Potential of Bots for Encyclopedia

Abstract

The wide range of applications and the capability to process and understand natural languages made chatbots very popular. Besides that, many applications chatbots are also used as information retrieval tools. Chatbots are changing the way users search for information. This document focuses on a chatbot that is used as an information retrieval tool. The chatbot enables information search in natural language in a geography domain. In case of a large number of search results, the chatbot engages users with clarification questions. It also provides support to users when uploading multimedia content to the website.

Full text


Abstract: The wide range of applications and the
capability to process and understand natural
languages made chatbots very popular. Besides
that, many applications chatbots are also used as
information retrieval tools. Chatbots are changing
the way users search for information. This
document focuses on a chatbot that is used as an
information retrieval tool. The chatbot enables
information search in natural language in a
geography domain. In case of a large number of
search results, the chatbot engages users with
clarification questions. It also provides support to
users when uploading multimedia content to the
website.
Index Terms: chatbots, information retrieval tool,
information search
1. INTRODUCTION
ustria Forum is an online encyclopedia, an
online collection, that provides Austria
related information. The content of Austria Forum
is divided in several categories and written in
English. The category of interest for this work is
the geography category. It offers information
about all countries of the world. Each country
page includes general information and links to
category pages. Each category page stores data
presented in form of text, tables or pictures. The
“Community Contributions” category provides
forms for uploading interesting pictures, video
and audio clips.
It is known that online encyclopedias provide a
large amount of information. The information
search on an online encyclopedia can be
illustrated in two scenarios. The first one is to
navigate through the website using links between
individual pages. The second, more common
way, is to use the search engine integrated in the
website. The search engines are mainly based
on keyword matching algorithms and provide a
list of results when supplied with an input. In
order to find desired information, the user needs
to browse the list. No matter which of the ways is
chosen, the information search is time
consuming. The relevance of the information
retrieved is questionable. As stated in [1], finding
Manuscript received May 4, 2004. (specify the date on which
you submitted your paper for review.) This work was supported in
part by the U.S. Department of Defense under Grant DR456 T. C.
Author is with the Electrical Engineering Department, University of
Tsukuba, Japan (e-mail: author@nxdfrim.go.jp).
relevant information has always been an issue
since the first search engines were built.
This publication is about design, architecture
and development of the chatbot prototype for
Austria Forum that can be used as an information
retrieval tool. The purpose of the chatbot is, on
the one hand, to enable information search in
natural language and, on the other, to guide
users when uploading content to the website.
Using natural language, users are able to
express their needs better and more accurately;
and using natural language processing and
understanding, the chatbot is able to understand
what users are searching for. Users participate
actively in information search providing answers
to clarification questions and thus contribute to
the relevance of search results. The chatbot
requires additional information in cases of
ambiguous questions or a large number of
search results. In terms of upload of content, the
chatbot engages users with a finite number of
questions in order to gather needed information.
The main focus of this work will be on
improving the relevance of the search results and
a faster access to information.
2. ABOUT CHATBOTS
Chatbots are software programs that enable
users to chat, communicate, and interact with
them in natural languages. They are also called
dialog-based systems, virtual assistants,
conversational agents or machine conversation
systems depending on the area of deployment.
[2]
Chatbots were primarily built to amuse and
entertain the users. ELIZA was the first
conversational system that was developed by
Joseph Weizenbaum in 1966. The system was
programmed with scripts and based on pattern
matching algorithms. The goal of ELIZA was to
imitate human conversation. Later in 1972 the
psychiatrist Kenneth Colby developed a chatbot
called PARRY, which used to simulate a
paranoid individual. In 1995 a chatbot named
ALICE was developed and used to entertain
users. ALICE is an open source software and can
be used to build customized bots. [2], [3], [4], [5]
Many chatbots developed in the last 50 years
were inspired by ELIZA. Though ELIZA was
limited and did not provide real understanding,
Potential of Bots for Encyclopedia
Saracevic, Mirhet; Ebner, Markus; and Ebner, Martin
A
Originally published in: Saracevic, M., Ebner, M & Ebner, M (2020)
Potential of Bots for Encyclopedia. In: Special issue - "Digital Heritage
and Related Tools“, Maurer, H. (ed.). IPSI BgD Transactions. 16(1). pp.
54-60. http://ipsitransactions.org/journals/papers/tir/2020jan/
fullPaper.pdf

users still wanted to communicate. It gave them
the feeling that they were talking to a real person.
This was the reason for the increased
development of chatbots. In the 90s the Leobner
Prize, a contest where chatbots performed the
Turing Test, was founded. The Turing Test is a
method proposed by Alan M. Turing in 1950 for
measuring intelligence of a computer system.
The Leobner Prize had a big impact on
developing chatbots and the Turing Test has
increased the interest in the area of artificial
intelligence (AI). [6], [7]
In 2015 the attraction to develop chatbots
intensified because the use of messaging
applications surpassed the use of social
networks. Big companies launched platforms for
bot development and integration. An increased
amount of data on the internet and improvements
in data processing and machine learning
enhanced artificial intelligence. It is possible to
develop more complex chatbots that can execute
multiple tasks. Additionally, the range of chatbot
applications has widened. Nowadays chatbots
are used for customer service, marketing,
finance, human resources, e-commerce or
entertainment. [8]
The type of a chatbot depends on various
parameters. The categorization can be performed
based on knowledge domain, service, goal or
methods for input processing and response
generation [6]. The conversation length is also a
possible parameter to perform the classification
of chatbots. Some chatbots tend to engage users
and have long conversations. These chatbots are
able to recall earlier conversations and determine
the context of the current conversation. Short
conversations are characteristic for chatbots
which provide some sort of information when
supplied with a question.
The information retrieval is also a field where
chatbots find their application. The chatbots to
question answering (QA) systems and the
chatbots to frequently asked questions (FAQs)
attract attention. QA systems tend to provide
answers when receiving a query in contrast to
search engines that deliver search results in form
of lists [3], [9]. As stated in [9] a chatbot can be
used as an interface to an open domain QA
system. Bayan Abu Shawar presented a chatbot
that is used as a natural web interface to QA
system [3]. ALICE bot was retrained to be able to
answer university related FAQs. The chatbot is
designed in such a way that it can be used to
provide answers to FAQs of any university [10].
Another example is the chatbot based on ALICE
that was trained on FAQs of the School of
Computing at the University of Leeds [4]. Natural
language systems were already built to provide
access to semi structured data of yellow pages
[11].
3. PROTOTYPE IMPLEMENTATION
The chatbot for Austria Forum is a standalone
application developed with Java 8 technologies
and runs on the Tomcat Apache server. The
knowledge base of the chatbot stores information
retrieved from the geography part of Austria
Forum. The chatbot architecture consists of a
client and a server side. The communication be-
tween the client and the server is enabled with
the help of Java API for Web Sockets. The server
side communicates also with a natural language
understanding (NLU) platform, called Dialogflow.
The following should clarify how the chatbot
basically works. When the chatbot is invoked in a
browser, a single client page is displayed. The
conversation starts when a user enters and
sends a question. The user input is forwarded to
the chatbot system running in background. The
chatbot then performs input processing and
searches for keywords and location information
within the input. Subsequently, the user input is
sent to Dialogflow for entity and intent
recognition. The response from the NLU platform,
which is sent in JSON string, is parsed. Having
the parsed information, the chatbot knows what
the user is asking for. It either creates a query
based on the parsed information and performs a
search, or engages users with questions and
gathers information needed for the upload. In the
case of a search, the search results are retrieved.
If the number of results is within a defined range
they are displayed to the user. Otherwise, the
chatbot activates a conversation context and
engages the user with clarification questions,
collects information, updates the query and
performs the search again. This procedure is
repeated until the results are displayed to the
user, or the chatbot cannot not find any
information in the knowledge base. In the case of
an upload, the chatbot activates a conversation
context and verifies if the entered information
corresponds to the requested format. If the
verification is successful the chatbot uploads the
information to the website. Otherwise, the chatbot
notifies the user about the failed verification and
is ready to process the next question.
3.1 Design
The design of the single client page was kept
as simple as possible. It was developed using
HTML, CSS and JavaScript. The client page, as
can be seen in Figure 1, includes an input text
field, a send button, and a conversation history
field. The important thing is that a user does not
need additional knowledge in order to
communicate with the chatbot. The design of the
client page is similar to commonly used
conversational interfaces in messenger
applications.

Fig. 1. Single Client Page
The knowledge domain represents an essential
part of the chatbot. As stated in [6] there are two
kinds of chatbots when considering the
knowledge base: open domain and closed
domain. With an open domain chatbot, the
conversation can go in any direction. Closed
domain chatbots are limited and can provide
answers regarding one specific topic [3]. Since
the chatbot for Austria Forum uses the
geography website as an information source, it
can be considered as a closed domain chatbot.
The information on the geography website is
accessed over API and retrieved in JSON format.
The JSON objects differ in content, structure
depth and nested fields. Therefore, an auxiliary
tool had to be developed to structure JSON
objects. Each JSON object represents a chunk of
information. This step was made in order to
facilitate analyzing, searching, and editing of the
knowledge base.
The natural language understanding (NLU)
platform called Dialogflow is used for input
understanding. Previously, the platform was
launched under the name api.ai. Dialogflow is an
artificial intelligence platform based on machine
learning. It is owned by Google and includes
built-in agents, which can be seen as modules
required for natural language understanding. The
agent “geo search” was created in order to
handle text input forwarded from the chatbot
system. The agent functions based on entity and
intent concepts. The entities represent a chunk of
information in a user input. In addition to built-in
entities (e.g. date, number, city, country), custom
entities were defined (category, continent). An
intent can be seen as a mapping between a user
input and possible responses. It helps to
understand what users are searching for. The
“geo search” agent has seven intents categorized
in two groups, the “search” and the “upload”
group. The intent groups help the chatbot
understand when users want to search and when
to upload information. The platform provides user
interface for agent training process. It is possible
to provide conversation examples and, in this
way, improve the agent understanding. Of
course, the agent is able to learn with the time
and become more intelligent.
3.2 Architecture
Each chatbot follows a defined flow that begins
with a user input and ends with displaying of
answers. The chatbot for Austria Forum follows
the general pipeline [6]. The first stage of the
pipeline is concerned with input processing,
followed with input understanding where named
entity and intent recognition is performed. The
last two stages deal with information retrieval for
the response or candidate response generation,
and selection or generation of the response. The
architecture of the chatbot system for Austria
Forum includes four components as illustrated in
Figure 2: natural language processing (NLP),
natural language understanding (NLU), search,
and logical component.
Fig. 2. Architecture of the chatbot
The Natural Language Processing Component
is based on the Stanford CoreNLP software
library. User inputs are passed through the
pipeline of different tools that perform tokenizing,
sentence splitting, part of speech (POS) tagging,
parsing, named entity recognition (NER), and
lemmatization [12]. The main task of the NLP
component is to extract keywords and location
tags. The results of the POS tagger are used for
keyword extraction. This tool provides over thirty
different tags. The main focus was on location
tag extraction since the knowledge base includes
geographical information. Location extraction was
performed with the help of named entity
recognizer. For English, NER recognizes person,
location, and organization entities. How these two

tools work is shown in an example question in
Figure 3 and Figure 4.
Fig. 3. Part of speech tagger
Fig. 4. Named entity recognizer
During the time of the research an evaluation
of the Stanford CoreNLP and the Apache
OpenNLP libraries was conducted. The number
of the provided tools is almost the same. Apache
OpenNLP uses different models for the POS
tagger and NER tools, whereby CoreNLP needs
only one model for its tools. The POS tagger of
CoreNLP delivers better accuracy and takes less
time than the POS tagger of OpenNLP. [13]
The Natural Language Understanding
Component is used to extract meaning from the
user input and makes it understandable for the
chatbot. The component communicates with
Dialogflow where entity and intent matching is
performed. The response from Dialogflow is
retrieved in JSON format containing detected
entities, matched intents, actions, and
parameters. Having this information in addition to
keywords and location information, it is possible
to create complex and efficient queries and
improve relevance of the search results. Several
NLU platforms were considered during the
research. Almost all platforms provide capabilities
for intent and entity recognition. Dialogflow and
wit.ai can be used free of charge while others are
for commercial use only. [6]
The Search Component provides search
functionalities and is based on Lucene Core.
Lucene Core is part of Apache Lucene, which is
an open-source full-text search library. Apache
Lucene can be used for creating search engines
[14]. The library is based on an indexing and
searching concept. The index of the search
component consists of documents. Each
document represents a mapping of a JSON
object. Each key-value pair of the JSON object is
stored in a text type field. The searching process
begins with the query generation. Each query is
created based on the information provided by
NLP and NLU components. The search
component implements methods for generation
and update of different query types (e.g. Boolean,
Term, Phrase, Range, Wildcard) and for running
them on a single or multiple text fields. For
example, a location query generated by a search
component is a term query and is executed on
the country document field. The search method
executes queries on the defined index and the
retrieve method retrieves them in form of
documents. The search results are listed based
on document score. To avoid large amounts of
results and to i-prove relevance, a threshold had
to be defined. Each search result is maximal
three hundred characters long and includes a link
to the page from where it was retrieved. An
evaluation of the Solr search server was also
considered during the research. Solr is built on
top of Lucene and provides REST-like API for
querying and retrieving of documents [15]. In
contrast to Solr which is used for enterprise and
content management system, the Lucene Core is
suitable for programming prototypes, because it
provides full control over internal processes.
The Logical Component communicates with
the client side and is responsible for interaction
and information flow between the components. It
also manages the conversation flow and context,
and forwards results to the client. The logical
component performs particular actions and sets
particular contexts depending on the intent
matched on the NLU platform. Five different
contexts can be activated. They are also grouped
in “search” and “upload” contexts. In case of a
large amount of search results, the logic
component activates one of the “search” contexts
(continent, country or category) and requests
additional input from the user. The context stays
active until the results are forwarded to the client.
Since the chatbot also supports the user while
uploading multimedia content, the logical
component provides methods for acquiring and
verification of the entered information. In this
case one of the “up-load” contexts is activated. If
the requested information is entered and its
verification is successful, the information is
uploaded to the website.
4. TESTING THE CHATBOT PROTOTYPE
In order to test the chatbot, several example
questions have been defined as can be seen in
Table 1. The aim is to show how the chatbot
behaves when it receives questions that include
different information.
Question
Text
Q1 Can you provide some information about
Nigeria?
Q2
Give me some information about energy in India
Q3
How many airports does Croatia have?
Q4
I need information about population
Q5
What are the most common natural hazards?
Q6
Can you provide information about energy?
Q7
I would like to upload a video
Table 1. Example questions

Since the knowledge base consists of
geography information, the main focus was on
location tags within a question. If the location is
found and the number of search results is within
a defined range, the chatbot displays the search
results as illustrated in Figure 5. The location
information is also present in question 1 and
question 2. Because of the large number of
results, the chatbot would pose clarification
questions. In this way it would gather additional
information and perform the search again. The
procedure is repeated until the number of results
is within the defined range.
Fig. 5. Search results for Q3
Questions 4, 5 and 6 do not provide any
information about the location. The chatbot would
perform a search and in case of a large number
of results, require a location information (country
name). Once the chatbot receives a location tag,
it would either proceed with follow up questions
or display the search results. Figure 6 shows the
conversation flow for question 5.
The NER tool of the used NLP library showed
shortcomings. Some of the continents were
recognized as a country or as a city. This caused
the chatbot to fail to answer questions containing
the misinterpreted information.
Fig. 6. Search results for Q5
As already mentioned, the chatbot is able to
support users when uploading pictures, video or
audio clips. This scenario occurs if a user poses
questions similar to question 7. The chatbot
would activate upload context and gather
information from user, verify it and in case of
successful verification upload the information to
the website.
Since the chatbot is a closed domain chatbot, it
does not provide answers to every user question.
The knowledge base determines the intelligence
of the chatbot. If a user searches for information
which is not related to geography and does not
exist in the knowledge base, the chatbot will
answer with one of the pre-defined answers. The
chatbot does not generate new answers but
retrieves information and makes it available to
users.
In addition to the search agent that was
created, the built-in Small Talk agent was
activated on Dialogflow. The Small Talk agent is
able to match Small Talk intents and extends the
number of user inputs that can be handled. This
capability contributes to the improvement of user
experience.
In contrast to search engines, the chatbot
accepts queries in form of full sentences in
natural language as well as keywords. In most
cases the choice of the question form affects the
search results. The search results contain a
chunk of text and a link to the page containing
the relevant information. The number of search
results to be displayed, as well as the content
length of each result, can be configured.
5. CONCLUSION
The goal of this work was to develop a chatbot
standalone application that can be used as an
information retrieval tool in a geography domain.
The chatbot should be used for information
search, as well for upload of information.
At the beginning of this paper, its context and
the problem of the relevance were introduced.
Definitions of chatbots and drivers behind
increased interest in development were
discussed. Types and areas of application were
mentioned with the focus on information retrieval
field.
The knowledge domain represents the brain of
chatbots. It was shown how to design and
structure a semi-structured data. In the future this
step should be considered to create a relational
database and use the chatbot as a natural
language interface. The setting up of an agent on
an NLU platform was described.
The architecture of the chatbot system, its
components and libraries used were discussed. It
was shown how a natural language
understanding component in an information
retrieval chatbot system can enable the
generation of qualitative and complex queries

and in this way improve the relevance of search
results.
At the end of this work, the chatbot prototype
was tested on several questions. The results
showed that the chatbot provides satisfactory
answers. The chatbot has potential as an
information retrieval tool and could be used as an
alternative to an integrated search engine in a
closed domain.
REFERENCES
[1] W. Bruce, Croft & Donald, Metzler & Trevor, Strohman.
(2010). Search Engines: Information Retrieval in
Practice. Pearson. ISBN: 978-0136072249
[2] Shawar, Bayan & Atwell, Eric. (2007). Chatbots: Are they
Really Useful?. LDV Forum. 22. 29-49.
[3] Shawar, Bayan. (2011). A Chatbot as a Natural Web
Interface to Arabic Web QA. International Journal of
Emerging Technologies in Learning. 6.
10.3991/ijet.v6i1.1502.
[4] Shawar, Bayan & Atwell, Eric & Roberts, Andrew.
(2005). FAQchat as an Information Retrieval System.
[5] Joseph Weizenbaum. 1966. ELIZA—a computer
program for the study of natural language
communication between man and machine. Commun.
ACM 9, 1 (January 1966), 36-45.
DOI=http://dx.doi.org/10.1145/365153.365168
[6] Nimavat, Ketakee & Champaneria, Tushar. (2017).
Chatbots: An overview. Types, Architecture, Tools and
Future Possibilities.
[7] A. M. Turing, I.—Computing Machinery and Intelligence,
Mind, Volume LIX, Issue 236, October 1950, Pages
433–460, https://doi.org/10.1093/mind/LIX.236.433
[8] DALE, R. (2016). The return of the chatbots. Natural
Language Engineering, 22(5), 811-817.
doi:10.1017/S1351324916000243
[9] Quarteroni, S., & Manandhar, S. (2007). A Chatbot-
based Interactive Question Answering System.
[10] Ranoliya, Bhavika & Raghuwanshi, Nidhi & Singh,
Sanjay. (2017). Chatbot for university related FAQs.
1525-1530. 10.1109/ICACCI.2017.8126057.
[11] Kruschwitz, Udo & De Roeck, Anne & Scott, Paul &
Steel, Sam & Turner, Raymond & Webb, Nick. (1999).
Natural Language Access to Yellow Pages. 34-37.
10.1109/KES.1999.820113.
[12] Manning, Christoper & Surdeanu, Mihai & Bauer, John &
Finkel, Jenny & Bethard, Steven & McClosky, David.
(2014). The Stanford CoreNLP Natural Language
Processing Toolkit. Proceedings of 52Nd Annual
Meeting of the Association for Computational Linguistics:
System Demonstrations. 10.3115/v1/P14-5010.
[13] Nanavati, Jay & Ghodasara, Yogesh. (2015). A
comparative study of Stanford NLP and Apache Open
NLP in the view of POS tagging. International Journal of
Soft Computing and Engineering (IJSCE).
[14] Balipa, Mamatha & Ramasamy, Balasubramani. (2015).
Search Engine using Apache Lucene. International
Journal of Computer Applications. 127. 27-30.
10.5120/ijca2015906476.
[15] Kumar, Vikash & Brawal, P.N.. (2016). Implementation of
highly optimized search engine using Solr. International
Journal of Innovative Research in Science, Engineering
and Technology.
Mirhet Saracevic works as a software developer at B&R
Industrial Automation in Graz.
Markus Ebner is currently working as a Junior Researcher at
the Department Educational Technology at the Graz
University of Technology.
Martin Ebner is head of the Department Educational
Technology at Graz University of Technology. He also works
as a Senior Researcher at the Institute of Interactive Systems
and Data Science.