smartlearning logo

AI Writers in Language Learning

Af Henrik Køhler Simonsen fra SmartLearning

Denne artikel blev første gang præsenteret ved konferencen: IEEE 2st International Conference on Advanced Learning Technologies i juli 2021, samt publiceret i en udgivelse af samme navn.

Abstract—Artificial intelligence (AI) is fast becoming a key instrument in several fields and is also transforming the way we communicate. Consequently, we need to understand the role of AI and augmented writing (AW) in language learning (L2 text production) and to understand how AI writers and language learners work together. The objective of this paper is to examine selected AI writers and discuss theoretical considerations on how AI writers can enhance language learning. The paper draws on data from a structured test and analysis of 39 different AI writers [7, 8, 9] and small experiments in different classes at a Danish university. One aspect from the data was that students found that AI writers were useful, but that they need to learn how to work with an AI writer when producing L2 texts. In conclusion, AI writing technologies seem to be highly relevant in language classes just as the calculator was in math classes 50 years ago.

 

I.     Introduction

Within the field of technology-enhanced language learning we need to understand how humans and AI work together [1] and, in particular, how humans may use AI in language learning [4, 5, 7, 8 and 9]. We need AI in language learning and we need to utilize the way AI can complete or augment tasks and at the same time learn from the AI writer. AI writers will not make language teachers obsolete; they will make their jobs easier and more interesting.

AI is increasingly getting smarter [2, 6 and 10] and some AI writers not only allow users to prompt them with specific sentences but also allow users to edit the AI-generated text during and after text generation. Therefore, we need to decide in advance, who or what is responsible for which tasks. Language learners need to learn how to effectively prompt or feed the AI and learn to be critical about the AI-generated content and learn how to edit the text before, during and after. Using AI will be preparing future generations for learning with AI or learning for AI [10].

The question is how do we integrate AI technologies in language classes?

 

II.     Research questions

This paper focuses on AI in language learning, more specifically on L2 text production in higher education and presents three models. The research objectives of the paper are to present selected cases of AI writers and to develop models for using AI writers in language learning.

 

III.    Method, data and delimitations

This discussion draws on empirical data from a structured test and analysis of 39 different AI and AW technologies [7]. Five cases have been selected to illustrate the applicability of the models developed and are shown in Table 1 below. The paper focuses on language learning in higher education and on L2 text production, but the models developed may be applied in many other language learning situations. The paper only focuses on AI writers, which in [6] are defined, as “the next evolution will be about leveraging smart tools that can help journalists augment their own writing. This means AI-powered interfaces capable of providing context to topics in real time”.

 

IV.     Five examples of AI writers

 

AI Writers in L2 Text Production
Name of AI Writer Human Prompt  Training and Editing
Sassbook AI Writer Yes – topic selector

Yes - prompt of 15-30 words available

Yes – mid-training of AI possible

Yes – mid- and post-editing possible

Zyro AI Content Generator Yes – topic category and sub-category

Yes - prompt of 2-3 sentences

No – mid-training of AI not possible

No – mid-editing not possible

AX Semantics Yes – topic domain selector

Yes – upload of own data

Yes – pre- and mid-training possible

No – mid-editing not possible

Jenni Yes – target search query

No – selection and paraphrasing of existing content

Yes – scrapes for new texts

Yes – partly manual mid-editing possible

KafKai Yes – niche article or general article selector

Yes – paragraph of text

Yes – mid-training is possible through seeding

Yes – manual mid-editing possible


Table 1: Five examples of AI Writers

The structured test and analysis of the five selected AI writers listed in Table 1 show that the five AI writers in varying degrees allow for human prompting or seeding and mid-training or mid-editing. Some of the AI writers even let you upload your own texts to build your own data model. All examples enable the student to post-edit the text.


V.     Theoretical discussion

The starting point of this discussion is taken in [5], who devised the DIKIW model. The DIKIW model discusses the concepts of data, information, knowledge, intelligence and wisdom and can to some extent be used to illustrate the challenges of working with an AI. Figure 1 below shows how these concepts are divided (dotted lines and vertical text are my additions).

HKS AI writer Fig 1 DIKIW

Fig. 1.    DIKIW (with my additions)

As Figure 1 illustrates, data and information is already abundantly accessible and is automatically scraped by AI Writers. Steps 3 and 4 in the DIKIW model are to some extent generated by the AI writer, but the output of most AI writers need to be quality-assured or augmented by humans. This is what many AI writers do by providing user prompts or mid-editing and post-editing functionalities.

Human-only text production processes are a black box and are characterized by several steps, which the human writer goes through. Every human writer is unique and it is difficult to describe precisely the human writing behaviour. However, research shows that, even though writing comes naturally and is a highly dynamic process for most of us, writing can be divided into a number of steps or phases see also [1]. Typically, there is a prewriting phase, a writing phase, a revision phase, an editing phrase and a publication phase. All this needs to be reconsidered in an era where students work with AI writers and exploit the collaborative intelligence of both humans and AIs [10]. Many of the human-only writing phases are still valid, but we need to realize that AI writers can help students and professionals write repetitive parts of texts, and perhaps most importantly, help users overcome writer’s block. Language students should focus on steps 3 and 4 in the DIKIW model and offer world knowledge to the text. Working with an AI in L2 text production processes is a challenging but rewarding task so the question is how do we implement that in our classes?

The models developed and presented below are based on the assumption that the student should first work with the AI alone, then work with the AI together with a peer, and finally consult the teacher. Such an approach presupposes that learning objectives prepare students for higher order thinking.

Building on the data discussed in [7] and on insights from experiments in text production classes with AI writers, the following models for how to use AI writers in L2 text production classes are presented. Clearly, the student should be instructed in advance in how to prompt or seed the AI, in how to reprompt or reseed the AI and finally in how to edit the text generated by the AI. This is a difficult phase for the student, because how can the student spot false friends. Therefore, future language learning using AI writers should involve a focused effort on teaching students to plan the structure of the text and to post-edit the text, see also [1].

Figure 2 below thus illustrates how a student can work with an AI on his own.

HKS AI writer Fig 2 level 1

Fig. 2.    Level 1 (Student and AI)

Figure 3 below shows how a student proceeds. First, the student interacts with a peer to discuss text version C of his text. After having consulted with a peer student the AI is then reprompted and an entirely new text is generated. The student then post-edits the peer student-backed and reprompted AI-generated text and creates a new version of the text version E.

HKS AI writer Fig 3 level 2

Fig. 3    Level 2 (Student, Peer and AI)

Not until the student has worked with the L2 text, production task on his own with an AI writer and having discussed the text with a peer, the teacher is involved. Figure 4 below illustrates how that process might take place. The student discusses the significantly improved text with a teacher and the teacher offers insights based on his experience and world knowledge. This enables the student to reprompt the AI, which then generates a new version of the text. As in levels 1 and 2, the student post-edits the text and ends up with text version G.

HKS AI writer Fig 4 level 3

Fig. 4    Level 3 (Student, Teacher and AI)

So what are the implications of such an approach? As pointed out above, working with an AI demands a lot from the students as they will have to spend much more time on high-cognitive processes such as pre-editing, mid-editing and post-editing texts. Such an approach is to some extent already suggested by [4], which discusses a model for post-editing Google Translate output. Learning languages with an AI to some extent changes what should be focused on in language teaching. Up until now, we have mainly focused on the student’s ability to write correct and coherent texts in text production classes, but we will have to change that focus in our future curricula to prepare students for an AI-intensive world. In other words, the learning objectives of our courses need to be changed as the curricula need to be updated to include texts and exercises on how to work with an AI and how to evaluate the AI-generated content. Therefore, we should focus on teaching our students to plan their texts, to spot the different types of errors that AI writers make and post-edit these texts.

So what are the practical applications of AI writers in L2 language learning? Based on the cases described above and a number of experiments with different AI writers and AW technologies in different language classes, it is argued that AI writers already now can be used when teaching SoMe content writing, blog content writing, marketing content writing, website content writing and thesis writing. This list shows how AI writers can be used to enhance language learning in both professional and academic discourses. When teaching language for professional purposes with AI writers it might be suitable to focus on how to write repetitive content for different SoMe channels, on how to write creative marketing texts for both offline and online channels and on how to write online content for websites etc. However, AI writers may also be used in more general language learning classes to enhance academic writing including academic report writing or thesis writing where AI writers both may offer automated content inspiration and content creation. Finally, it is also argued that AI writers could be used when teaching students grammar, spelling and textual elements such as coherence and cohesion, which in fact, are very important when working with an AI in language learning classes.

VI.     Conclusions and perspectives

This paper discussed AI writers in language learning and found that AI writers may augment language. The analysis of the AI writers showed that students to a very high extent could prompt AI writers and in some instances even train the AI with specific texts. The analysis also showed that students could dynamically edit the AI-generated text both during and after. Admittedly, the AI writer does require a lot from the language student, but it is argued that an AI writer significantly enhances language learning. The empirical data and the theoretical discussion led to the formulation of three models that can be used when designing AI-based language classes with a special focus on text production. The models are based on a three-level approach. The first model can be used by the student and the AI writer alone to produce an L2 text. The second model can be used by the student, a peer and the AI and is thus based on a peer-to-peer approach. The third model involves the student, the teacher and the AI and is designed to be used for high-cognitive tasks or when the student has already used models 1 and 2. AI writers in language learning not only enhance the student’s ability to edit a text before, during and after content generation but also prepare the student for a labour market with AI where the ability to critically reflect on and assess AI-generated content is crucial. AI writers are improving and their impact on language learning is already here. However, there is still much to do. We need to explore human and virtual tutors and AI writers and Intelligent Tutoring Systems and AI etc.

References:

[1]     Abas, I., Abd A., and N. Hashima, “Model of the Writing Process and Strategies of EFL Proficient Student Writers: A Case Study of Indonesian Learners”, Pertanika Journal of Social Science and Humanities. 26, 2018, pp.1815-1842.

[2]     Banks, C., “What is an Augmented Writing Platform?” 2019, in: https://medium.com/swlh/what-is-an-augmented-writing-platform-b28fa588a1c5.

[3]     Colson, E., “What AI-Driven Decision Making Looks Like” 2019, in: https://hbr.org/2019/07/what-ai-driven-decision-making-looks-like.

[4]     Leroyer, P. and Simonsen, H. K., “Google Translate som trussel eller redning for oversættelsesordbøger”, LexicoNordica 26, 2019.

[5]     Liew, A., “DIKIW: Data, Information, Knowledge, Intelligence, Wisdom and their Interrelationships”, Business Management Dynamics. Vol. 2, Issue 10, April 2013, 49-62.

[6]     Marconi, F. “NiemanLab. Predictions for Journalism (2017): The Year of Augmented Writing”, 2017 in https://www.niemanlab.org/2016/12/the-year-of-augmented-writing

[7]     Simonsen, H. K., “Augmented Writing Needs Lexicography”, Gavriilidou, Z, Mitsiaki, M, Fliatouras, A., 2020, Proceedings of XIX EURALEX Congress: Lexicography for Inclusion, Vol. I, Democritus University of Thrace, pp. 509-514.

[8]     Simonsen, H. K., ”Når Augmented Writing og leksikografi går hånd i hånd”, LEDA-nyt nr. 69 - april 2020, pp. 3-13.

[9]     Simonsen, H. K., ”Augmented Writing: nye muligheder og nye teorier”, Nordiska Studier i Lexikografi 15, 2020, Rapport från 15 konferensen om lexikografi i Norden – Helsingfors 4–7 juni 2019, pp. 307-315.

[10]   Zandan, N., “The Future of Human Communication: How Artificial Intelligence Will Transform the Way We Communicate”, 1997, https://www.quantifiedcommunications.com/blog/artificial-intelligence-in-communication