Topic modelling. Variation on themes and the Holy Grail

Massive amount of data is collected and stored by companies in the search for the “Holy Grail”. One crucial component is the discovery and application of novel approaches to achieve a more complete picture of datasets provided by the local (sometimes global) event-based analytic strategy that currently dominates a specific field.

Bringing qualitative data to life is essential since it provides management decisions’ context and nuance. An NLP perspective for uncovering word-based themes across documents will facilitate the exploration and exploitation of qualitative data which are often hard to “identify” in a global setting.  NLP can be used to perform different analysis mapping drivers. 

Broadly speaking, drivers are factors that cause change and affect institutions, policies and management decision making. Being more precise, a “driver” is a force that has a material impact on a specific activity or an entity, which is contextually dependent, and which affects the financial market at a specific time. (Litterio, 2018). Major drivers often lie outside the immediate institutional environment such as elections or regional upheavals, or non-institutional factors such as Covid or climate change.  In Total global strategy: Managing for worldwide competitive advantage, Yip (1992) develops a framework based on a set of four industry globalization drivers, which highlights the conditions for a company to become more global but also reflecting differentials in a competitive environment.  In The lexicons: NLP in the design of Market Drivers Lexicon in Spanish, I have proposed a categorization into micro, macro drivers and temporality and a distinction among social, political, economic and technological drivers. Considering the “big picture”, “digging” beyond usual sectors and timeframes is key in state-of-the-art findings.

Working with qualitative data.
There is certainly not a unique “recipe” when applying NLP strategies. Different pipelines could be used to analyse any sort of textual data, from social media and reviews to focus group notes, blog comments and transcripts to name just a few when a MetaQuant team is looking for drivers.

Generally, being textual data the source, it is preferable to avoid manual task on the part of the analyst, though sometimes, depending on the domain, content, cultural variables, etc. it might be required.  If qualitative data is the core, then the preferred format is .csv. because of its plain nature which typically handle written responses better. Once the data has been collected and exported, the next step is to do some pre-processing. The basics include normalisation, morphosyntactic analysis, sentence structural analysis, tokenization, lexicalization, contextualization. Just simplify the data to make analysis easier.

Topic modelling.
Topic modelling refers to the task of recognizing words from the main topics that best describe a document or the corpus of data. LAD (Latent Dirichlet Allocation) is one of the most powerful algorithms with excellent implementations in the Python’s Gensim package.

The challenge: how to extract good quality of topics that are clear and meaningful. Of course, this depends mostly on the nature of text pre-processing and the strategy of finding the optimal number of topics, the creation of a lexicon(s) and the corpora. We can say that a topic is defined or construed around the most representative keywords. But are keywords enough? Well, there are some other factors to be observed such as:
1. The variety of topics included in the corpora.
2. The choice of topic modelling algorithm.
3. The number of topics fed to the algorithm.
4. The algorithms tuning parameters.

As you probably have noticed finding “the needle in the haystack” is not that easy. And only those who can use creatively NLP will have the advantage of positioning for global success.