Deep Learning For NLP: When Not To Use It
Deep Learning (a.k.a neural networks) enjoyed a significant uptake in the recent years. While it has a wide applicability in general ML/AI field it is important to understand its natural limitations.
Non-determinism (or imprecision)
The nature of deep learning (and neural network) is their non-deterministic nature. In plain English — they can only give at best an approximate answer by design. This leads to imprecise results and potential compounding imprecision problem (especially in unsupervised ML applications).
This is a significant limitations which eliminates deep learning as a apparatus for a certain class of problems. For example, while computer vision, most of speech recognition, sentiment analysis, categorization, and supervised ML can tolerate certain imprecision of the results — descriptive data analytics, control systems, business intelligence and most forms of data-driven decision making can not.
Let me explain. Let’s look at sentiment analysis vs. descriptive data analytics.
It doesn’t really matter if your twitter feed is 65% or 68% positive on a given day — all it matters in real-life application is the trending of this value and its relation to similar values in a target group. In fact, calculating the precise sentiment value would yield almost no tangible business value as it would effect neither trending nor relation of this value in any significant way.
In case of data analytics the imprecise result, however, is simply not acceptable. If user asks for an average sales revenue from a particular AdWords campaign — she’s expecting an exact dollar amount — approximation in data reporting simply doesn’t work.
Cognitive Analytics is one area where deep learning has rather limited applicability. Cognitive Analytics is often characterized as a marriage between a fully deterministic processing of a free-form language and advanced data analytics. It requires a completely new class of NLP algorithms to deliver precise and deterministic results.
At DataLingvo we’ve developed a patent-pending Human Curated Linguistics (HCL) technology that provides highly effective computational linguistic parser, real time human curation for language pragmatics and multi-layered supervised learning. All in all — technology like HCL (and similar efforts by Viv, for example) for the first time provides a fully deterministic free-form language comprehension, precise results in real time and operational efficiency to make it a viable business model.
May 2018 Update: check out new DataLingvo 2.0