This article presents the functionalities of the Natural Language Processing (NLP) used in industry, such as text prediction, there entity recognition, L'relationship extraction, there automatic translation, L'sentiment analysis and the answers to questions. He also mentions use cases for chatbots and spam filters.
After having introduces Natural Language Processing (NLP) to our blog, we will see in this article some features of NLP, which can be used in the industrial world.
Let's go !
Automatic suggestions
This feature is used to speed up the writing of emails, messages and other texts. Very “trendy” in recent years, it is increasingly integrated into our mailing and word processing tools. It is based on the prediction of the next word most suited to the sentence being written. This technology is based on probabilistic or word vectorization models who learned how words fit together in a sentence.
Named entity recognition
This technique is mainly used to identify and classify named entities (words or groups of words) in unstructured text into predefined categories such as:
- organizations
- names of people
- geographical areas
- quantities
- prices
- durations
- percentages…
To achieve this, we can use linguistic resources (dictionaries) but also deep learning algorithms which have learned to identify each entity.
Extracting relationships
Here, we seek to extract the semantic relationships between the entities identified in the text or speech such as “is located in”, “is married to”, “is employed by”, “lives in”, etc…
This use case is handled mainly through linguistic relationships such as “part-of-speech” (verb, adverb, noun, etc.) or lexical relationships (synonyms, homonyms, etc.).
Here is an example of what such an approach can automatically understand.
Text analyzed (extracted from Cnetfrance.fr) [1]
SpaceX launched the 16th batch of his Starlink satellites Since Cape Canaveral in Florida (United States), establishing a new record for his Falcon 9 rockets.
Extracted relationships
Subject | Relationship | Object |
Starlink satellites | Belong to | SpaceX |
Falcon 9 rockets | Belong to | SpaceX |
Cape Canaveral | Is located in | Florida |
Automatic translation
One of the most widespread uses of NLP is machine translation. Thanks to this, translations are done much more quickly, despite the need to check certain turns of phrase retrospectively. The added value is very important for businesses that need to regularly translate content like product reviews, regulatory documents and emails. The most famous applications for machine translation are Google Translate, Amazon Translate And DeepL. Typically, these systems rely on networks such as Recurrent Neural Networks (RNN), Sequential Networks (Seq2Seq), and more recently, Transformers (which we will discuss in our next article).
Sentiment analysis
This feature addresses people's perceptions on certain topics or services. This makes this option very useful for many businesses (customer service, communications, strategic marketing, etc.).
EIts purpose is to verify that goods or services will satisfy customers and to create surveys for brands and even political candidates. This not only helps businesses gain knowledge about how customers perceive these goods or services, but also helps improve concepts, products, marketing and advertising while reducing the level of dissatisfaction.
This is a classification problem and can be treated in a “classical” way by machine learning approaches such as linear classifiers (logistic regression, Bayesian classifier, etc.), k-nearest neighbors, decision, or neural networks such as recurrent neural networks (e.g. LSTM) or transformers (e.g. BERT).
Answer to questions
Most questions asked by humans can be answered using an NLP feature called Question Answering. The model will first analyze the question, using in particular the named entity recognition function, then will formulate an answer in return, according to its knowledge base (which can be very broad!). Siri, OK Google, and Virtual Assistants are examples of question answering apps.
VShis feature is developed using machine learning methods that learn to understand language features without human supervision. Here are two examples of language features that can be used:
- Check if the sentence structure is correct (eg: subject, verb, object)
- Understand the meaning of the sentence (for example: “what is the distance between the earth and the moon?” It is therefore a question of giving an answer which contains the value of the distance)
Statistical methods have paved the way for this approach. This is broken down into 2 main stages:
- Training: a knowledge base is created from the analysis of a corpus of annotated text (model training data set),
- Prediction: identification of the best answer to a new corpus of text (user question).
The context is generally identified by searching in a taxonomy (classification of different words into categories and subcategories) which has been created using the named entities and the classifier. Neural network architectures are also very useful for this use case. They allow textual context to be mapped into logical representations which are then used for response prediction. Among these neural networks, it can be noted that recurrent neural networks (RNN) have shown their effectiveness in processing this type of task.
These features are the basis of what NLP can do.
Here are two examples of application cases that you have most likely heard of: Chatbots and spam filters.
Chatbots
Unlike the “Answer to questions” part where the answer exists in a given corpus, chatbots generate their own answers.
Chatbots are very effective for both businesses and consumers. They already make it possible to answer many questions.
However, companies are now pushing further in the development of chatbots to be able to communicate on a human level with all its complexity.
Chatbots are useful for businesses not only when it comes toimprove customer experience and satisfaction but also to respond to the numerous employee questions.
Chatbots use 2 different types of approaches:
- NLU (Natural Language Understanding) to understand the meaning of the question.
- NLG (Natural Language Generation) to generate a response easily understandable by a human.
Spam filters
Email filters are a common use case for NLP. This feature helps block unwanted emails. These are identified as such by extracting the meaning and frequency of certain words in the body of an email.
These features are integrated into machine learning approaches or neural network algorithms, which will make it possible to improve thanks to user feedback.
All these features can be used in different fields of activity. This is what we will talk about in our next blog post. So stay tuned!
References