DeLTA 2022 Abstracts

Area 1 - Big Data Analytics

Short Papers

Paper Nr:	30
Title:	Structural Damage Localization via Deep Learning and IoT Enabled Digital Twin
Authors:	Marco Parola, Federico A. Galatolo, Matteo Torzoni, Mario A. Cimino and Gigliola Vaglini
Abstract:	Structural Health Monitoring (SHM) of civil structures using IoT sensors is a major emerging challenge. SHM aims to detect and identify any deviation from a reference condition, typically a damage-free baseline, to keep track of the relevant structural integrity. Machine Learning (ML) techniques have recently been employed to empower vibration-based SHM systems. Supervised ML can provide more information than unsupervised ML, but it requires human intervention to appropriately label data describing the nature of the damage. However, labelled data related to damage conditions of civil structures are often unavailable. To overcome this limitation, a key solution is a Digital Twin relying on physics-based numerical models to simulate the structural response in terms of the vibration recordings provided by IoT devices during the events of interest, such as wind or seismic excitations. This paper presents such comprehensive approach to address the damage localization task by exploiting a Convolutional Neural Network (CNN). Early experimental results related to a pilot application involving a sample structure, show the potential of the proposed approach and the reusability of the trained system in presence of varying loading scenarios.
Download

Area 2 - Computer Vision Applications

Full Papers

Paper Nr:	25
Title:	RoSELS: Road Surface Extraction for 3D Automotive LiDAR Point Cloud Sequence
Authors:	Dhvani Katkoria and Jaya Sreevalsan-Nair
Abstract:	Road surface geometry provides information about navigable space in autonomous driving. Ground plane estimation is done on “road” points after semantic segmentation of three-dimensional (3D) automotive LiDAR point clouds as a precursor to this geometry extraction. However, the actual geometry extraction is less explored, as it is expensive to use all “road” points for mesh generation. Thus, we propose a coarser surface approximation using road edge points. The geometry extraction for the entire sequence of a trajectory provides the complete road geometry, from the point of view of the ego-vehicle. Thus, we propose an automated system, RoSELS (Road Surface Extraction for LiDAR point cloud Sequence). Our novel approach involves ground point detection and road geometry classification, i.e. frame classification, for determining the road edge points. We use appropriate supervised and pre-trained transfer learning models, along with computational geometry algorithms to implement the workflow. Our results on SemanticKITTI show that our extracted road surface for the sequence is qualitatively and quantitatively close to the reference trajectory.
Download

Short Papers

Paper Nr:	11
Title:	Creating an Automatic Road Sign Inventory System using a Fully Deep Learning-based Approach
Authors:	Gabriele Galatolo, Matteo Papi, Andrea Spinelli, Guglielmo Giomi, Andrea Zedda and Marco Calderisi
Abstract:	Some road sections are a veritable forest of road signs: just think how many indications you can come across on an urban or extra-urban route, near a construction site or a road diversion. The automatic recognition of vertical traffic signs is an extremely useful task in the automotive industry for many practical applications, such as supporting the driver while driving with an in-car advisory system or the creation of a register of signals for a particular road section to speed up maintenance and replacement of installations. Recent developments in deep learning have brought huge progress in the image processing area, which triggered successful applications like traffic sign recognition (TSR). The TSR is a specific image processing task in which real traffic scenes (images or frames from videos taken from vehicle cameras in uncontrolled lighting and occlusion conditions) are processed in order to detect and recognize traffic signs within it. Traffic Sign Recognition is a very recent technology facilitated by the Vienna Convention on Road Signs and Signals of 1968: during that international meeting, it was decided to standardize traffic signs so that they could be recognised more easily abroad. Finally, this work summarizes our proposal of a practical pipeline for the development of an automatic traffic sign recognition software.
Download

Paper Nr:	27
Title:	A Lightweight Machine Learning Pipeline for LiDAR-simulation
Authors:	Richard Marcus, Niklas Knoop, Bernhard Egger and Marc Stamminger
Abstract:	Virtual testing is a crucial task to ensure safety in autonomous driving, and sensor simulation is an important task in this domain. Most current LiDAR simulations are very simplistic and are mainly used to perform initial tests, while the majority of insights are gathered on the road. In this paper, we propose a lightweight approach for more realistic LiDAR simulation that learns a real sensor’s behavior from test drive data and transforms this to the virtual domain. The central idea is to cast the simulation into an image-to-image translation problem. We train our pix2pix based architecture on two real world data sets, namely the popular KITTI data set and the Audi Autonomous Driving Dataset which provide both, RGB and LiDAR images. We apply this network on synthetic renderings and show that it generalizes sufficiently from real images to simulated images. This strategy enables to skip the sensor-specific, expensive and complex LiDAR physics simulation in our synthetic world and avoids oversimplification and a large domain-gap through the clean synthetic environment.
Download

Paper Nr:	9
Title:	Real-time Distance Measurement in a 2D Image on Hardware with Limited Resources for Low-power IoT Devices (Radar Control System)
Authors:	Jurij Kuzmic and Günter Rudolph
Abstract:	This paper presents an approach for real-time distance measurement in a 2D image on hardware with limited resources without a reference object. Additionally, different approximated functions for distance measurement are presented. Here, we focus on an approach to develop real-time distance detection for hardware with limited resources in the field of the Internet of Things (IoT). Also, our distance measurement system is evaluated with simulated data, real data from model making area and data from a real vehicle from real environment. In the beginning, related work of this paper is discussed. The data acquisition of the different simulated and real data sets is also discussed in this paper. Additionally, dissimilar resolutions for distance measurement are compared in accuracy and run time to find the better and faster system for distance measurement in a 2D image on hardware with limited resources for low-power IoT devices. Through the experiments described in this paper, the comparison of the run time depending on different IoT hardware is presented. Here, the idea is to develop a radar control system for self-driving cars from model making area and vehicles from real environment. Finally, future research and work in this area are discussed.
Download

Paper Nr:	13
Title:	RRConvNet: Recursive-residual Network for Real-life Character Image Recognition
Authors:	Tadele Mengiste, Birhanu H. Belay, Bezawork Tilahun, Tsiyon Worku and Tesfa Tegegne
Abstract:	Variations in fonts, styles, and ways to write a character have been the major bottlenecks in OCR research. Such problems are swiftly tackled through advancements in deep neural networks (DNNs). However, the number of network parameters and feature reusability are still the issues when applying Deep Convolutional Neural networks(DCNNs) for character image recognition. To address these challenges, in this paper, we propose an extensible and recursive-residual ConvNet architecture (RRConvNet) for real-life character image recognition. Unlike the standard DCCNs, RRConvNet incorporates two extensions: recursive-supervision and skip-connection. To enhance the recognition performance and reduce the number of parameters for extra convolutions, layers of up to three recursions are proposed. Feature maps are used after each recursion for reconstructing the target character. For all recursions of the reconstruction method, the reconstruction layers are the same. The second enhancement is to use a short skip-connection from the input to the reconstruction output layer to reuse the character features maps that are already learned from the prior layer. This skip-connection could be also used as an alternative path for gradients where the gradient is too small. With an overall character recognition accuracy of 98.2 percent, the proposed method achieves a state-of-the-art result on both publicly available and private test datasets.
Download

Paper Nr:	28
Title:	Analysis of Ensemble of Neural Networks and Fuzzy Logic Classification in Process of Semantic Segmentation of Martian Geomorphological Settings
Authors:	Kamil Choromański, Joanna Kozakiewicz, Mateusz Sobucki, Magdalena Pilarska-Mazurek and Robert Olszewski
Abstract:	Deep learning analysis of multisource Martian data (both from orbiter and rover) allows for the separation and classification of different geomorphological settings. However, it is difficult to determine the optimal neural network model for unambiguous semantic segmentation due to the specificity of Martian data and blurring of the boundary of individual settings (which is its immanent property). In this paper, the authors describe several variants of multisource deep learning processing system for Martian data and develop a methodology for semantic segmentation of geomorphological settings for this planet based on the combination of selected solutions output. Network ensemble with use of the weighted averaging method improved results comparing to single network. The paper also discusses the decision rule extraction method of individual Martian geomorphological landforms using fuzzy inference systems. The results obtained using FIS tools allow for the extraction of single geomorphological forms, such as ripples.
Download

Paper Nr:	29
Title:	Automatic UML Defects Detection based on Image of Diagram
Authors:	Murielle S. Lokonon and Vinasetan R. Houndji
Abstract:	Unified Modeling Language (UML) is a standardized modeling language used to design software systems. However, software engineering learners often have difficulties understanding UML and often repeat the same mistakes. Several solutions automatically correct UML diagrams. These solutions are generally restricted to the modeling tool used or need teachers’ intervention for providing exercises, answers, and other rules to consider for diagrams corrections. This paper proposes a tool that allows the automatic correction of UML diagrams by taking an image as input. The aim is to help UML practicers get automatic feedback on their diagrams regardless of how they have represented them. We have conducted our experiments on the use case diagrams. We have first built a dataset of images of the most elements encountered in the use case diagrams. Then, based on this dataset, we have trained some machine learning models using the Detectron2 library developed by Facebook AI Research (FAIR). Finally, we have used the model with the best performances and a predefined list of errors to set up a tool that can syntactically correct any use case diagram with relatively good precision. Thanks to its genericity, the use of this tool is easier and more practical than the state-of-the-art UML diagrams correction systems.
Download

Area 3 - Models and Algorithms

Full Papers

Paper Nr:	15
Title:	Calculating the Credibility of Test Samples at Inference by a Layer-wise Activation Cluster Analysis of Convolutional Neural Networks
Authors:	Daniel Lehmann and Marc Ebner
Abstract:	A convolutional neural network model is able to achieve high classification performance on test samples at inference, as long as those samples are drawn from the same distribution as the samples used for model training. However, if a test sample is drawn from a different distribution, the performance of the model decreases drastically. Such a sample is typically referred to as an out-of-distribution (OOD) sample. Papernot and McDaniel (2018) propose a method, called Deep k-Nearest Neighbors (DkNN), to detect OOD samples by a credibility score. However, DkNN are slow at inference as they are based on a kNN search. To address this problem, we propose a detection method that uses clustering instead of a kNN search. We conducted experiments with different types of OOD samples for models trained on either MNIST, SVHN, or CIFAR10. Our experiments show that our method is significantly faster than DkNN, while achieving similar performance.
Download

Paper Nr:	24
Title:	Bridging the Gap between Real and Synthetic Traffic Sign Repositories
Authors:	Diogo Lopes da Silva and António R. Fernandes
Abstract:	Current traffic sign image repositories for classification purposes suffer from scarcity of samples due to the compiling and labelling images being mainly a manual process. Thus, researchers resort to alternative approaches to deal with this issue, such as increasing the model architectural complexity or performing data augmentation. A third approach is the usage of synthetic data. This work addresses the data shortage issue by building a synthetic repository proposing a pipeline to build synthetic samples introducing previously unused image operators. Three use cases for synthetic data usage are explored: as a standalone training set, merging with real data, and ensembling. The first option provides results that not only clearly surpass any previous attempt on using synthetic data for traffic sign recognition but are also encouragingly placing the obtained accuracies closer to results with real images. Merging real and synthetic data in a single data set further improves those results. Due to the different nature of the datasets involved, ensembling provides a boost in accuracy results. Overall we got results in three different datasets that surpass previous state of the art results: GTSRB (99:85%), BTSC (99:76%), and rMASTIF (99:84%). Finally, cross testing amongst the three datasets hints that our synthetic datasets have the potential to provide better generalization ability than using real data.
Download

Short Papers

Paper Nr:	3
Title:	Structural Extensions of Basis Pursuit: Guarantees on Adversarial Robustness
Authors:	Dávid Szeghy, Mahmoud Aslan, Áron Fóthi, Balázs Mészáros, Zoltán Á. Milacski and András Lőrincz
Abstract:	While deep neural networks are sensitive to adversarial noise, sparse coding using the Basis Pursuit (BP) method is robust against such attacks, including its multi-layer extensions. We prove that the stability theorem of BP holds upon the following generalizations: (i) the regularization procedure can be separated into disjoint groups with different weights, (ii) neurons or full layers may form groups, and (iii) the regularizer takes various generalized forms of the `1 norm. This result provides the proof for the architectural generalizations of (Cazenavette et al., 2021) including (iv) an approximation of the complete architecture as a shallow sparse coding network. Due to this approximation, we settled to experimenting with shallow networks and studied their robustness against the Iterative Fast Gradient Sign Method on a synthetic dataset and MNIST. We introduce classification based on the `2 norms of the groups and show numerically that it can be accurate and offers considerable speedups. In this family, linear transformer shows the best performance. Based on the theoretical results and the numerical simulations, we highlight numerical matters that may improve performance further. The proofs of our theorems can be found in the supplementary material a .
Download

Paper Nr:	16
Title:	Blanket Clusterer: A Tool for Automating the Clustering in Unsupervised Learning
Authors:	Konstantin Bogdanoski, Kostadin Mishev and Dimitar Trajanov
Abstract:	We propose a generic hierarchical clustering algorithm - named Blanket Clusterer, which allows researchers to examine their data and verify the results gained from other machine learning techniques. We also integrate a three-dimensional visualization plugin that provides better understanding of the clustering results. We verify the tool on a specific use-case, i.e., measuring the clustering techniques performances on a textual dataset based solely on ICD-9 descriptions encoded using the Word2Vec distributed representations. The verification shows that Blanket Clusterer provides an efficient pipeline for evaluating and interpreting the most frequently used clustering methods in unsupervised learning.
Download

Area 4 - Natural Language Understanding

Short Papers

Paper Nr:	23
Title:	Open-domain Conversational Agent based on Pre-trained Transformers for Human-Robot Interaction
Authors:	Mariana F. Fernandes and Plinio Moreno
Abstract:	Generative pre-trained transformers belong to the breakthroughs in Natural Language Processing (NLP), allowing Human-Robot Interactions (e.g. the creation of an open-domain chatbot). However, a substantial amount of research and available data are in English, causing low-resourced languages to be overlooked. This work addresses this problem for European Portuguese with two options: (i) Translation of the sentences before and after using the model fine-tuned on an English-based dataset, (ii) Translation of the English-based dataset to Portuguese and then fine-tune this model on it. We rely on the DialoGPT (dialogue generative pre-trained transformer), a tunable neural conversational answer generation model that learns the basic skills to conduct a dialogue. We use two sources of evaluation: (i) Metrics for text generation based on uncertainty (i.e. perplexity), and similarity between sentences (i.e. BLEU, METEOR and ROUGE) and (ii) Human-based evaluation of the sentences. The translation of sentences before and after of the modified DialoGPT model, using the Daily Dialogue dataset led to the best results.
Download

Area 5 - Machine Learning

Full Papers

Paper Nr:	6
Title:	Using Machine Learning for Classification of Cancer Cells from Raman Spectroscopy
Authors:	Lerina Aversano, Mario L. Bernardi, Vincenzo Calgano, Marta Cimitile, Concetta Esposito, Martina Iammarino, Marco Pisco, Sara Spaziani and Chiara Verdone
Abstract:	Since cancer represents one of the leading causes of death worldwide, the development of approaches capable of discerning healthy from diseased cells would be of fundamental importance to support diagnostic and screening techniques. Raman spectroscopy is the most effective molecular analysis technique currently available and provides information on the molecular composition, bonds, chemical environment, phase, and crystalline structure of the samples under examination. This work exploits a combination of Raman spectroscopy and machine learning models to discriminate patients’ liver cells between tumor and non-tumor. The research uses real patient data, provided by the Center for Nanophotonics and Optoelectronics for Human Health (CNOS), which analyzed the cells of a patient with liver cancer. Specifically, the dataset has been built through a long data collection process, which first involved the analysis of the cells with Raman spectroscopy and then the training of two classifiers, Decision Tree and Random Forest. The results show good performance for the trained classifiers, especially those relating to the Random Forest, which reaches an accuracy of 90%.
Download

Paper Nr:	12
Title:	Neural Networks for Indoor Localization based on Electric Field Sensing
Authors:	Florian Kirchbuchner, Moritz Andres, Julian von Wilmsdorff and Arjan Kuijper
Abstract:	In this paper, we will demonstrate a novel approach using artificial neural networks to enhance signal processing for indoor localization based on electric field measurement systems Up to this point, there exist a variety of approaches to localize persons by using wearables, optical sensors, acoustic methods and by using Smart Floors. All capacitive approaches use, to the best of our knowledge, analytic signal processing techniques to calculate the position of a user. While analytic methods can be more transparent in their functionality, they often come with a variety of drawbacks such as delay times, the inability to compensate defect sensor inputs or missing accuracy. We will demonstrate machine learning approaches especially made for capacitive systems resolving these challenges. To train these models, we propose a data labeling system for person localization and the resulting dataset for the supervised machine learning approaches. Our findings show that the novel approach based on artificial neural networks with a time convolutional neural network (TCNN) architecture reduces the Euclidean error by 40% (34.8cm Euclidean error) in respect to the presented analytical approach (57.3cm Euclidean error). This means a more precise determination of the user position of 22.5cm centimeter on average.
Download

Short Papers

Paper Nr:	5
Title:	A Faster Converging Negative Sampling for the Graph Embedding Process in Community Detection and Link Prediction Tasks
Authors:	Kostas Loumponias, Andreas Kosmatopoulos, Theodora Tsikrika, Stefanos Vrochidis and Ioannis Kompatsiaris
Abstract:	The graph embedding process aims to transform nodes and edges into a low dimensional vector space, while preserving the graph structure and topological properties. Random walk based methods are used to capture structural relationships between nodes, by performing truncated random walks. Afterwards, the SkipGram model with the negative sampling approach, is used to calculate the embedded nodes. In this paper, the proposed SkipGram model converges in fewer iterations than the standard one. Furthermore, the community detection and link prediction task is enhanced by the proposed method.
Download

Paper Nr:	17
Title:	xAMR: Cross-lingual AMR End-to-End Pipeline
Authors:	Maja Mitreska, Tashko Pavlov, Kostadin Mishev and Monika Simjanoska
Abstract:	Creating multilingual end-to-end AMR models requires a large amount of cross-lingual data making the parsing and generating tasks exceptionally challenging when dealing with low-resource languages. To avoid this obstacle, this paper presents a cross-lingual AMR (xAMR) pipeline that incorporates the intuitive translation approach to and from the English language as a baseline for further utilization of the AMR parsing and generation models. The proposed pipeline has been evaluated via the cosine similarity of multiple state-of-the-art sentence embeddings used for representing the original and the output sentences generated by our xAMR approach. Also, BLEU and ROUGE scores were used to evaluate the preserved syntax and the word order. xAMR results were compared to multilingual AMR models’ performance for the languages experimented within this research. The results showed that our xAMR outperforms the multilingual approach for all the languages discussed in the paper and can be used as an alternative approach for abstract meaning representation of low-resource languages.
Download

Paper Nr:	21
Title:	Recommender System using Reinforcement Learning: A Survey
Authors:	Mehrdad Rezaei and Nasseh Tabrizi
Abstract:	Recommender systems are rapidly becoming an integral part of our daily lives. They play a crucial role in overcoming the overloading problem of information by suggesting and personalizing the recommended items. Collaborative filtering, content-based filtering, and hybrid methods are examples of traditional recommender systems which had been used for straightforward prediction problems. More complex problems can be solved with new methods which are applied to recommender systems, such as reinforcement learning algorithms. Markov decision process and reinforcement learning can take part in solving these problems. Recent developments in applying reinforcement learning methods to recommender systems make it possible to use them in order to solve problems with the massive environment and states. A review of the reinforcement learning recommender system will follow the traditional and reinforcement learning-based methods formulation, their evaluation, challenges, and recommended future work.
Download

Paper Nr:	22
Title:	Active Data Collection of Health Data in Mobile Devices
Authors:	Ana S. Machado, Heitor Cardoso, Plinio Moreno and Alexandre Bernardino
Abstract:	This paper aims to develop an intelligent notification system to help sustain user engagement in mHealth applications, specifically those that support self-management. We rely on Reinforcement Learning (RL), an approach where agent learns by exploration the most opportune time to perform a questionnaire, throughout their day, only from easily obtainable non-sensitive data and usage history. This history allows the agent to remember how the user reacts or has reacted in the past to its actions. We consider several options on algorithm, state representation and reward function under the RL umbrella (Upper Confidence Bound, Tabular Q-learning and Deep Q-learning). In addition, a simulator was developed to mimic the behavior of a typical user and utilized to test all possible combinations with users experiencing distinct lifestyles. We obtain promising promising results, which still requiring further testing to be fully validated. We demonstrate that an efficient and well-balanced notification system can be built with simple formulations of an RL problem and algorithm. Furthermore, our approach does not require to have access to sensitive user data. This approach diminishes privacy issues that might concern the user and limits sensor and hardware concerns, such as lapses in collected data or battery drainage.
Download

Paper Nr:	31
Title:	Identifying Users’ Emotional States through Keystroke Dynamics
Authors:	Stefano Marrone and Carlo Sansone
Abstract:	Recognising users’ emotional states is among the most pursued tasks in the field of affective computing. Despite several works show promising results, they usually require expensive or intrusive hardware. Keystroke Dynamics (KD) is a behavioural biometric, whose typical aim is to identify or confirm the identity of an individual by analysing habitual rhythm patterns as they type on a keyboard. This work focuses on the use of KD as a way to continuously predict users’ emotional states during message writing sessions. In particular, we introduce a time-windowing approach that allows analysing users’ writing sessions in different batches, even when the considered writing window is relatively small. This is very relevant in the field of social media, where the exchanged messages are usually very small and the typing rhythm is very fast. The obtained results suggest that even very short writing windows (in the order of 30”) are sufficient to recognise the subject’s emotional state with the same level of accuracy of systems based on the analysis of larger writing sessions (i.e., up to a few minutes).
Download

Paper Nr:	2
Title:	Forecast of Dengue Cases based on the Deep Learning Approach: A Case Study for a Brazilian City
Authors:	Luiz Sérgio de Souza, Solange N. Alves-Souza, Lucia L. Filgueiras, Leandro R. Velloso, Mailson Fontes de Carvalho, Luciano A. Garcia, Marcia Ito, Johne M. Jarske, Tânia D. Santos, Henrique M. Fernandes, Gabriela M. Araújo and Wesley L. Barbosa
Abstract:	According to the World Health Organization (WHO), dengue is an endemic disease in more than 100 countries, with about 50 million people infected each year and 2.5 billion living in risk areas. Dengue requires a major research effort in countries affected by the disease, as its incidence is strongly determined by non-linear local processes, such as climatic conditions, social characteristics and habits of populations (Falcón-Lezama, 2016). In this scenario, forecasting models can be important tools for outbreak control, allowing health institutions to anticipate the mobilization of resources. In this article, we use deep learning, including long and short-term memory (LSTM) and dense layers of perceptrons to implement a forecast model of dengue cases for 5 epidemiological weeks ahead with a mean accuracy of 93%.
Download

Paper Nr:	14
Title:	An Ensemble-based Dimensionality Reduction for Service Monitoring Time-series
Authors:	Farzana Anowar, Samira Sadaoui and Hardik Dalal
Abstract:	Our work introduces an ensemble-based dimensionality reduction approach to efficiently address the high dimensionality of an industrial unlabeled time-series dataset, intending to produce robust data labels. The ensemble comprises a self-supervised learning method to improve data quality, an unsupervised dimensionality reduction to lower the ample feature space, and a chunk-based incremental dimensionality reduction to further increase confidence in data labels. Since the time-series dataset is massive, we divide it into several chunks and evaluate each chunk’s quality using time-series clustering method and metrics. The experiments reveal that clustering performances increased significantly for all the chunks after performing the ensemble approach.
Download

Paper Nr:	18
Title:	Linguistic Feature-based Classification for Anger and Anticipation using Machine Learning
Authors:	Kalaimagal Ramakrishnan, Vimala Balakrishnan and Kumanan Govaichelvan
Abstract:	Growing number of online discourses enables the development of emotion mining models using natural language processing techniques. However, language diversity and cultural disparity alters the sentiment orientation of words depending on the community and context. Therefore, this study investigates the impacts of linguistic features, namely lexical and syntactic, in predicting the presence two emotions among Malaysian YouTube users, anger and anticipation. Term Frequency–Inverse Document Frequency (TF-IDF), Unigrams, Bigrams and Parts-of-Speech Tags were used as features to observe the classification performance. The dataset used in this study contains 2500 YouTube comments by Malaysian users on 46 Covid-19 related videos. Comments were extracted from three prominent Malaysian-centric English news channels: Channel News Asia (CNA), The Star News, and New Strait Times, ranging from 16 March 2020 – 30 April 2020 (i.e., first lockdown phase). Random Forest, Support Vector Machine, Logistic Regression, Decision Tree, K-Nearest Neighbour and Multinomial Naïve Bayes were the six classification algorithms tested, with results indicating Support Vector Machine with TF-IDF provided the best performance, achieving accuracy of 76% and 73% for anger and anticipation, respectively.
Download