virtual teams participated in an exercise where they completed an information-sharing task ten times as a team. The results suggest that, contrary to one-shot, ad hoc virtual teams, longer-lived virtual teams follow a sequential group development process. Virtual team development appears to differ from face-to-face teams because the use of computer-mediated communication heightens pressure to conform when a virtual team is first formed, meaning trust is most strongly linked with feeling that the team was accomplishing the task appropriately. As the virtual teams developed, trust in peers was more strongly linked with goal commitment. Once the teams were working together effectively, accomplishing the task appropriately was the strongest link with trust in peers. I suggest that virtual team managers should cultivate virtual workspaces that are similar to those proven to work in face-to-face contexts: (1) teams should have clear, specific goals, (2) members should be encouraged or even required to communicate with each other, and (3) team members should feel that they might work with the other team members again.     

Adaptive coordination development in student anaesthesia teams: a longitudinal study

Although adaptive coordination has been highlighted by several studies, research dealing with how adaptive coordination develops is still rare. Thus, the aim of this study was to investigate the development of coordination mechanisms and their task-related adaptation in a longitudinal observation of medical simulation-based training of final year students. We recorded six anaesthesia teams during a sequence of four task scenarios, and each scenario comprised of a routine and a complication phase. After trained observers rated sub-tasks within each scenario for explicit and implicit coordination, statistical analysis revealed a statistically significant effect of previous scenarios on coordination development in the routine phases. While the amount of explicit coordination decreased, implicit coordination increased, revealing adaptive coordination as a skill developed through repeated group interaction. We conclude that anaesthesia training should consider cost- and patient safety-benefits of implicit and explicit coordination and focus on adaptive coordination. PRACTITIONER SUMMARY: Group coordination is crucial to anaesthesia team performance. Results of this longitudinal observation of six anaesthesia teams during four medical simulation-based training scenarios document that teams develop adaptive patterns of coordination. This study also demonstrates that adaptive coordination is a trainable skill within crisis resource management training.