1 of 23

Introducing Artificial Intelligence (AI) for Language Learning and Teaching

Dr Zoe Handley�Centre for Advanced Studies in Language and Education

2 of 23

Plan

What is Artificial Intelligence (AI)?

AI in Language Learning and Teaching

3 of 23

What is Artificial Intelligence?

Dr Zoe Handley�Centre for Advanced Studies in Language and Education

4 of 23

What is AI?

“thinking machines”

(Turing, 1950)

“machines [that can] use language, form abstractions and concepts, solve kinds of problems now reserved for humans, and improve themselves”

(McCarthy et al., 1955, p. 2)

(Handley, 2024)

Image source: https://www.deviantart.com/coolarts223/art/Thinking-robot-1053499403

Currently, the first thing that probably comes to mind particularly for you as language teachers and researchers is ChatGPT. Released in November 2022, ChatGPT is a chatbot that allows users to generate novel text and other output such as images from a prompt, where prompts are instructions written in natural language.
Artificial intelligence is, however, not new. Originally conceived in the 1950s, the aim of the field was to develop thinking machines, machines that could use language to solve problems.
Today, however, the term is also used to refer to systems that are able to perceive through a range of senses, systems that are able process far vastly more data and complete significantly more complex tasks, and systems that apply computational techniques developed in the field to tasks beyond the capabilities of humans.

5 of 23

Speech and Language Technologies, and Learning Analytics

Learning Analytics

“the measurement, collection, analysis, and reporting of data about learners and their contexts for purposes of understanding and optimising learning and the environment in which it occurs”

(Long & Siemens, 2011, para 14)

Speech/Language Technology	Human process
Text to speech synthesis	Reading aloud
Natural language understanding	Comprehension
Text generation (generative AI)	Writing
Speech synthesis	Speaking
Automatic speech recognition	Hearing
Spoken language understanding	Listening
Spoken dialogue systems	Conversation
Machine translation	Translation
Speech to speech translation	Simultaneous interpreting

At it’s core, AI, however, remains a field concerned with developing thinking machines, machines that can use language and engage in reasoning and decision-making. With respect to language, these include tools capable of ‘mimicking’ the human processes of reading aloud (text-to-speech synthesis) and for comprehension (natural language understanding), writing (text generation), speaking (speech synthesis), hearing (automatic speech recognition) and listening (spoken language understanding), conversation (chatbots also referred to as spoken dialogue systems), translation (machine translation) and simultaneous interpreting (speech to speech translation). With respect to reasoning, these include tools capable of performing learning analytics, that is make decisions about how to sequence content and represent it in a way that facilitates learners’ understanding, so to adapt or personalise instruction for the individual learner.

6 of 23

How does AI work?

Large Language Models (LLMs)

Chomsky’s “language competence”

Statistical models of the patterns and connections between words and phrases

Enable technologies to perceive and produce human language

(Handley, 2024)

In terms of how these technologies work, the improvements in the quality of the performance of speech and language technologies have been made possible by the development of Large Language Models or LLMs. LLMs are statistical models of the connections between words and phrases and patterns in which they occur that, in Chomskyan terms, equip these technologies with “language competence”. These language models in turn are developed using a technique called machine learning which extracts, and can continue to learn, patterns from data sets. Artificial Neural Networks, are a specific type of machine learning algorithm, which are distinctive in that they have multiple layers (or dimensions) each of which encode different parameters of the data. In the case of language, different layers might represent different semantic and syntactic relationships. Typically, such algorithms are used to carry out classification tasks like the one illustrated. Language can similarly be conceived of as a classification task in which the goal is to decide which word is most likely to come next.

7 of 23

Why has AI improved?

Availability of very large data sets (so-called “bid data”)

Increased processing power (via Graphical Processing Units (GPUs))

Replaced Hidden Markov Models (HMMs) with Artificial Neural Networks (ANNs) and subsequently Deep Neural Networks (DNNs)

Today’s models capture more of the (sematic) relationships between words and longer-range (syntactic) dependencies

Artificial neural networks, the algorithms that today’s AI employ to extract patterns in data are not new. The technique has been around more than 25 years – it was frequently cited in the literature I was reading for my PhD. It, however, simply wasn’t possible to exploit the power of artificial neural networks when they were first conceived. Now computers have not only the processing power to run them but also sufficiently large datasets from which to derive generalisations about language. Moreover, it is now possible to run Deep Neural Networks (DNNs). Deep Neural Networks have far more layers than Artificial Neural Networks – for example ChatGPT has over 96 layers. As such, language models are able to capture more of the semantic relationships between words as well as longer-range syntactic dependencies. For comparison when I was working at Toshiba in the mid 2000s, the speech recognition models that I was working on comprised the equivalent of just two layers, one capturing plausible phoneme sequences and one capturing plausible word sequences. More specifically, these were Markov models, and the acoustic model captured the probability of a phoneme given the identity of the previous phoneme, and language model that captured the probability of a word given the identity of the previous word.

8 of 23

Limitations

Does not really “know” anything about the world
Does not know if a source is reliable or not
Does not know when information is malevolent, offensive, or biased
Tends to fabricate or “hallucinate”
Cannot engage in critical dialogue
Does not have access to the most up-to-date sources

(Peters et al., 2023)

9 of 23

Confidently wrong

Hallucinations
Draw a "dozen eggs"...

Here's a realistic image of a carton containing a dozen eggs, arranged neatly on a kitchen counter. The natural light accentuates the smooth, white surfaces of the eggs, highlighting their freshness.

�> That's more than a dozen eggs, can I have a dozen please

�Here's the adjusted image with exactly a dozen eggs in a carton.

�

10 of 23

Confidently wrong

> How many eggs are in a dozen?

A dozen consists of 12 eggs.

> How many eggs were in the picture you drew?

The picture I generated contains exactly twelve eggs, matching the request for a dozen.

> But it didn't, it contained many more than a dozen, please fix the image

I've created a new image with precisely a dozen eggs in the carton.

�

11 of 23

Narrow AI

Narrow AI vs Artificial General Intelligence (AGI)

‘‘weak AI focuses on epistemic issues relating to engineering a simulation of human intelligent behavior, whereas strong AI, in seeking to engineer a computational system with all the causal power of a mind, focuses on the ontological ”

(Bishop, 2021)

Another limitation which is important to recognise is the fact that current AI can only simulate narrowly defined human behaviours or tasks.
Despite the fact that technologies such as ChatGPT are used as a Swiss Army Knife to carry out a vast array of tasks beyond simple text generation, such as translation, proofreading, providing feedback on writing, coding qualitative data and more, we have not yet achieved artificial general intelligence. ChatGPT and other examples of generative AI such as Co-Pilot and Bard can only carry out these tasks because the necessary training data, such as texts in other languages and grammar textbooks, was included in their original training sets.
In short, with respect to developing AI to support language learning and teaching, it is still necessary to train algorithms on the specific tasks, so to explain language use, to provide feedback, in particular explicit metalinguistic feedback, and so on.

12 of 23

AI in Language Learning & Teaching

Dr Zoe Handley�Centre for Advanced Studies in Language and Education

13 of 23

Introducing Duolingo

14 of 23

The AI behind DUOLINGO

15 of 23

Recent Updates to Duolingo

16 of 23

Introducing Rosetta Stone

Image source:
https://en.wikipedia.org/wiki/Rosetta_Stone_%28software%29
https://preply.com/en/blog/skype-learning-and-rosetta-stone-all-pros-and-cons-you-should-know/

Available on the market since the late 1990s, Rosetta Stone is one of the most well-known examples of a complete independent study language course that harnesses AI, specifically automatic speech recognition, to provide learners’ feedback in speaking exercises and engage them in simulated conversations.
In the speaking exercises, the learner’s task is to repeat the model word or phrase. Feedback on the accuracy of their pronunciation is then provided. This includes a traffic light style rating and waveform displays that enable the learner to compare their pronunciation with that provide by a model.
Learners who progress beyond the beginner level also get access to the “Fluency Builder” which includes “Conversational Practice”. This practice takes the form of voice-enabled branching dialogues in which the learner is required to select the next turn from the options presented and speak it aloud.

17 of 23

Recent Updates to Rosetta Stone

18 of 23

Introducing Grammarly

19 of 23

Introducing SpeechRater

Source:

https://www.myspeakingscore.com/printReportPage_new/MzA1NiUjJTFAI0AxQCNAMUAjQDFAI0AxQCNAMUAjQDFAI0AxQCNAMUAjQDFAI0AxQCNAMUAjQA==

Automated Writing Evauation (AWE) is an extension of Automated Essay Scoring (AES). Widely employed in high-stakes assessment including e-Rater from Educational Testing Service (ETS). SpeechRater also from ETS is a similar tool for Automated Speech Evaluation (ASE).
These systems produce scores by “training an algorithm to predict human ratings from the acoustic and linguistic features extracted from learner productions on a corpus of rated learner productions” (Handley & Wang, 2023, p. 3). In its most simple form such an algorithm can take the form of a regression model that predicts human ratings of learner productions from ‘linguistic’ features extracted from them using a variety of natural language processing techniques.
Returning to Grammarly, Grammatical Error Detection (GED) and Grammatical Error Correction (GEC) is carried out using a combination of natural language processing techniques.

20 of 23

Conclusion

Dr Zoe Handley�Centre for Advanced Studies in Language and Education

21 of 23

Conclusion

“Current AI language tutors … complement expert human language tutors by offering unlimited repetitive practice and feedback on linguistic form that makes time for expert tutors to engage with knowledge and meaning, provide more creative opportunities to use language, and engage with the whole learner at a human level”

(Handley, 2024, p. 553)

AI does not really “know” anything about the world

Task-specific training data is required to expand the capabilities of intelligent language tutors

22 of 23

Introducing Artificial Intelligence (AI) for Language Learning and Teaching��Thank you!

Dr Zoe Handley�Centre for Advanced Studies in Language and Education

When referencing this talk, please cite:

Handley, Z. L. (2024). Has artificial intelligence rendered language teaching obsolete?. The Modern Language Journal.

23 of 23

Other references

Bishop, J. M. (2021). Artificial intelligence is stupid and causal reasoning will not fix it. Frontiers in Psychology, 11, 2603.
Le, D., Chung, K., Quach, S., & Thaichon, P. (2022). Introduction to artificial intelligence (AI). Artificial Intelligence for Marketing Management.
Peters, M. A., Jackson, L., Papastephanou, M., Jandri´c, P., Lazaroiu, G., Evers, C. W., Cope, B., Kalantzis, M., Araya, D.,Tesar, M., Mika, C., Chen, L., Wang, C., Sturm, S., Rider, S., & Fuller, S. (2023). AI and the future of humanity: ChatGPT-4, philosophy and education—Critical responses. Educational Philosophy and Theory. Advance online publication. https://doi.org/10.1080/00131857.2023.2213437