Wyniki wyszukiwania - BazTech

1

A comparative study of CNN, LSTM, BiLSTM, AND GRU architectures for tool wear prediction in milling processes

Zegarra Fabio C., Vargas-Machuca Juan, Coronado Alberto M.

Journal of Machine Engineering

|

2023

|

Vol. 23, No. 4

122--136

EN

Accurately predicting machine tool wear requires models capable of capturing complex, nonlinear interactions in multivariate time series inputs. Recurrent neural networks (RNNs) are well-suited to this task, owing to their memory mechanisms and capacity to construct highly complex models. In particular, LSTM, BiLSTM, and GRU architectures have shown promise in wear prediction. This study demonstrates that RNNs can automatically extract relevant information from time series data, resulting in highly precise wear models with minimal feature engineering. Notably, this approach avoids the need for excessively large window sizes of data points during model training, which would increase model complexity and processing time. Instead, this study proposes a procedure that achieves low prediction errors with window sizes as small as 100 data points. By employing Bayesian hyperparameter optimization and two preprocessing techniques (detrend and offset), RMSE errors consistently fall below 10. A key difference in this study is the use of boxplots to provide a better representation of result variability, as opposed to solely reporting the best values. The proposed approach matches more complex state of-the-art. methods and offers a powerful tool for wear prediction in engineering applications.

2

Urban traffic crash analysis using deep learning techniques

Sobhana Mummaneni, Vemulapalli Nihitha, Mendu Gnana Siva Sai Venkatesh, Ginjupalli Naga Deepika, Dodda Pragathi, Subramanyam Rayanoothala Bala Venkata

Informatyka, Automatyka, Pomiary w Gospodarce i Ochronie Środowiska

|

2023

|

T. 13, nr 3

56--63

EN

Road accidents are concerningly increasing in Andhra Pradesh. In 2021, Andhra Pradesh experienced a 20 percent upsurge in road accidents. The state's unfortunate position of being ranked eighth in terms of fatalities, with 8,946 lives lost in 22,311 traffic accidents, underscores the urgent nature of the problem. The significant financial impact on the victims and their families stresses the necessity for effective actions to reduce road accidents.This study proposes a framework that collects accident data from regions, namely Patamata, Penamaluru, Mylavaram, Krishnalanka, Ibrahimpatnam,and Gandhinagar in Vijayawada(India)from 2019 to 2021. The dataset comprises over 12,000 records of accident data. Deep learning techniquesare applied to classify the severity of road accidents into Fatal, Grievous, and Severe Injuries. The classification procedure leverages advanced neural network models, including the Multilayer Perceptron, Long-Short Term Memory, Recurrent Neural Network, and Gated Recurrent Unit. These modelsare trained on the collected data to accurately predict the severity of road accidents. The project study to make important contributions for suggesting proactive measures and policies to reduce the severity and frequency of road accidents in Andhra Pradesh.

PL

Liczba wypadków drogowych w Andhra Pradesh niepokojąco rośnie. W 2021 r. stan Andhra Pradesh odnotował 20% wzrost liczby wypadków drogowych. Niefortunna pozycja stanu, który zajmuje ósme miejsce pod względem liczby ofiar śmiertelnych, z 8946 ofiarami śmiertelnymiw 22311 wypadkach drogowych, podkreśla pilny charakter problemu. Znaczący wymiar finansowy dla ofiari ich rodziny podkreśla konieczność podjęcia skutecznych działań w celu ograniczenia liczby wypadków drogowych. W niniejszym badaniu zaproponowano system gromadzenia danych o wypadkachz regionów Patamata, Penamaluru, Mylavaram, Krishnalanka, Ibrahimpatnam i Gandhinagar w Vijayawada (India) w latach 2019–2021. Zbiór danych obejmuje ponad 12 000 rekordów danych o wypadkach. Techniki głębokiego uczenia są stosowane do klasyfikowania wagi wypadków drogowychna śmiertelne, poważne i ciężkie obrażenia. Procedura klasyfikacji wykorzystuje zaawansowane modele sieci neuronowych, w tymwielowarstwowy perceptron, pamięć długoterminową i krótkoterminową, rekurencyjną sieć neuronową i Gated Recurrent Unit. Modele te są trenowane na zebranych danych w celu dokładnego przewidywania wagi wypadków drogowych. Projekt ma wnieść istotny wkład w sugerowanie proaktywnych środków i polityk mających na celu zmniejszenie dotkliwości i częstotliwości wypadków drogowych w Andhra Pradesh.

3

Influence of modelling phase transformations with the use of LSTM network on the accuracy of computations of residual stresses for the hardening process

Wróbel Joanna, Kulawik Adam

Bulletin of the Polish Academy of Sciences. Technical Sciences

|

2023

|

Vol. 71, nr 4

art. no. e145681

EN

Replacing mathematical models with artificial intelligence tools can play an important role in numerical models. This paper analyses the modeling of the hardening process in terms of temperature, phase transformations in the solid state and stresses in the elastic-plastic range. Currently, the use of artificial intelligence tools is increasing, both to make greater generalizations and to reduce possible errors in the numerical simulation process. It is possible to replace the mathematical model of phase transformations in the solid state with an artificial neural network (ANN). Such a substitution requires an ANN network that converts time series (temperature curves) into shares of phase transformations with a small training error. With an insufficient training level of the network, significant differences in stress values will occur due to the existing couplings. Long-Short-Term Memory (LSTM) networks were chosen for the analysis. The paper compares the differences in stress levels with two coupled models using a macroscopic model based on CCT diagram analysis and using the Johnson-Mehl-Avrami-Kolmogorov (JMAK) and Koistinen-Marburger (KM) equations, against the model memorized by the LSTM network. In addition, two levels of network training accuracy were also compared. Considering the results obtained from the model based on LSTM networks, it can be concluded that it is possible to effectively replace the classical model in modeling the phenomena of the heat treatment process.

4

Gray-level co-occurrence matrix of Smooth Pseudo Wigner-Ville distribution for cognitive workload estimation

Mirzaeian Rezvan, Ghaderyan Peyvand

Biocybernetics and Biomedical Engineering

|

2023

|

Vol. 43, no. 1

261--278

EN

Automatic, cost-effective, and reliable cognitive workload estimation (CWE) is one of the important issues in diagnosis and treatment of neurocognitive diseases, cognitive performance improvement and error preventive strategies. To address this issue, this paper has proposed a novel and robust CWE method by detecting the time–frequency (TF) changes of electrodermal activities (EDA). Firstly, the local and global properties of the time-variant characteristics of EDA have been presented using Smooth Pseudo WignerVille distribution with enhanced TF resolution. Then, the transient changes in TF images of EDA signals have been quantified using a set of textural features based on Gray Level Co-occurrence Matrix descriptor (GLCM). Several static and dynamic classifiers, such as support vector machine, K- k-nearest neighbor, cascade forward neural network, and recurrent neural network have been explored. A real EDA data experiment recorded during arithmetic task with different workload levels have been used to evaluate the performance of the proposed method. The obtained results have confirmed that it can achieve a high estimation performance of 97.71% using contrast feature for discrimination of three workload levels. Further analysis has also suggested that the model is robust to GLCM parameters and classifiers and can provide a better tradeoff between computational complexity and high performance using minimum number of textural features in comparison with previous studies.

5

Combining homomorphic filtering and recurrent neural network in gait signal analysis for neurodegenerative diseases detection

Saljuqi Masume, Ghaderyan Peyvand

Biocybernetics and Biomedical Engineering

|

2023

|

Vol. 43, no. 2

476--493

EN

Automatic, cost-effective, and reliable detection of neurodegenerative diseases (NDs) is one of the important issues in clinical practice. The main idea of the proposed method in this study is to utilize the advantages of gait time series that can provide low-cost and noninvasive measures, homomorphic filtering that can effectively separate muscle activity from body dynamic and recurrent neural network or cascade forward neural network that can learn sequential time-varying data. Experimental results on gait time series of 16 healthy control subjects, 13 patients with amyotrophic lateral sclerosis, 15 patients with Parkinson’s disease and 20 patients with Huntington’s disease have demonstrated high detection performance with an accuracy rate of 100 % using K-fold cross validation for all three types of NDs outperforming other existing methods. The results have also indicated that the use of real cepstral coefficients, oscillation components, or basic statistics feature set has improved the detection performance.

6

Robust zeroing neural networks with two novel power-versatile activation functions for solving dynamic Sylvester equation

Zhou Peng, Tan Mingtao

Bulletin of the Polish Academy of Sciences. Technical Sciences

|

2022

|

Vol. 70, nr 3

art. no. e141307

EN

In this work, two robust zeroing neural network (RZNN) models are presented for online fast solving of the dynamic Sylvester equation (DSE), by introducing two novel power-versatile activation functions (PVAF), respectively. Differing from most of the zeroing neural network (ZNN) models activated by recently reported activation functions (AF), both of the presented PVAF-based RZNN models can achieve predefined time convergence in noise and disturbance polluted environment. Compared with the exponential and finite-time convergent ZNN models, the most important improvement of the proposed RZNN models is their fixed-time convergence. Their effectiveness and stability are analyzed in theory and demonstrated through numerical and experimental examples.

7

Predicting hospital emergency department visits with deep learning approaches

Zhao Xinxing, Lai Joel Weijia, Ho Andrew Fu Wah, Liu Nan, Ong Marcus Eng Hock, Cheong Kang Hao

Biocybernetics and Biomedical Engineering

|

2022

|

Vol. 42, no. 3

1051--1065

EN

Overcrowding in emergency department (ED) causes lengthy waiting times, reduces adequate emergency care and increases rate of mortality. Accurate prediction of daily ED visits and allocating resources in advance is one of the solutions to ED overcrowding problem. In this paper, a deep stacked architecture is being proposed and applied to the daily ED visits prediction problem with deep components such as Long Short Term Memory (LSTM), Gated Recurrent Units (GRU) and simple Recurrent Neural Network (RNN). The proposed architecture achieves very high mean accuracy level (94.28–94.59%) in daily ED visits predictions. We have also compared the performance of this architecture with non-stacked deep models and traditional prediction models. The results indicate that deep stacked models outperform (4–7%) the traditional prediction models and other non-stacked deep learning models (1–2%) in our prediction tasks. The application of deep neural network in ED visits prediction is novel as this is one of the first studies to apply a deep stacked architecture in this field. Importantly, our models have achieved better prediction accuracy (in one case comparable) than the state-of-the-art in the literature.

8

Monitoring regenerative heat exchanger in steam power plant by making use of the recurrent neural network

Niksa-Rynkiewicz Tacjana, Szewczuk-Krypa Natalia, Witkowska Anna, Cpałka Krzysztof, Zalasiński Marcin, Cader Andrzej

Journal of Artificial Intelligence and Soft Computing Research

|

2021

|

Vol. 11, No. 2

143--155

EN

Artificial Intelligence algorithms are being increasingly used in industrial applications. Their important function is to support operation of diagnostic systems. This paper presents a new approach to the monitoring of a regenerative heat exchanger in a steam power plant, which is based on a specific use of the Recurrent Neural Network (RNN). The proposed approach was tested using real data. This approach can be easily adapted to similar monitoring applications of other industrial dynamic objects.

9

An optimized parallel implementation of non-iteratively trained recurrent neural networks

El Zini Julia, Rizk Yara, Awad Mariette

Journal of Artificial Intelligence and Soft Computing Research

|

2021

|

Vol. 11, No. 1

33--50

EN

Recurrent neural networks (RNN) have been successfully applied to various sequential decision-making tasks, natural language processing applications, and time-series predictions. Such networks are usually trained through back-propagation through time (BPTT) which is prohibitively expensive, especially when the length of the time dependencies and the number of hidden neurons increase. To reduce the training time, extreme learning machines (ELMs) have been recently applied to RNN training, reaching a 99% speedup on some applications. Due to its non-iterative nature, ELM training, when parallelized, has the potential to reach higher speedups than BPTT. In this work, we present Opt-PR-ELM, an optimized parallel RNN training algorithm based on ELM that takes advantage of the GPU shared memory and of parallel QR factorization algorithms to efficiently reach optimal solutions. The theoretical analysis of the proposed algorithm is presented on six RNN architectures, including LSTM and GRU, and its performance is empirically tested on ten time-series prediction applications. Opt- PR-ELM is shown to reach up to 461 times speedup over its sequential counterpart and to require up to 20x less time to train than parallel BPTT. Such high speedups over new generation CPUs are extremely crucial in real-time applications and IoT environments.

10

Forecasting future values of time series using the lstm network on the example of currencies and WIG20 companies

Mróz Bartosz, Nowicki Filip

Zeszyty Naukowe. Telekomunikacja i Elektronika / Uniwersytet Technologiczno-Przyrodniczy w Bydgoszczy

|

2020

|

nr 24

13--30

EN

The article presents a comparison of the RNN, GRU and LSTM networks in predicting future values of time series on the example of currencies and listed companies. The stages of creating an application which is a implementation of the analyzed issue were also shown – the selection of networks, technologies, selection of optimal network parameters. Additionally, two conducted experiments were discussed. The first was to predict the next values of WIG20 companies, exchange rates and cryptocurrencies. The second was based on investments in cryptocurrencies guided solely by the predictions of artificial intelligence. This was to check whether the investments guided by the predictions of such a program have a chance of effective earnings. The discussion of the results of the experiment includes an analysis of various interesting phenomena that occurred during its duration and a comprehensive presentation of the relatively high efficiency of the proposed solution, along with all kinds of graphs and comparisons with real data. The difficulties that occurred during the experiments, such as coronavirus or socio-economic events, such as riots in the USA, were also analyzed. Finally, elements were proposed that should be improved or included in future versions of the solution – taking into account world events, market anomalies and the use of supervised learning.

PL

W artykule przedstawiono porównanie sieci RNN, GRU i LSTM w przewidywaniu przyszłych wartości szeregów czasowych na przykładzie walut i spółek giełdowych. Przedstawiono również etapy tworzenia aplikacji będącej realizacją analizowanego zagadnienia – dobór sieci, technologii, dobór optymalnych parametrów sieci. Dodatkowo omówiono dwa przeprowadzone eksperymenty. Pierwszym było przewidywanie kolejnych wartości spółek z WIG20, kursów walut i kryptowalut. Drugi opierał się na inwestycjach w kryptowaluty, kierując się wyłącznie przewidywaniami sztucznej inteligencji. Miało to na celu sprawdzenie, czy inwestowanie na podstawie przewidywania takiego programu pozwala na efektywne zarobki. Omówienie wyników eksperymentu obejmuje analizę różnych ciekawych zjawisk, które wystąpiły w czasie jego trwania oraz kompleksowe przedstawienie relatywnie wysokiej skuteczności proponowanego rozwiązania wraz z wszelkiego rodzaju wykresami i porównaniami z rzeczywistymi danymi. Analizowano również trudności, które wystąpiły podczas eksperymentów, takie jak koronawirus, wydarzenia społeczno-gospodarcze czy zamieszki w USA. Na koniec zaproponowano elementy, które należałoby ulepszyć lub uwzględnić w przyszłych wersjach rozwiązania, uwzględniając wydarzenia na świecie, anomalie rynkowe oraz wykorzystanie uczenia się nadzorowanego.

11

A Study of Correlation between Fishing Activity and AIS Data by Deep Learning

Shen K. Y., Chu Y. J., Chang S. J., Chang S. M.

TransNav : International Journal on Marine Navigation and Safety of Sea Transportation

|

2020

|

Vol. 14. no. 3

527--531

EN

Previous researches on the prediction of fishing activities mainly rely on the speed over ground (SOG) as the referential attribute to determine whether the vessel is navigating or in fishing operation. Since more and more fishing vessels install Automatic Identification System (AIS) either voluntarily or under regulatory requirement, data collected from AIS in real time provide more attributes than SOG which may be utilized to improve the prediction. To be specific, the ships' trajectory patterns and the changes in course become available and should be considered. This paper aims to improve the accuracy in the identification of fishing activities. First, we do feature extraction from the AIS data of coastal waters around Taiwan and build a Recurrent Neural Network (RNN) model. Then, the activity data of fishing vessels are divided into fishing and non-fishing. Finally, based on the testing by feeding various fishing activity data, we can identify the fishing status automatically.

12

Predykacja zajętości kanału LTE za pomocą rekurencyjnej sieci neuronowej

Wasilewska Małgorzata, Bogucka Hanna

Przegląd Telekomunikacyjny + Wiadomości Telekomunikacyjne

|

2020

|

nr 7-8

269--273, CD

PL

W artykule przedstawiono metody detekcji i przewidywania obecności transmisji sygnału LTE. Zastosowano algorytmy uczenia maszynowego, takie jak algorytm k najbliższych sąsiadów, drzewo decyzyjne, sieć neuronową oraz rekurencyjną sieć neuronową. Za pomocą eksperymentów wykazano, że wymienione algorytmy, a w szczególności sieć rekurencyjna osiągają wysokie wartości prawdopodobieństwa poprawnej detekcji zarówno dla detekcji w czasie rzeczywistym, jak i dla przewidywania obecności sygnałów w przyszłości.

EN

In the paper, the methods of LTE spectrum detection and future state predictions have been presented. Machine learning algorithms have been implemented for spectrum sensing, namely k-nearest neighbors, decision tree, neural network and recurrent neural network. Conducted experiment has shown that these algorithms reach high values of probability of correct detection for current moment as well as for future prediction.

13

Single-ended quality measurement of a music content via convolutional recurrent neural networks

Organiściak Kamila, Borkowski Józef

Metrology and Measurement Systems

|

2020

|

Vol. 27, nr 4

721--733

EN

The paper examines the usage of Convolutional Bidirectional Recurrent Neural Network (CBRNN) for a problem of quality measurement in a music content. The key contribution in this approach, compared to the existing research, is that the examined model is evaluated in terms of detecting acoustic anomalies without the requirement to provide a reference (clean) signal. Since real music content may include some modes of instrumental sounds, speech and singing voice or different audio effects, it is more complex to analyze than clean speech or artificial signals, especially without a comparison to the known reference content. The presented results might be treated as a proof of concept, since some specific types of artefacts are covered in this paper (examples of quantization defect, missing sound, distortion of gain characteristics, extra noise sound). However, the described model can be easily expanded to detect other impairments or used as a pre-trained model for other transfer learning processes. To examine the model efficiency several experiments have been performed and reported in the paper. The raw audio samples were transformed into Mel-scaled spectrograms and transferred as input to the model, first independently, then along with additional features (Zero Crossing Rate, Spectral Contrast). According to the obtained results, there is a significant increase in overall accuracy (by 10.1%), if Spectral Contrast information is provided together with Mel-scaled spectrograms. The paper examines also the influence of recursive layers on effectiveness of the artefact classification task.

14

Learning OWL 2 Property Characteristics as an Explanation for an RNN

Potoniec J.

Bulletin of the Polish Academy of Sciences. Technical Sciences

|

2020

|

Vol. 68, nr 6

1481--1490

EN

We propose an approach to indirectly learn the Web Ontology Language OWL 2 property characteristics as an explanation for a deep recurrent neural network (RNN). The input is a knowledge graph represented in Resource Description Framework (RDF) and the output are scored axioms representing the characteristics. The proposed method is capable of learning all the characteristics included in OWL 2: functional, inverse functional, reflexive and irreflexive, symmetric and asymmetric, transitive. We report and discuss experimental evaluation on DBpedia 2016-10, showing that the proposed approach has advantages over a simple counting baseline.

15

Human activity recognition using improved complete ensemble EMD with adaptive noise and long short-term memory neural networks

Altuve Miguel, Lizarazo Paula, Villamizar Javier

Biocybernetics and Biomedical Engineering

|

2020

|

Vol. 40, no. 3

901--909

EN

The recognition of human activities is a topic of great relevance due to its wide range of applications. Different approaches have been proposed to recognize human activities, ranging from the comparison of signals with thresholds to the application of deep and machine learning techniques. In this work, the classification of six human activities (walking, walking downstairs, walking upstairs, standing, sitting, and lying down) is performed using bidirectional LSTM networks that exploit intrinsic mode function (IMF) representation of inertial signals. Records with inertial signals (accelerometer and gyroscope) of 2.56 s, available at the UCI Machine Learning Repository, were collected from 30 subjects using a smartphone. First, inertial signals were standardized to take them to the same scale and were decomposed into IMF using the improved complete ensemble empirical mode decomposition with adaptive noise (ICEEMDAN). IMF were then segmented (split) into nine segments of 1.28 s with 12.5% overlap and introduced to a first network with four outputs to identify the dynamic activities and the statics as a single class called ‘‘statics’’, giving 98.86% accuracy. Then, the non-segmented IMF of the records assigned to the statics class were introduced to a second network to classify their three activities, giving an accuracy of 88.46%. In total, 92.91% accuracy was obtained to classify the six human activities. This performance is because ICEEMDAN allowed the extraction of information that was embedded in the signal, and the segmentation of the IMF allowed the network to discriminate between static and dynamic activities.

16

A novel deep LSTM network for artifacts detection in microelectrode recordings

Hosny Mohamed, Zhu Minwei, Gao Wenpeng, Fu Yili

Biocybernetics and Biomedical Engineering

|

2020

|

Vol. 40, no. 3

1052--1063

EN

Microelectrode recording (MER) signals are world-widely used for validating the planned trajectories in the procedure of deep brain stimulation (DBS) surgery to obtain accurate position of electrodes inside the brain structure. Besides, MER signals are important source for studying extracellular neuronal activity and DBS biomarkers, such as, spike clustering and sorting. However, MER signals are prone to several artifacts derived from electrical equipment in the operating room, electrode movement and patient activities, etc., which reduce the signal-to-noise ratio of the MER signals. Therefore, in this paper, we propose a novel deep learning architecture based on long short-term memory (LSTM) network for automatic artifact detection in MER signals. Frequency and time-domain features were extracted from the raw MER signals and fed to the deep LSTM network. A manually annotated MER database obtained from 17 Parkinson's disease (PD) patients were used to validate the proposed architecture. The proposed architecture achieved promising results of 97.49% accuracy, 98.21% sensitivity and 96.87% specificity on an unseen test set. To our best knowledge, this is the first study to use LSTM network for artifacts detection in MER signals. The MER data will be available at http://homepage.hit.edu.cn/wpgao.

17

Short-term air pollution forecasting based on environmental factors and deep learning models

Arsov Mirche, Zdravevski Eftim, Lameski Petre, Corizzo Roberto, Koteli Nikola, Mitreski Kosta, Trajkovik Vladimir

Annals of Computer Science and Information Systems

|

2020

|

Vol. 21

15--22

EN

The effects of air pollution on people, the environment, and the global economy are profound - and often under-recognized. Air pollution is becoming a global problem. Urban areas have dense populations and a high concentration of emission sources: vehicles, buildings, industrial activity, waste, and wastewater. Tackling air pollution is an immediate problem in developing countries, such as North Macedonia, especially in larger urban areas. This paper exploits Recurrent Neural Network (RNN) models with Long Short-Term Memory units to predict the level of PM10 particles in the near future (+3 hours), measured with sensors deployed in different locations in the city of Skopje. Historical air quality measurements data were used to train the models. In order to capture the relation of air pollution and seasonal changes in meteorological conditions, we introduced temperature and humidity data to improve the performance. The accuracy of the models is compared to PM10 concentration forecast using an Autoregressive Integrated Moving Average (ARIMA) model. The obtained results show that specific deep learning models consistently outperform the ARIMA model, particularly when combining meteorological and air pollution historical data. The benefit of the proposed models for reliable predictions of only 0.01 MSE could facilitate preemptive actions to reduce air pollution, such as temporarily shutting main polluters, or issuing warnings so the citizens can go to a safer environment and minimize exposure.

18

Shallow, Deep, Ensemble models for Network Device Workload Forecasting

Liu Cenru

Annals of Computer Science and Information Systems

|

2020

|

Vol. 21

101--104

EN

Reliable prediction of workload-related characteristics of monitored devices is important and helpful for management of infrastructure capacity. This paper presents 3 machine learning models (shallow, deep, ensemble) with different complexity for network device workload forecasting. The performance of these models have been compared using the data provided in FedCSIS'20 Challenge. The R2 scores achieved from the cascade Support Vector Regression (SVR) based shallow model, Long short-term memory (LSTM) based deep model, and hierarchical linear weighted ensemble model are 0.2506, 0.2831, and 0.3059, respectively, and was ranked 3rd place in the preliminary stage of the challenges.

19

Future Graduate Salaries Prediction Model Based On Recurrent Neural Network

Siłka Jakub, Wieczorek Michał, Woźniak Marcin

Annals of Computer Science and Information Systems

|

2020

|

Vol. 21

427--430

EN

Prediction models are widely applied in several fields. In this study we present a discussion on using Recurrent Neural Network as predictor for salaries of future graduates. The model is based on feature analysis which leads to input values of the predictor. We have analyzed several compositions and ideas. As a result we have selected Recurrent Neural Network to be the most accurate. Presented results confirm this selection and show high precision.

20

Character-based recurrent neural networks for morphological relational reasoning

Mogren Olof, Johansson Richard

Journal of Language Modelling

|

2019

|

Vol. 7, No. 1

139--170

EN

We present a model for predicting inflected word forms based on morphological analogies. Previous work includes rule-based algorithms that determine and copy affixes from one word to another, with limited support for varying inflectional patterns. In related tasks such as morphological reinflection, the algorithm is provided with an explicit enumeration of morphological features which may not be available in all cases. In contrast, our model is feature-free: instead of explicitly representing morphological features, the model is given a demo pair that implicitly specifies a morphological relation (such as write: writes specifying infinitive:present). Given this demo relation and a query word (e.g. watch), the model predicts the target word (e.g. watches). To address this task, we devise a character-based recurrent neural network architecture using three separate encoders and one decoder. Our experimental evaluation on five different languages shows that the exact form can be predicted with high accuracy, consistently beating the baseline methods. Particularly, for English the prediction accuracy is 94.85%. The solution is not limited to copying affixes from the demo relation, but generalizes to words with varying inflectional patterns, and can abstract away from the orthographic level to the level of morphological forms.