Machine learning using textual data is increasingly important where people interact with events on social media in real-time. Topic modeling is a technique that looks to identify topics characterized by word usage.
Improved topic modeling would assist during these current unprecedented times of the Covid-19 pandemic to allow countries and organizations to timeously track changing public concerns around the pandemic over time. This research does a comparative study between a traditional LDA (Latent Dirichlet Allocation) algorithm implemented using Python sklearn for topic modeling and a knowledge graph-based approach implemented using Python NetworkX. LDA is implemented using both CountVectoriser and Gensim. LDA accuracy is measured using the Topic Coherence metric.
As topic modeling procedures are very valuable and compelling in natural language processing to semantic mining and latent discovery in a text corpus. It is significant since Machine learning using textual data is very important during the current time where people interact regarding every happening in social media.
| Corpus | Contains a large structured/pre-processed text |
| LSI | Latent Semantic Indexing is the earliest topic modeling algorithm used by search engines nowadays |
| LDA | Latent Dirichlet Allocation is a widely used statistical topic modeling algorithm |
| NetworkX | Python package which aids in the creation and study of complex networks and graph objects. |
| BigARTM | Novel technique known for Additive Regularization |
| Anaconda | Open-source distribution for python programming, it can isolate different libraries its versions/packages hence improves stability |
| Jupyter | IDE for Python programming with the support of file navigation and integrations. |
Road map of research
The research writing is separated into different chapters as mentioned below:
The Chapter One of the document is about the basic idea and motivation behind the research. Chapter two focus on literature review since the need for adequate knowledge from previous research papers were properly identified. Literature review in the first part is aligned towards the history behind topic modeling, where the initial topic modeling technique started Latent Semantic Indexing method.
The literature review in the second part is about the need for topic modeling and its progress over the period is studied where the most interesting areas where topic modeling is implemented are discovered. It is followed by the literature review in the third part regarding the most popular approach in topic modeling called the Latent Dirichlet Allocation, this widely used algorithm is studied understanding the mathematical equation behind since it is often considered as the base of modern topic modeling architectures that are applied on this dissertation like BigARTM & NetworkX
Literature review in the later part is regarding the comparison between the traditional topic modeling methods and ones which are widely used, this review helps to identify the factors that are improving and have the potential to be improved, the advancements that happened over the wide transition to use LDA for topic modeling, and the areas where still the traditional model is applied.
The literature review in the fifth part is regarding the variants of LDA in the libraries Sklearn and Gensim.
The sixth part of literature review is then conducted on the latest topic model introduced based on the Additive Regularization of topic models known as the BigARTM which is an opensource tool to infer topic models and review of recent papers on NetworkX is also conducted.
The seventh literature review concept talks about graph databases that can influence topic modeling. The 8th literature concept is regarding the covid-19 outbreak and spread which is important since the textual data considered is regarding covid19.
Chapter three comprises of the Modulus Operandi regarding how the conducted the research starting from choosing the appropriate interest domain evolved from answering the call for advanced evolution in the topic modeling progressing to the data collection methods, text pre-processing, studying the data – understanding facts and insights, text visualization using word cloud, and application of different topic modeling techniques to evaluate them based on scales which are identified to suit the best.
Chapter four encompasses analysis and discussion based on the implementation of different models and their comparison. Discussion is done on the proposed idea of implementation of graph networks for topic modeling. Discussion is also done on applying the standard text-pre-processing methods and comparing them with novel methods of pre-processing and their evaluation.
Whilst chapter five consists of conclusions made from the metrics derived and it consists of proposals for problems or researches requiring implementation of topic modeling or natural language processing needing better accuracy on results. It also consists of the hierarchy of procedures taken that simplifies the research question. It also discusses the future potential developments based on this research.
If you are interested to read more please send me a message
