Language representations for generalization in reinforcement learning
- Authors: Goodger, Nikolaj , Vamplew, Peter , Foale, Cameron , Dazeley, Richard
- Date: 2021
- Type: Text , Conference paper
- Relation: 13th Asian Conference on Machine Learning, Virtual, 17-19 November 2021, Proceedings of The 13th Asian Conference on Machine Learning Vol. 157, p. 390-405
- Full Text:
- Reviewed:
- Description: The choice of state and action representation in Reinforcement Learning (RL) has a significant effect on agent performance for the training task. But its relationship with generalization to new tasks is under-explored. One approach to improving generalization investigated here is the use of language as a representation. We compare vector-states and discreteactions to language representations. We find the agents using language representations generalize better and could solve tasks with more entities, new entities, and more complexity than seen in the training task. We attribute this to the compositionality of language
Lateral cyclic performance of hybrid fabricated beam‐columns
- Authors: Javidan, Fatemeh
- Date: 2021
- Type: Text , Conference paper
- Relation: 9th European Conference on Steel and Composite Structures, Sheffield, UK, 1-3 September 2021, ce/papers Eurosteel 2021 Sheffield Vol. 4, p. 1657-1662
- Full Text: false
- Reviewed:
- Description: In this paper, a high‐capacity fabricated hollow section is proposed as a lightweight replacement for concrete‐filled composite members which are commonly used as beam‐columns in seismic zones. The lateral cyclic performance of the fabricated hollow beam‐column comprising of mild steel plates (grade 250) and high strength steel (HSS) tubes (grade 750) is investigated. A finite element model is developed and validated against available experimental data. To incorporate the cyclic hardening and softening of both steel materials, a combined plasticity model obtained from low‐cycle material tests is incorporated in the finite element model. A comprehensive sensitivity analysis is undertaken to propose an optimum slenderness design for the thin‐walled steel elements with ductile failure. While the high strength steel tube elements can significantly increase the axial and lateral capacity of the hybrid section, the cyclic failure mechanism is governed by mild steel plate elements, rather than a sudden failure in the high strength steel tubes.
Machine learning based biosignals mental stress detection
- Authors: Al-Jumaily, Adel , Matin, Nafisa , Hoshyar, Azadeh
- Date: 2021
- Type: Text , Conference paper
- Relation: 6th International Conference on Soft Computing in Data Science, SCDS 2021 Vol. 1489 CCIS, p. 28-41
- Full Text: false
- Reviewed:
- Description: Mental Stress can be defined as a normal physiological and biological reaction to an incident or a situation that makes a person feel challenged, troubled, or helpless. While dealing with stress, some changes occur in the biological function of a person, which results in a considerable change in some bio-signals such as, Electrocardiogram (ECG), Electromyography (EMG), Electrodermal Activity (EDA), respiratory rate. This paper aims to review the effect of mental stress on mental condition and health, the changes in biosignals as an indicator of the stress response and train a model to detect stressed states using the biosignals. This paper delivers a brief review of mental stress and biosignals correlation. It represents four Support Vector Machine (SVM) models trained with ECG and EMG features from an open access database based on task related stress. After performing comparative analysis on the four types of trained SVM models with chosen features, Gaussian Kernel SVM is selected as the best SVM model to detect mental stress which can predict the mental condition of a subject for a stressed and relaxed condition having an accuracy of 93.7%. These models can be investigated further with more biosignals and applied in practice, which will be beneficial for the physician. © 2021, Springer Nature Singapore Pte Ltd.
MAM : a metaphor-based approach for mental illness detection
- Authors: Zhang, Dongyu , Shi, Nan , Peng, Ciyuan , Aziz, Abdul , Zhao, Wenhong , Xia, Feng
- Date: 2021
- Type: Text , Conference paper
- Relation: 21st International Conference on Computational Science, ICCS 2021 Vol. 12744 LNCS, p. 570-583
- Full Text:
- Reviewed:
- Description: Among the most disabling disorders, mental illness is one that affects millions of people across the world. Although a great deal of research has been done to prevent mental disorders, detecting mental illness in potential patients remains a considerable challenge. This paper proposes a novel metaphor-based approach (MAM) to determine whether a social media user has a mental disorder or not by classifying social media texts. We observe that the social media texts posted by people with mental illness often contain many implicit emotions that metaphors can express. Therefore, we extract these texts’ metaphor features as the primary indicator for the text classification task. Our approach firstly proposes a CNN-RNN (Convolution Neural Network - Recurrent Neural Network) framework to enable the representations of long texts. The metaphor features are then applied to the attention mechanism for achieving the metaphorical emotions-based mental illness detection. Subsequently, compared with other works, our approach achieves creative results in the detection of mental illnesses. The recall scores of MAM on depression, anorexia, and suicide detection are the highest, with 0.50, 0.70, and 0.65, respectively. Furthermore, MAM has the best F1 scores on depression and anorexia detection tasks, with 0.51 and 0.71. © 2021, Springer Nature Switzerland AG.
MATLAB/Simulink Modelling of Multi-junction PV Cell for Conversion Efficiency Improvement using Maximum Power Point Tracking Method
- Authors: Dave, Malvika , Das, Narottam , Islam, Syed
- Date: 2021
- Type: Text , Conference paper
- Relation: 2021 IEEE PES Innovative Smart Grid Technologies - Asia, ISGT Asia 2021, Brisbane, 5-8 December 2021
- Full Text: false
- Reviewed:
- Description: This research focuses the modelling of multi-junction solar cells (MJSCs) for conversion efficiency improvement considering maximum power point tracking (MPPT) utilizing MATLAB/Simulink software. The conversion efficiency of a single-junction photovoltaic (PV) cell is comparatively low than the MJSCs (GaInP/GaInAs/Ge). The cell conversion efficiency is improved by forming MJSCs, where two or more single junctions are stacked, and higher conversion efficiency is achieved using the MPPT method in MATLAB/Simulink environment to maximize the output power of PV panel. This research considered different parameters of MJSCs, such as lattice matching, bandgap matching, current matching to improve the conversion efficiency. The maximum performance of MJSCs also obtained through the I-V and P-V characteristics by varying different environmental factors, such as temperature and irradiance in MATLAB/Simulink environment. perturb and observe (P&O) MPPT algorithm which is the most suitable for PV applications. Finally, two different Pulse Width Modulation techniques are compared and analyzed with respect to Perturb and observe by applying Quasi Z source inverter (QZSI) within the MATLAB/Simulink environment to enhance the conversion efficiency of triple MJSC. © 2021 IEEE
Melanoma classification using efficientnets and ensemble of models with different input resolution
- Authors: Karki, Sagar , Kulkarni, Pradnya , Stranieri, Andrew
- Date: 2021
- Type: Text , Conference paper
- Relation: 2021 Australasian Computer Science Week Multiconference, ACSW 2021, Virtual, Online, 1-5 February 2021, ACM International Conference Proceeding Series
- Full Text:
- Reviewed:
- Description: Early and accurate detection of melanoma with data analytics can make treatment more effective. This paper proposes a method to classify melanoma cases using deep learning on dermoscopic images. The method demonstrates that heavy augmentation during training and testing produces promising results and warrants further research. The proposed method has been evaluated on the SIIM-ISIC Melanoma Classification 2020 dataset and the best ensemble model achieved 0.9411 area under the ROC curve on hold out test data. © 2021 ACM.
Mining outlying aspects on healthcare data
- Authors: Samariya, Durgesh , Ma, Jiangang
- Date: 2021
- Type: Text , Conference paper
- Relation: 10th International Conference on Health Information Science, HIS 2021, Melbourne, 25-28 October 2021, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) Vol. 13079 LNCS, p. 160-170
- Full Text: false
- Reviewed:
- Description: Machine learning and artificial intelligence have a wide range of applications in medical domain, such as detecting anomalous reading, anomalous patient health condition, etc. Many algorithms have been developed to solve this problem. However, they fail to answer why those entries are considered as an outlier. This research gap leads to outlying aspect mining problem. The problem of outlying aspect mining aims to discover the set of features (a.k.a subspace) in which the given data point is dramatically different than others. In this paper, we present an interesting application of outlying aspect mining in the medical domain. This paper aims to effectively and efficiently identify outlying aspects using different outlying aspect mining algorithms and evaluate their performance on different real-world healthcare datasets. The experimental results show that the latest isolation-based outlying aspect mining measure, SiNNE, have outstanding performance on this task and have promising results. © 2021, Springer Nature Switzerland AG.
Mitigation of Power Quality Issues with Solar PV Penetration into LV/MV Distribution System
- Authors: Patel, Bhoomi , Das, Narottam , Islam, Syed
- Date: 2021
- Type: Text , Conference paper
- Relation: 2021 IEEE PES Innovative Smart Grid Technologies - Asia, ISGT Asia 2021, Brisbane, 5-8 December 2021
- Full Text: false
- Reviewed:
- Description: Over the last decades, solar photovoltaic (PV) system on the low and medium voltage distribution network is growing dramatically. The power grids are still facing the experience of more penetration to satisfy electricity demand of the consumer. Another reason for solar PV penetration is government subsidies for the customers, decreasing the capital cost of solar PV, feed-in-tariff and to balance the CO2 emission from the atmosphere. The integration of PV system in low voltage or medium voltage ((LV/MV) distribution networks is raising new challenges in terms of power quality (PQ). In the LV/MV distribution networks, PQ problems are voltage sag and swell are the most serious concern, and it occurs when residential renewable energy sources (i.e., solar PV) produce more energy than the local demand. As per the PQ studies, voltage sag/swell issues in the LV/MV power grid considered to be the most common type of PQ issue. Impact of this type of issue on sensitive loads is harsh. Hence, different types of solution have been used to protect sensitive loads from such voltage problems, but the dynamic voltage restorer is to be considered, effective and efficient solution. Dynamic Voltage Restorer (DVR) is a series connected power electronic device which is used to alleviate the voltage sag/swell problem in the distribution system and restore the load voltage to the required value. Therefore, the main objective of this research project is to investigate and understand the impact of PQ issues in LV/MV distribution system and aim is to alleviate them. MATLAB/Simulink software is used to solve the PQ issues by designing battery supported DVR with energy storage systems. The performance of battery supported DVR is designed, simulated in MATLAB/Simulink software environment, and analysed in detail. © 2021 IEEE
Multi-objective optimisation to manage trade-offs in water quality and quantity of complex water resource system
- Authors: Dey, Sayani , Barton, Andrew , Bagirov, Adil , Kandra, Harpreet , Wilson, Kym
- Date: 2021
- Type: Text , Conference paper
- Relation: Hydrology and Water Resources Symposium 2021, HWRS 2021: Digital Water: Hydrology and Water Resources Symposium 2021, Virtual online, 31 August-1 September 2021, HWRS 2021: Digital Water: Hydrology and Water Resources Symposium 2021 p. 465-480
- Full Text: false
- Reviewed:
- Description: Water of adequate quality and quantity is the key to health and integrity of the environment and fundamental to good water supply. Achieving water quality and quantity objectives can conflict and has become more complicated with challenges like, climate change, growing populations and changed land uses. Therefore, a multi-objective optimisation strategy is required for achieving optimal water quality and quantity outcomes from a water resources system. This study uses a multi-objective optimisation approach to illustrate the trade-offs occurring when water quantity and quality in a reservoir system are optimised. Taylors Lake, part of the Grampians Reservoir System in Western Victoria, Australia was chosen as the case study for this research as it is quite complex and includes many contemporary water resources challenges seen around the world, such as high turbidity and salinity. The objective functions are set in a way to maximise the water quantity available for supply, while minimising the deviation of quality parameters from the accepted limits. The water system is modelled using eWater Source® modelling platform, while optimisation is undertaken using NSGA-II optimisation technique. Daily time step data over a ten-year period was used in this work. Various optimisation runs were performed with different population sizes and generations to seek out the best trade-off curve. The optimisation results indicate trade-offs between salinity, turbidity, and quantity. Key findings for this case study show that through optimisation, stored water never exceeded 19,000 ML even though the storage capacity was 27,000 ML indicating a significant loss of water to improve quality, or alternatively, a potential asset re-design opportunity.
Multimodal sensor selection for multiple spatial field reconstruction
- Authors: Nguyen, Linh , Thiyagarajan, Karthick , Ulapane, Nalika , Kodagoda, Sarath
- Date: 2021
- Type: Text , Conference paper
- Relation: 16th IEEE Conference on Industrial Electronics and Applications, ICIEA 2021 p. 1181-1186
- Full Text: false
- Reviewed:
- Description: The paper addresses the multimodal sensor selection problem where selected colocated sensor nodes are employed to effectively monitor and efficiently predict multiple spatial random fields. It is first proposed to exploit multivariate Gaussian processes (MGP) to model multiple spatial phenomena jointly. By the use of the Matérn cross-covariance function, cross-covariance matrices in the MGP model are sufficiently positive semi-definite, concomitantly providing efficient prediction of all multivariate processes at unmeasured locations. The multimodal sensor selection problem is then formulated and solved by an approximate algorithm with an aim to select the most informative sensor nodes so that prediction uncertainties at all the fields are minimized. The proposed approach was validated in the real-life experiments with promising results. © 2021 IEEE.
Multivariate versus univariate sensor selection for spatial field estimation
- Authors: Nguyen, Linh , Thiyagarajan, Karthick , Ulapane, Nalika , Kodagoda, Sarath
- Date: 2021
- Type: Text , Conference paper
- Relation: 16th IEEE Conference on Industrial Electronics and Applications, ICIEA 2021 p. 1187-1192
- Full Text: false
- Reviewed:
- Description: The paper discusses the sensor selection problem in estimating spatial fields. It is demonstrated that selecting a subset of sensors depends on modelling spatial processes. It is first proposed to exploit Gaussian process (GP) to model a univariate spatial field and multivariate GP (MGP) to jointly represent multivariate spatial phenomena. A Matérn cross-covariance function is employed in the MGP model to guarantee its cross-covariance matrices to be positive semi-definite. We then consider two corresponding univariate and multivariate sensor selection problems in effectively monitoring multiple spatial random fields. The sensor selection approaches were implemented in the real-world experiments and their performances were compared. Difference of results obtained by the univariate and multivariate sensor selection techniques is insignificant; that is, either of the methods can be efficiently used in practice. © 2021 IEEE.
Non-invasive smartphone use monitoring to assess cognitive impairment
- Authors: Thang, Nguyen , Oatley, Giles , Stranieri, Andrew , Walker, Darren
- Date: 2021
- Type: Text , Conference paper
- Relation: 13th International Conference on Computer and Automation Engineering, ICCAE 2021 p. 64-67
- Full Text: false
- Reviewed:
- Description: Background: There are many tests for the early detection of Mild Cognitive Impairment (MCI) to prevent or delay the development of dementia, particularly amongst the elderly. However, many tests are complex and are required to be performed repeatedly. Cognitive assessment apps for a smartphone have emerged, but like other tests, they require the user to perform complex tasks like drawing time on a clock. Few studies have explored non-invasive ways of tracking and assessing MCI without having the user perform specific tests. Objective: This research ultimately aims to develop an app that runs in the background and collects smartphone usage data that correlates well with MCI test results. The focus of this preliminary study was to develop an app that collects usage data and common MCI questionnaires to see if usage data between people varied, and to establish associations between phone usage and cognitive tests results. Method: An android application was developed to gather data over three weeks by three volunteers (authors). Usage data collected included: SMS and call log, accelerometer, location, app usage, self-report. Cognitive tests implemented were Stroop, Go/No Go tests and absent-mindedness questionnaires. Due to the small sample size and Covid-19 restrictions (October 2020), location data was not reliable. SMS text was not collected for privacy reasons. Results: App categories can differentiate people, but the app usage cannot be used to distinguish people. © 2021 IEEE.
Open banking and electronic health records
- Authors: Stranieri, Andrew , McInnes, Angelique , Hashmi, Mustafa , Sahama, Tony
- Date: 2021
- Type: Text , Conference paper
- Relation: 2021 Australasian Computer Science Week Multiconference, ACSW 2021, Virtual, Online, 1-5 February 2021, ACM International Conference Proceeding Series
- Full Text:
- Reviewed:
- Description: The Open Banking model is a data sharing model emerging in financial services sector that involves technological and regulatory innovations designed to facilitate access to banking records by third party providers such as payment service providers. The model is predicted to disrupt financial services and encourage a wave of new third-party providers offering innovative services that will change the relationship between customers and banks. This article juxtaposes the Open Banking model against models for electronic health records. Providers that could supply innovative third party services with health record data if an Open Banking model were adopted in the health care industry are advanced. © 2021 ACM.
Oscillations and periodic solutions in a two-dimensional differential delay model
- Authors: Ivanov, Anatoli , Dzalilov, Zari
- Date: 2021
- Type: Text , Conference paper
- Relation: International Conference on Applied Mathematics, Modeling and Computational Science, AMMCS 2019 Vol. 343, p. 59-70
- Full Text:
- Reviewed:
- Description: A class of two-dimensional differential systems with delay and overall negative feedback is considered. Sufficient conditions for the existence of periodic solutions are established. The instability of the unique equilibrium together with the one-sided boundedness of one of the two nonlinearities lead to the existence of periodic solutions. Systems of this type appear in various applications in engineering and natural sciences, in particular in mathematical biology and physiology as models of circadian rhythm generator and glucose-insulin regulation models in humans. © 2021, Springer Nature Switzerland AG.
Predicting mental health problems with personality, behavior, and social networks
- Authors: Zhang, Dongyu , Guo, Teng , Han, Shiyu , Vahabli, Sadaf , Naseriparsa, Mehdi , Xia, Feng
- Date: 2021
- Type: Text , Conference paper
- Relation: 2021 IEEE International Conference on Big Data, Big Data 2021, virtual online, 15-18 December 2021, Proceedings - 2021 IEEE International Conference on Big Data, Big Data 2021 p. 4537-4546
- Full Text:
- Reviewed:
- Description: Mental health is an integral part of human health and well-being. Unhealthy mentality leads to serious consequences such as self-mutilation and suicide, especially for college students. While the literature focused on analysing the relationship between mental health and a single factor such as personality or behavior, accurate prediction is yet to be achieved due to the lack of cross-dimensional analysis and multi-dimensional joint prediction. To this end, this work proposes leveraging multiple factors from three crucial dimensions of mental health: behaviors, personality, and social networks. We recruited 490 college students, and collected their behavioral records from smart cards. In addition, we extracted their psychological traits from questionnaires, and social networks by conducting the survey on the nominating community members. We created a neural network-based model to integrate behavioral, psychological, and social network factors to predict mental health problems. The experimental results verify the efficacy of the proposed model, and demonstrate that the classification model of various factors effectively predicts the students' mental issues. © 2021 IEEE.
Preparation for post-school careers in rural and peri-urban Australia : connections with employers and labour markets
- Authors: Smith, Erica , Foley, Annette
- Date: 2021
- Type: Text , Conference paper
- Relation: 6th International Conference on Employer Engagement: Preparing Young People for the Future, 1-2 July 2021, Virtual online
- Full Text: false
- Reviewed:
Rethinking IS Graduates Work-readiness: Employers' perspectives
- Authors: Faisal, Nadia , Chadhar, Mehmood , Goriss-Hunter, Anitra , Stranieri, Andrew
- Date: 2021
- Type: Text , Conference paper
- Relation: 27th Annual Americas Conference on Information Systems (AMCIS)
- Full Text: false
- Reviewed:
- Description: Being a significant stakeholder in the graduates' employment outcomes, it is vital to understand employers' perceptions of graduates' work-readiness. However, existing information systems (IS) literature focuses mainly on the perceptions of students or universities. This paper aims to fill this gap by analysing scoping interviews conducted with graduate recruiters and industry experts in Australia regarding attributes that can improve graduates' employment prospects in the information and communication technology industry. A preliminary investigation based on grounded theory identified three emergent themes from the data: behaviors, skills, and knowledge levels. Based on the findings, this study proposes an IS graduate work-readiness framework that can help universities to develop academic programs aimed at enhancing desirable skills and attitudes among IS graduates' employment.
Smartcolor : automatic web color scheme generation based on deep learning
- Authors: Feng, Zhitao , Hou, Mingliang , Liu, Huiyang , Kaur, Achhardeep , Febrinanto, Falih
- Date: 2021
- Type: Text , Conference paper
- Relation: 12th International Conference on Information and Communication Systems, ICICS 2021 p. 285-290
- Full Text: false
- Reviewed:
- Description: The color scheme plays an important role in different aspects of our everyday lives, such as web design and human-computer interaction. The generation of color scheme requires a long-term accumulation of design experience and advanced knowledge of color matching. However, there is little work focusing on the automatic generation of color scheme based on learning capabilities. In this work, we propose a novel color scheme designer, SmartColor, which incorporates deep learning methods with knowledge of color psychology. The Generative Adversarial Network (GAN) is used to learn experienced insights from widely recognized color schemes obtained from online color matching websites. Color schemes based on various themes are transformed as statistical constraints in the construction of the objective function of GAN. SmartColor is both data-driven and knowledge-driven. In contrast to current color scheme solutions. SmartColor will automatically create color schemes based on the input theme. Experimental results show that SmartColor was successful in creating color schemes for websites. © 2021 IEEE. **Please note that there are multiple authors for this article therefore only the name of the first 5 including Federation University Australia affiliate “Achhardeep Kaur and Falih Febrinanto” is provided in this record**
Soil organic carbon in rehabilitated coal mine soils as an indicator for soil health
- Authors: Baumgartl, Thomas , Chan, J. , Bucka, F. , Pihlap, E.
- Date: 2021
- Type: Text , Conference paper
- Relation: 14th International Conference on Mine Closure, Mine Closure 2021 Vol. 2021-August
- Full Text: false
- Reviewed:
- Description: Rehabilitation intends to provide a safe, stable and sustainable environment. Soil health is often used as a parameter, which describes the success of reclamation, the performance of the soil and its associated soil system functions. Reclaimed and therefore young soils are in general deprived of soil organic carbon. They are not in equilibrium with their environment and undergo changes over time, faster than natural and developed soils. Carbon content as a summarising criterion for soil health status can be used as an indicator as it reflects the performance of important soil processes, like water holding capacity, drainage and aeration potential and nutrient supply and storage. It is well established that carbon content affects soil functions like hydraulic conductivity by creating structural elements through aggregation processes. Increasing carbon content leads to increased water infiltration, reduced surface runoff and erosional risks, and increases the exchange rate of gases and improves aeration and has in general positive consequences for microbiological activity. Results from a study of soil covers of different ages emphasise this consequence. The evaluation of the rehabilitation success in coal mining using carbon content is complicated due to the difficulty distinguishing between carbon forms: Organic carbon naturally formed by decomposition vs. carbon originating from coal as coal dust or charred material. Furthermore, the assessment of the performance of rehabilitated soils is strongly affected by climatic conditions, which affect the production, decomposition and translocation of organic matter. Litter and dead organic matter from plants are decomposed on the soil surface and incorporated through organisms into the soil profile. Dissolvable organic constituents may be transported with infiltrating water down the profile. Consequently, carbon is primarily concentrated close to the surface. In semi-arid environments the accumulation depth may only be centimetres as was found from a study on the performance of the carbon pool of rehabilitated soils across sites up to 35 years of age. This has implications for the sampling strategy and the assessment of the performance of soils over time. As the carbon content in soils at coal mines can be affected by precipitation and incorporation of coal dust into the soil, the content of new organic carbon representing soil health status can be misleadingly interpreted. Therefore, separation of carbon fractions is necessary to identify the "green" carbon pool (carbon originating from plant litter and residue) as best as possible and extract the correct fraction to assess the performance of soil development. A method has been developed and is presented which allows the separation between the various carbon pools. From the presented study, the following conclusions were drawn: 1) Soil carbon is an easy to measure indicator for the assessment of the performance of soil health of rehabilitated soils; 2) soil functional properties are affected by carbon content and age and hence change with soil development; 3) only green carbon represents soil health and appropriate methodology has to be in place to exclude other carbon pools; 4) carbon storage in rehabilitated soils of semi-arid environments of Australia is below that of natural soils. © 2021, Qualified Mining Consultants LLC,.All right reserved.
Solving ESL sentence completion questions via pre-trained neural language models
- Authors: Liu, Qiongqiong , Liu, Tianqiao , Zhao, Jiafu , Fang, Qiang , Ding, Wenbiao , Wu, Zhongqin , Xia, Feng , Tang, Jiliang , Liu, Zitao
- Date: 2021
- Type: Text , Conference paper
- Relation: 22nd International Conference on Artificial Intelligence in Education, AIED 2021, Virtual, Online, 14-18 June 2021, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) Vol. 12749 LNAI, p. 256-261
- Full Text:
- Reviewed:
- Description: Sentence completion (SC) questions present a sentence with one or more blanks that need to be filled in, three to five possible words or phrases as options. SC questions are widely used for students learning English as a Second Language (ESL) and building computational approaches to automatically solve such questions is beneficial to language learners. In this work, we propose a neural framework to solve SC questions in English examinations by utilizing pre-trained language models. We conduct extensive experiments on a real-world K-12 ESL SC question dataset and the results demonstrate the superiority of our model in terms of prediction accuracy. Furthermore, we run precision-recall tradeoff analysis to discuss the practical issues when deploying it in real-life scenarios. To encourage reproducible results, we make our code publicly available at https://github.com/AIED2021/ESL-SentenceCompletion. © Springer Nature Switzerland AG 2021.