Walking through the challenges and the use cases of NLP

challenges and the use cases of NLP

In our previous articles we went over the various approaches to NLP. However, NLP is not without its complexities. In this article, we'll understand why there are many challenges in implementing mainstream NLP in multiple scenarios. We will look at some of the key reasons. 

Challenges with NLP:

The first one is overfitting on small dataset. Deep learning models have more parameters than traditional machine learning models. This enables the model to fine-tune results. However, this comes with the curse of needing a sufficiently large amount of training datasets. With a smaller dataset, deep learning models tend to overfit, and this leads to low generalization, which ultimately leads to degraded performance and production.

There's an interesting concept of Occam's razor, which suggests that, given all other conditions equal, it is always preferable to choose the simplest solution. So for use cases where we have less data initially, it is advisable to select machine learning approach in place of deep learning. The next is domain adaptation. When we train models on general dataset and then try to apply the model to a new domain that is different to common domain, the model tends to have degraded performance.

This holds real true in the case of deep learning models as well. For instance, if we have a model trained on social media post and then applied a trained model in a domain such as health care, where there is a difference in both syntax and semantics, there will be noticeable drop in performance. We need specialized models to encode domain knowledge. The last one is interpretable models. Deep learning models have high predictability and low interpretability.

Even though this works great when it comes to doing NLP, business needs interpretable results. At the end of the day nods is the results, but the reasoning behind the result needs to be explained to the customer or the end user. For instance, even though LSTM model will yield a more accurate result for sentiment analysis, a traditional model such as Naïve Bayes is easier to interpret.

Let's continue the challenges that we face in terms of real-world application of deep learning for NLP. The next one is few-shot learning and synthetic data generation. Few-shot learning is the model's ability to learn from very few training samples. This has been achieved in computer vision application, thus resulting in wider adoption of problem-solving in industrial scenarios. We have not yet seen the same for NLP. Next is common sense and world knowledge.

The reality is that as human beings, we perform multi-step reasoning much faster and intuitively than we realize. A sentence such as I like burgers, can be interpreted as the user like burgers and feel happy when they eat a burger, whereas a relatively complex reasoning example can be: if Alice walks out of a kitchen and goes to the garden, then Alice is not in the kitchen anymore and her current location is the garden. This might seem trivial to human, but this multi-step reasoning needs to be encoded into machine. That is currently an ongoing effort to collect common sense events and similar logical rules. And the last is cost and on-device deployment.

For lot of real-world use-cases, building deep learning solution for NLP task can be very expensive. The associated costs both in terms of money and time can be huge as DL and ML models are data guzzlers and hardware-intensive. Getting label data for lot of use-cases involves manual tagging which takes both time and money. Models are useful when we can deploy them.

When we have a model that is extremely large and the target device is on mobile phone, with a limited memory and a power on mobile device, it becomes grudgingly hurdle to overcome, even if it is for task of machine translation. These are some of the real challenges faced with the application of NLP. In the next article we learn about the current tool ecosystem of NLP.

Use cases of NLP:

Now that we have a foundational understanding of NLP, we'll go over detail industry use-case. To recollect, some general application of NLP is translation, autocorrect, autocomplete, conversational AI, speech recognition, text summarization and language models. Let us now look at use-cases from four prominent domains, the first one being healthcare.

As one of the oldest industries, healthcare is one of the domains that has most to gain from the strides made in natural language processing. Some of the NLP use-cases in healthcare are: clinical documentation, computational phenotyping, clinical diagnosis and virtual therapist.

The next domain is finance. Application and processes in finance are much more manual than we think. Automating finance application with NLP can be done in the following ways: insurance claim management, financial reporting and auditing, fraud detection and stock price prediction.

Let's look at couple of more industry use-cases. The first one being retail and e-commerce. An industry that has lot of scope for automation, retail and e-commerce is one of the fastest growing industries in terms of technology.

A few of the use-cases includes: customer service chatbot, market intelligence, review analysis, customer sentiment analysis, and product recommendation. The next domain is human resource. A horizontal of a company that is people-intensive without the benefit of directly impacting the profits, HR tends to have a restriction on budget and never enough people.

A great way to assist HR is to build more efficient system and this can be done in the following ways with the help of NLP: resume evaluation, employee sentiment analysis, exit interview analysis, profile and skill recommendation engine. That was an overview of some of the latest industry use-cases in NLP. With that, we're almost at the end of the course. In the next article, let's summarize what we have learned so far.