Interactive Attention Model Explorer for Natural Language Processing Tasks with Unbalanced Data Sizes

Dong, Zhihang; Wu, Tongshuang; Song, Sicheng; Zhang, Mingrui

doi:10.1109/pacificvis48177.2020.1031

Cited by 10 publications

(6 citation statements)

References 15 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…We observe heterogeneous visual encodings adopted to support feature attributions. The most used ones are heatmaps for local post hoc feature attributions on images (Figure 7b) [vdBCR*20, HSL*21, HJZ*21, ZZM16, HCC*20, WGZ*19, SW17, CBN*20, JVW20] and text [CHS20, CGR*17, ŠSE*21, JTH*21]; matrices [WONM18, DWB21, JKV*22, PCN*19, LLL*19], node‐link diagrams [JTH*21, Vig19, JCM20, LLL*19] and custom Sankey diagrams [DWSZ20, PCN*19, HSG20, MSHB22] for self‐explainable attentive models; bar charts[WWM20, PCN*19] or averaged inputs [WGYS18, WGSY19] for global feature attribution; and enhanced line [SMM*19, CWGvW19, LYY*20, SWJ*20, ŠSE*21], area chart [KCK*19] or bar chart [MXC*20, KCK*19, SWJ*20, WWM20] for post hoc approaches to sequential data. Among them, systems that support the analysis of attentive models, and in particular of Transformers, employ the most complex and novel visualization techniques (Figure 7a) such as radial layouts [WTC21, DWB21] or grid ones [WTC21, DWB21, ŠSE*21]: these systems must show the flow of attention weights across multiple layers simultaneously to help the user understand the most important features.…”

Section: Papers Categorizationmentioning

confidence: 99%

“…Examples of added information are the attribution scores’ magnitude, which is encoded using colours, size [JCM20], opacity [WSP*21] or just its value, or bounding boxes [HSL*21, JKV*22, CBN*20], which highlight the most important region over images. While sorting [CHS20, MXC*20, KCK*19, Vig19, PCN*19] and filtering [WONM18, DWSZ20, JKV*22, JTH*21] by attribution scores capabilities are quite common to ease the data exploration and reduce the visual clutters, some works provide additional tools for a deeper understanding. For example, feature removal interactions guided by local attributions [HSL*21, HJZ*21, KCK*19] (e.g.…”

Section: Papers Categorizationmentioning

confidence: 99%

See 1 more Smart Citation

State of the Art of Visual Analytics for eXplainable Deep Learning

Rosa

Blasilli

Bourqui

et al. 2023

Computer Graphics Forum

View full text Add to dashboard Cite

The use and creation of machine-learning-based solutions to solve problems or reduce their computational costs are becoming increasingly widespread in many domains. Deep Learning plays a large part in this growth. However, it has drawbacks such as a lack of explainability and behaving as a black-box model. During the last few years, Visual Analytics has provided several proposals to cope with these drawbacks, supporting the emerging eXplainable Deep Learning field. This survey aims to (i) systematically report the contributions of Visual Analytics for eXplainable Deep Learning; (ii) spot gaps and challenges; (iii) serve as an anthology of visual analytical solutions ready to be exploited and put into operation by the Deep Learning community (architects, trainers and end users) and (iv) prove the degree of maturity, ease of integration and results for specific domains. The survey concludes by identifying future research challenges and bridging activities that are helpful to strengthen the role of Visual Analytics as effective support for eXplainable Deep Learning and to foster the adoption of Visual Analytics solutions in the eXplainable Deep Learning community. An interactive explorable version of this survey is available online at https:// awarediag-sapienza.github.io/ VA4XDL.

show abstract

Section: Papers Categorizationmentioning

confidence: 99%

Section: Papers Categorizationmentioning

confidence: 99%

State of the Art of Visual Analytics for eXplainable Deep Learning

Rosa

Blasilli

Bourqui

et al. 2023

Computer Graphics Forum

View full text Add to dashboard Cite

show abstract

“…In this respect, tedious labeling of data set will be eliminated; thus, large unlabeled bug repositories can be utilized to train models in an unsupervised fashion for improved generalization. Another model that could be used to improve imbalance learning is the attention mechanism (Dong, Wu, Song and Zhang, 2020).…”

Section: Bug Localizationmentioning

confidence: 99%

Deep Bug Reports Processing (DBRP): A Systematic Literature Review

Ahmad

Kholief

et al. 2023

Preprint

View full text Add to dashboard Cite

Many software projects utilize Bug Tracking System (BTS) to manage and process bug reports. Over the years, the number of bug report submissions has increased exponentially with some projects receiving as many as about a hundred submissions daily. Bug report processing (BRP) consists of five key processes: duplicate detection, severity prediction, fix-time prediction, bug triage, and bug localization. Previously, traditional machine learning (ML) algorithms -based models were proposed to automate BRP tasks. However, in recent years deep BRP (DBRP) models were proposed to exploit the ever-increasing bug repositories for automatic extraction of semantic and contextual features of bug reports. Although, some papers reviewed related literature from many perspectives of software maintenance, no existing work comprehensively reviewed the state of DBRP. In this paper, we review the state of DBRP models in the five key BRP tasks. For this, we collect papers from four international databases published between 2015 and 2021. We evaluate the papers with regard to the deep neural networks, text representation models, and bug report features utilized. Finally, we present findings and prospects for future research works in DBRP. Our review analyzes how word embedding, CNNs, LSTMs, attention mechanism are used to represent bug report and source code, model performance and challenges such as data imbalance, feature utilization and complexity.

show abstract

“…Existing visualization methods for visually understanding machine learning models can be categorized into two classes: domain irrelevant [17], [18] and domain specific [19], [20], [21], [22], [23], [24]. Our work is in the second category with a focus in the NLP domain, which enables an in-depth understanding of the working mechanism of the model training process.…”

Section: Visualization For Understanding Nlp Modelsmentioning

confidence: 99%