Automated assessment, including the use of Artificial Intelligence (AI), is one of the latest education tech solutions. It speeds up exam marking times, removes human biases, and is as accurate and at least as reliable as human examiners. As innovations go, this one is a real game-changer for teachers and students.
However, it has understandably been met with a lot of questions and sometimes scepticism in the ELT community – can computers really mark speaking and writing exams accurately?
The answer is a resounding yes. Students from all parts of the world already take AI graded tests. PTE Academic and Versant tests – for example – provide unbiased, fair and fast automated scoring for speaking and writing exams – irrespective of where the test takers live, or what their accent or gender is.
This article will explain the main processes involved in AI automated scoring and make the point that AI technologies are built on the foundations of consistent expert human judgements. So, let’s clear up the confusion around automated scoring and AI and take a look into how it can help teachers and students alike.
AI versus traditional automated scoring
First of all, let’s distinguish between traditional automated scoring and AI. When we talk about automated scoring, generally we mean scoring items that are either multiple choice or cloze items. You may have to reorder sentences, choose from a drop down list, insert a missing word- that sort of thing. These question types are designed to test particular skills and automated scoring ensures that they can be marked quickly and accurately every time.
While automatically scored items like these can be used to assess receptive skills such as listening and reading comprehension, they cannot mark the productive skills of writing and speaking. Every student response in writing and speaking items will be different, so how can computers mark them?
This is where AI comes in.
We hear a lot about how AI is increasingly being used in areas where there is a need to deal with large amounts of unstructured data, effectively and 100% accurately – like in medical diagnostics, for example. In language testing, AI uses specialized computer software to grade written and oral tests.
How AI is used to score speaking exams
The first step is to build an acoustic model for each language that can recognize speech and convert it into waveforms and text. While this technology used to be very unusual, most of our smartphones can do this now.
These acoustic models are then trained to score every single prompt or item on a test. We do this by using human expert raters to score the items first, using double marking. They score many hundreds of oral responses for each item, and these ‘Standards’ are then used to train the engine.
Next, we validate the trained engine by feeding in many more human marked items, and check that the machine scores are very highly correlated to the human scores. If this doesn’t happen for any item, we remove the item, as it is essential to match the standard set by human markers. We expect a correlation of between .95-.99. That means that tests will be marked between 95-99% exactly the same as human marked samples.
This is incredibly high compared to the reliably of human marked speaking tests. In essence, we use a group of highly expert human raters to train the AI engine, and then their standard is replicated time after time.
How AI is used to score writing exams
Our AI writing scoring uses a technology called latent semantic analysis. LSA is a natural language processing technique that can analyze and score writing, based on the meaning behind words – and not just their superficial characteristics.
Similarly to our speech recognition acoustic models, we first establish a language specific text recognition model. We feed a large amount of text into the system, and LSA uses artificial intelligence to learn the patterns of how words relate to each other and are used in, for example, the English language.
Once the language model has been established, we then train the engine to score every single written item on a test. As in speaking items, we do this by using human expert raters to score the items first, using double marking. They score many hundreds of written responses for each item, and these ‘Standards’ are then used to train the engine. We then validate the trained engine by feeding in many more human marked items, and check that the machine scores are very highly correlated to the human scores.
The benchmark is always the expert human scores. If our AI system doesn’t very closely match the scores given by human markers, we remove the item, as it is essential to match standard set by human markers.
AI’s ability to mark multiple traits
One of the challenges human markers face in scoring speaking and written items is assessing many traits on a single item. For example, when assessing and scoring speaking, they may need to give separate scores for content, fluency and pronunciation.
In written responses, markers may need to score a piece of writing for vocabulary, style and grammar. Effectively, they may need to mark every single item at least three times, and maybe more. However, once we have trained the AI systems on every trait score in speaking and writing, they can then mark items on any number of traits instantaneously – and without error.
AI’s lack of bias
A fundamental premise for any test is that there should be no advantage or disadvantage given to any candidate. In other words, there should be no positive or negative bias. This can be very difficult to achieve in human-marked speaking and written assessments. In fact, candidates often feel that they may have received a different score if someone else had heard them or read their work.
Our AI systems remove the issue of bias completely. This is done by ensuring our speaking and writing AI systems are trained on a very wide range of human accents and writing types.
We don’t want perfect native speaking accents or writing styles to train our engines. We use representative non-native samples from across the world. When we initially set up our AI systems for speaking and writing scoring, we trialled our items and trained our engines using millions of student responses, and continue to do this now as new items are developed.
The benefits of AI automated assessment
There is nothing wrong with hand-marking homework tests and exams. In fact, it is essential for teachers to get to know their students and provide personal feedback and advice. However, manually correcting hundreds of tests, daily or weekly can be repetitive, time-consuming, not always reliable and takes time away from working alongside students in the classroom. The use of AI in formative as well as summative assessments can increase assessed practice time for students and reduce the marking load for teachers.
Language learning takes time, lots of time to progress to high levels of proficiency. The blended use of AI can:
- Address the increasing importance of formative assessment to drive personalized learning and diagnostic assessment feedback
- Allow students to practice and get instant feedback inside and outside of allocated teaching time
- Address the issue of teacher workload
- Create a virtuous combination between human and machines, taking advantage of what humans do best and what machines do best.
- Provide fair, fast and unbiased summative assessment scores in high stakes testing.
We hope this article has answered a few burning questions about how AI is used to assess speaking and writing in our language tests. An interesting quote from Fei-Fei Li, Chief scientist at Google and Stanford Professor describes AI like this:
“I often tell my students not to be misled by the name ‘artificial intelligence’ — there is nothing artificial about it, A.I. is made by humans, intended to behave [like] humans and, ultimately, to impact human lives and human society.”
AI in formative and summative assessments will never replace the role of teachers. AI will support teachers, provide constant opportunities for students to improve, and provide a solution to slow, unreliable and often unfair high stakes assessments.
Examples of AI assessments in ELT
At Pearson we have developed a range of assessments using AI technology.
The Versant tests are a great tool to help establish language proficiency benchmarks in any school, organization or business. They are specifically designed for placement tests to determine the appropriate level for the learner.
Find out more about how the Versant tests can help you automate scoring processes and learn why it’s important to place university students in the right English program.
The Pearson Test of English Academic is aimed at those who need to prove their level of English for a university place, a job or a visa. It uses AI to score tests and results are available within five days.
Find out more on the PTE Academic website.
English Benchmark is also scored using the same automated assessment technology. This test, which is taken on a tablet, is aimed at young learners and takes the form of a fun, game-like test. Covering the skills of speaking, listening, reading and writing, it not only measures the student’s ability, but also gives suggestions on follow up activities and next teaching steps.
Read more in our article: Introducing English Benchmark: An innovative test for young learners.