The culture and heritage of India depicted in the ancient manuscripts are unique in its linguistic and traditional diversity. However, the paucity of skill and expertise in present-day research posed a threat to the exploration of this textual legacy. In this paper, we aim to create a visual narrative of one of the major epics of Indian Literature 'Ramayana', by summarizing the major topics and linking them with the characters and locations using Artificial Intelligence (AI). Using this research, any person interested in studying these manuscripts can visualize the tenor of the entire script, without an intensive study. 'Ramayana' was originally written in Sanskrit, but modern versions have Sanskrit text with the explanation in Hindi (as most people in India are well versed with Hindi). In this paper, we have divided the Hindi and Sanskrit text and considered only Hindi text for our further research. We have used existing scientific models (that are trained on the Hindi Language) to find events/topics, summaries, characters and locations that are later used to produce a visual narrative of the data. For the evaluation of our results, we have tried to review the understanding of our summaries and topics. We achieved this by providing a part of our input text and its summary as well as topics/events created by our data pipeline, to 30 people (who are well versed with the Hindi Language). From the survey, it was found that 70%of the respondents understood the summarized text, while 56%of the respondents understood the topics clearly, that is generated from our model.
Modeling and Named Entity Recognition
At present, there is a growing enthusiasm among historians and humanitarians for exploring the quintessence of ancient Indian manuscripts. As there exists physical evidence of the events occurring in most of these scriptures, we can gain information about the demography and cultural aspects of ancient
architecture, medicine, engineering, beliefs, etc. can be studied through these manuscripts.
underlying notion behind this research is to create a pipeline, where-in any ancient manuscript can be provided as an input, ACI’21:
Workshop
Advances in
Computational
2020 Copyright for this paper by its authors. 1 https://archive.org/details/in.ernet.dli.2015.345471/page/n1/mode/2up history ’Ramayana’.
exists
around300 versions of ’Ramayana’ throughout
the world [
triumphed over the evil forces of Ravana, the Demon King of Lanka.Not only the characters in this scripture are considered as gods in India, also there exists physical evidence of the events stated in these scriptures, that can be traced back to present-day locations in India and Sri Lanka. Our downloaded data consists of the original Sanskrit Shlokas followed by its interpretation in Hindi.
From this data we want to produce a visual
narrative so that any individual can understand
the gist of the entire text without having to read
the entire text. Our pseudo-pipeline can be used
to reproduce similar results for any Indian
manuscript. Our designed pipeline can be
divided into five major components. These
components are:
1. Input data (Digitize the data for further
processing) The main aim of this process is
to digitize the given input data that is in pdf
format. Firstly, we need to fetch each page
and convert it into PNG format. Then these
images are fed into an OCR engine that will
convert the images into machine readable
text. The OCR engine used for our study is
the Py-Tesseract OCR2 for Devanagari
Scripts.
2. Basic Pre - Processing (Preparing the
data) After getting the data in machine
readable text, we need to clean the data, as
the OCR’d text might be wrongly read, and
if this wrong text is further processed in our
pipeline, then the results can be hugely
impacted. The first step in this process is to
remove the headers from the text document,
as they do not provide any value to the data.
Now, we have the data with both Sanskrit
and Hindi text. We decided to use only
Hindi text for further processing, as the
Hindi text contains the description of the
Sanskrit Sargas that should provide better
results for our Summarization and Topic
Modeling models (due to considerable
amount of data as compared to Sanskrit
text). Hence, the Sanskrit and Hindi texts we
reseparated (using some keyword
identifiers) and stored in 2 different files.
The Hindi text was then divided into 26
parts based on a word limit, that would help
in creating small summaries and would
make it easier to find events from the
divided subtext. All these summaries and
topics/events can then be used to find the
quintessence of the entire data. From this
step, the data was simultaneously sent into 2
different processes, i.e., one for Text
Summarization and NER Tagging, and other
for Topic Modeling.
3. Summarization and NER Tagging
(creating a concise summary and then
finding locations/persons involved in those
summaries using NER Tagging) This step
is divided into 2 processes as described
below:
• Summarization (summing up the most
important or relevant information from the
entire text) The main process of the Text
Summarization producing a concise and
fluent summary of the text while preserving
key information content and overall
meaning [
In one of the dashboards, the locations 4 https://github.com/stanfordnlp/stanfordnlp (retrieved using the NER Tags) were plotted on the map of Indian Subcontinent and when hovered over a mapped location, the summary linked to that location is displayed. Using this, anyone can be able to create a mental image of the story that was described in its summary. The other dashboard displays the characters (retrieved using the NER Tags) distributed according to their corresponding summary. And, when hovered each character, the topics/events associated with that character are displayed. Using both these dashboards, anyone can easily find the quintessence of the whole script without doing intensive study on it.
In this section, we discuss the existing
papers from which the basic scientific models
used for our study, have been influenced.
Richman, Paula [
Qi, Peng et al. [
Pre-processing is the most important part of
the text processing system. In the preprocessing
pipeline, the removal of functional words (stop
words) is important in in sense of performance
of text processing. Jha, Vandana et.al.[
In the journal paper by Allahyari, Mehdi et
al., text summarization and its techniques are
explained in detail, which was very useful to get
a background knowledge of text summarization
and its types. The survey of Text
Summarization for Indian and foreign language
Dhawale, Apurva et al. [
Athvale, Vinayak et al.in their paper”
Towards deep learning in Hindi NER: An
approach to tackle the labelled data scarcity
„provide describe an end-to-end Neural Model
for Named Entity Recognition (NER) which is
based on Bi Directional RNN-LSTM. The
authors claim state of the art performance in
both English and Hindi without the use of any
morphological analysis or without using
gazetteers of any sort. Sharma, Rajesh et.al.[
The paper by Zhou Tong and Haiyi, Zhang
[
For the creation of the visual narrative, there was a need for text mining and preparation of the Valmiki Ramayana data set. The steps involved in attaining the same were identified as: creation of input files, text preprocessing, text Summarization, Named Entity Recognition (NER) tagging and topic modeling. The output from these steps were used as inputs for the visualization process. The complete workflow of the text processing and mining is shown in Figure 1: Workflow of Text preparation for the Visual Narrative
The first part of the workflow was about
creating the input files in the format of
machine-readable texts for further processing.
The single pdf file of Valmiki Ramayana was
converted into 394 images in PNG format with
300 dpi resolution using a free image utility
software called ImageMagick8. For converting
the images into editable text documents,
Optical Character Recognition (OCR) process
was used. Tesseract-OCR [
The output of Py-tesseract i.e., the OCR’d text documents contained some data such as page headers like book’s name (e.g. श्रीवाल्मीकि रामायण) on odd numbered pages and chapter’s name (e.g.सुन्दर िण्ड ) on even numbered pages, page numbers; that were not required for text mining. These unwanted texts were captured using regular expressions and were cleaned out from the text corpus. The documents also consisted of the multi-lingual texts: Sanskrit and Hindi. It was decided to continue further processing using only Hindi texts considering the factors such as the good number of resources available for Hindi language w.r.t. text processing and mining, the relatively greater length of the Hindi text documents than the Sanskrit texts in our case 10 https://github.com/amjha/hindiExtraction and the popularity of Hindi over Sanskrit. Hindi and Sanskrit texts were separated using” मूल” and” टीिा ” keywords from the all the documents, respectively.
There was a fork in further processing of the
Hindi texts: one or Text Summarization and
another for Topic Modeling. Atopic model is a
probabilistic model for finding out the abstract
topics that appear in a collection of text
documents. Topic Modeling is the most used
text mining tool for discovering latent semantic
structures in textual data. For the topic
Modeling branch in our workflow, additional text
preprocessing was required. To extract topics, the
text documents were tokenized into words. In
the text documents, the most commonly
appearing words were the articles, prepositions,
helping verbs, etc., known as the stop words.
These stop words could affect the topic model
as it is generally based on the frequency of
words occurring in a document. So, they needed
to be removed. As there was not out of the box
Hindi stop words removal functionality
available, a list of Hindi stop words10 was
created and this was used to remove stop words
from the tokenized list of words.
Lemmatization is the text processing step of
grouping together the different forms of a word
so they can be analyzed as a single word. This
also helps in reducing redundancy of the same
root word in the topics extracted. Stanford NLP
[
The other branch of the fork is Text
Summarization. Text Summarization refers to
the process of compacting a large text. The
reason behind this is to create a comprehensible
and expressive summary having only the main
points outlined in the text. There are two main
methods to summarize the text in NLP,
Extraction based summarization and
Abstraction based summarization. We are using
extraction-based summarization for our
research. Summarization of the text is based on
the ranks of text sentences using a variation of
the Text Rank algorithm. Text Rank is an
automatic summarization technique, is
implemented in two different ways in our
pipeline, the Gensim python open-source
library [
Named entity recognition (NER) plays an
important role to complete the narrative model.
Using this concept, the persons/characters, as
well as the location, can be extracted from the
story text [
Our complete narrative model’s aim was to have a story of events. So, for our final step of the pipeline we built two dashboards using Tableau. The output from both; the summarization as well as the topic Modeling branch were fed into tableau and the events were tagged with a topic, a summary and NER tags to help pick out characters and locations. In the first dashboard, satellite view of the map is used for plotting, the identified locations from the scripture. For plotting the story on the map, a validation dataset of names and places of the events is manually created. This data is matched with the words which are tagged with B-LOCATION and I-LOCATION NER tags. After identification of the matched location, they are associated with the respective latitude and longitude values so that it can be plotted on maps correctly. The lines between two places on the map shows the sequence of the mentioned locations in the narrated story. A line between two places on the map is created using Tableau’s spatial functions Make Line and Make Point. As seen in the Fig. 2, the narration comprises location “मैकिली” (birth place of Sita) “कित्र िूट ” (Forest in which Ram and Sita went for staying ), “अःतपुर ” (A village in Kishkindha where Ram met Hanuman and his friends), ”महेन्दर ” (Mahendragiri is a name of a mountain from where Hanuman jumped towards Shri Lanka in order to search for Sita), “महासागर” (The ocean between India and Shri Lanka), “पवत ” (Trikoot parvat where Hanuman landed after jumping from Mahendra Giri Moutain), “अशोिवैन ” (Ashok Vatika garden where Sita was kept as a captive by Ravana). On hovering over each line origin city with its corresponding destination city and summary of the text is shown in the tooltip. Dealing with such kind of ancient geo spatial data is some of crucial. Gettty thesaurus playa a significant role in this. Getty is an opensource Geo data base where all possible occurrences of every Geo name include its ancient name can tagged with a unique number. Through this number we can visualize that geo name wit longitude and latitude. Getty plays a crucial role in our project as well for geo data visualization. We used Geety for Geo data visualization. characters in the symbol chart unfolds along with their corresponding summary in the Fig. 3. All the symbols and characters names are shown as legends. Each symbol is manually designed as per its characteristics in the Ramayana epic. On hovering over each character, the related topics for the given summarized is shown in the tool tip.
The source code for our research can be found on GitHub11
In the second dashboard, symbol chart is used to represent different characters of the Ramayana. The sequence of the story from 0 to 25 is plotted from left to right and the occurrence of every character is plotted as per its reference in the text associated with that sequence. The validation dataset of the character names in the Ramayana are matched with Hindi words tagged with B-PERSON and I-PERSON NER tags. The matched names do not identify the synonyms of the same name. In Ramayana, each person has been associated with various names, e.g., “सीात ” is also known by the names “जानिी ” or “जनिपुत्री ” in the same story. So, the result of plotting such data lead to plotting three different points for the same Symbol. To avoid this, grouping of such names under one name is performed in Tableau. As Ramayana is a story and stories are narrated in sequence, Tableau’s page control functionality is used to make the dashboard dynamic.The top topic and the summary of text is shown below the symbol chart. As the sequence changes in the page control the 11 https://github.com/rajrohan/ramayanaocr
For the evaluation of our model, we conducted a survey in which 30 persons participated. The respondents were chosen based on the criteria, that they should be competent in understanding Hindi text. They were provided with the input text and the summary as well as the topics created by our data pipeline. The two metrics, Text Summarization, and Topic Modeling were evaluated in the survey. There are four options to choose from and each option has a 20% bucket size; 80100%, 6080%, 4060% or less than 40%. Selecting the first option i.e., 80100%, implies that the individual has understood the summarized text, and the same goes for the second and third options. Anyone choosing the fourth option i.e., less than 40%, be understood. The results of the survey can be seen in the following chart at Fig.4. It can be inferred from the Fig. 4 that 70% of the respondents were able to understand the summarized text whereas 30% of the respondents were not able to understand the summarized text. From Fig. 5 around 56% of the respondents understood the topic clearly and around 44% of the respondents were not able to understand the topic.
The major objective of the envisioned pipeline
has been achieved but the pipeline can be
further improved. Due to drawbacks in the
process’s in pipeline (due to unavailability of
proficient models for Hindi textual data), we
could not achieve the optimal result. We were
not able to find a better-quality input file which
could have helped the OCR model to identify
the characters in a much better way.
This reduced the quality of NER tags that we
obtained after running our NER model. At first,
we were not able to build a NER Models we
were not able to procure the Hindi NER tag data
and creating the data from scratch was not
possible in the given timeframe of the research.
After failing to find Hindi NER tagged corpus
online, we were helped by AU-KBC research
Centre [
For our input dataset, Ramayana, it is observed that our model reaches the score of more than 70 percent in explaining the Summarized text and about 70 percent in explaining the topic/events generated from the script. It can also be concluded that the NER done on the summarized text generates better results than when it is performed on the topics/ events. Taking usability into account we were successful in making a pipeline for the visual narration of Ramayana. Using our visualization someone even with very little knowledge on Ramayana can easily understand the whole summary of the script. The demo is built on an image-based input and can be later extended to other sources and languages too. However, for its application to all other Devanagari languages, their respective models should be available. We can prove the physical evidence of the locations mentioned in the script by plotting the coordinates of the present day location along with the events that took place on these locations.