You are chatting on WhatsApp and exchanging hilarious GIFs here and there. Have you noticed that those matching the subject of your recent conversation helpfully line up first? Have you ever wondered why?
If humans are able to reconstruct thoughts from words, could computers do this as well?
Source: Towards Data Science
The gist of NLP lies in providing means of how computers can analyse and learn from human languages, and ultimately generate human-like responses. Recent years have seen specialised machine learning algorithms evolve by leaps and bounds capitalising on decades of research.
The research has come a long way and transitioned from philosophy into a subfield of artificial intelligence. Check this article about NLP history to see this fascinating subject in the context of linguistics and other disciplines.
As a result, NLP has become widespread and often irreplaceable. It is at the core of virtual assistants, spell checks, translation services, auto suggestions, spam or hate speech filters, customer care services and the list goes on. It helps us daily, oftentimes in subtle and entertaining ways.
Speech Recognition
Let me start with arguably the most apparent use case – voice driven commands. Virtual assistants like Siri or Alexa continually listen to, interpret and act upon verbal instructions. It’s not just them though. Speech recognition has established itself as a full-blown user to machine interaction in the last few decades. NLP only made it better and allowed for new unconventional uses.
- Diminished language barriers through automatic real-time translations. As NLP matures real-time machine translation become an integral of chatting apps of all kinds. Current examples include Microsoft’s Skype Translator, Google’s Translate Voice, iTranslate and others. They are all limited in terms of the number of supported languages, but at the same time they won’t go away anytime soon. Speech recognition has become accessible to for custom development. For instance, it has become an integral part of Android SDK. In a hands-free mode the device captures voice commands without requiring any other sort of interaction (a touch or a visual check).
- Increased efficiency and convenience through voice interactions. Speaking as opposed to typing creates new opportunities across industries. Voice-driven searches bear additional information that can be used by marketeers. At the same time it’s for the users and cuts on their screen time. Many simple, but important, tasks lend themselves well to off-screen voice interaction. Example include checking for account balance, booking an appointment, search for specific keywords (disease symptoms, financial indicators etc.).
- Education and language learning benefits a lot from speech recognition. Sadly, I’m not talking about a virtual assistant standing in for a native speaker. When it comes to low-cost conversation services available round the clock is still reality to-be. This article explains in detail why aren’t we quite there yet. However, pronunciation checks and scoring have become a common enhancement.
Key Points:
- Cognitive computing , as in machine’s capability of simulating human’s brain, has become more broadly accessible. Moore’s Law accelerated the evolution of machine learning algorithms which led to a wider adoption of voice controlled assistants of all kinds.
- Voice controlled assistants and devices have come a long way to better align technology with our life styles.
Machine Translation
Machine translation is an obvious case for NLP (Google Translate and similar). Rather than hand-crafted word for word substitution, companies like Google deploy statistical models backed by a large corpora based on search queries and other sources. Think entire libraries of books, contents scraped from websites and all other imaginable sources a search giant can possibly get their hands on. Check this short article on Google’s recent improvements to their translation technology.
Key takeaways:
- Whole sentences are translated at a time instead of word for word translations.
- Zero-shot translations as the ability to translate text never seen before.
Recommendations
Text Summarization
Transfer Learning
Traditional Use
Question answering is another task NLP can handle really well. Given a piece of text a computer is able to answer (unforeseen) questions about the contents. Have fun with this demo of a Q&A system backed by deep learning. Check this blog post to learn more about it.
Categorisation and topic mining help us make sense of large swathes of articles, news, Twitter feeds etc. How is that useful? A mailbox organiser is a good example. Services like Clean Mail, Google Unsubscribe, Mailstorm crunch through your mailbox and reliably discern (and discard) marketing campaigns from the rest of your communication.
Sentiment analysis looks at what people say on Twitter or other channels and evaluates their feelings. It gave a rise to recommender systems. As a result, public reviews are enough to rate a hotel on TripAdvisor or a restaurant on Google maps.
Text and speech generators deal with autocompletion and prediction of the most suitable response in general. This article shares experience with testing of an autonomous text generator developed by OpenAI.
Digital Footprint Analysis
Digital footprint analysis adds a creepy twist to it. Imagine your personal credit score derived from your interactions on social networks and browsing activity. We all leave a huge virtual footprint, unintentionally revealing facts about our personality and private matters. NLP provides means of capturing the essence about who we are. I’ve tried this demo provided by the University of Cambridge. The results weren’t exactly shocking, also due to the fact that the Twitter data upload no longer works (hey folks, make it open source and I help you fix it!), but even with limited data the results weren’t too bad at all.
