Available via license: CC BY-NC-ND 4.0
Content may be subject to copyright.
Digital Object Identifier
Consumer Document analytical
Accelerator Hardware
ASWANI RADHAKRISHNAN1, DIBYASHA MAHAPATRA1, and ALEX JAMES,1
1Digital University Kerala, Kerala, India (e-mail: apj@ieee.org)
Corresponding author: Alex James (e-mail: apj@ieee.org).
The grant support with technical and database provided by Clootrack Pvt Ltd is acknowledged.
ABSTRACT Document scanning devices are used for visual character recognition, followed by text
analytics in the software. Often such character extraction is insecure, and any third party can manipulate
the information. On the other hand, near-edge processing devices are restrained by limited resources and
connectivity issues. The primary factors that lead to exploring independent hardware devices with natural
language processing (NLP) capabilities are latency during cloud processing and computing costs. This paper
introduces a hardware accelerator for information retrieval using memristive TF-IDF implementation. In
this system, each sentence is represented using a memristive crossbar layer, with each column containing a
single word. The number of matching scores for the TF and IDF values was implemented using operational
amplifier-based comparator accumulator circuits. The circuit is designed with a 180nm CMOS process,
Knowm Multi-Stable Switch memristor model, and WOx device parameters. We compared its performance
with that of a standard benchmark dataset. Variability and device-to-device related issues were also taken
into consideration in the analysis. This paper concludes with implementing TF-IDF score calculation for
applications such as information retrieval and text summarization.
INDEX TERMS Natural language processing, TF-IDF, hardware accelerator, memristive systems, mem-
ristor, analog computation.
I. INTRODUCTION
REAL-TIME document scanning with natural language
processing systems that use translations, emotion
recognition, and sentiment analysis involves raw text data,
which can reveal the individual’s context, identity, or be-
havior. However, such analysis can threaten the trust and
confidentiality of the information collected by individuals.
With the asset of an on-chip document scanner, encrypting
the data post-processing before storing it can help resolve the
confidentiality and privacy of the information.
Many traditional classifications and deep learning models
deployed in cloud servers are currently used in most Natu-
ral language processing implementations. Regardless, many
document scanners and scanning applications are available in
the market. Document scanners that can perform text analysis
on edge devices still need to be developed. However, cloud-
dependent model implementation and devices impede the
process and lag after a specific capacity and data load. The
processor power, memory, and communication bandwidth
supported by the edge device determine the processing speed
and responsiveness of the end user. However, these factors
might pose a severe obstacle to delivering dependable real-
time performances. We look at standard searching tech-
niques, such as TF-IDF based information retrieval systems,
for implementing low-power edge Natural Language Pro-
cessing (NLP) edge solutions. The traditional method of cal-
culating the TF-IDF is modified into an application specific to
incorporate the requirements. The literature shown in Table
1 provides insights into the above. Apart from the TF-IDF
technique, there are numerous to extract features from text
data in language processing tasks; a few are listed in Table
2. Studies like Keyword Extraction have been done with
unsupervised learning methods using TextRank, RAKE, and
PositionRanking on research documents Table 2 [1], which
also contributes to the many NLP applications. The extrac-
tive summarization technique used for text summarization
work better along with the TF-IDF score-based keywords
retrieval Table 2 [2]. Previous research shows that TF-IDF
based extraction combined with classification and clustering
techniques boosts the model’s accuracy, as shown in Table
2 [3] [4] [2]. In [3], after comparing F1 score values, the
TF-IDF outperforms the Okapi BM25. After considering
the TF-IDF metric’s strengths and weaknesses, we used
it as an information retrieval method for hardware-driven
VOLUME , 2022 1
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3237463
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
FIGURE 1. Various document analytic accelerator applications ranging from
healthcare to power and energy sectors.
applications. The proposed applications can scan the doc-
ument and perform text summarization, sentiment analysis,
and information retrieval in real time, significantly aiding
in diminishing computation power over the head. Various
applications of document analysis accelerators are shown in
Figure 1.
There are several attacks which influence the NLP process
to retrieve the information. Such major attacks to the NLP
applications are Adversarial attacks, Eavesdropping attacks,
File injection attacks, Pattern reconstruction attacks, Key-
word inference attacks etc. All mentioned attacks are focus-
ing on malicious modifications of information using keyword
extraction or replacement of words using probability or any
semantic approaches. In hardware implementation the char-
acters are mapped to conductance values which is unable
to attack due to programming complexity. The character
mapping and writing into the memristive crossbars difficult
for reverse engineering. In our model we have developed ana-
log hardware approach to perform the information retrieval
measures. There are hardware based security systems such
as PUF to avoid external attacks on real time systems.
Current existing document scanners rely on software tools
to perform NLP processes for computation. This will in-
crease the transportation delay as well as generate security-
related issues. Whereas hardware-based NLP systems are
more reliable in terms of power and speed. IBML Image-
Trac Series 6000 is a hardware friendly document scanner
with data analysis facility. SpAtten’s architecture from MIT
research labs shows an SRAM and DRAM-based hardware-
software solution for information extraction applications in
NLP. Memristors are emerging resistive storing element that
contributes to several neuromorphic applications. Memristors
are non-volatile 2 terminal devices that reduce the area and
power consumption. Each document represents sentences
using a layered crossbar that contains memristive cross-
bar nodes. The words in the document are converted into
columns, where characters represent a single crossbar node.
Crossbar nodes with memristive conductance show character
representations.
The most common edge-computing hardware comprises
general-purpose microprocessors, FPGA, or GPUs for im-
plementation. Nevertheless, these general-purpose solutions
consume enormous amounts of power and are limited by
battery lifetimes. Meanwhile, application-specific integrated
chips (ASICs) in digital or analog domain implementations
can be designed with meager power but with high design
complexity and costs. In-memory processing is a growing ap-
proach for developing reliable, low-power, and low-on-chip
area edge processing systems among custom-made hardware
solutions.
In this paper, we present an analog custom integrated on-
chip document scanner solution for implementing the TF-
IDF system, which is further used for document analysis
involving information extraction, sentiment analysis, and text
categorization. The workflow of the general-purpose docu-
ment scanner is illustrated in Figure 2. Figure 3 represents
the TF-IDF workflow. TF-IDF(term frequency-inverse doc-
ument frequency) [1] score is a prevalent method used for
page ranking in search engines, text summarization [2], text
similarity computation, and web data mining. TF-IDF is the
most widely used term-weight algorithm. The following sec-
tion explains the background information regarding the TF-
IDF and the necessary hardware requirements. The results
and discussion are presented in Section III, followed by the
conclusion.
II. BACKGROUND
A. TF-IDF ALGORITHM
•Bag-of-words The disorganized nature of text makes it
difficult to model, and methods such as machine learn-
ing algorithms favor inputs and outputs with clearly
specified set lengths. A bag-of-words is a textual illus-
tration that shows where words appear in a manuscript.
This entails measuring the presence of recognized
words and having a lexicon of such terms. In simple
terms, if the word appears more frequently, it is an
important word, and we can use a vector to represent
the bag-of-words model. In term frequency - inverse
document frequency (TF-IDF), each term is weighted
by dividing the term frequency by the number of doc-
uments in the corpus containing the word, as opposed
to expressing a term in a document by its primary
frequency (number of occurrences) or relative frequency
(term count divided by document length) [12]. The
ultimate aim of this weighting is to prevent a common
issue with text processing: words commonly used in one
document are frequently used in all texts. Contrarily,
terms with the best TF-IDF scores stand out as being
particularly prevalent in a document when compared
with other texts.
2VOLUME , 2022
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3237463
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
FIGURE 2. Workflow of document analytic accelerator. Different document formats are scanned and digital and paper documents are processed separately.
Prepossessed data are used for various applications like information retrieval, Key word extraction and Sentimental analysis.
TABLE 1. TF-IDF based information retrieval techniques with tranditional and modifed TF-IDF approaches.
Work area Dataset Algorithm Evaluation parameters
Text categorisation-
Modified TF-IDF Term
weighting strategies [5]
20 newssgroups text dataset
SVM, AdaBoost, Decision Trees,
Gaussian Naive-Bayes(G-NB),
Multinomial Naive-Bayes(M-NB)
Binomial Naive-Bayes(B-NB)
F-meaures and Accuracy,
difference in modified and
traditional TF-IDF.
Text classification-
improved TF-IDF [6]
Natural Language Processing
Group of the International Database
Center of the Department of
Computer Information and
Technology of Fudan University
and thetext corpora of the Sogou Lab
TF-IDF, Word2vec model
Precision, recall, f-measures
among original TF-IDF and
improved TF-IDF
Research paper classification-
TF-IDF and LDA schemes [7]
Titles and abstracts of the papers
published on Future Generation
Compute Systems(FGCS)
journal from 1984-2017
TF-IDF, Latent Dirichlet
allocation (LDA), K-Means
F-scores on three methods where
K=10,20,30 respectively
Key word extraction-
comparison of TF-IDF and
Luhn significant words [8]
Documents are obtained from the
IEEE Xplore research database these
are Technology-enhanced
learning(TEL) processed 40TEL
articles, published no earlier than 2010
TF-IDF
TF-IDF approcah is proven to be
better than Luhn’s significant
words approach as it finds the
more important words from
different catafories tha the later
Research and realization of
internet public opinion analysis
based on improved TF-IDF
algorithm. [8]
Data is crawled by web crawler
4288 buyers reviwes on "Taobao"
shopping site
TF-IDF and improved TF-IDF-NL
Comparison is done based
on the Original TF-IDF score
to the modified in different
catagories of reviews with
different key words
Text summarization [9] Collection of research papers
ANN(Multilayer perceptron,
Probabilistic Neural Network,
Time Delay Neural Network)
Error-values(RMSE,NRMSE,
Max and Min abs error etc)
Keyword Generation Model [10]
Chinas largest scientific and
technological publications databases
and largest publicaly availabale
keyword generation dataset KP20k
Neural Topic Model and
Title-Guide Hierarchial Encoder
F1-measures of KGM-TT to other
models such as Seq2seq, TF-IDF,
CopyRNN, CopyCNN,
TextRank and KEA
•Term Frequency
tf(t, d) = N(t, d)/||D|| (1)
tf(t,d) is the term-frequency for a term ‘t’ in the docu-
ment ‘d’, N(t,d) represents the number of times a term
‘t’ occurs in document ‘d’, ‘D’ gives us the total number
of words in a document.
•Document Frequency
df(t) = N(t)(2)
idf(t) = N/df(t)(3)
N(t) is the amount of documents that the phrase ’t’, and
df(t) reflects the Document Frequency of the term ‘t’.
idf(t) = log(N/df (t)) + 1 (4)
As the N/df(t) value increased, the slope of the logarith-
mic function decreased. This indicates that drastically
increasing N will not have as much of an impact on
the TF-IDF score after a certain point, mimicking actual
life. Dissimilarity does not matter much after a certain
point.
VOLUME , 2022 3
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3237463
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
TABLE 2. Feature Extraction Methods Used in Natural Language Processing other than TF-IDF
Work Area Dataset Algorithm Evaluation Parameters
Comparison between BM25
and TF-IDF techniques
for feature extraction [3]
2,196 tweets from twelve general
subcategories are filtered based on
location New York and Toronto
TF-IDF and Okapi BM25 Classification metrics
is used for evaluation
Automatically unsupervised
extracting the keywords
and keyphrases from published
research documents. [1]
titles, abstracts, and introductions
of research documents PositionRank, TextRank, and RAKE
Evaluation was based on
the maximum prediction
of accurate keyword or
Keyphrase extraction
Keyword extraction from
cross-language Documents
with Entropy TextRanking [4]
English and Chinese Confucius
Institute 6420 web text news report.
Information Entropy and TextRank
with Machine Learning approach
Latent Dirichlet Allocation(LDA)
Precision, Recall, and
f1-score values were
used to evaluate the models
Weight-based extraction of text
summarization with abstractive
and extractive summarization [11]
Contents 5 documents with
20 sentences each
A statistical novel approach
based on Abstractive and
Extractive text summarization
The proposed work was
compared with MSword
and manual-based Analysis
Text summarization LSA
topic modeling along with BERT
and TF-IDF keyword Extractor [2]
Kaggle dataset for text
summarization
Extractive Text summarization
with Latent Semantic Analysis (LSA)
along with BERT and TF-IDF
Keyword Extractor
The evaluation metric for
the proposed method was
Recall-oriented understudy
for Gisting Evaluation(ROUGE-1
and ROUGE-L ) with Precision,
Recall and F1- score
TABLE 3. Conductance variability analysis for various percentages (c) of variability ratios. RCE denotes Relative Current Error. Case 1, 3 and 4 represent three
categories of variability conditions which are explained under variability in Section III
Document Case 1 Case 3 Case 4
10% 30% 50% 10% 30% 50% 10% 30% 50%
1 0.007 0.117 0.216 0.004 0.006 0.01 0.005 0.06 0.22
2 0.004 0.16 0.18 0.003 0.035 0.02 0.006 0.05 0.19
3 0.004 0.10 0.20 0.002 0.005 0.016 0.008 0.05 0.18
FIGURE 3. TF-IDF algorithm workflow
TABLE 4. Top ten TF-IDF scores of 3 documents
Document 1 Document 2 Document 3
pill 0.041612 taste 0.03939 design 0.040571
home 0.027741 different 0.02626 filter 0.040571
stuff 0.027741 liver 0.02626 feed 0.040571
product 0.027741 variety 0.02626 lid 0.027047
pet 0.023582 rice 0.02626 lift 0.027047
replace 0.023582 herb 0.02626 fish 0.027047
cat 0.023582 herbal 0.02626 drop 0.027047
good 0.018343 tea 0.02626 food 0.027047
surei 0.013871 drank 0.02626 way 0.022993
park 0.013871 starchy 0.02626 door 0.022993
•TF-IDF Score
tf −idfscore =tf(t, d)Xidf (t)(5)
The TF-IDF score informs us of a word’s significance in
a collection or corpus of documents. A word’s frequency
in the corpus offsets the proportionate increase in the
importance that a word experiences in a document.
B. HARDWARE ARCHITECTURE
The Hardware TF-IDF-based architectures are shown in fig-
ure 4. Each document represents sentences using a layered
crossbar [13] that contains memristive crossbar nodes. The
words in the document are converted into columns, where
characters represent a single crossbar node. Crossbar nodes
with memristive conductance [14], [15] show character rep-
resentations. Each character was encoded as a conductance
value. Multistate memristors [16], [17], each with 26 states,
are mapped into appropriate characters in order. In the pro-
posed design for the TF-IDF implementation, we considered
the number mas the number of sentences and j as the
number of words in each document. The dynamic nature
of the layered crossbar representation makes it useful for
general-purpose architectures. If the number of words is less
than j, the remaining column lines are bypassed; hence, the
energy loss can be minimized. In the ignored column lines,
the conductance values were assigned as zero. The remaining
conductance values represent each characteristic. The current
measured at the end of the single crossbar lines were used to
compute each word in the TF equation.
The bag of words that contains unique words is arranged
in a separate column fashion as a search word(swi). This
can be used to count the number of occurrences of a term
in a particular document, and each term is compared with
all words in the layer. Word is represented as the current
value in the analog circuit architecture. A searching word
4VOLUME , 2022
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3237463
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
FIGURE 4. TF-IDF hardware accelerator architecture (a) Crossbar array for searching word(sw1and sw2(b) TF Crossbar array with comparator for 2 document
sentences and two terms (c) IDF hardware structure from tf values using logarithmic circuit
column is shown in figure 4(a), and the ideal current value of
each search word is compared with the crossbar layer column
using a comparator. The comparator output is zero if the two
input values are identical. However, practically, this is not the
case. There should be some minor current out of the column
line due to the leaky current flow. Leakage current can be
neglected. This makes it more advantageous to calculate TF
values without directly applying the TF formula. Connect
resistors out of each comparator and accumulate all currents.
The averaged circuit output provides a similar representation
of the TF value. The IDF score value can be determined using
this equation. As we take the number of sentences as d, we
can modify that equation to
idft1dj=log(N/df(t)) + 1 (6)
Where N=d;
idft1dj=log(d)−log(df(t)) + 1 (7)
Lets take log(d)as a constant value C.
idft1dj=C−log(df (t)) + 1 (8)
log(df (t)) this expression can be determined by accumulat-
ing all TF values together into a logarithmic op-amp.
III. RESULTS AND DISCUSSION
A. DATA SET DESCRIPTION AND EXPERIMENTAL
SETUP
Several known labeled datasets have been used in different
text classification studies. This study, namely the Amazon
product review dataset [18], contains reviews of different
categories, such as health, personal care, toys, games, beauty,
pet supplies, baby products, and grocery gourmet food. This
study’s dataset for checking the model’s performance is a
subset taken from 200k Amazon reviews. The boundary
resistance values for mapping the characters were 1k and 10k.
The multistate memory element can be programmed into 26
states, each for different characters. For example, to represent
a ‘spectrogram’, the resistance list can be [7480, 6400, 2440,
1720, 7840, 7120,3160, 7120, 1000, 5320]. Simulations were
performed using LTSPICE. 180 nm high-k PTM models are
used as transistor switches, and the memristor model used
here is the Knowm MSS (Multi-Stable Switch) memristor
model and WOx device parameters. Verification of the text
categorization of documents was performed using the Python
platform.
B. AREA AND POWER
Crossbar nodes are designed with 1T1M(One transistor, one
memristor) structure. Each layer represents a sentence line
in a document. Here, we take the number of column lines
as m for ease of calculation. The number of characters
varied depending on character size. Let us take the maximum
number of characters in a word as i. Characters are mapped
to the rows of a single crossbar layer. Hence, if there are m
columns and irows, the total number of characters in the
sentence line can be m×i. Therefore, each layer will have
m×icrossbar nodes. The area occupied by a single crossbar
node was 0.14µm2. The comparator and logarithmic circuits
were designed using OpAmp. The area taken for a single
OpAmp was 2801.76µm2and had a power conception of
39.8mW .
C. VARIABILITY
The output from the comparator is the current value, rep-
resenting the unique words associated with each sentence.
VOLUME , 2022 5
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3237463
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
TABLE 5. Sentence scores for three documents
Document
No. Sentence Score
1Cats love them! You will never have to force down a pill again!!! Place 1/2 tablet per 1 pill pocket 0.3917055011
I like to clean the cat box and replace the litter also to reduce the possibility of re-infestation 0.4700466013
2Not much of a "taste" but just kind of a starchy flavor 0.0376493637
I guess this herb is supposed to help clean your liver but who can really say how much benefit it’s having?
I don’t have any specific liver problems,
but I just like to try a variety of different herbal formulations to get the
benefits of items that don’t get consumed in the normal daily diet
0.1129480911
3great! They tellyou to replace the filter every week or few weeks,
so they make a nice little door on the cover forthe purpose 0.03877884461
It seemed like a good design in the store 0.1163365338
Ideally, the search word and crossbar column containing
the keyword should have the same current values. However,
owing to the nonidealities [19] associated with the memristor
element, the conductance value changes. This change in
conductance value leads to inappropriate character mapping
and alters the current value. The changes in conductance
variability [20] can be divided into four categories. Such as
Case 1:
GON +A×GON and GOF F +A×GO F F (9)
Case 2:
GON +A×GON and GOF F −A×GOF F (10)
Case 3:
GON −A×GON and GOF F +A×GOF F (11)
Case 4:
GON −A×GON and OF F −A×GO F F (12)
GON and GOF F represent the maximum and minimum
conductance, respectively. The conductance variability ratio
is denoted by ’A’. Ideally, the peak values affect the conduc-
tance variability more than other conductance states. How-
ever, when the variability exceeds a threshold level, the other
conductance states also diminish. As in character mapping,
only 26 states are required to represent any word. Memory
elements with 26 or more conductance states are ideal for
one-to-one character mapping. If the equivalent conductance
state is less than 26, the characters cannot be distinguished
between the words. More conductance states were also least
affected by variability.
The maximum and minimum resistances were 100kΩand
10kΩ, respectively. Variability can lead to a change in the
equivalent conductance levels. When the upper conductance
value increases and the lower conductance value decreases
(Case 2) owing to external factors, the equivalent available
conductance count will not affect the character one-to-one
mapping. All other cases can change the character mapping
because of a reduction in the number of conductance states.
The change in the relative current error due to the conduc-
tance variability for the three different conditions is shown
in Table 2. Here, three documents with single sentences were
considered.
D. APPLICATIONS
1) Information retrieval
Retrieving information from text or documents computation-
ally condenses a set of data into a subset that best captures the
most crucial or pertinent information in the original material.
For information retrieval, keywords [21] from the document
or text were required with the help of the TF-IDF score. Table
3 describes the most relevant words from each document
with their TF-IDF scores. Document 1 consists of a review
of cat products, document 2 is a review of herbal tea, and
document 3 consists of fish aquarium features. The highest
TF-IDF score for each document indicates the document’s
most important words. Hence, by using the TF-IDF formula,
we can extract the outline of the text by the scores of each
word and select a few words with the highest score among
other terms.
2) Text summarization
By removing less important information from the text, a text
summary makes it more concise and accessible for the reader
to find the information they need [22]. Table 5 consists of two
sentences from each document with high TF-IDF scores. The
sentence score was computed by summing the scores of each
unique term present in that sentence. Sentences with higher
scores [23] depict a summary of the entire document. There-
fore, the TF-IDF score is valuable for text summarization.
IV. CONCLUSION
The proposed system consists of a hardware implementation
for consumer document analysis. This document analytics
accelerator hardware performs natural language processing-
based tasks using the TF-IDF equation. The multilayered
architecture can include multiple documents with more than
ten sentences, and each word is stored in a single-column
mode. Memristor analog crossbar structures with compara-
tors and logarithmic circuits were designed to mimic TF-IDF
operation. The developed model was used for information
retrieval and text summarization applications. The proposed
circuit can be integrated with edge-embedded devices to
perform various real-time NLP tasks without compromising
confidentiality or latency reduction. We conclude that there
can be further motivation for our work, and this proposed
idea can be further applied to other NLP applications. Due
6VOLUME , 2022
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3237463
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
to vast amounts of data, NLP applications requires large
computation power to execute complex information-handling
processes. The amount of data can significantly impact the
system performance. Huge data handling require ample space
for storage and it slows down the computation process.
Proper hardware- aware extraction techniques and optimiza-
tion models are essential for a sufficient preprocessing stage.
This will reduce the time delay and storage requirements.
REFERENCES
[1] M. G. Thushara, T. Mownika, and R. Mangamuru, “A comparative study
on different keyword extraction algorithms,” in 2019 3rd International
Conference on Computing Methodologies and Communication (ICCMC),
2019, pp. 969–973.
[2] H. Gupta and M. Patel, “Method of text summarization using lsa and sen-
tence based topic modelling with bert,” in 2021 International Conference
on Artificial Intelligence and Smart Systems (ICAIS), 2021, pp. 511–517.
[3] A. I. Kadhim, “Term weighting for feature extraction on twitter: A com-
parison between bm25 and tf-idf,” in 2019 International Conference on
Advanced Science and Engineering (ICOASE), 2019, pp. 124–128.
[4] X. Zhang, Y. Wang, and L. Wu, “Research on cross language text keyword
extraction based on information entropy and textrank,” in 2019 IEEE 3rd
Information Technology, Networking, Electronic and Automation Control
Conference (ITNEC), 2019, pp. 16–19.
[5] R. K. Roul, J. K. Sahoo, and K. Arora, “Modified tf-idf term weighting
strategies for text categorization,” in 2017 14th IEEE India Council Inter-
national Conference (INDICON), 2017, pp. 1–6.
[6] C.-z. Liu, Y.-x. Sheng, Z.-q. Wei, and Y.-Q. Yang, “Research of text clas-
sification based on improved tf-idf algorithm,” in 2018 IEEE International
Conference of Intelligent Robotic and Control Engineering (IRCE), 2018,
pp. 218–222.
[7] Kim, Sang-Woon, Gil, and Joon-Min, “Research paper classification
systems based on tf-idf and lda schemes,” in Human-centric Computing
and Information Sciences, 2019, pp. 2192–1962. [Online]. Available:
https://doi.org/10.1186/s13673-019-0192-7
[8] A. Rahmah, H. B. Santoso, and Z. A. Hasibuan, “Exploring technology-
enhanced learning key terms using tf-idf weighting,” in 2019 Fourth
International Conference on Informatics and Computing (ICIC), 2019, pp.
1–4.
[9] N. Abdelaleem, H. Abd elkader, R. Salem, D.-D. Salama, and A. Elmi-
naam, “Extractive text summarization using neural network,” 11 2020.
[10] J. C. J. Z. Y. Zhang, Dalin Ma, “A novel keyword generation
model based on topic-aware and title-guide,” in Computational
Intelligence and Neuroscience, 2022, pp. 1687–5265. [Online]. Available:
https://doi.org/10.1155/2022/1787369
[11] J. Madhuri and R. Ganesh Kumar, “Extractive text summarization using
sentence ranking,” in 2019 International Conference on Data Science and
Communication (IconDSC), 2019, pp. 1–3.
[12] H. C. Wu, R. W. P. Luk, K. F. Wong, and K. L. Kwok,
“Interpreting tf-idf term weights as making relevance decisions,”
ACM Trans. Inf. Syst., vol. 26, no. 3, jun 2008. [Online]. Available:
https://doi.org/10.1145/1361684.1361686
[13] D. N. Yadav, K. Datta, and I. Sengupta, “Analyzing fault tolerance be-
haviour in memristor-based crossbar for neuromorphic applications,” in
2020 IEEE International Test Conference India, 2020, pp. 1–9.
[14] J. Lin, Z. Zhu, Y. Wang, and Y. Xie, “Learning the sparsity for reram:
Mapping and pruning sparse neural network for reram based accelerator,”
in Proceedings of the 24th Asia and South Pacific Design Automation
Conference. Association for Computing Machinery, 2019, p. 639–644.
[15] A. P. James and L. O. Chua, “Analog neural computing with super-
resolution memristor crossbars,” CoRR, vol. abs/2105.04614, 2021.
[Online]. Available: https://arxiv.org/abs/2105.04614
[16] Y. Zhang, X. Wang, and E. G. Friedman, “Memristor-based circuit design
for multilayer neural networks,” IEEE Transactions on Circuits and Sys-
tems I: Regular Papers, vol. 65, no. 2, pp. 677–686, 2018.
[17] C. Sung, H. Hwang, and I. Yoo, “Perspective: A review on memristive
hardware for neuromorphic computation,” Journal of Applied Physics, vol.
124, p. 151903, 10 2018.
[18] “Hierarchical text classification, url =
https://www.kaggle.com/datasets/kashnitsky/hierarchical-text-
classification, urldate =Accessed: 2022-11-17.”
[19] P. Pouyan, E. Amat, and A. Rubio, “Reliability challenges in design
of memristive memories,” in 2014 5th European Workshop on CMOS
Variability (VARI), 2014, pp. 1–6.
[20] C.-Y. Chen and K. Chakrabarty, “Pruning of deep neural networks for
fault-tolerant memristor-based accelerators,” in 2021 58th ACM/IEEE
Design Automation Conference (DAC), 2021, pp. 889–894.
[21] N. Firoozeh, A. Nazarenko, F. Alizon, and B. Daille, “Keyword extraction:
Issues and methods,” Natural Language Engineering, vol. 26, no. 3, p.
259–291, 2020.
[22] H. Christian, M. Agus, and D. Suhartono, “Single document automatic text
summarization using term frequency-inverse document frequency (tf-idf),”
ComTech: Computer, Mathematics and Engineering Applications, vol. 7,
p. 285, 12 2016.
[23] C. Fang, D. Mu, Z. Deng, and Z. Wu, “Word-sentence co-ranking for auto-
matic extractive text summarization,” Expert Systems with Applications,
vol. 72, pp. 189–195, 2017.
ASWANI A.R. is a PhD student at the School of
Electronics Systems and Automation, Digital Uni-
versity of Kerala. Aswani’s research area focuses
on neural network accelerators. Aswani received
a Master of Technology in Embedded Systems
in 2020 and Bachelors degree in Electronics and
Communication in 2017. Aswani was working as
a Research Assistant (2020-2021) and is currently
involved in a few projects related to hardware
based neural network implementation. Aswani has
published several papers related to both hardware and software neural
network optimization problems including CNN, LSTM, etc. Aswani’s areas
of interest include neural networks, analog circuits, natural language pro-
cessing, and neuromorphic computing systems. She is a graduate student
member of IEEE. She is a reviewer and author of several top IEEE journals
and conferences.
DIBYASHA MAHAPATRA is an MSc student at
the Digital University of Kerala. She received her
BSc degree in Mathematics. Dibyasha works in
the areas of data analytics and natural language
processing. Her research interests include the ap-
plications of machine learning and data analysis
for consumer technology and real-time product
design.
VOLUME , 2022 7
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3237463
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
ALEX JAMES received the Ph.D. degree from the
Queensland Micro and Nanotechnology Center,
Griffith University, Brisbane, QLD, Australia. He
works as a professor of AI hardware at the School
of Electronic Systems and Automation and Dean
(Academic) at Digital University Kerala. He works
in a broad range of brain-inspired systems, mem-
ristive systems, intelligent semiconductor devices,
analog circuits, and imaging systems. He is a prof-
in-charge of Maker Village that supports over 80+
hardware startups. He is the chief investigator at the Center for Intelligent
IoT Sensors and the India Innovation Centre for Graphene, developing new
products to the market. Dr. James was the founding chair of IEEE Kerala
Section Circuits and Systems Society. He is a member of the IEEE CASS
Technical Committee on Nonlinear Circuits and Systems, IEEE CASS
Technical Committee on Cellular Nanoscale Networks and Memristor Array
Computing. IEEE Consumer Technology Society Technical Committee on
Quantum in Consumer Technology (QCT), Technical Committee on Ma-
chine Learning, Deep Learning and AI in CE (MDA), and Member of BCS
Fellows Technical Advisory Group (F-TAG). He was an editorial member
of the Information Fusion, Elsevier, and is an Associate Editor for Frontiers
in Neuroscience (Section: Neuromorphic Systems), IEEE ACCESS, IEEE
Transactions on Emerging Topics in Computational Intelligence (2017-18)
(guest associate editor) and IEEE Transactions on Circuits and Systems 1
(2018-present). He is a Senior Member of IEEE, Life Member of ACM,
Senior Fellow of HEA, Fellow of the British Computer Society (FBCS), and
Fellow of IET (FIET).
8VOLUME , 2022
This article has been accepted for publication in IEEE Access. This is the author's version which has not been fully edited and
content may change prior to final publication. Citation information: DOI 10.1109/ACCESS.2023.3237463
This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 License. For more information, see https://creativecommons.org/licenses/by-nc-nd/4.0/
