chunk stringlengths 20 498 |
|---|
q
Check for
updates
BigQA: A Software Reference Architecture
for Big Data Question Answering Systems
Leonardo Mauro Pereira Moraes 2(B , Pedro Calciolari Jardim' O&,
and Cristina Dutra Aguiar' &
! University of Sao Paulo, Sao Carlos, SP, Brazil
? Amaris Consulting, Vernier, Geneva, Switzerland
fleonardo.mauro,pedro... |
Abstract. Querying massive and heterogeneous text data is challenging, tran-
scending different business domains. Our study outlines the BigQA architecture,
which is specifically designed to support text data queries on Big Data systems
using natural language. The architectural design comprises several layers that are
... |
tecture remains unclear. In this study, we showcase the versatility and adaptability
of BigQA by offering a comprehensive set of guidelines and three practical imple-
mentation pipelines. In addition, we performed 60 experiments on four different
datasets and compared the recall results of three popular algorithms: BM2... |
Keywords: Question answering * Big data + Software reference architecture -
Design principles + Pipelines
1 TIntroduction |
Nowadays, businesses across various industries are generating an unprecedented vol-
ume of textual data. This data surge has created challenges and opportunities for orga-
nizations seeking to harness the power of big data analytics. Textual data, including
websites, social media posts, customer reviews, financial repo... |
this vast amount of textual data has become essential for gaining competitive advantage,
understanding market share, making better business decisions, and driving innovation. |
The magnitude of textual data introduces storage, processing, and analysis chal-
lenges. For instance, the time spent finding specific information can not impact business
decisions and activities. Furthermore, recovered information needs to be updated, pos-
itively impacting decision-making. Another requirement is the ... |
can not adequately handle the enormous scale and complexity of textual data. |
! We use the terms query and question as synonymous throughout the paper.
(O The Author(s), under exclusive license to Springer Nature Switzerland AG 2024
J. Filipe et al. (Eds.): ICEIS 2023, LNBIP 518, pp. 42-65, 2024.
https://doi.org/10.1007/978-3-031-64748-2 3 |
BigQA: Big Data Question Answering Architecture 43 |
A suitable solution to overcome the aforementioned challenges is to use a Big Data
Question Answering system, which should offer the functionalities described as follows.
The system should store and process several documents in different formats, such as
web-based, PDF and ODF files. These documents should refer to mul... |
and accurate way to query documents and get answers to these queries. Furthermore,
because a Big Data Question Answering system should scale effectively, it must support
the volume, velocity, and variety characteristics inherent to the concept of big data [16]. |
Big Data Question Answering systems can understand questions written in natural
language sentences. They also manage context, question intention, among other charac-
teristics [3, 13,42]. Thus, these systems tend to be more efficient and accurate than tra-
ditional search engines like Google [3], introducing advantages... |
search engine for managing pharmaceutical or legal documents and providing informa-
tion for decision-making questions. |
A distinctive characteristic of Big Data Question Answering systems is using Ques-
tion Answering (QA) algorithms to query documents. These algorithms usually com-
prise two steps: Document Retriever and Document Reader. The Document Retriever
analyzes the questions issued in natural language by the users and retrieves... |
use Natural Language Processing (NLP) techniques. According to [13,27,32], current
NLP techniques recall factual knowledge without any fine-tuning, demonstrating their
potential as unsupervised open-domain algorithms. |
To the best of our knowledge (as outlined in Sect. 2), little attention has been devoted
to the problem of investigating how to build a Big Data Question Answering system. To
face this gap, we must define design principles and specify a software reference archi-
tecture for these systems. Design principles encompass gu... |
features. On the other hand, a reference architecture serves as a software blueprint,
outlining the structures, components, and relations to provide a concrete system archi-
tecture for a particular application or a group of software systems [6,8, 14]. |
In this paper, we introduce BigQA, a software reference architecture explicitly
designed for Big Data Question Answering. The architecture comprises six layers
to provide a powerful query engine. BigQA collects structured, semi-structured, and
unstructured data from various sources through the Input Layer. Data can exi... |
Storage Layer 1s applied at the bottom to prepare high-quality data for all kinds of ana-
lytical demands required by the upper layers. The Big Querying Layer is responsible for
processing the users questions sent from the Connection Layer. Finally, all the layers
are securely connected by the Security Layer and produc... |
Our contributions are six-fold, as described as follows.
l. A set of design principles based on Business, Data, and Technical requirements,
guiding for designing effective Big Data Question Answering systems. |
44 L. M. Pereira Moraes et al.
2. BigQA, a software reference architecture designed to adhere to the design principles
and address the challenges of Big Data and Question Answering.
3. A set of guidelines to streamline the developing and deploying of BigQA.
4. Examples of pipelines to implement BigQA using currently... |
6. Experiments to compare three well-established algorithms, namely BM25, TF-IDF,
and Dense Passage Retriever, to identify the most suitable algorithm for implement-
ing the Document Retriever component of BigQA.
& |
A preliminary version of this work was presented in [23]. Here, we update and pro-
vide more details for the discussions throughout the paper. We provide a more detailed
analysis of related studies. We also include a better explanation of the design principles
and a detailed description of the BigQA architecture contai... |
sive set of guidelines and pipelines to assist organizations in building and implementing
the BigQA architecture. Moreover, we improve the experiments comparing the Docu-
ment Retriever algorithms by using a new and larger dataset and investigating more
parameters. |
This paper is organized as follows. Section 2 reviews related work. Section 3 intro-
duces the design principles. Section 4 describes BigQA. Section 5 outlines the proposed
guidelines and exemplifies some pipelines. Section 6 validates BigQA in real-world sce-
narios. Section 7 describes experiments for investigating t... |
Big Data Question Answering
Characteristics Architectures Architectures BigQA
[43] [18] [2] [5] [40] [[39] 25] [42] [26] [13]
(c.1) Reference Architecture || VV V Á A| ÃO V Í f
(c.2) Big Data qA ANA A Í f
(c.3) Design Principles q N TN á
(c.4) Question Answering q ÃN á f
(c.5) Open Domain í V á
(c.6) Guidelines NAA á á... |
In this section, we review related architectures, including research proposals (Sect. 2.1)
and proprietary technologies (Sect. 2.2). |
BigQA: Big Data Question Answering Architecture 45
2.1 . Research Proposals
In Table 1, we compare related research proposals and our work considering the main
features and requirements of Big Data Question Answering architectures. We consider
the characteristics described as follows.
(c.1) Provides a software refer... |
(c.3) Defines design principles to improve the overall quality of the architecture.
(c.4) Implements a Question Answering solution as querying engine.
(c.5) Can query text data from different domains to improve decision-making.
(c.6) Specifies guidelines to aid users to implement the architecture.
(c.7) Introduces ... |
(c.8) Employs the architecture to real-world cases to showcase its applicability.
We discuss the surveyed Big Data and Question Answering architectures by catego-
rizing them into two groups. Group (1) consists of general-purpose software reference
architectures explicitly designed for data analysis and applications w... |
Regarding Group (1), Big Data architectures, the studies detailed in [2,5,18,40,43]
fulfil Big Data requirements and provide software reference architectures. A five-layer
big data processing and analytics architecture is proposed in [43] to collect, manage,
process, and analyse the vast volume of both static data and ... |
methodology composed of 16 steps to implement Big Data solutions, which is based
on CRISP-DM (cross-industry standard process for data mining). The software refer-
ence architecture proposed in [2] is based on events and microservices. It allows for
developing scalable and flexible big data software. The Sigma [5] big ... |
cases. Also inspired by the Lambda architecture, Yousfi et al. (2021) designed a big
data processing for multimodal data, such as images, video and audio. |
However, the constraints and layer configurations of the architectures in Group (1)
make 1t challenging to adapt them to the specific requirements of Question Answering
needs. For instance, offer a large data storage solution and a scalable query engine
for textual documents. Furthermore, the only two studies that pres... |
Considering Group (11), Question Answering architectures, the primary focus of the
studies outlined in [13,25,26,39,42] is to introduce advanced QA algorithms. Sucuna
et al. (2010) proposes a question answering architecture for repository of documents
using ontologies, taxonomies and knowledge bases. Nielsen et al. (20... |
combines a question answering system, search engines, and instant communication to
answer queries about Chinese marketplace products. Natural Language Preprocessing |
46 L. M. Pereira Moraes et al.
Architecture [26] is a framework based on BDP4J (Big Data Pipelining For Java) to pre-
process textual data, using data pipelines, for common NLP tasks like text classification
and semantic analysis. Karpukhin et al. (2020) proposes a question answering Docu-
ment Retriever called Dense ... |
The primary limitation of the studies in Group (11) is that the proposed architectures
lack generality and flexibility. They rely on specific algorithms, which determine the
architecture components and functionalities. In addition, the group does not consider
broad design principles as they only focus on technical aspe... |
The studies described in [26,39] are the closest to our work. The architecture intro-
duced in [39] uses an outdated algorithm and a fixed schema for query processing,
unlike our proposed BigQOA, which uses modern NLP algorithms. The architecture
described in [26] mainly focuses on text processing. It is not designed t... |
Our proposed BigQA architecture overcomes the shortecomings presented by the
studies in Groups (1) and (11) and fulfills all the characteristics depicted in Table 1.
BigQA 1s a cutting-edge software reference architecture designed to seamlessly address
the challenges posed by Big Data and open-domain Question Answering... |
ming language, QA algorithm, and technology. Furthermore, we establish design prin-
ciples rooted in Business, Data, and Technical requirements. Moreover, we introduce
guidelines to implement BigQA, exemplify some pipelines, and validate BigQA using a
real-world case study. |
2.2 Proprietary Technologies
IBM Watson Discovery?, Amazon Kendra?, and Sinch AskFranki are examples of pri-
vate technologies that are capable of complying with the requirements of Big Data
Question Answering systems. We don't cover these technologies in detail here because
they are proprietary. Furthermore, to the b... |
In addition, OpenAI has recently launched ChatGPTS, an innovative chat question-
answering technology. It leverages the information retrieval experience, enabling high-
level context understanding and answer generation. ChatGPT possesses an impressive
ability to generate natural and human-like responses to intricate qu... |
results, 1t faces common NLP issues found in large models, such as hallucination and
providing misleading answers [10], and other drawbacks detailed as follows. |
? IBM Watson Discovery - https://www.ibm.com/cloud/watson-discovery.
? Amazon Kendra - https://aws.amazon.com/en/kendra/.
* Sinch AskFrank - https://askfrank.ai/home.
* ChatGPT - https://openai.com/blog/chatept/. |
BigQA: Big Data Question Answering Architecture 47
— Static data is usually used to train large models. Thus, ChatGPT effectiveness is
limited by its inability to work with real-world and dynamic data.
— The last version of ChatGPT was designed to access live data from the internet,
enabling more dynamic experiences.... |
— Itis unclear whether the answers generated by ChatGPT are factual or hallucinated,
as 1t does not offer any references to 1ts source information.
— Incorporating business rules into ChatGPT reasoning is unclear. As a result, filtering
or enhancing its answers becomes challenging. |
In summary, ChatGPT does not provide answers relying on real-time data or current
events. Furthermore, the knowledge of ChatGPT currently derives from static Internet
data that goes up until 2021. Consequently, users may discover outdated information
or receive unhelpful answers. Moreover, it fails to integrate data fr... |
aforementioned challenges by allowing the insertion and updating of data. In addition,
there are currently no available public research papers that thoroughly describe and
detail the logic and implementation of ChatGPT system. |
3 Design Principles
Designing a suitable Big Data Question Answering system still remains unclear, despite
the vast literature dedicated to software architectures. From our perspective, the system
should be designed based on principles that take into account the business, data, and
technical aspects. We highlight the ... |
In this section, we present a collection of design principles for Big Data Ques-
tion Answering systems. These principles are derived from distinct sources, including
business models [24,36], the agile manifesto [21], and the characteristics of the Big
Data [16]. We define each principle as follows. |
Business (B). Business principles are foundational procedures that shape the user capa-
bilities within the system. These principles serve as a framework for system features.
The system should be efficient, easy to use, and secure. It also should protect user data
and scale to attend to user queries properly. |
B1: The system must accurately respond to user questions. The answer may be unclear
due to missing information or a potential misinterpretation of the QA algorithm
employed. The user must be notified in these situations.
B2: The system must grant access only to authorized documents. Thus, the system
must provide the i... |
B3: The system must support queries using natural language. Thus, it should compre-
hend the context, subject, and intention of the question.
B4: The system must concisely summarize the information found in the documents
related to the question. Typically, the system can offer two answer formats: a
concise FAQ respons... |
48
L. M. Pereira Moraes et al.
Data (D). Data principles refer to the rules and procedures that govern how the system
manipulates data. These principles are important for ensuring data accuracy, consis-
tency, and security. Furthermore, it 1s essential to ensure that the system effectively
supports the characteristic... |
D1:
D?2:
D3:
DA4:
D5:
The system must persist the raw documents. To this end, the system should store
raw documents in a Data Lake or a similar repository, even 1f only a small por-
tion of data is used. The advances in NLP algorithms and document processing
techniques motivate the need for storing raw documents a... |
The system must process documents from various data sources. For instance, doc-
uments can be obtained from data systems, databases, websites, and web-based
collaborative platforms. |
The system must support documents with different features: format, size, and
organization. The documents can be of different formats, including web pages,
PDF, ODF, Word, and JSON. Furthermore, the document sizes vary depending on
the number of pages each document contains. Moreover, a document can contain
structured, ... |
The system must be able to extend its functionalities according to the volume
characteristc of Big Data since the amount of produced documents and texts writ-
ten can easily reach large volumes of data.
The system must ensure that the data 1s readily available as soon as new or updated
documents are added. Hence, the ... |
Technical (T). Technical principles are essential practices for effectively implementing
a system. These principles are crucial for ensuring the system reliability, efficiency, and
scalability.
T1: Modularity. The system must encompass specialized components, each providing
specific functionalities and working as a s... |
T2: Flexibility. The system must easily handle new components with their particu-
lar complexity and functionality. In addition, the system must support employing
different programming languages and technologies.
T3: Analytic. The system must store data, metadata, and usage information for ana-
lytical analysis. This ... |
T4: Security. The system must provide security artifacts to protect its integrity. Exam-
ples include user authentication systems, data governance policies, and encrypted
connections between components. |
Design principles provide a framework for creating effective and user-friendly sys-
tems. They can help to ensure that the system is easy to use, efficient, and secure.
Following the design principles can help improve the user experience, provide the right
system functionalities, and increase team productivity. |
BigQA: Big Data Question Answering Architecture 49
4 BisQA Architecture
In this section, we introduce BigQA, a pioneering reference architecture for Big Data
Question Answering. It compromises six layers, as depicted in Fig. 1.
Manager
Input Layer Security Layer Insights Layer
( =m
e ZA
Data Data ... |
Big Data Storage Layer
— ==
= '
Document Metadata Document Document
Store Repository Retriever Reader
uNEs=
&) o
lh
A
Authentication Reporting
System Tools
Horizontal Layers
Fig. 1. BigQA, the pioneering Big Data Question Answering architecture. |
There are two types of layers: horizontal and vertical. Horizontal layers are layers
that have a direct connection between them. For instance, there is a connection between
the Input and the Big Data Storage layers. Vertical layers are those that can connect
to any other layer within the architecture. For instance, the... |
We outline the functionality of each layer as follows. We first describe the horizontal
layers, 1.e., Input (Sect. 4.1); Big Data Storage (Sect. 4.2); Big Querying (Sect. 4.3); and
Communication (Sect. 4.4). Then, we detail the vertical layers, 1.e., Security (Sect. 4.5);
and Insights (Sect. 4.6). In Sect. 4.7, we disc... |
Before detailing the layers, we introduce in Example | a business application that
requires the implementation of the BigQA architecture. We employ this pharmaceutical
case as a running example throughout this section. |
Example I. Consider a leading pharmaceutical company that provides several health-
care products and handles a huge number of essential documents, such as pharmaceu-
tical leaflets, product reports, financial records, and contractual agreements. The com-
pany requires a centralized knowledge base for employees and cust... |
this requirement, the company implemented a system using BigQA as reference archi-
tecture. |
50 L. M. Pereira Moraes et al.
4.1 Input Layer
The Input Layer ingests documents into the system. The layer gathers documents from
different data sources, like company files, project reports, and trusted websites. There-
fore, these data sources can contain a variety of document formats, including Word,
JSON, PDF, an... |
Example 2. Data providers use the input layer to add documents to the system. In
Example 1, data providers are the employees who create content on web-based col-
laborative platforms and external sources like market research documents.
4.2 Big Data Storage Layer |
The Big Data Storage Layer manages data storage. The layer stores raw documents in
the Data Lake, that were received from the Input Layer. The system performs differ-
ent transformations for each document. First, it identifies the document type. Second,
it employs processing techniques to clean and convert the document... |
Finally, the system integrates metadata about the new text document into the Metadata
Repository. |
There are three repositories for data storage. The Data Lake stores raw data. Every
managed document is stored in this repository. The Document Store contains converted
data generated by the extraction, transformation, and cleanness processes. Converted
data are pieces of data from the raw documents, such as sentences ... |
authors, and the creation date. Example 3 illustrates the use of the Big Data Storage
Layer. |
Example 3. Employees from the pharmaceutical industry are documenting a new med-
ication using a Word file. This file is stored in the Data Lake as a raw document. Then,
the Big Data Storage Layer identifies that the document format is Word and uses Word
specific processing techniques to clean and convert the data. Onc... |
The Big Data Storage Layer is also responsible for data processing. This functional-
ity requires using a big data infra infrastructure and a distributed and parallel processing
framework (e.g., Apache Spark [41]). The objective is to manage the massive volume
of documents, which can reach huge sizes.
4.3 Big Queryin... |
4.3 Big Querying Layer
The Big Querying Layer is the core of the architecture. It is the query engine responsi-
ble for processing, interpreting, and producing answers to the user questions. This layer
obtains documents from the Big Data Storage, receives queries from the Communica-
tion Layer, and sends back answers... |
BigQA: Big Data Question Answering Architecture 51 |
Example 4. A pharmacist needs information about a new medicine. The pharmacist
accesses a web page through the Communication Layer and sends the following ques-
tion: “What are the ingredients of the new medicine?”. The Big Querying Layer
retrieves sentences about the new medicine from the Document Store and processes
... |
Communication Layer. |
The Big Querying Layer consists of two components. The Document Retriever
finds most valuable text data that might have the answer to the question. The text
data is retrieved from the Document Store using a big data search engine like Apache
Lucene [20]. The second component, named Document Reader, analyzes the data an... |
4.4 / Communication Layer
The Communication Layer is the interface where users ask questions and get answers,
as depicted in Example 5. This layer can include multiple components. In Fig. 1, we
illustrate two of them. Front-end and API are two components that allow users and
applications to send questions to the syste... |
Example 5. The Communication Layer receives the question “What are the ingredients
of the new medicine?” from the pharmacist web page. It also presents the answer “it
contains metamizole” to the pharmacist on the web page.
4.5 Security Layer |
4.5 Security Layer
The Security Layer addresses security issues like network connection, credentials, and
data governance. As a vertical layer, it can apply security artifacts to all BigQA lay-
ers. Example 6 illustrates how to use the Security Layer to provide data security for
information access. |
Example 6. Consider the pharmacist that issued the question “What are the ingredients
of the new medicine?”. The Data Governance component should guarantee that this
pharmacist is authorized to access the sentences that were extracted from the raw Word
document about the new medicine. |
The Security Layer includes but is not limited to the following components. Authen-
tication System aims to ensure the secure authentication of user operations. Credentials
& Permissions provides suitable credentials and permissions for the network connection
among the components. Finally, Data Governance refers to man... |
52 L. M. Pereira Moraes et al.
4.6 Insights Layer
The Insights Layer encompasses the process of data analysis. This layer handles and
analyzes data from all other layers. Example 7 demonstrates the Insights Layer usage
by system managers. |
Example 7. AÀ manager uses a data mining tool to discover patterns in the text of the
documents and categorize them. For instance, the tool identifies the word “metami-
zole” and “analgesic” in our current example and categorizes the document as a “drug”
document. Moreover, the tool finds the words “hospitalization” in... |
to generate tables and charts to achieve insights about the number of documents within
each category. |
Components in this layer include: (1) Reporting tools used to generate reports sum-
marizing data, such as tables, graphs, and visuals; (11) Data Analysis tools used to
explore and analyze data, often to find relationships between variables and test hypothe-
ses; and (111) Data Mining tools used to discover hidden patt... |
4.7 Architecture Discussion
BigQA focuses on providing features that align with design principles for a Big Data
Question Answering system, rather than being limited to specific technologies or use
cases. Hence, every development team can select the technology, software programming
language, and QA algorithm that bes... |
Furthermore, BigQA is built upon the design principles outlined in Sect. 3. Table2
highlights the layers that cover the design principles.
Table 2. The Design Principles and the BigQA layers.
Design Principles | BigOA Layers
B1-B4 Communication, Big Querying, Security
D1-D5 Input, Big Data Storage
T1... |
T4 Security
5 Guidelines and Pipelines
In this section, we introduce guidelines and exemplify pipelines to guide organizations
and development teams in implementing their Big Data Question Answering systems |
BigQA: Big Data Question Answering Architecture 53
following the BigQA architecture. By providing clear and concise instructions on per-
forming tasks, guidelines can help reduce errors and improve productivity. On the other
hand, pipelines can showcase different BigQA instantiations using current and up-to-
date tool... |
Section 5.1 details the proposed guidelines. Section 5.2 showcases different exam-
ples of implementing the pipelines. Together, guidelines and implementation pipelines
can help organizations achieve their goals more effectively.
5.1. Guidelines for Architecture Implementation |
Implementing a Big Data Question Answering system involves the selection of appro-
priate components to satisfy the requirements introduced in Sects.3 and 4. Here, we
present a comprehensive set of guidelines to support development teams. The guide-
lines focus on the BigQA core components and include various technolog... |
Lake and the Document Store. Then we introduce guidelines for the Big Data Storage,
Big Querying, and Insights layers. |
End of preview. Expand in Data Studio
README.md exists but content is empty.
- Downloads last month
- 2