Publication Date: 2024
Multimedia Tools and Applications (13807501)83(35)pp. 83275-83309
The atmosphere is one of the game elements that can significantly influence player's emotions. However, creating an immersive atmosphere that effectively influences player emotions poses several challenges, necessitating the utilization of various elements, such as audio-visual coordination and gameplay design. This paper introduces a general framework for procedurally generating dungeons with joyful and horror atmospheres in games, providing an abstract perspective to address these challenges. The proposed framework introduces a categorization system for game elements based on their role within the game. Leveraging this categorization, the Comprehensive Arrangement of Game Elements (CAGE) pattern is introduced, which facilitates the appropriate placement of elements within the dungeon environment. Subsequently, the General Framework for Generating Dungeons with Atmosphere (GFGDA) is employed to procedurally create the dungeon using the Feasible–Infeasible Two-Population (FI-2Pop) algorithm. To enhance gameplay experience, similar elements in the dungeon environment that impact gameplay are grouped and their coordination is evaluated by creating a graph based on the CAGE pattern. The transition and coordination of audio-visual elements along the path between these impactful elements are assessed in order to generate an immersive atmosphere within the dungeon. To ensure diversity, examining the variety of dungeons generated over 100 runs demonstrates that our method consistently produces distinct results in each iteration. Moreover, two comparative studies were conducted, one with 51 volunteers and another with 10 volunteers. In the first study, the Game Experience Questionnaire (GEQ) was utilized to assess the emotional impact of dungeons generated by our method. These were compared to dungeons created using a uniform random approach, alongside relevant research. The results suggest that our method significantly influences player emotions across the four components of the GEQ—sensory and imaginary immersion, flow, negative effects, and challenge—when compared to dungeons generated by the uniform random approach and another researched method. In another study, the emotional impact of two dungeons, one generated with joyful elements and the other with eerie elements, was evaluated using the GEQ. The findings indicate significant differences between the two components of the GEQ—tension and positive effects—when players interacted with the level containing joyful elements compared to the one with eerie elements. © The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2024.
The tourism industry has undergone a significant shift towards data-driven strategies in recent years. As a means of improving the quality of their service and performance, service providers are analyzing feedback from their customers to increase the number of tourists they attract. Negative feedback also provides valuable insights into the factors that detract from a location's appeal. Datasets that gather information on people's experiences and opinions of tourist destinations can be analyzed to extract valuable information. However, there are currently few existing datasets that specifically capture user reviews about historical and tourist attractions in Iran. To fill this gap, users have shared their travel experiences on various websites, and sentiment analysis can be employed to extract insights from this data. Effective sentiment analysis requires a suitable approach for data extraction, pre-processing, and storage. This study provides a framework for the user review dataset preparation, including data collection, ETL, data storage, and evaluation phases. A rich dataset containing user reviews about 178 Iran's historical and tourist attractions was prepared through the proposed framework in which automated crawlers were developed to collect data from Tripadvisor platforms. Data labelling was achieved using the DistilBERT-base-uncased language model for sentiment analysis and human evaluators for final annotations. A total of approximately 25 thousand samples were included in the dataset, and positive user comments outnumbered negative user comments by a wide margin. This high percentage of positive comments suggests that the locations were of a satisfactory standard, making it likely that users would return in the future. The findings of this study can help providers to improve the overall quality of their services by analyzing user reviews. The proposed framework and achieved dataset can also guide future efforts to leverage data for improved performance and customer satisfaction in the tourism industry by identifying areas that need improvement. © 2023 IEEE.
Publication Date: 2022
Applied Soft Computing (1568-4946)122
Despite the empirical success of Genetic programming (GP) in various symbolic regression applications, GP is not still known as a reliable problem-solving technique in this domain. Non-locality of GP representation and operators causes ineffectiveness of its search procedure. This study employs semantic schema theory to control and guide the GP search and proposes a local GP called semantic schema-based genetic programming (SBGP). SBGP partitions the semantic search space into semantic schemas and biases the search to the significant schema of the population, which is gradually progressing towards the optimal solution. Several semantic local operators are proposed for performing a local search around the significant schema. In combination with schema evolution as a global search, the local in-schema search provides an efficient exploration–exploitation control mechanism in SBGP. For evaluating the proposed method, we use six benchmarks, including synthesized and real-world problems. The obtained errors are compared to the best semantic genetic programming algorithms, on the one hand, and data-driven layered learning approaches, on the other hand. Results demonstrate that SBGP outperforms all mentioned methods in four out of six benchmarks up to 87% in the first set and up to 76% in the second set of experiments in terms of generalization measured by root mean squared error. © 2022 Elsevier B.V.
Publication Date: 2020
Road Materials and Pavement Design (14680629)21(3)pp. 850-866
Fatigue cracking is the most important structural failure in flexible pavements. The results of a laboratory study evaluating the fatigue properties of mixtures containing precipitated calcium carbonate (PCC) using indirect tensile fatigue (ITF) test were investigated in this paper. The hot mix asphalt (HMA) samples were made with four PCC contents (0%, 5%, 10%, and 15%), and tested at three different testing temperatures (2°C, 10°C and 20°C) and stress levels (100, 300, and 500 kPa). Due to the complex behaviour of asphalt pavement materials under various loading conditions, pavement structure, and environmental conditions, accurately predicting the fatigue life of asphalt pavement is difficult. In this study, genetic programming (GP) is utilised to predict the fatigue life of HMA. Based on the results of the ITF test, PCC improved the fatigue behaviour of studied mixes at different temperatures. But, the considerable negative effect of the increase of the temperature on the fatigue life of HMA is evident. On the other hand, the results indicate The GP-based formulas are simple, straightforward, and particularly valuable for providing an analysis tool accessible to practicing engineers. © 2018, © 2018 Informa UK Limited, trading as Taylor & Francis Group.
Publication Date: 2018
Applied Intelligence (0924669X)48(6)pp. 1442-1460
Semantic schema theory is a theoretical model used to describe the behavior of evolutionary algorithms. It partitions the search space to schemata, defined in semantic level, and studies their distribution during the evolution. Semantic schema theory has definite advantages over popular syntactic schema theories, for which the reliability and usefulness are criticized. Integrating semantic awareness in genetic programming (GP) in recent years sheds new light also on schema theory investigations. This paper extends the recent work in semantic schema theory of GP by utilizing information based clustering. To this end, we first define the notion of semantics for a tree based on the mutual information between its output vector and the target and introduce semantic building blocks to facilitate the modeling of semantic schema. Then, we propose information based clustering to cluster the building blocks. Trees are then represented in terms of the active occurrence of building block clusters and schema instances are characterized by an instantiation function over this representation. Finally, the expected number of schema samples is predicted by the suggested theory. In order to evaluate the suggested schema, several experiments were conducted and the generalization, diversity preserving capability and efficiency of the schema were investigated. The results are encouraging and remarkably promising compared with the existing semantic schema. © 2017, Springer Science+Business Media, LLC.
Publication Date: 2018
Soft Computing (14327643)22(10)pp. 3237-3260
A considerable research effort has been performed recently to improve the power of genetic programming (GP) by accommodating semantic awareness. The semantics of a tree implies its behavior during the execution. A reliable theoretical modeling of GP should be aware of the behavior of individuals. Schema theory is a theoretical tool used to model the distribution of the population over a set of similar points in the search space, referred by schema. There are several major issues with relying on prior schema theories, which define schemata in syntactic level. Incorporating semantic awareness in schema theory has been scarcely studied in the literature. In this paper, we present an improved approach for developing the semantic schema in GP. The semantics of a tree is interpreted as the normalized mutual information between its output vector and the target. A new model of the semantic search space is introduced according to semantics definition, and the semantic building block space is presented as an intermediate space between semantic and genotype ones. An improved approach is provided for representing trees in building block space. The presented schema is characterized by Poisson distribution of trees in this space. The corresponding schema theory is developed for predicting the expected number of individuals belonging to proposed schema, in the next generation. The suggested schema theory provides new insight on the relation between syntactic and semantic spaces. It has been shown to be efficient in comparison with the existing semantic schema, in both generalization and diversity-preserving aspects. Experimental results also indicate that the proposed schema is much less computationally expensive than the similar work. © 2017, Springer-Verlag GmbH Germany.
Publication Date: 2016
Soft Computing (14327643)20(5)pp. 2031-2045
Determining suitable mesh density for complicated finite element analysis, e.g., laser forming process, has always been the main concern of analytical engineers because of its high computation time and costs. Few works addressed the application of optimization methods for finite element analysis of linear path laser scan; however, no study has yet considered optimum finite element analysis of circular path laser forming. The main objective of this article is to develop a method for determining optimum mesh density to estimate the deflection caused by laser beam circular path scan considering analysis time and forming accuracy. Optimum ranges of mesh densities are investigated first and then a deflection estimating process based on adaptive-network-based fuzzy inference system has been introduced. The proposed model was finally optimized using genetic algorithm considering accuracy and time. The numerical analysis results were finally confirmed by the conducted experimental results. © 2015, Springer-Verlag Berlin Heidelberg.
Publication Date: 2016
Applied Intelligence (0924669X)44(1)pp. 67-87
Schema theory is the most well-known model of evolutionary algorithms. Imitating from genetic algorithms (GA), nearly all schemata defined for genetic programming (GP) refer to a set of points in the search space that share some syntactic characteristics. In GP, syntactically similar individuals do not necessarily have similar semantics. The instances of a syntactic schema do not behave similarly, hence the corresponding schema theory becomes unreliable. Therefore, these theories have been rarely used to improve the performance of GP. The main objective of this study is to propose a schema theory which could be a more realistic model for GP and could be potentially employed for improving GP in practice. To achieve this aim, the concept of semantic schema is introduced. This schema partitions the search space according to semantics of trees, regardless of their syntactic variety. We interpret the semantics of a tree in terms of the mutual information between its output and the target. The semantic schema is characterized by a set of semantic building blocks and their joint probability distribution. After introducing the semantic building blocks, an algorithm for finding them in a given population is presented. An extraction method that looks for the most significant schema of the population is provided. Moreover, an exact microscopic schema theorem is suggested that predicts the expected number of schema samples in the next generation. Experimental results demonstrate the capability of the proposed schema definition in representing the semantics of the schema instances. It is also revealed that the semantic schema theorem estimation is more realistic than previously defined schemata. © 2015, Springer Science+Business Media New York.
Zojaji, Z.,
Jahani, M.,
Zojaji, Z.,
Montazerolghaem, A.,
Palhang, M.,
Ramezani, R.,
Golkarnoor, A.,
Safaei, A.A.,
Bahak, H.,
Saboori, P.,
Halaj, B.S. Publication Date: 2025
Journal Of Medical Signals And Sensors (22287477)15(1)
Background: The pharmaceutical industry has seen increased drug production by different manufacturers. Failure to recognize future needs has caused improper production and distribution of drugs throughout the supply chain of this industry. Forecasting demand is one of the basic requirements to overcome these challenges. Forecasting the demand helps the drug to be well estimated and produced at a certain time. Methods: Artificial intelligence (AI) technologies are suitable methods for forecasting demand. The more accurate this forecast is the better it will be to decide on the management of drug production and distribution. Isfahan AI competitions-2023 have organized a challenge to provide models for accurately predicting drug demand. In this article, we introduce this challenge and describe the proposed approaches that led to the most successful results. Results: A dataset of drug sales was collected in 12 pharmacies of Hamadan University of Medical Sciences. This dataset contains 8 features, including sales amount and date of purchase. Competitors compete based on this dataset to accurately forecast the volume of demand. The purpose of this challenge is to provide a model with a minimum error rate while addressing some qualitative scientific metrics. Conclusions: In this competition, methods based on AI were investigated. The results showed that machine learning methods are particularly useful in drug demand forecasting. Furthermore, changing the dimensions of the data features by adding the geographic features helps increase the accuracy of models. © 2025 Journal of Medical Signals & Sensors.
Publication Date: 2025
Expert Systems with Applications (0957-4174)275
Entity search websites enable consumers to purchase products and entities. These websites commonly provide predefined search filters and contain entity information and user-created reviews. Consumers commonly exploit filters to eliminate irrelevant entities. Consequently, they explore reviews about the purified entities to identify an entity centered on their preferences. Inspired by this consumer-centric process, this paper introduces a mechanized model for entity search. In the presented model, consumer preferences are considered as queries comprising configurations for predefined filters and a textual part containing more explanations and constrains. Previous works on entity retrieval have not employed capabilities of structured entity information, and complete potential of insights in user reviews. In contrast, the presented model, considering this limitations, exploits structured entity information of predefined_slots to filter out many irrelevant entities. Consequently, it ranks entities by comparing text_part of consumer preferences with the entire core content of user reviews, indicating the order of corresponding entities centered on consumer preferences. The entity search model presented in this paper employs cutting-edge natural language processing (NLP) methods. Moreover, this paper introduces a model to curate entity search dataset, considering structured entity information and objective information in user reviews and exploit it to create a dataset by extracting and labeling restaurant information from Tripadvisor. The created entity search model incorporates a monoBERT-based text_ranker, fine-tuned employing the created training dataset, and evaluations indicate perceptible improvements in mean reciprocal rank (MRR), mean average precision (MAP), and normalized discounted cumulative gain (nDCG). © 2025 Elsevier Ltd
Publication Date: 2025
PLoS ONE (19326203)20(5 May)
The extraction of subjective comparative relations is essential in the field of question answering systems, playing a crucial role in accurately interpreting and addressing complex questions. To tackle this challenge, we propose the SCQRE model, specifically designed to extract subjective comparative relations from questions by focusing on entities, aspects, constraints, and preferences. Our approach leverages multi-task learning, the Natural Language Inference (NLI) paradigm, and a specialized adapter integrated into RoBERTa_base_go_emotions to enhance performance in Element Extraction (EE), Compared Elements Identification (CEI), and Comparative Preference Classification (CPC). Key innovations include handling X- and XOR-type preferences, capturing implicit comparative nuances, and the robust extraction of constraints often neglected in existing models. We also introduce the Smartphone-SCQRE dataset, along with another domain-specific dataset, BrandsCompSent-19-SCQRE, both structured as subjective comparative questions. Experimental results demonstrate that our model outperforms existing approaches across multiple question-level and sentence-level datasets and surpasses recent language models, such as GPT-3.5-turbo-0613, Llama-2-70b-chat, and Qwen-1.5-7B-Chat, showcasing its effectiveness in question-based comparative relation extraction. © 2025 Babaali et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Publication Date: 2025
Neurocomputing (09252312)616
Open Information Extraction (Open IE) is the task of identifying structured and machine-readable information from natural language text within an open domain context. This research area has gained significant importance in the field of natural language processing (NLP), attracting considerable attention for its potential to extract valuable information from unstructured textual data. Previous investigations heavily relied on manual extraction patterns and various NLP tools. While these methods often produce errors that accumulate and propagate throughout the systems, ultimately affecting the accuracy of the results. Moreover, recent Open IE studies have focused on extracting binary relations involving two entities. However, these binary approaches occasionally lead to the omission of essential information in the text, preventing a deeper comprehension of the content. This limitation arises from the fact that real-world relations often involve multiple entities, but binary approaches may oversimplify these relations and miss additional details crucial for a thorough understanding of text. To address these challenges, our study introduces an innovative system called “Open N-ary Information EXtraction (ONIEX).” This system incorporates two novel techniques: multihead relation attention mechanism and relation embedding. Multihead relation attention, in combination with relation embedding, enables the system to focus on relations extracted through the SpanBERT model and accurately identify associated entities for each relation. The ONIEX system's superior performance is substantiated through extensive experiments conducted on the OpenIE4 and LSOIE datasets, benchmark datasets for Open n-ary Information Extraction (Open n-ary IE). The results demonstrate the superiority of the ONIEX system over the existing state-of-the-art systems. © 2024 Elsevier B.V.
Publication Date: 2025
Information Sciences (0020-0255)692
Open relation extraction is a critical task in natural language processing aimed at automatically extracting relations between entities in open-domain corpora. Most existing systems focus on extracting binary relations (relations between two entities) while extracting more complex n-ary relations (involving more than two entities) remains a significant challenge. Additionally, many previous systems rely on hand-crafted patterns and natural language processing tools, which result in error accumulation and reduced accuracy. The current study proposes a novel approach to open n-ary relation extraction that leverages recent advancements in deep learning architectures. This approach addresses the limitations of existing open relation extraction systems, particularly their reliance on hand-crafted patterns and their focus on binary relations. It utilizes SpanBERT to capture relational patterns from text data directly and introduces entity embedding vectors to create distinct representations of entities within sentences. These vectors enhance the proposed system's understanding of the entities within the input sentence, leading to more accurate relation extraction. Notably, the proposed system in the present study achieves an F1-score of 89.79 and 92.67 on the LSOIE-wiki and OpenIE4 datasets, outperforming the best existing models by over 12% and 10%, respectively. These results highlight the effectiveness of the proposed approach in addressing the challenges of open n-ary relation extraction. © 2024 Elsevier Inc.
Publication Date: 2024
PLoS ONE (19326203)19(5 May)
In the domain of question subjectivity classification, there exists a need for detailed datasets that can foster advancements in Automatic Subjective Question Answering (ASQA) systems. Addressing the prevailing research gaps, this paper introduces the Fine-Grained Question Subjectivity Dataset (FQSD), which comprises 10,000 questions. The dataset distinguishes between subjective and objective questions and offers additional categorizations such as Subjective-types (Target, Attitude, Reason, Yes/No, None) and Comparison-form (Single, Comparative). Annotation reliability was confirmed via robust evaluation techniques, yielding a Fleiss's Kappa score of 0.76 and Pearson correlation values up to 0.80 among three annotators. We benchmarked FQSD against existing datasets such as (Yu, Zha, and Chua 2012), SubjQA (Bjerva 2020), and ConvEx-DS (Hernandez-Bocanegra 2021). Our dataset excelled in scale, linguistic diversity, and syntactic complexity, establishing a new standard for future research. We employed visual methodologies to provide a nuanced understanding of the dataset and its classes. Utilizing transformer-based models like BERT, XLNET, and RoBERTa for validation, RoBERTa achieved an outstanding F1-score of 97%, confirming the dataset's efficacy for the advanced subjectivity classification task. Furthermore, we utilized Local Interpretable Model-agnostic Explanations (LIME) to elucidate model decision-making, ensuring transparent and reliable model predictions in subjectivity classification tasks. Copyright: © 2024 Babaali et al.
Question-answering systems, characterized by their fundamental functions of question classification, information retrieval, and answer selection, demand refinement to enhance precision in retrieving exact answers. Question classification, a cornerstone task, anticipates the probable answer to a posed query. However, the performance of question classification algorithms is hampered, particularly in agglutinative languages with complex morphology like Persian, where linguistic resources are limited. In this study, we propose a novel multi-layer Long-short-term memory (LSTM) Attention Convolutional Neural Network (CNN) (LACNN) classifier, tailored to extract pertinent information from Persian language contexts. Notably, this model operates autonomously, obviating the need for prior knowledge or external features. Moreover, we introduce UIMQC, the first medical question dataset in Persian, derived from the English GARD dataset. The inquiries within UIMQC are inherently intricate, often pertaining to rare diseases necessitating specialized diagnosis. Our experimental findings demonstrate a notable enhancement over baseline methods, with a 9% performance increase on the UTQC dataset, and achieving 67.08% accuracy on the UIMQC dataset. Consequently, we advocate for the adoption of the LACNN model in various morphological analysis tasks across low-resource languages, as in Question Answering systems it improves the performance for retrieving accurate answers to the users’ queries. ©2024 IEEE.
Large language models excel in various natural language processing tasks but often struggle with knowledge-intensive queries, particularly those involve rare entities or require precise factual information. This paper presents a novel framework that enhances capabilities of an LLM-based question answering system by incorporating structured knowledge from knowledge graphs. Our approach employs entity extraction, semantic similarity scoring, and adaptive graph exploration to efficiently navigate and extract relevant information from knowledge graphs. The core of the presented solution is a knowledge graph-enhanced language model process that iteratively refines subgraph exploration and answer generation, complemented by a fallback mechanism for robustness across diverse question types. Experiments on location-based questions from the Entity Questions dataset demonstrate significant improvements in the quality of responses. Using the Gemini 1.5 Flash model, our system achieved an accuracy increase from 36% to 71% for partially correct answers and from 22% to 69% for exactly correct answers, as evaluated by human assessors. This approach offers a promising direction for developing more reliable and accurate question answering systems, particularly for queries involving long-tail entities or specific factual knowledge. © 2024 IEEE.
Retrieval augmented generation (RAG) models, which integrate large-scale pre-trained generative models with external retrieval mechanisms, have shown significant success in various natural language processing (NLP) tasks. However, applying RAG models in Persian language as a low-resource language, poses distinct challenges. These challenges primarily involve the preprocessing, embedding, retrieval, prompt construction, language modeling, and response evaluation of the system. In this paper, we address the challenges towards implementing a real-world RAG system for Persian language called PersianRAG. We propose novel solutions to overcome these obstacles and evaluate our approach using several Persian benchmark datasets. Our experimental results demonstrate the capability of the PersianRAG framework to enhance question answering task in Persian. © 2024 IEEE.
Publication Date: 2023
Expert Systems with Applications (0957-4174)229
Question Classification is the very first and pivotal step in Question Answering(QA) systems. It maps a given question into a predefined class. Classification of Consumer Health Questions contributes to answering medical queries of non-expert users. It plays a principal role in the performance of Consumer Health Question Answering systems. There are several approaches to the classification of questions. However, they do not consider the nature of Consumer Health Questions. In this study, we perform experiments focusing on the question representation methods to classify Consumer Health Questions accurately. These experiments provide an intuition of practical considerations for Consumer Health Question classification. The questions were represented through certain manually-designed features, word embedding, sentence embedding, and finetuning-based techniques. We create question classifiers that involve designing models for representing and categorizing questions. Based on the result, the fine-tuned Bidirectional Encoder Representations from Transformers(BERT)-large-based model with an accuracy of 86.00% on Genetic and Rare Diseases (GARD) and 80.40% on Yahoo! Answers-based datasets outperform previous works. In addition, the method based on A Robustly Optimized BERT Pretraining Approach(RoBERTa)-large model with 86.20% accuracy surpasses fine-tuned BERT-large model on the GARD dataset. Examining the outcomes of the models gives insight into decisive considerations for representing and classifying Consumer Health Questions. © 2023 Elsevier Ltd
Publication Date: 2023
Engineering Applications of Artificial Intelligence (09521976)126
COVID-19 is a huge global challenge, on which numerous people and resources all over the world have been focused over the last three years to slow down its spread. A screening scale for the early detection of COVID-19 cases is highly important to diagnose patients before they transmit the disease to others. Using the data collected from Sept. 2020 to Nov. 2020 from 12,123 employees of a company including a weekly self-report questionnaire and information on their demographics, chronic disease, work, and commute, we developed a short, optimized screening scale for the early detection of COVID-19. To develop the scale, we have modified an interpretable machine learning (ML) method which uses integer programming techniques, and consequently reduced its computational cost significantly and made it robust to erroneous data and protective of sensitive information without loss of performance. The developed scale is highly interpretable and easily scored to the extent that it does not require a computer or even a calculator. However, we showed that the operating characteristics of the screening scale on the test dataset are excellent (area under the ROC curve = 82% and area under the Precision–Recall curve = 27%) and comparable to other well-known ML models such as penalized logistic regression and random forest when predicting positive PCR test results within 7 days from the date the self-report questionnaire was filled out. From Dec. 2020 to Nov. 2021, the developed scale has been used in the company weekly to monitor and assess its employees. Although different variants of COVID-19 have emerged since the development of the scale, the results show that the scale has performed well in the early detection of COVID-19, which helped the company take appropriate actions to slow down the spread of COVID-19 among its employees successfully. We expect our findings in this paper to help researchers have a better understanding of COVID-19 and control the spread of the disease. © 2023 Elsevier Ltd
Kazemi, A.,
Zojaji, Z.,
Malverdi, M.,
Mozafari, J.,
Ebrahimi, F.,
Abadani, N.,
Varasteh, M.R.,
Nematbakhsh, M.A. Publication Date: 2023
Information Retrieval Journal (13864564)26(1)
Nowadays, a considerable volume of news articles is produced daily by news agencies worldwide. Since there is an extensive volume of news on the web, finding exact answers to the users’ questions is not a straightforward task. Developing Question Answering (QA) systems for the news articles can tackle this challenge. Due to the lack of studies on Persian QA systems and the importance and wild applications of QA systems in the news domain, this research aims to design and implement a QA system for the Persian news articles. This is the first attempt to develop a Persian QA system in the news domain to our best knowledge. We first create FarsQuAD: a Persian QA dataset for the news domain. We analyze the type and complexity of the users’ questions about the Persian news. The results show that What and Who questions have the most and Why and Which questions have the least occurrences in the Persian news domain. The results also indicate that the users usually raise complex questions about the Persian news. Then we develop FarsNewsQA: a QA system for answering questions about Persian news. We developed three models of the FarsNewsQA using BERT, ParsBERT, and ALBERT. The best version of the FarsNewsQA offers an F1 score of 75.61%, which is comparable with that of QA system on the English SQuAD dataset made by the Stanford university, and shows the new Bert-based technologies works well for Persian news QA systems. © 2023, The Author(s), under exclusive licence to Springer Nature B.V.
Publication Date: 2023
Expert Systems with Applications (0957-4174)211
This study addresses the problem of relation detection for answering single-relation factoid questions over knowledge bases (KBs). In this kind of questions, the answer is obtained from a single KB fact in the form of subject-predicate-object. Conventional fact extraction methods have two steps: entity linking and relation detection, in which the output of the entity linking is used by the relation detection step to first find candidate relations, and then choose the best relation from candidate relations. Such methods have difficulties with the relation detection if there is an error or ambiguity in the entity linking step. This paper explores the relation detection task without the entity-linking step utilizing the hierarchical structure of relations and an out-of-box POS tagger. As relations are of different levels of abstraction, the proposed solution uses multiple classifiers in pipeline, each of which uses separate BiGRU neural networks fed with questions embedded with one-hot encoding at the character level. Besides, to increase the accuracy of the proposed model and to avoid the need for large amounts of training data, after each word of the question, its POS tag is inserted before feeding the network. The experimental results show that the accuracy of the proposed solution for the direct relation detection is 89.5%. In addition, the proposed solution can be used for the indirect relation detection whose accuracy is 96.3%, which is higher than state-of-the-art relation detection techniques. Finally, the positive effects of using POS tags have been examined. © 2022 Elsevier Ltd
Publication Date: 2022
Knowledge-Based Systems (0950-7051)235
One of the effective approaches for answering natural language questions (NLQs) over knowledge graphs consists of two main stages. It first creates a query graph based on the NLQ and then matches this graph over the knowledge graph to construct a structured query. An obstacle in the first stage is the need to build question interpretations with candidate resources, even if some implicit phrases exist in the sentence. In the second stage, a serious problem is to map diverse NLQ relations to their corresponding predicates. To overcome these problems, in this paper, we propose a novel sequential word parsing-based method to construct and refine an uncertain question graph that is disambiguated directly over the knowledge graph. Instead of relying on the syntactic dependency relations and some predefined rules that recognize the relations and their arguments, we consider the identified entities and variables in the NLQ as well as their corresponding place in the structure of a query graph pattern to build question triples. First, by leveraging the ordered dependency tree of an NLQ, sentence words are reordered. Then the question graph structure is constructed by parsing the new sequence backward, starting from the identified items. Subsequently, the question graph is refined by eliminating the useless elements. Additionally, to improve the relation similarity measure in the graph similarity process, we exploit the knowledge hidden in a relation pattern taxonomy. Experimental studies over several benchmarks demonstrate that our proposed approach is effective as it achieves promising results in answering the complex NLQs. © 2021
Dialogue state tracking is one of the main components in task-oriented dialogue systems whose duty is to track the user's goal during the conversation. Due to the diversity in natural languages and existing utterances, the user requests may include unknown values at different turns in these systems. However, predicting the actual values of the user requests is necessary for completing the intended task. In existing studies, these values are determined using span-based methods to predict a span in utterances or previous dialogues. However, the slots are not correctly filled when values are multi-word. In addition, in some scenarios, the slot values in a given turn may depend on previous dialogue states. However, due to the limitation of the input length of language models, it is impossible to access all the previous dialogue states. This study proposes a new approach that uses a span-tokenizer and adds the Bi-LSTM layer on top of the BERT model to predict the exact span of multi-word values. This approach uses parameters like user utterances, important dialogue histories, and all dialogue states as input to decrease the length of the sequences. The results show that this strategy has led to a 1.80% improvement in the joint-goal accuracy and 0.15% improvement in the slot accuracy metrics over the MultiWOZ 2.1 dataset compared to the SAVN model. © 2022 IEEE.
Publication Date: 2022
IEEE Access (21693536)10pp. 26045-26057
Developing Question Answering systems (QA) is one of the main goals in Artificial Intelligence. With the advent of Deep Learning (DL) techniques, QA systems have witnessed significant advances. Although DL performs very well on QA, it requires a considerable amount of annotated data for training. Many annotated datasets have been built for the QA task; most of them are exclusively in English. In order to address the need for a high-quality QA dataset in the Persian language, we present PersianQuAD, the native QA dataset for the Persian language. We create PersianQuAD in four steps: 1) Wikipedia article selection, 2) question-answer collection, 3) three-candidates test set preparation, and 4) Data Quality Monitoring. PersianQuAD consists of approximately 20,000 questions and answers made by native annotators on a set of Persian Wikipedia articles. The answer to each question is a segment of the corresponding article text. To better understand PersianQuAD and ensure its representativeness, we analyze PersianQuAD and show it contains questions of varying types and difficulties. We also present three versions of a deep learning-based QA system trained with PersianQuAD. Our best system achieves an F1 score of 82.97% which is comparable to that of QA systems on English SQuAD, made by the Stanford University. This shows that PersianQuAD performs well for training deep-learning-based QA systems. Human performance on PersianQuAD is significantly better (96.49%), demonstrating that PersianQuAD is challenging enough and there is still plenty of room for future improvement. PersianQuAD and all QA models implemented in this paper are freely available. © 2013 IEEE.
Publication Date: 2022
Journal of Supercomputing (15730484)78(1)pp. 596-615
Today, social networks have created a wide variety of relationships between users. Friendships on Facebook and trust in the Epinions network are examples of these relationships. Most social media research has often focused on positive interpersonal relationships, such as friendships. However, in many real-world applications, there are also networks of negative relationships whose communication between users is either distrustful or hostile in nature. Such networks are called signed networks. In this work, sign prediction is made based on existing links between nodes. However, in real signed networks, links between nodes are usually sparse and sometimes absent. Therefore, existing methods are not appropriate to address the challenges of accurate sign prediction. To address the sparsity problem, this work aims to propose a method to predict the sign of positive and negative links based on clustering and collaborative filtering methods. Network clustering is done in such a way that the number of negative links between the clusters and the number of positive links within the clusters are as large as possible. As a result, the clusters are as close as possible to social balance. The main contribution of this work is using clustering and collaborative filtering methods, as well as proposing a new similarity criterion, to overcome the data sparseness problem and predict the unknown sign of links. Evaluations on the Epinions network have shown that the prediction accuracy of the proposed method has improved by 8% compared to previous studies. © 2021, The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature.
Publication Date: 2022
Computational Intelligence and Neuroscience (discontinued) (16875273)2022
Question answering (QA) systems have attracted considerable attention in recent years. They receive the user's questions in natural language and respond to them with precise answers. Most of the works on QA were initially proposed for the English language, but some research studies have recently been performed on non-English languages. Answer selection (AS) is a critical component in QA systems. To the best of our knowledge, there is no research on AS for the Persian language. Persian is a (1) free word order, (2) right-to-left, (3) morphologically rich, and (4) low-resource language. Deep learning (DL) techniques have shown promising accuracy in AS. Although DL performs very well on QA, it requires a considerable amount of annotated data for training. Many annotated datasets have been built for the AS task; most of them are exclusively in English. In order to address the need for a high-quality AS dataset in the Persian language, we present PASD; the first large-scale native AS dataset for the Persian language. To show the quality of PASD, we employed it to train state-of-the-art QA systems. We also present PerAnSel: a novel deep neural network-based system for Persian question answering. Since the Persian language is a free word-order language, in PerAnSel, we parallelize a sequential method and a transformer-based method to handle various orders in the Persian language. We then evaluate PerAnSel on three datasets: PASD, PerCQA, and WikiFA. The experimental results indicate strong performance on the Persian datasets beating state-of-the-art answer selection methods by 10.66% on PASD, 8.42% on PerCQA, and 3.08% on WikiFA datasets in terms of MRR. © 2022 Jamshid Mozafari et al.
