Natural language processing (NLP) and computational linguistics in education involve the understanding of speech and text by software and algorithms to improve, scale, and broaden the reach of education in society.
You may be an educator or tech professional with an idea that improves the learning process for an individual — student, teacher, parent, or tutor — or streamlines a process in mass-market educational settings like universities, testing agencies, or government departments. Whatever your idea is, making yourself aware of the possibilities of NLP in education can help you refine the idea and even imagine entirely new ideas.
Pedagogy recognizes that the learning process is different for each student at any level, from early education to post-graduation. The pace of student learning is different for each student and subject. Some have special needs that require different teaching methods.
Although we know that personalized learning is ideal, unfortunately, there aren’t enough teachers or time to deliver it to each student. This is where your education idea can use NLP to provide a personalized learning environment.
Summarization and paraphrasing are two features you can include to help with the reading comprehension of any learning material you produce. Summarization conveys a concise version of a chapter, essay, or lecture containing just its high-level ideas and information. Paraphrasing conveys a summary using a different set of words chosen for their reading level, simplicity, accuracy, geographical variation, dialect, or cultural sensitivity. Both features are useful for specialized fields like medicine and law too where students have to process insane volumes of information in a limited time.
These are possible thanks to transformer neural networks. The transformer architecture is a type of deep neural network with a specialized set of neuron layers called attention layers that enable it to embed long-range context dependencies and higher-level intangible aspects like semantics, tone, and emotions.
Companies like Google and OpenAI have published pre-trained transformer models like BERT, GPT-3, RoBERTa, and T5 that are typically trained on web-scraping corpora. All human linguistic, cultural, and social norms that are visible on the web get latently embedded into these models.
Software libraries like spaCy and Hugging Face have made these pre-trained models ridiculously easy to use and tune.
Tasks like summarization are called downstream tasks because they start with one of these pre-trained models, add task-specific neurons to them, and retrain the entire assemblage using task-specific datasets. This process of adjusting a generic pre-trained model to perform a specific task is called transfer learning or fine-tuning.
A subject-specific language model is created by fine-tuning a pre-trained model on subject-specific datasets to learn their terms, lingo, and norms. For example, BioBERT is a biomedical NLP model trained using biomedical research papers. LEGAL-BERT is a model for legal NLP tools trained on legal case studies.
Summarization and paraphrasing start by running your learning material through a language model to obtain internal representations called embeddings. The tasks then add a sequence-to-sequence layer and fine-tune the entire thing using a summarizing or paraphrasing dataset.
Essay-type tests, quizzes, and flashcards are immensely popular among students and teachers but preparing them manually can be daunting to already overworked teachers.
NLP can help out here by generating questions and answers for your learning material. It can also evaluate how similar a student’s typed responses are to the information in the learning material. It can customize the questions as well as the assessment to a student’s personalized learning, like their reading level and learning pace. It can evaluate an essay on grammar, structure, semantics, and reasoning, a lifesaver for busy teachers.
A question-answering model is created on similar lines as a summarization model by adding a sequence-to-sequence layer to a pre-trained transformer model. For a question, which is the input sequence, it should output another sequence which is the answer. The assemblage of pre-trained base model and sequence-to-sequence layer is fine-tuned using subject-specific question-answer datasets.
This Hugging Face biology question-answering model demonstrates its capabilities by answering questions about prions. A cloud service like Amazon Kendra provides an alternate easier implementation path for question-answering.
Chatbots are yet another feature you can look into because they can act as robotic educators that explain subjects and answer questions through conversational NLP.
A speech recognition sequence-to-sequence model is created just like the one for question-answering. It first converts the student’s spoken questions or answers to text. A question-answering model then generates answers for those questions or evaluates the student’s replies. The results are spoken back to the student using text-to-speech synthesis.
Intangible aspects like enthusiasm, kindness, warmth, friendliness, and accent can be baked into the synthesized voice to help the student feel welcome and comfortable.
Cloud services like Amazon Lex and Google Dialogflow can be used to implement chatbots easily.
Writing on topics helps students distill their understanding. That’s why essay assignments and quizzes are popular assessment methodologies at all education levels. Feedback and prompts while writing can compel a student to explore a topic more deeply and broadly than they’d normally have done.
However, educators don’t always have the time to provide detailed feedback. NLP can help here with real-time and post-completion feedback for each and every piece of writing. It evaluates all the metrics that we humans evaluate. At the lowest level are things like spelling and grammar. Then we see sentence structure and readability. We next look into the correctness and logic of claims. In higher education fields like law and science, information and claims have to be presented using accurate words without ambiguities. Finally, we judge the structure according to accepted norms in that subject. Science papers are expected to have a particular structure. Legal writing has its own structure. NLP can evaluate all this in milliseconds in real-time while the student is writing.
If your idea is for organizations like testing agencies that evaluate a large number of students, you should definitely look into NLP to improve the quality of feedback and scale automated assessment. Even an individual student, parent, or teacher can benefit from educational applications that offer such feedback.
You saw how transformer-architecture models are used in the learning environment. It should come as no surprise that the same models are used for writing feedback and assessment too.
For automated essay scoring tasks, basic rules of spelling, grammar, and sentence structure are already embedded in the model. Given a fragment of text written by the student, you use a sequence-to-sequence model to paraphrase it. The output will automatically be grammatically correct. Your app can compare the model’s output against the written text, highlight differences, and give suggestions.
In higher education fields like science, each research paper builds on the shoulders of many past papers. The citations in a paper are not just for etiquette but also to form a pyramid of claims on which the new claims rest. The sentences in each paper are all semantically connected to one another and form a web of specialized knowledge, also known as a knowledge graph.
NLP techniques include graph transformers that are designed to process knowledge graphs for tasks like accuracy checking. Graph transformer networks can embed long-range dependencies between sentences that are quite distant on the knowledge graph. If a student makes a logical mistake or misunderstands the claims in another paper, your app can detect it in real-time and provide immediate feedback.
The third top use of NLP is in the niche and popular field of language learning.
At basic levels, language learning involves learning to read, write, and speak the words of a second language by associating them with words in your native language or with pictures of objects. More advanced language skills require immersing yourself in a language's script, literature, culture, and society.
If your idea produces learning material, it’s likely that you are targeting only English language learners. In a few cases, it’s probably translated manually to a handful of other languages. Targeting hundreds of languages is impractical but by not doing so, you are restricting your idea to a particular geography and probably missing out on good business opportunities. By incorporating NLP’s language translation capabilities, and making them fun through gamification, you can aid students and help teachers in many more geographies with minimal effort.
NLP techniques can also mimic some of the activities of advanced language learning:
A machine translation system is essentially a sequence-to-sequence model that, given a sentence in the first language, outputs a sentence in the second language.
But first, we have to capture the syntax and semantics of each language. This is done using the same pre-trained language models like BERT or T5 we saw earlier. Given a sentence in its respective language, each model outputs an internal representation — an embedding.
With these two language models, we then train a sequence-to-sequence model on a translation dataset like Wikipedia. On Wikipedia, the same articles are written in multiple languages by native speakers. They may not be perfect dictionary-to-dictionary translations of one another but they contain the same semantics. Semantics make their way into the embeddings. The sequence-to-sequence model learns to associate the embeddings of one language with the embeddings of the other language. Its output is a sentence in the second language.
You can use this approach in your idea to geographically broaden your customer base with minimal effort.
There seems to be a scarcity of artificial intelligence and natural language processing in education.
YCombinator’s startup directory shows only three companies are using machine learning to solve problems in the education system. That's unfortunate in a world where the student population is increasing but the teaching population is not keeping up. The gap can be bridged if individuals, companies, and governments in the education industry start using technologies like NLP for positive impacts.
If you have an idea like this, we can help you build it with a free consultation. Let’s teach the world a thing or two about learning.
