基于神经网络的端到端的事件指代消解研究

事件作为文本信息的关键语义组件,对篇章的理解具有重要意义。由于事件具有自身包含信息丰富、表达方式多样,以及在文本中分布稀疏等特点,使得事件指代消解成为自然语言处理领域的一个难点任务。在以往的事件指代消解任务中,多借助人工提取词匹配和句法结构等信息,再基于这些抽取的特征进行消解,然而这些特征并不能有效地处理具有复杂语义的事件任务。该文提出一种端到端的神经网络模型,通过多种词表征、双向循环神经网络和注意力机制来学习上下文的语义内容,从而完成事件的指代消解任务。在KBP2015、2016数据集上进行的事件指代消解实验验证了该文给出模型的有效性。实验结果表明,该神经网络模型能很好地理解文本语义信息,捕捉关键单词,提高事件指代消解任务的性能,最终端到端的事件指代消解性能在CoNLL评测标准下达到39.9%的F1值。     

多尺度模态感知在文本指代实例分割中的研究与应用

文本指代实例分割(RIS)任务是解析文本描述所指代的实例,并在对应图像中分割出该实例,是计算机视觉与媒体领域中热门的研究课题。当前,大多数RIS方法基于单尺度文本/图像模态信息的融合,以感知指代实例的位置和语义信息。然而,单一尺度模态信息很难同时涵盖定位不同大小实例所需的语义和结构上下文信息,阻碍了模型对任意大小指代实例的感知,进而影响模型对不同大小指代实例的分割。对此,设计多尺度视觉-语言交互感知模块和多尺度掩膜预测模块：前者增强模型对不同尺度实例语义与文本语义之间的融合与感知;后者通过充分捕捉不同尺度实例的所需语义和结构信息提升指代实例分割的表现。由此,提出了多尺度模态感知的文本指代实例分割模型(MMPN-RIS)。实验结果表明,MMPN-RIS模型在RefCOCO,RefCOCO+和RefCOCOg3个公开数据集的oIoU指标上均达到了前沿性能;针对文本指代不同尺度实例的分割,MMPN-RIS模型有着较好的表现。     

名词短语事件指代消解研究

随着篇章理解、信息融合要求的日益迫切,提高事件指代消解系统的性能逐渐成为一个研究热点。事件指代的指代词指代的是事件、事实或者命题等实践性、抽象性对象。该文对名词短语的事件指代消解进行研究,使用平面特征、结构化句法特征和语义特征等,根据SVM机器学习的方法进行英文事件的指代消解,通过在计算事件语义相似度的元组（语义角色）中加入时间和地点元素改进语义特征来提高事件指代消解系统的性能;并且单独使用每种特征对语料进行实验,分析每种特征单独使用时对系统的影响;OntoNotes 4.0语料库上的实验结果显示,引入改进的语义特征后,与基准系统相比,系统的准确率和F值均有所提高。由此来看,在语义特征中加入时间和地点元素对事件指代消解具有的正向作用。     

基于自然语言处理的车牌匹配技术

研究车牌的准确匹配识别问题。马路上车辆日益增多,车牌个性化的趋势也在蔓延,由于个性化车牌与常规车牌在字体、大小、颜色等特性存在差异,像素识别特征不明显,传统车牌识别方法无法有效解决个性化车牌与匹配库中车牌像素特征的差异,在库中根据车牌信息进行匹配识别存在识别准确率不高的问题。为了解决上述问题,提出了一种基于图像和自然语言处理相结合的车牌匹配识别方法。通过将车牌中的数字信息准确提取,转化成自然语言环境下的语义特征,由于语义特征不受像素因素影响,解决了像素特征不一致造成的漏识别问题,实现对车牌的准确识别。实验表明,利用改进算法实现了个性化车辆图像的正确识别,取得令人满意的效果。     

内容结构保持的图像风格迁移方法

现有的风格迁移算法大多是实现艺术作品到真实场景,如照片等的转换任务。在这种任务设定下,图像内容的边界等结构信息可能会出现一定程度的变化,然而这种变化不利于下游任务的完成,因而不适用于大部分工业场景。为此提出一种新的结构保持的风格迁移方法,该方法分为频域约束的图像生成模块和基于memory bank机制的语义匹配模块。前者用以保证转换前后图像的整体语义结构一致性,而后者保证了图像的语义与风格匹配,从而物体内部结构一致。为了验证该方法在工业场景中的迁移效果,采用光伏板数据集,在虚拟到真实的迁移方向上,所提出的风格迁移算法能在完成风格转换的同时较好地解决图像结构形变问题,进而满足后续任务的需求。     

基于局部深度匹配的行人再识别

针对行人再识别精度低的难题进行研究,提出了一种新的基于分块匹配的行人再识别方法。首先,引入带人体结构信息的人体DPM模型对行人外观进行分割,得到的带语义信息的身体部件作为匹配识别的基本单元;其次,基于深度神经网络模型提取各部件的深度特征作为匹配依据;再次,基于余弦距离判断各身体部件与目标行人对应部件的相似性;最后,融合所有身体部件的识别结果得到最终的再识别结果。实验结果表明,跟已有方法相比,本文方法具有更好的鲁棒性,在识别精度上有较明显的优势。     

基于模型的笔划抽取和匹配的手写体汉字识别方法

本文提出了一种基于模型结构匹配的手写汉字识别方法（HCCR）。这种方法可以得到可靠的笔划对应信息及结构解释。在这个模型中，每个类别的标准字符用属性形状描述，输入字符用特征点及线段描述，字符笔划与输入笔划间的关系直到被一个标准字符匹配后才能确定，这个结构的匹配分两个阶段：修选字符的提取和一致性匹配，所有匹配的修选字符先通过跟踪提取然后通过启发式搜索进行一致性匹配，同时一些结构的后处理操作应用于改善笔划对比的识别中，识别实验是在一个称为“KAIST”的字符数据库中进行的，并给出了结果。     

基于曲率匹配的细胞几何形状特征提取方法

细胞轮廓的几何形状是细胞学涂片判读的重要参考,对研究宫颈病变的计算机辅助诊断具有重要意义。针对现有基于形状模板匹配的几何形状识别方法鲁棒性较差的问题,提出了基于曲率匹配的几何形状特征提取方法,通过比较模板轮廓和待识别轮廓的曲率,计算曲率曲线之间的相似度,进而得到细胞轮廓的形状特征,并采用依次旋转轮廓选取最佳匹配的方法来解决轮廓方向不一致的问题,采用以面积等效圆的半径比作为放大比率进行轮廓缩放的方法来解决轮廓大小不一致的问题。通过相关实验证明了该方法所提取的几何形状特征具有尺度不变性和旋转不变性,并与改进Hausdorff距离进行了实验对比,结果表明提取的形状特征能更加准确地识别出细胞轮廓的几何形状。     

基于谱对称的形状配准方法

针对现有三维形状配准方法中存在左右翻转的错误匹配问题,提出了基于内蕴对称特征检测的高效形状配准算法。首先,通过热核与几何约束构建模型的内蕴自对称点对;其次,基于谱嵌入特征空间分析提取模型的内蕴对称平面,并依据模型表面法向量有效识别模型的左右结构属性;然后,根据内蕴对称点对获得模型的一致性谱对称结构描述;最后,引入一致性点漂移算法(CPD),实现基于谱对称的非刚性模型的形状配准,有效避免了模型配准中的左右结构翻转问题。实验进一步论证了这种方法不仅有效提高了模型匹配的效率,而且能有效识别同类模型的结构特征,对于非刚性模型的配准具有较强的鲁棒性。     

基于改进典型形状上下文特征的形状识别方法

针对形状上下文特征难以解决大规模样本的形状识别问题,提出一种利用角点典型形状上下文特征进行快速形状识别的方法.该方法仅以少数角点作为代表点生成直方图,对目标形状关键特征进行描述,通过减少匹配的特征数目降低了采样点匹配时间;在此基础上提出了局部约束匹配的方法,能够快速实现形状匹配并解决特征旋转不变性的问题,最终通过结合快速剪枝和精确匹配完成形状的识别.对形状数据进行仿真实验的结果证明,文中方法能够快速、有效地实现大规模数据的形状识别和检索.     

Matching performance of vehicle icons in graphical and textual formats

The current research classified 82 vehicle icons into seven categories (image-related, concept-related, semi-abstract, arbitrary, abbreviation, word, and combined) for their matching accuracy, matching sequence, and matching time. These data can be compared and used as a framework for future icon development. Forty participants, all with a university degree, took part in this experiment. Half of the participants had intensive driving experience, while the other half never driven a car. The results indicated that on average, word icons had a significantly greater matching accuracy than the other icon formats; ranging from 4.7 to 20.8% difference. Regarding the matching sequence, participants matched image-related icons before other icon formats. Arbitrary and combined icons took significantly longer to match than other icon formats by 1.4–6.2 s. Based on the high matching accuracy (86.3%) and high ratings on subjective design features, word format can be used for functions describable using simple English for users with English reading ability. Confusion matrices showed that 63.2% of the misunderstandings were caused by similarity in format or function.     

Development and validation of icons varying in their abstractness

Icons are used widely in human-computer interfaces. The level of abstractness-concreteness of an icon and its effect upon performance is of widespread interest. The authors have devised a quantitative measure for abstractness based on the complexity of the icon. They test their metric against subjective judgments of abstractness as identified by two different groups of subjects. After ranking two sets of ‘abstract’ and ‘concrete’ icons, the authors examined how well the icons were matched to the Pascal constructs that they represented. Further experiments were conducted using different groups of subjects to check whether correct matching of the icons with constructs was influenced by context. In summary the authors found that their metric was a good match for subjective measures of abstractness-concreteness. They also found that there is a better identification of concrete icons than abstract icons. Finally, it was shown that context does affect the correct identification of icons.     

Is visual content in textual search interfaces beneficial to dyslexic users?

Dyslexia is a learning disability characterised by problems with accurate or fluent word recognition, poor decoding, and poor spelling abilities. Although several studies have addressed dyslexia and Web accessibility, less is known about how dyslexia affects information search. This study investigated whether the inclusion of icons in search user interfaces enhances performance among dyslexics. A total of 21 dyslexics and 21 controls completed 52 search tasks in 4 conditions: icons only, words only, and both icons and words in a grid layout and a list layout, while eye movements were recorded. Dyslexics took significantly longer than controls to locate targets in tasks containing text, but not in the icon-only condition. Dyslexics had longer fixation durations than controls in both icon and text based search arrays, suggesting higher mental load associated with search tasks generally. The addition of words to icon arrays led to faster search times within controls, but not dyslexics. Dyslexics also exhibited more fixations on dual-modality tasks, and longer scanpaths than controls in list layout. Both groups were fastest searching the list layout, with icons and words listed in columns. Results are discussed in terms of the design of accessible search interfaces for dyslexic users, taking into account mental load of dual-modality information display, and the arrangement of search items. Empirical data is provided for the design of accessible search results interfaces for dyslexics.     

An experimental research study on the effect of pictorial icons on a user-learner’s performance

This experimental counterbalanced repeated measures study on the effect of three types of icons (abstract, pictorial and drawing) on the learning and performance of two types of learners (abstract and concrete) was conducted during the Spring 2000–Spring 2001 semesters at the University of Central Florida. The counterbalanced lesson followed by a counterbalanced and completely randomized quiz on three levels was field tested in the Spring–Summer 2000 on 37 UCF graduate students and final tested in the Fall 2000–Spring 2001 on 53 UCF graduate student subjects. A general linear model repeated measures ANOVA revealed that icon type had a significant effect on the learning and performance of both types of learners. Pictorial icons resulted in best scores. No significant interaction was found between icon type and learner type even though on a matched condition, abstract learners with abstract icon resulted in better scores than when concrete learners were matched with concrete (pictorial and drawing) icons. Implications for future research are drawn as well as practical applications identified in the field of teaching, learning, training and performance.     

EvIcon: Designing High-Usability Icon with Human-in-the-loop Exploration and IconCLIP

Interface icons are prevalent in various digital applications. Due to limited time and budgets, many designers rely on informal evaluation, which often results in poor usability icons. In this paper, we propose a unique human-in-the-loop framework that allows our target users, that is novice and professional user interface (UI) designers, to improve the usability of interface icons efficiently. We formulate several usability criteria into a perceptual usability function and enable users to iteratively revise an icon set with an interactive design tool, EvIcon. We take a large-scale pre-trained joint image-text embedding (CLIP) and fine-tune it to embed icon visuals with icon tags in the same embedding space (IconCLIP). During the revision process, our design tool provides two types of instant perceptual usability feedback. First, we provide perceptual usability feedback modelled by deep learning models trained on IconCLIP embeddings and crowdsourced perceptual ratings. Second, we use the embedding space of IconCLIP to assist users in improving icons' visual distinguishability among icons within the user-prepared icon set. To provide the perceptual prediction, we compiled IconCEPT10K, the first large-scale dataset of perceptual usability ratings over 10,000 interface icons, by conducting a crowdsourcing study. We demonstrated that our framework could benefit UI designers' interface icon revision process with a wide range of professional experience. Moreover, the interface icons designed using our framework achieved better semantic distance and familiarity, verified by an additional online user study.     

An algorithm for the learning of weights in discriminationfunctions using a priori constraints

We introduce a learning algorithm for the weights in a very common class of discrimination functions usually called “weighted average.” The learning algorithm can reduce the number of free variables by simple but effective a priori criteria about significant features. Here we apply our algorithm to three tasks of different dimensionality all concerned with face recognition     

Influence mechanism of icon semantics on visual search performance: Evidence from an eye-tracking study

Explicit icon semantics can reduce the difficulty of understanding complex visual information. Optimizing the icon semantics and text semantics of icons can effectively improve the cognitive performance of digital interfaces. This paper adopts visual search tasks to study the effects of different combinations of icon semantic familiarity and the presence or absence of text on icon search performance under horizontal and vertical layouts. The behavioral experiment results show that under two layouts: 1. The main effect of icon semantics is significant, and the search performance increases with the increase of semantic familiarity. 2. The main effect of text is significant, and the search performance is negatively correlated with the addition of text. The eye movement experiment found that the semantic familiarity of icons had a significant impact on average fixation time. Furthermore, the number of fixation points changed significantly after the text variable was added. Therefore, there was no significant difference in the number of fixation points in the horizontal layout, and icon semantics was the main influencing factor in visual search. In the vertical layout, there was no significant difference in average fixation time, and text was the main influencing factor of visual search. The results show that the semantic familiarity of icons and different combinations with or without text significantly affect visual search performance in horizontal and vertical layouts. This paper provides a theoretical reference for the combination of icons and text in interface design.     

Learning and retention of associations between auditory icons and denotative referents: implications for the design of auditory warnings

OBJECTIVE: This study examined the way in which the type and preexisting strength of association between an auditory icon and a warning event affects the ease with which the icon/event pairing can be learned and retained. BACKGROUND: To be effective, an auditory warning must be audible, identifiable, interpretable, and heeded. Warnings consisting of familiar environmental sounds, or auditory icons, have potential to facilitate identification and interpretation. The ease with which pairings between auditory icons and warning events can be learned and retained is likely to depend on the type and strength of the preexisting icon/event association. METHOD: Sixty-three participants each learned eight auditory-icon/denotative-referent pairings and attempted to recall them 4 weeks later. Three icon/denotative-referent association types (direct, related, and unrelated) were employed. Participants rated the strength of the association for each pairing on a 7-point scale. RESULTS: The number of errors made while learning pairings was greater for unrelated than for either related or direct associations, whereas the number of errors made while attempting to recall pairings 4 weeks later was greater for unrelated than for related associations and for related than for direct associations. Irrespective of association type, both learning and retention performance remained at very high levels, provided the strength of the association was rated greater than 5. CONCLUSION: This suggests that strong preexisting associations are used to facilitate learning and retention of icon/denotative-referent pairings. APPLICATION: The practical implication of this study is that auditory icons having either direct or strong, indirect associations with warning events should be preferred.     

构建基于OBE的立体化制图教学新体系

以机械工程大类的“机械制图”课程为载体,以成果导向教育(OBE)理念为牵引,以 课内教学为主线,整合和挖掘课外教学资源,依托“互联网+”的技术手段,搭建了基于主动学习 能力培养的“线上线下融合式教学平台”;构建了“课内与课外相结合、线上与线下相结合、必修 课与开放实验课相结合、教师与学生相结合、理论与实践相结合”的全方位立体化制图教学新体 系。经过 7 年的教学实践及逐步完善,适度解决了课内学时有限、生师比高、实践教学欠缺、 课外辅导力量不足、学生学习兴趣不高等长期困扰工程类基础课教学中的难题,收到了良好的 教学效果,为工程类基础课程教学提供改革思路。     

Asymptotic model matching for nonlinear systems

This paper investigates the problem of designing a compensating control law for a square invertible nonlinear plant so that the response of the closed-loop system asymptotically matches that of a prescribed, driven, nonlinear model. A set of necessary conditions for achieving asymptotic model matching is established. One of these involves the stability properties for subdynamics which are common to every model matching closed loop. These subdynamics, called the “fixed dynamics,” are intrinsically characterized. Based on these results, a new set of sufficient conditions for achieving asymptotic model matching is given. The interrelation between the minimum-phase and vector relative degree properties of a plant and a matchable model are studied