IRJET- Determining Document Relevance using Keyword Extraction

INTERNATIONAL RESEARCH JOURNAL OF ENGINEERING AND TECHNOLOGY (IRJET) E-ISSN: 2395-0056
VOLUME: 06 ISSUE: 11 | NOV 2019 WWW.IRJET.NET P-ISSN: 2395-0072
© 2019, IRJET | Impact Factor value: 7.34 | ISO 9001:2008 Certified Journal | Page 2374
Determining Document Relevance using Keyword Extraction
Vinay Patil1, Sachin Jadhav2, Pawan Lokapur3, Akash Mhatre4, Prof. Tushar Ghorpade5
1,2,3,4Student, Dept of computer Engineering, Ramrao Adik Institute of Technology
5Assistant Professor, Dept. Of Computer Engineering, Ramrao Adik Institute of Technology
----------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - This paper lies in the data analysis domain
describing about the system which attempts to search for
a relevant document from a large set of documents, or
more specifically to fetch a summary of answer for a given
query from one of the selected relevant document. In few
of the organizations involving large data set in form of
documents, where clients and customers need to retrieve
some of these documents frequently, the process of
searching this document becomes a very hectic task. For
instance, the educational institute like Mumbai University
which has large corpus of documents. So to overcome this
manual searching approach we have proposed a system
which successfully fetches the desired documents to user
based on query provided to system. This is being done by
priory extracting the documents at time of uploading and
storing the necessary stats required for search algorithm.
For document extraction we use the TF-IDF algorithm.
And during search we analyse the TF-IDF weight of
keywords in search algorithm to fetch the desired set of
documents to user. There will also be a feedback
mechanism for user to interact with system through which
user can upvote or downvote for a particular document
thus making the system to learn and improve its search in
future. The system is supposed to deliver accurate results
for every query given by user combined with less
processing time. The system contains three operational
elements i.e. keyword extraction, search module and topic
selection.
General Terms: Data Mining, Data Analysis, Keyword
Extraction.
Keywords: TF-IDF, QnA, Document Search, Artificial
Feedback, Keyword Extraction.
1. INTRODUCTION
In recent world of digitalization, data is coming in huge
amounts from many sources like news, social media,
banking and education. And hence because of this
unregulated and unordered growth in data, there is a
need of automated information retriever which will help
users to retrieve relevant information by searching in
piles of unstructured data. But there exists challenges to
implement one, such as retrieving correct sense of
information. Information retrievers is emerged as an
important research area in recent past. In this regard,
study of existing work is useful to carry on further
research. Using keywords to predict document contents
is accurate and fast method. Keywords can be used as
entry points into an index which will then help to
identify files, records, texts or any unstructured data.
However, a large number of data is not marked using
keyword and giving it to a human for tagging will be
difficult and also needs large amount of time. Thus, there
is need of an algorithm which will tag every document or
data by its relevant keywords. Hence in this paper, a
keyword extraction algorithm is presented also a faster
method of searching through those extracted keyword is
proposed. An algorithm named “TF-IDF (Term
Frequency- Inverse Document Frequency)” is one of the
simple method which serves the purpose of retrieving
keywords with high accuracy. The working of TF-IDF
algorithm is based on the number of times a word
appeared in a sentence and the number of times it
appeared in whole corpus. More the occurrences in a
document, more is the weightage and contradictorily
more the occurrences in corpus, less is the weightage.
The keywords extracted by this algorithm can be used in
searching algorithm which we proposed. This searching
algorithm make use of combination concept and sets to
give list of documents which contains the search query.
It also discusses about different databases structures
used for text extraction and a quick question answering
bot which can be used in different domains. Finally, it
discusses briefly about issues and research challenges
faced by us along with future direction.
2. LITERATURE REVIEW
In LDA based Paper [1], system uses generative
probabilistic model on Chinese corpus data and
categories documents by different topic names, but the
major drawback is that it does not remove stop-words
and also the topic names or subject needs to be known in
advance. hence limiting the corpus contents.
In TF-IDF based paper [2], a system is developed which
will search social media platforms to get the latest trends
which then can be used for advertisement purpose.
Paper presented an algorithm which will monitor
Instagram accounts for 50 recently posted photos and its
captions. These captions are then analysed to get latest
trends and fashion accessories which will be used for
marketing those product effectively. But limitations of
this system is it is concentrated only on 20 most
followed users on Instagram and hence the scope of this
system is narrow down to 20 users only.
Modified IDF paper [3] focuses more on ambiguous word
sensing by using KNN approach which has less worth

from document relevance checking point of view, since
word ambiguity won’t affect much in large corpus of
documents.
Delta TF-IDF paper [4] deals with positive and negative
set of classes for efficiently weighing the word scores
before pushing them into classification. So, it requires
positive and negative training set to work prior to
classification.
TF-IDF to Determine Word Relevance in Document
Queries paper [8] where the TF-IDF algorithm was not
able to equate one of the word (e.g. drum) with its plural
form (drums) slightly decreasing the TF-IDF weight of
that particular keyword. For larger corpus of documents
this could present an escalated problem
3. EXPLANATION
The system contains three modules i.e. topic extraction,
document search and question answering bot. All three
modules will be mutually functioning with the help of
global and local database schemas. So before
understanding the module functioning let us dive into
the structure of databases and relationship between the
tables.
3.1 Database
Database module has two modes Private and
Organization. These modes will then be used to serve
different purposes.
Private Mode: This database will provide private
workspace for every user. Hence users can upload their
private files in this workspace which then can only be
visible and accessible to its owner only. but for
operations like search and QnA will work the same, just
depending on the mode the respective database will be
altered. Hence can be act as private Document Manager.
Organization Mode: This mode is designed for
organizations to manage their files which should only be
available to their employees. In this mode, files will be act
as public and only authorized person can upload the data
from this workspace.
Mode selection: These options will be available to users
which will allow them to select mode while uploading or
searching files.
Database will include small corpus of documents which
contains only stop-words so as to reduce the impact of
higher TF-IDF weights of those words during initial
document extraction.
Fig 3.1: Database structure
3.2 Topic Extraction
This is the very first module with which user will interact
by uploading all the documents which he/she desires.
Once the documents are uploaded, document details are
stored in a doc_details table. All words of a document are
extracted then stop words are removed for reducing
memory overhead and to improve the power of
searching. Also, instead of the keywords plural form its
root word is inserted into the database so that TF-IDF
weight for each word is computed correctly without
discriminating between singular and plural form. Term
frequencies are calculated [6] for each word and send for
the database operation. There are two threads for
simultaneously processing the database operations. The
first thread will do two tasks. It will increment a count of
a total number of documents in a doc_count table after
that IDF (Inverse Document Frequency) is calculated.
TF: Term Frequency, which measures how frequently a
word occurs in a document. TF(t) = (Number of times
word w appears in a document) / (Total number of
words in the document).
IDF: Inverse Document Frequency, which measures how
important a word is.
IDF(t) = log_e(Total number of documents / Number of
documents with word w in it).

3.2.1 Equations
Variables:
Table 1. Symbols with Description
Description Symbol
Number of times word w appears in a
document
Wi
Total number of words in the document Tw
Total number of documents N
Number of documents with word w in it Di
Equation No. 1: Term Frequency
Equation No. 2: Inverse Document Frequency
idf(i)=loge ( )
Equation No. 3: Term Frequency-Inverse Document
Frequency
tf
TF-IDF weight of each word is calculated and stored in a
tfidf table before which fuzzy inference system is
implemented so as to threshold number of keywords
which will be inserted in database for searching purpose
and hence all the stop words [7] are eliminated.. The
second thread for updating search_list and search_doc
tables to improve searching.
Fig 3.2: Keyword Extraction
3.3 Document Search
This is the core module of the system which runs the
search algorithm and fetches the appropriate set of
documents to the user depending on the query given to
it. The algorithm primarily relies on the extracted data
stored in the database for its processing. It will first sort
the keywords in query and determine which keywords
are required for searching. Accordingly, a list of such
keywords is prepared and is passed as parameter to
algorithm. The algorithm works on principle of “Highest
priority to Largest intersection” for which the
combinational logic has also been implemented to
systematically form the sets of doclist of keywords. Also,
the ordering in combination is taken care of on the basis
of rank which is taken as a feedback from user. This
artificial feedback will help to rank the documents and
make the right choice in future.
Fig 3.3: Searching module
3.4 Question and Answering bot
A natural language processing (NLP)[5] gives capability
to computer which allows communication to happen
between user and computer using human natural
languages. There are three stages to understand natural
language i.e. parsing, semantic interpretation, and
summarization. In parsing phase, two actions are
performed on input provided from user. First, the main
linguistic words are identified. Second, using TF-IDF
values of words in input query for selected document the
sum is calculated. Then in next phase i.e. summarization,
the sentences are collected from selected document for
paragraph summarization.

Main words are retrieved from sentence by using
sentence semantics. These keywords are then searched
in database which in turn gives list of documents sorted
by its rating taken from users which might content the
answer to the question. Later, the top document is
tokenized and values for tf-idf are calculated for every
sentence. And those sentences having greater values
than tf-idf value of the query are selected for
summarization. Resultant answers are processed to get a
most popular answer or most accepted answer using
ranking algorithm which will then be shown as output to
users. Ranking Algorithm works on feedback taken from
users. This feedback asks users whether the answer was
helpful or not? If user thinks it was helpful will mark yes
and the document rank is upgraded otherwise it will be
degraded.
Fig 3.4: Question and Answering bot Module
4. Conclusions
In order to get a relevant document or specific answer
from a corpus of a large collection of organizational
documents, this paper proposes a document extraction
system which has following modules first is keyword
extraction in this user will upload all the documents
which he/she desires. Once the documents are uploaded,
the system will internally start the extraction of
keywords by using the TF-IDF algorithm. From each
document, significant keywords are selected these
keywords will occupy their positions in database which
will be used during the document searching. Second is
document search is responsible for fetching the relevant
document to the user by making the use of weighting of a
combination of keywords and uses indexed database
which contains the list of keywords along with their
respective list of documents.
Further on the basis of user satisfaction, a feedback will
be taken up by the user to check the correctness of the
generated output. If the user is not satisfied then the
next set of documents will be fetched. This system also
has Question and Answering bot. which will provide user
to have active conversation with system using simple
human language(English). QnA also includes an artificial
feedback taken from user which will help improve the
answer quality and relevance regarding to question
asked.
This method helps people to locate relevant documents
from the corpus of a large collection of organizational
documents. On the other hand, an artificial feedback
mechanism can rank the documents and make the right
choice in future.
5. Future Work
The System currently doesn’t support odt and
image file format. So, support for these files can be added
in future. Semantic analysis can be used for getting
context wise meaning of a word to improve the power of
searching. Comments can be used with artificial feedback
to improve searching. The correctness of result is
checked by artificial feedback which currently just
suggests the alternate ranked documents. But in future,
this feedback technique can be replaced with views
hence can be used for ranking the document. So, it will
be more accurate as most of the times users don’t vote
up although it helped them. Users can set the expiry date
of documents which they needed for short period of
time. After expiry date that document is deleted hence it
will improve the searching and reduce the overhead of
database.
6. References
[1] Qihua Liu, ”A novel Chinese text topic extraction
method based on LDA”, International Conference
on Computer Science and Network Technology,
Nanchang, 2015.
[2] Bernardus Ari Kuncoro, Bambang Heru Iswanto,
“TF-IDF method in ranking keywords of Instagram
users' image captions”, TF-IDF method in ranking
keywords of Instagram users' image captions
Information Technology System and Innovation,
Bali, November 2015.
[3] Zun May Myint, May Zin Oo, “Analysis of modified
inverse document frequency Variants for word
sense disambiguation”, International Journal of
Advanced Computational Engineering and
Networking, Myanmar, Aug.-2016.
[4] Justin Martineau, Tim Finin,“Delta TFIDF: An
Improved Feature Space for Sentiment Analysis”,

Third AAAI International Conference on Weblogs
and Social Media, San Jose CA, May 2009.
[5] Bayu Setiaji, Ferry Wahyu Wibowo, “Chatbot Using
A Knowledge in Database Human-to-Machine
Conversation Modeling”, 2016 7th International
Conference on Intelligent Systems, Modelling and
Simulation, Yogyakarta, Indonesia, April 2016.
[6] Aizhang Guo,Tao Yang, “Research and
Improvement of feature words weight based on
TFIDF Algorithm ”, International Conference on
Computer Science and Network
Technology,Jinan250353, China,May 2016.
[7] Rajaraman, A., Ullman J., "Data Mining: Mining of
Massive Datasets”, India, 2011.
[8] Juan Ramos, “Using TF-IDF to Determine Word
Relevance in Document Queries”, Department of
Computer Science, Rutgers University, Piscataway,
NJ.

IRJET- Determining Document Relevance using Keyword Extraction

More Related Content

What's hot (17)

Similar to IRJET- Determining Document Relevance using Keyword Extraction (20)

More from IRJET Journal (20)

Recently uploaded (20)

IRJET- Determining Document Relevance using Keyword Extraction