chunk
stringlengths
20
498
Set of Guidelines 1. Related to Implementing the Data Lake. The Data Lake is a centralized data repository, regardless of format or structure. It is often used to store large amounts of raw data that can be used for analytics, machine learning, and other data-driven applications. When implementing the Data Lake for storing documents, 1t is important to consider the following guidelines.
— Use a Schema-Less Approach. Documents can be semi-structured or unstructured. Using a schema-less repository allows storing documents in their original format without a rigid structure, as discussed in [35].
— Choose a Distributed File System. The Data Lake can handle large amounts of doc- uments and is accessed by multiple users at the same time. In this context, a dis- tributed file system, such as the Hadoop Distributed File System (HDFS) [37], 1s an appropriate choice for storing raw documents.
— Employ a metadata management system. It is usually difficult to understand the meaning of documents without some context [35]. A metadata management system can assist to track and manage related data in the Data Lake.
In BigQA, we employ the Data Lake to store the raw documents inserted into the system. Examples of current technologies to store these documents include NoSQL databases like MongoDB, Apache CouchDB, Azure Cosmos DB, and Oracle NoSQL Database, as well as Blob services like Amazon S3, Azure Blob Storage, and Google Cloud Storage. Another possibility is to employ a Data Lakehouse [1] like Databricks and Delta Lake.
Set of Guidelines 2. Related to Implementing the Document Store. The Document Store is a database that stores textual data converted from documents. Most text data are either unstructured or semi-structured. The following guidelines should be considered when implementing the Document Store.
54 L. M. Pereira Moraes et al. — Specify the Schema Design. The schema is the physical implementation of the data model. It defines how to store data in the Document Store. An example is keeping the text in a text field, the meta-information in a JISON field, and the relationships in a graph database. It is defined as follows. e Text. Refers to textual data usually stored as plain text.
e Meta-Information. Concerns to textual data storing meta-information, such as authors, title, source identifier and collection date. e Relationships. Refers to textual data storing relationships between other data. For instance, when two converted data are part of the same book. — Choose a Data Indexing. Each Document Store tool has a unique method of indexing data. Choosing an appropriate data indexing guarantees efficient retrieval of docu- ments based on the schema employed.
— Employ a Metadata Management System. Similar to the Data Lake, a metadata man- agement system can help track and manage related data in the Document Store. The Data Lake and the Document Store typically share the same metadata management system.
The Document Store can be implemented using NoSQL databases [15] like Elastic- search and OpenSearch and vector databases [12] like Milvus, Pinecone, Qdrant, and FAISS. Some factors to analyze when choosing the technology include the size and type of data to store and the level of scalability required by the Big Data Question Answering system to develop.
Set of Guidelines 3. Related to Building the Big Data Storage Layer. The Big Data Storage Layer is responsible for storing and processing large volumes of data. Further- more, 1t must be scalable and fault-tolerant. The layer performs multiple processing, cleaning, and transformation steps for each document to prepare data to store in the Document Store. The guidelines to observe to implement the Big Data Storage Layer are detailed as follows.
— Operate with Batch Processing Pattern. Converting the raw data from the Data Lake to insert in the Document Store involves loading, cleaning, transforming, and ana- lyzing many documents. Thus, this layer must support batch processing to clean and transform the documents into converted text data, using batch processing patterns such as Lambda [11] and Sigma [5].
— Operate with Stream Processing Pattern. Stream processing operations are a set of operations used to process text data in near real-time. Developers optionally imple- ment stream processing. These operations clean and transform documents before storing related converted data in the Document Store, using stream processing pat- terns such as Kappa [38] and Sigma [5].
Set of Guidelines 4. Related to Building the Big Querying Layer. In the Big Query- ing Layer, the Document Retriever retrieves a set of documents relevant to a given query. The Document Reader reads the retrieved documents and extracts the answers to the query. When building the Big Querying Layer, it 18 essential to consider the guide- lines described as follows.
BigQA: Big Data Question Answering Architecture 55
— Specify Separate or Combined Components. The Document Retriever and the Doc- ument Reader can be implemented separately. In this context, it can be implemented using different technologies, but making 1t harder to maintain and integrate. On the other hand, it can be combined using close solutions. Combining components makes integration tighter and easier. However, separated solutions offer a greater level of customization than combined solutions.
— Choose the Appropriate Document Reader. When implemented separately, the Doc- ument Retriever can use different techniques to rank documents. These techniques include sparse methods like TF-IDF [34] and BM25 [31], as well as dense methods like DPR [13]. — Choose the Appropriate Document Retriever. When implemented separately, the Document Reader typically uses a variety of robust NLP techniques to extract the answers, such as GPT [29], BERT [7] and RoBERTa [19].
— Choose the Appropriate Combined. Using combined solutions, choose a solution that reflects the software used in the company and its use case needs, examples of private solutions that combine these two components are IBM Watson Discovery, Amazon Kedra and Sinch AskFrank.
— Implementation Scalability. The Document Reader and Retriever should scale to handle large amounts of data. Examples of parallel computing strategies to imple- ment these components are Kubernetes [28], Docker Swarm, Red Hat Openshift, and Amazon Elastic Container Service.
Set of Guidelines 5. Related to Building the Insights Layer. The Insights Layer is responsible for providing insights from the data. Indeed, organizations can achieve valu- able insights to improve their business performance and decision-making by using the right tools. When building the Insights Layer, it is essential to follow the guidelines described as follows.
— Choose a Reporting Tool. Reporting tools are used to generate reports that summa- rize the data. They can track trends and support decisions. There are many reporting tools available, such as Google Data Studio, Microsoft Power BI, Tableau, Pentaho, Metabase, Apache Superset, and Kibana.
— Choose a Data Analysis Tool. Data analysis tools are used to analyze the data in more detail. They find patterns, correlations, and outliers. Python, R, and Julia are pro- gramming languages used for data analysis. SAS, Dataiku, Jupyterlab, and Apache Zeppelin are examples of data analysis tools.
— Choose a Data Mining Tool. Data mining tools are used to discover hidden data patterns and relationships. These tools can support the investigation of new insights that would not be visible from the data alone. Examples of data mining tools include WEKA, KNIME, RapidMiner, and Orange. 5.2 Pipelines for Architecture Instantiation
5.2 Pipelines for Architecture Instantiation In this section, we provide three examples of pipelines for the architecture proposed in Sect.4, using the guidelines introduced in Sect. 5.1. These pipelines demonstrate the practical implementation of the proposed BigQA architecture and showocase its imple- mentation versatility and adaptability.
56 L. M. Pereira Moraes et al.
The first pipeline, depicted in Fig.2, uses open-source technologies. Regarding the Big Data Storage Layer, we employ Delta Lakeº as the Data Lake component and Elas- ticsearch' as the Document Store. Delta Lake is an open-source storage tool that man- ages large-scale, constantly evolving data lakes with support of big data processing frameworks, such as Apache Spark. Furthermore, Elasticsearch is an open-source big
data text search engine designed to be distributive, scalable, and near real-time capa- ble [15].
In the Big Querying layer, we use the following open-source algorithms: BM25 [31] as the Document Retriever and ROBERTàa [19] as the Document Reader. As the pipeline is independent of data input and output, we can employ any components for implement- ing the Input and Communication Layers. In Fig.2, the Input Layer receives documents as ISON files, and the Communication Layer supports users interacting through a API component.
By incorporating open-source technologies into BigQOA, we harness the power of collaborative development and community support, fostering innovation and adaptabil- ity for the instantiated architecture. Furthermore, BigQA ensures accessibility and flex- ibility in its implementation. Input Layer Big Data Storage Layer Big Querying Layer Comr::;:auon Delta Lake Elasticsearch &D JSON documents Document Document Document Store Retriever Reader <' º '> A J BngÉ RoBERTa
<' º '> A J BngÉ RoBERTa Fig. 2. BigQA pipeline using open-source technologies.
Figure 3 depicts the second pipeline, which uses paid private technologies. We employ the AWS Cloud Service. We use AWS S3 and Amazon OpenSearchº to imple- ment the Data Lake and the Document Store, respectively. Moreover, we employ Ama- zon Kendra, a Platform-as-a-Service (PaaS) solution, as the Big Querying Layer, com- bining the functionalities of the Document Retriever and the Document Reader compo- nents into one core query engine. Communication Layer Y ; r aus D)
Communication Layer Y ; r aus D) Document Document Retriever Reader Input Layer Big Data Storage Layer Big Querying Layer OpenSearch Documents Document Store Front-end Fig. 3. BigQA pipeline using paid private technologies. 6 Delta Lake - https://delta.io/. ? Elasticsearch - https://www.elastic.co/. $ Amazon OpenSearch - https://aws.amazon.com/opensearch-service/.
BigQA: Big Data Question Answering Architecture 57 By incorporating paid proprietary technologies into BigQA, we can obtain tailored support, specialized tools and services. Furthermore, BigQA can provide enterprise- grade features to enhance its performance, robustness, and reliability while meeting specific business requirements.
Each Big Data Question Answering system may have different requirements, so the implementation of BigQA does not have to include every layer shown in its architecture (Sect. 4) Developers should choose the appropriate components and layers according to their specific requirements. For instance, the pipelines described in Figs. 2 and 3 contain only four layers and a few components.
However, incorporating data mining and data analysis tools into pipelines is required to support decision-making. The third pipeline, depicted in Fig.4, shows how the open-source pipeline illustrated in Fig. 2 can be extended with open-source mon- itoring and reporting tools. The Insights layer uses Metricbeat for metrics, Logstash for logging, and Kibana for reporting. Metricbeat collects metrics from systems and
applications, such as memory and disk usage. Logstash collects, processes, and stores data from various sources, such as metric and logging sources. It collects metrics from Metricbeat and stores them in a database format. Kibana is a web-based visualization tool that can visualize the formatted data from Logstash. It creates charts, graphs, and dashboards to help managers monitor the system and extract data insights.
Communication Input Layer Layer Big Data Storage Layer Big Querying Layer Delta Lake Elasticsearch o <' º '> A |3M2.ÉE RoBEReTÉ) - JSON Document Document Document documents Store Retriever Reader Insights Layer Metricbeat Logstash , Kibana . .-.JLJKT Metric Logging Reporting Manager Tool Tool Tool Fig. 4. BIigQA pipeline extending Fig.2 to encompass the Insights Layer.
We strongly recommend using innovative approaches and modular components in the implementation process. Every BigOA component has a specific purpose and can operate autonomously. As stated by the Agile manifesto [21], the components should be independently and evolutionarily developed. Furthermore, it is needed promote col- laboration across times, location, and organizational boundaries to have agile teams to simplify the development process [36]. 6 CaseStudy: Pharmaceutical Company
6 CaseStudy: Pharmaceutical Company In this section, we present a case study to show how to deploy BigQA to enable a knowledge base containing real-world documents. Our goal is not to perform an exten-
58 L. M. Pereira Moraes et al. sive analysis of the architecture components. Instead, we implement a real-world case to assess the architecture purpose, following the design principles and guidelines pro- posed in Sects. 3 and 5, respectively. Section 6.1 describes how to instantiate BigQA. Section 6.2 details the queries. 6.1 . Architecture Instantiation
6.1 . Architecture Instantiation Figure 2 depicts the BigOA components and layers instantiated in the case study. The Input Layer contains ISON documents obtained from the training sets of two real-world datasets: (1) the Stanford Question Answering Dataset (SQuAD) v1.1 [30]; and (11) the COVID-QA [22]. Besides other open-source components.
SQuAD is a valuable question answering dataset for a wide range of domains. There are more than 18k different converted data. They also include over 87k questions and answers about Wikipedia articles. The content covers several topics, including phar- macy, software testing, TV series, car companies, and geology.
There are more than 2k questions and answers in COVID-QA, carefully annotated by biomedical experts. These experts have reviewed 147 scientific articles specifically focused on COVID-19. This dataset is not open-domain. We incorporated it in Query 3 (Sect. 6.2) as data augmentation to demonstrate how BigQA can effectively incorporate data from various data source and formats.
The Data Lake, in this particular case, does not retain raw documents. The dataset source has already cleaned and processed the text data, which is now available as ISON documents. As a result, the ISON documents are just transformed into records that are directly stored in the Document Store.
We employed the Elasticsearch tool to implement the Document Store. Further- more, we used the Haystack? tool to create the Big Querying and Communication layers. Haystack is an open-source Python framework that supports different search engines and includes many state-of-the-art NLP models. We used the well-established QA algorithms BM25 [31] and RoBERTa [19] as Document Retriever and Document Reader, respectively. Regarding BM25, it was the best Document Retriever algorithm
experimented in Sect.7. The code was written in Python using Jupyter Notebooks and is available on GitHub'º.
6.2 Real-World Queries We showcase three queries that can be executed on the instantiated architecture described in Sect. 6.1. We analyzed real-world applications by issuing various queries, focusing on different aspects. We formulate the queries based on the pharmaceutical company described in Example 1.
The Document Retriever was designed to return the top 20 documents with useful information for each query. As for the Document Reader, it was set to return the 3 most probable answers. Thus, there are three possible answers for each question. Data Reader provides a probability score for each answer. Higher scores indicate more confidence in ? Haystack - https://haystack.deepset.ai/. 1 BigQA codes - https://github.com/leomaurodesenv/big-ga-architecture.
BigQA: Big Data Question Answering Architecture 59 the prediction. Moreover, all queries returned answers because we did not check for a scenario with no answer.
Query 1. What law regulates drug marketing in the pharmaceutical industry? This query represents the interest of pharmacists, marketing, and legal employees in know- ing about regulatory laws on drug marketing. This query aims to find the name of a regulatory law, considering that only one document contains the correct answer. We executed Query 1 on the SQuAD dataset.
Table 3 shows the results of Query 1. The first two answers are about pharmaceutical industry documents, and the last one is about a legal penalty document in the United States. The answer with the highest probability score is the right and expected answer. We can conclude that the instantiated architecture could extract the answer to Query | with a score of about 76%. Table 3. Query 1: What law regulates drug marketing in the pharmaceutical industry?
Answer Document Score Prescription Drug Pharmaceutical 76.34% Marketing Act of 1987 | industry Food and Drug Pharmaceutical 1977% Administration (FDA) | industry Torture Capital punishment | 11.01% Regulation in the United States Note. Adapted from Moraes et al. [23].
Note. Adapted from Moraes et al. [23]. Query 2. When was the Luria-Delbriick? This query represents the interest of micro- biologists in extracting information about a bacterial experiment for antibiotics, which occurred in 1943. This query is date-oriented as it specifically searches for a particular year. It focuses on examining the architecture capability to recognize dates in docu- ments. We ran Query 2 on the SQuAD dataset.
Table4 depicts the results of Query 2. The first retrieved document is related to antibiotics, while the remaining documents refer to Arnold Schwarzenegger. The answer with the highest probability score is the right and expected one. Nevertheless, as the score is less than 50%, the Document Reader struggles to accurately extract the answer. Typically, when scores are below 50%, the algorithm fails to locate a solution.
In spite of this aspect, the instantiated architecture identified the answer to Query 2.
Query 3. What is the Novel Coronavirus? This query provides valuable information for pharmaceutical employees and the external public. We ran Query 3 on the SQUAD dataset, which we expanded with the COVID-QA dataset. This augmented query type explores the architecture ability to extract knowledge from new documents from differ- ent data sources and document formats.
Table 5 depicts the results of Query 3. Every returned document refers to the Coro- navirus and provides a score of over 70%. The first and third answers are the correct ones. After processing and inserting new documents into the Document Store, the archi- tecture can retrieve the augmented data to answer Query 3.
60 L. M. Pereira Moraes et al. Table 4. Query 2: When was the Luria-Delbrick? Answer | Document Score 1943 Antibiotics 29.89% 14 Arnold Schwarzenegger | 6.84% 14 Arnold Schwarzenegger | 3.06% Note. Adapted from Moraes et al. [23]. Table 5. Query 3: What is the novel Coronavirus? Answer Document | Score SARS-CoV-2 COVID-QA | 87.70% Prevention for 2019 | COVID-QA | 76.78% SARS-CoV-2 COVID-QA | 71.66% Note. Adapted from Moraes et al. [23].
Note. Adapted from Moraes et al. [23]. 6.3 Case Study Discussion
6.3 Case Study Discussion The case study showcased how BigQA can be successfully applied to real-world sce- narios, utilizing large datasets comprised of Wikipedia articles and FAQ questions and answers. As discussed in Sect. 6.1, we adapted the instantiated architecture to the application requirements by implementing only the appropriate layers and components. Finally, in Sect. 6.2, we presented distinct types of queries analyzing different aspects related to business applications.
The results showed that the system could respond to queries about the name of the law and augmented data. But, 1t struggled with the probabilities of the date-related question. In this context, the Document Reader should be fine-tuned with date question samples to optimize the performance of related questions. 7 Document Retriever Experiments
7 Document Retriever Experiments Because BigQA is agnostic, the Document Retriever can employ any QA algorithm. In this section, we conduct 60 experiments to evaluate three well-established QA algo- rithms and investigate their recall score. Our motivation is driven by the fact that employing a higher recall algorithm leads to better end-to-end querying and answer- ing performance [13]. Section7.1 details the experiment setup. Section 7.2 discusses the experiment results. 71 Experiment Setup
71 Experiment Setup We used the same instantiation described in Sect. 5.2 (Fig.2) in the experiments. The Input Layer received JSON files as dataset documents. The Document Store maintained these documents using Elasticsearch. We implemented the Document Retriever for eval- uation. The code was written in Python using Jupyter Notebooks and is available on GitHub (See footnote 10).
BigQA: Big Data Question Answering Architecture Table 6. Recall results of the Document Retriever algorithms investigated.
SQUAD AdversarialQA BM25 TF-IDF DPR BM25 TF-IDF DPR k=1 T1.15% 63.72% 48.59% 52.80% 45.92% 33.02% k=5 91.50% 86.50% 76.66% 69.51% 67.07% 56.81% k=10 94.43% 92.01% 85.72% 81.35% 81.81% 89.17% k=15 95.64% 94.46% 89.40% TI147% 78.30% 70.18% k = 20 96.29% 95.83% 91.38% 84.89% 85.56% 99.43% DuoRC QASports / Basketball BM25 TF-IDF DPR BM25 TF-IDF DPR k=1 TI1.37% 56.34% 21.46% 65.60% 51.87% 28.15% k = 88.83% 82.41% 36.41% 81.80% 74.52% 50.32% k=10 91.49% 87.47% 35.83% 87.04% 81.89% 59.65%
k=10 91.49% 87.47% 35.83% 87.04% 81.89% 59.65% k=15 90.59% 85.26% 48.42% 90.79% 86.71% 65.78% k = 20 93.76% 90.81% 44 . 78% 90.64% 87.51% 68.23%
We explored the following well-established QA algoritims: BM25 [31], TF-IDF [34], and Dense Passage Retriever (DPR) [13]. Furthermore, we used question- document pairs as input data from the following three open-domain real-world datasets with different data characteristics. — SQuAD v1.1 [30]: with 10,570 question-document pair samples. — AdversarialQA [4]: a QA dataset in which humans have created adverse and com-
plex questions, so the models cannot answer these questions easily. The dataset con- tains a total of 3k samples of question-document pairs. — DuoRC [33]: a dataset of movie plot questions and answers on articles from Wikipedia and IMDb, containing 12,845 question-document pair samples. — QASports [9]: a large dataset containing more than 1.5 million records encompass- ing various sports domains. We used 23,242 question-document pairs about basket- ball from this dataset. 61
The purpose of the experiments was to evaluate the performance of the QA algo- rithms to retrieve the correct document for a given question. To this end, we employed the recall measure. This measure calculates the number of times that a given algorithm correctly retrieves the desired document out of the total k documents retrieved. We var- ied the value of & in [1,5,10,15,20]. Literature typically utilizes a value of & that is
equal to or greater than 20. We used smaller values to provide fast answers without sacrificing performance, as business applications need to retrieve fewer documents and provide accurate answers quickly.
62 L. M. Pereira Moraes et al. 7.2 Experiment Results Table 6 depicts the recall results of the investigated algorithms. The results demonstrate that the recall score usually increases as the value of k also increases, indicating that retrieving more documents impacts the decisive probability.
In most cases, BM25 provided the best performance. BM25 is an extension of TF-IDF that incorporates a probabilistic information retrieval model, resulting in an enhanced recall score. BM25 is a sparser algorithm compared to the dense DPR algo- rithim. Dense algorithms can be quite costly regarding time and secondary memory usage. Based on these findings, we implemented BM25 as the Document Retriever for the case study described in Sect. 6.
DPR outperformed BM25 and TF-IDF in the AdversarialQA dataset, providing higher performance for k values of 10 and 20. DPR better understood the subject and context of the questions in these cases because dense algorithms are more effective over complex datasets. Despite these results, we recommend employing BM25 as the standard algorithm for the Document Retriever. 8 Conclusion
In this paper, we introduced a set of design principles for developing reliable and secure systems based on business, data, and technical aspects. Based on these principles, we proposed BigQA, the first Big Data Question Answering architecture. BigQA 1s a soft- ware reference architecture composed of six layers: (1) Input, for document ingestion; (11) Big Data Storage, for storing and processing textual data; (111) Big Querying, as
query engine; (1v) Communication, for the user interface; (v) Security, to provide secu- rity artifacts; and (iv) Insights, to assist with data analysis. The architecture 1s agnostic, Le., is independent of programming language, technology, and Question Answering algorithm.
We also outlined guidelines to support teams to develop and employ BigQA. The guidelines refer to good practices for implementing the Data Lake and the Document Store components of the Big Data Storage Layer. They also provide procedures for building the Big Data Storage, Big Querying, and Insights Layers. Furthermore, we showed three implementation pipelines to demonstrate BigQA versatility and applica- bility, focusing on using open-source and paid proprietary technologies and incorporat-
ing data mining and analysis tools.
Moreover, we validated BigQA by implementing a case study in the context of a pharmaceutical company. We used two real-world datasets: one with Wikipedia arti- cles and another with frequently asked questions about COVID-I19. We issued different queries, demonstrating the potential of BigQA in developing real-world applications. We implemented the BM25 algorithm as Document Retriever since 1t provided the best results according to our evaluation. In this evaluation, we conducted 60 experiments
over four datasets to compare the BM25, TF-IDF, and Dense Passage Retriever algo- rithms. All code is available on GitHub (See footnote 10).
We are currently conducting experiments to assess the performance of different algorithms to implement the Document Reader. Another future work involves studying technologies and algorithms to implement the Insights and Security layers. Finally, we
BigQA: Big Data Question Answering Architecture 63 plan to analyze new case studies to instantiate BigQA considering different real-world applications.
Acknowledgements. We thank Amaris Consulting, São Paulo Research Foundation (FAPESP), Brazilian Federal Research Agency CNPq, and Coordenação de Aperfeiçoamento de Pessoal de Nível Superior, Brazil (CAPES) [Finance Code 001] for supporting this work. P. C. Jardim has been supported by the grant t2023/08293-9, FAPESP. C. D. Aguiar has been supported by the grant t2018/22277-8, FAPESP. L. M. P. Moraes has been supported by Amaris Consulting. References 10. 11. 12. 13.
References 10. 11. 12. 13. . Armbrust, M., Ghodsi, A., Xin, R., Zaharia, M.: Lakehouse: a new generation of open plat- forms that unify data warehousing and advanced analytics. In: Proceedings of the Conference on Innovative Data Systems Research, vol. 8 (2021)
Ataei, P., Litehfield, A.: NeoMycelia: a software reference architecture for big data systems. In: Proceedings of the 28th Asia-Pacific Software Engineering Conference, pp. 452462 (2021). https://doi.org/10.1109/APSECS53868.2021.00052 . Athira, P., Sreeja, M., Reghura], P.: Architecture of an ontology-based domain-specific nat- ural language question answering system. Int. J. Web Semant. Technol. 4(4), article number 31 (2013). https://doi.org/10.48550/ARXIV.1311.3175
Bartolo, M., Roberts, A., Welbl, J., Riedel, S., Stenetorp, P.: Beat the AI: investigating adver- sarial human annotation for reading comprehension. Trans. Assoc. Comput. Linguist. 8, 662— 678 (2020). https://doi.org/10.1162/tacl.a 00338
Cassavia, N., Masciari, E.: Sigma: a scalable high performance big data architecture. In: Proceedings of the 29th Euromicro International Conference on Parallel, Distributed and Network-Based Processing, pp. 236—239 (2021). https://doi.org/10.1109/PDP52278.2021. 00044 Derras, M,., et al.: Reference architecture design: a practical approach. In: Proceedings of the 13th International Conference on Software Technologies, pp. 633-640 (2018). https://doi. org/10.5220/0006865006330640
Devlin, J., Chang, M.W., Lee, K., Toutanova, K.: BERT: pre-training of deep bidirectional transformers for language understanding. In: Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Lan- guage Technologies, pp. 4171—4186. Association for Computational Linguistics, Minneapo- lis (2019). https://doi.org/10.18653/v1/N19-1423
Galster, M., Avgeriou, P.: Empirically-grounded reference architectures: a proposal. In: Pro- ceedings of the Joint ACM SIGSOFT Conference - QoSA and Architecting Critical Systems, pp. 153—158 (2011). https://doi.org/10.1145/2000259.2000285 Jardim, P., Moraes, L.M.P., Aguiar, C.D.: QASports: a question answering dataset about sports. In: Proceedings of the Brazilian Symposium on Databases: Dataset Showcase Work- shop. SBC, Belo Horizonte (2023)
Ji, Z., et al.: Survey of hallucination in natural language generation. ACM Comput. Surv. (2022). https://doi.org/10.1145/3571730 John, T., Misra, P.: Data Lake for Enterprises. Packt Publishing Ltd. (2017) Johnson, J., Douze, M., Jégou, H.: Billion-scale similarity search with GPUs. IEEE Trans. Big Data 7(3), 535-547 (2019) Karpukhin, V,, et al.: Dense passage retrieval for open-domain question answering. arXiv preprint arXiv:2004.04906 (2020). https://doi.org/10.48550/ARXIV.2004.04906
64 14. 15. 16. 17. 18. 19. 20. 21. 22. 23. 24. 25. 206. 27. 28. 29. 30. 31. L. M. Pereira Moraes et al. Klein, J., Buglak, R., Blockow, D., Wuttke, T., Cooper, B.: A reference architecture for big data systems in the national security domain. In: Proceedings of the IEEE/ACM 2nd Inter- national Workshop on Big Data Software Engineering, pp. 51—57 (2016). https://doi.org/10. 1145/2896825.2896834
Kononenko, O., Baysal, O., Holmes, R., Godfrey, M.W.: Mining modern repositories with elasticsearch. In: Proceedings of the 11th Working Conference on Mining Software Reposi- tories, pp. 328—331 (2014).https://doi.org/10.1145/2597073.2597091 Laney, D., et al.: 3D data management: controlling data volume, velocity and variety. META Group Res. Note 6(70), 1 (2001)
Lepenioti, K., Bousdekis, A., Apostolou, D., Mentzas, G.: Prescriptive analytics: literature review and research challenges. Int. J. Inf. Manage. 50, 57—70 (2020) Li, Q., Xu, Z., Wei, H., Yu, C., Wang, S.: General big data architecture and methodology: an analysis focused framework. In: Debruyne, C., et al. (eds.) OTM 2019. LNCS, vol. 11878, pp. 33—43. Springer, Cham (2020). https://doi.org/10.1007/978-3-030-40907-4 .4
Liu, Y,, et al.: RoBERTa: a robustly optimized BERT pretraining approach. arXiv preprint arXiv:1907.11692 (2019). https://arxiv.org/abs/1907.11692 Lydia, E.L., Satyanarayan, S., Kumar, K.V., Ramya, D.: Indexing documents with reliable indexing techniques using Apache Lucene in Hadoop. Int. J. Intell. Enterp. 7(1—3), 203-214 (2020). https://doi.org/10.1504/IJIE.2020.104656
Misra, S., Kumar, V., Kumar, U., Fantazy, K., Akhter, M.: Agile software development prac- tices: evolution, principles, and criticisms. Int. J. Qual. Reliab. Manage. 29(9), 972-980 (2012) Móller, T., Reina, A., Jayakumar, R., Pietsch, M.: COVID-QA: a question answering dataset for Covid-19. In: Proceedings of the 1st Workshop on NLP for COVID-I19 at Association for Computational Linguistics, p. 1 (2020)
Moraes, L.M.P,., Jardim, P., Aguiar, C.D.: Design principles and a software reference archi- tecture for big data question answering systems. In: Proceedings of the 25th International Conference on Enterprise Information Systems, pp. 57-67. INSTICC, SciTePress (2023). https://doi.org/10.5220/0011842700003467
Múller, M., Vorraber, W., Slany, W.: Open principles in new business models for information systems. J. Open Innov.: Technol. Mark. Complexity 5(6), 1-13 (2019). https://doi.org/10. 3390/joitme5010006 Nielsen, R.D., et al.: An architecture for complex clinical question answering. In: Proceed- ings of the Ist ACM International Health Informatics Symposium, pp. 395-399 (2010). https://doi.org/10.1145/1882992.1883050
Novo-Loures, M., Pavon, R., Laza, R., Ruano-Ordas, D., Mendez, J.R.: Using natural lan- guage preprocessing architecture (NLPA) for big data text sources. Hindawi Sci. Program. 1—13, article id 2390941 (2020) Petroni, F., et al.: Language models as knowledge bases? In: Proceedings of the Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Con- ference on Natural Language Processing, pp. 2463—-2473 (2019). https://doi.org/10.18653/ v1/D19-1250
Poniszewska-Maraúda, A., Czechowska, E.: Kubernetes cluster for automating software pro- duction environment. Sens. J. 21(5), article number 1910 (2021). https://doi.org/10.3390/ s21051910 Radford, A., Narasimhan, K., Salimans, T., Sutskever, L., et al.: Improving language under- standing by generative pre-training (2018) Rajpurkar, P., Zhang, J., Lopyrev, K., Liang, P.: SQuAD: 100,000+ questions for machine comprehension of text. arXiv e-prints arXiv:1606.05250 (2016)
Robertson, S.E., Jones, K.S.: Relevance weighting of search terms. J. Am. Soc. Inf. Sci. 27(3), 129—146 (1976). https://doi.org/10.1002/as1.4630270302
32. 33. 34. 35. 36. 37. 38. 39. 40. 41. 42. 43. BigQA: Big Data Question Answering Architecture 65 Romualdo, A., Real, L., Caseli, H.: Measuring Brazilian Portuguese product titles similarity using embeddings. In: Proceedings of the 13th Brazilian Symposium on Information Tech- nology and Human Language, pp. 121—132. SBC (2021). https://doi.org/10.5753/stil.2021. 17791