Little formal language learning was taught in ancient days. Only the elite learned the writing systems necessary for studying through a translation method. Yet millions of tradespeople, migrants, border dwellers, and travellers successfully acquired 2nd languages both then and continue to do so now through the twin engines of need and comprehensible input. And while instrumental orientation has been the prime motivating factor for most SLA learners throughout history, recent studies have shown that both instrumental and strong integrative orientation can be critical drivers of SLA (DuLay & Burt (1977), Krashen (1981, 1982), Ellis (2005), et. al).
Today, hundreds of millions of people study second languages yet fail to achieve speaking proficiency, mainly due to a lack of comprehensible input, proper orientation or both. We believe that AI can play a profound role in delivering these two critical components. This paper explores current pedagogical considerations revolving around these specific issues and examine how “AI” has been and might be implemented to best address each one.
An AI Primer
AI is a powerful technology has begun to engage an entire generation of new language learners. Its potential for enhancing SLA is huge, but so is its potential for being wasted. It’s worth examining its antecedents, current usage, and potential future as an SLA enhancement tool, within the language-learning ecosystem.
In 1950 Alan Turing postulated a test to determine the criteria for a truly “intelligent” computer. The “Turing Test”, as it came to be known, requires that the natural language responses from a smart machine and a human being be indistinguishable, in response to queries put to each by a third party.
The very first widely publicized instance of a computer program “passing” the Turing test , albeit in a limited set of circumstances, was the “ELIZA” “Doctor” program, written by Joseph Weizenbaum (1965). The program used Rogerian Psychotherapy interviewing protocols and offered them up in response to queries by simulated first-time patients. ELIZA’s ability to generate appropriate “open ended” responses to these queries gave “patients” the impression they were being interviewed by another human being
AI has evolved since then along many parallel and sometimes diverging tracks. Let’s take a look at a few.
One branch of AI has focused on game theory and has used machine learning, on a massive scale to analyze and improve outcomes for games requiring player moves. This is epitomized by Big Blue’s chess victory over Gary Kasparov, and, more recently, Google’s “Alpha Go’s” defeat of the world’s top “Go” player.
A slice of these technologies are used today in online and mobile gaming and in other AI-fueled applications, which examines current user data and behavior (e.g. player position and motion) and adjusts NPC (non-player-character) position, content or environment to move its characters in the most efficient way so as to target players, focus and shoot its weapons, or to communicate with players at critical points.
The use of Games, whether off-the-shelf (COTS) or “Serious” games desired for learning and MALL (Mobile Assisted Language Learning) for SLA has become increasingly popular during the past 10 years, especially since the advent of smart phones. Learners clearly gravitate towards games and game-like simulations and environments, because they offer high engagement and potential rewards. Thus, there is a powerful instrumental motivation for learners to understand the various messages, instructions, and textual content displayed during each game session, as it has direct impact on the outcome.
Various studies have provided evidence that Game-based and MALL learning improves outcomes in a variety of SLA areas (Suh, Kim, Kim, 2010). Meta-studies have also examined MALL Burston (2015) and delved into differences in performance based on game type (“Serious” games vs, MMPORG, COTS, etc.), Yudintseva (2015)
Reinders (2016) looked at how gamers process and use instructions in a 2nd language and how this affects the brain and comprehension. He points out that the areas of the brain involved with gaming and SLA have considerable overlap. This is hardly surprising in that communication is an important aspect of both sets of behavior. What’s interesting is that there may be opportunities to provide content that stimulates an area of the brain (through gaming) that also opens the participant up for SLA by providing the lowest possible affective filter.
Another branch of AI has focused on analyzing detectable visual, environmental, and behavioral events and looked for actionable patterns within vast arrays of data—for example, world events that cause financial markets to move in certain ways. Big data’s ability to analyze and model commonalities in the structure and usage of various language corpuses may soon have real world applicability for SLA. For example, it could help develop conversational models for high school students, for example by developing (based on frequency and semantic context for usage) a normative distribution of social media chat by high school students, thus providing them with an immersive L2 social chat experience.
NLP and Chat Bots
Natural Language Processing (NLP) applies a complex set of rules for contextualizing, sequencing, assigning meaning, and weighting to lexical input. In providing a machine with sets of data, in the form of sentences the goal is to produce appropriate human-like output.
To do that, the machine is trained to parse and organize lexical data. It first, typically, examines individual lexical entities for, within the sentence and then performs a syntactic analysis, in which words are parsed and extracted, and rules are developed for creating appropriate output. Semantic and sentiment analyses are also performed.
The goal of this training is to be able to provide appropriate responses for any particular sentence, so long as it has trained on this material. So, when “How are you feeling today” the machine should be able to respond with something like “I am feeling OK”.
Far from being “intelligent” in a human sense, ELIZA was basically a list search function that identified keyword patterns in user-generated content, and activated canned responses. For simple queries with known keywords, it provided “intelligent” responses.
NLP has vastly improved the ability of computers to parse and analyze lexical content in recent years, and to provide syntactically and, in many cases, semantically appropriate responses. Still, they don’t really “think”—they train for months on a specific corpus of text, and attempt to provide lexically and semantically acceptable responses. Furthermore, they can only parse unambiguous input. This causes errors. “I had a piece of pie” might be analyzed as meaning I had prior possession of the pie, based on the primary meaning of “had”. Real world experience tells us otherwise. Humans know and understand things related to meaning that are difficult to fully describe, and which limits NLP’s ability to act as a true human-like chat companion.
One of the more interesting recent L2 chat bot projects has been CSIEC’s Chat bot Lucy (Jia, 2009). Chat Bot Lucy is designed to engage in both general chat, within certain parameters and guided chat. The guided chat system initiates conversations with a human using a branching script. The system then predicts potential human responses to its queries and, responds appropriately when it correctly predicts the response or very close. For example, if Lucy asked “Do you want to order now?” in a restaurant simulation it might look for any 4 or 5 expected responses. Lucy also has a life-like avatar and a human voice, designed to create engagement. Lucy also gives give learners open-ended gap filling quizzes, which promote creative L2 thinking and make grammar and spelling corrections. In classroom tests, students that worked with Lucy showed large effect sizes.
Chat bots, along the lines of Lucy have recently been implemented on several language learning websites, in ways where their inherent weaknesses are avoided. Duolingo.com, for example, for uses bots in narrowly demarcated situations, such as booking a room. These bots function best when their interactions with humans are very predictable, and where they can receive sufficient training to respond appropriately to a variety of queries.
The threshold question for bots regarding their use for SLA is this. How much long-term engagement can they produce. We believe we may have found a way to leverage NLP in a manner that creates more coherent and profitable conversations than bot chat and, simultaneously, developer far higher engagement and motivation over the long term.
The nexus of AI and SLA Pedagogy
How and where to deploy AI
AI can be employed in fully autonomous situations, including games, interactive lessons, or unique exploration activities that customize the experience based on user data or real time user behavior, be it click patterns, text queries, or voice. It can be used to:
Vary the visual, auditory, tactile, etc. experience of the user or its rules and/or content
Provide feedback based on individual user performance
Engage users with system or with each other, through conversation/interaction
Help the user to find information, images, or other content
Considering these capabilities one can imagine a whole host of games, social, exploration, and learning activities, and search-related activities designed to enhance SLA that might profitably incorporate AI. The very nature of AI itself—the ability to present itself as a thinking entity to a human, lends itself to a communicative approach. AI’s greatest asset, as it relates to SLA, is its ability to deliver engaging, low-stress, comprehensible conversation or exciting exploratory or gaming experiences conjoined with “mission critical” L2 messaging.
This squares well with Ellis’ (2004), who notes that:
“Given that it is implicit knowledge that underlies the ability to communicate fluently and confidently in an L2, it is this type of knowledge that should be the goal of any instructional programme….There is consensus that learners need the opportunity to participate in communicative activity to develop implicit knowledge.” (p. 36)
We think the best and highest use of AI is to provide what’s most lacking in an L2 teacher’s arsenal—extensive, quality L2 interaction, built around implicit learning, within a communicative context.
Language Hero Smart Chat
Intro and description of Smart Chat
The authors of this paper are both working to develop a new language-learning App called Language Hero, an AI-fueled video chat language practice app, which empowers students, who can’t speak each other’s language to have comfortable, comprehensible, varied conversations in it, leading to speaking proficiency.
L2 learners select from pre-translated content, which resides in a database. They listen to a corresponding audio clip, and repeat it. Their L2 speaking partner corrects pronunciation and chooses/speaks new content. This back and forth process creates a conversation “tree”, with ever branching, comprehensible content, fueling millions of potential low stress conversations. Learners may also select and build sentences via selecting from image strips or word lists, which gloss/reinforce keyword choices and processing.
Core loop for “Smart Chat”:
The challenge for Smart Chat is to keep users motivated and engaged. Simply reading dialogs degrades engagement rapidly. To insure the most natural conversations possible, we use unique coding architecture and AI that examines user data and behavior and syntactic and semantic content to provide interesting/engaging and somewhat unpredictable content. Smart Chat empowers learners to direct each conversation into areas of interest in order to create a strong sense of real-life communication.
Because Smart Chat competes for our learners’ attention with a plethora of autonomous and social apps, we try to create strong partner bonds through a dating style search function, that matches them by speaking proficiency, personal attributes, and interests.
To enhance motivation, learners take turns engaging in 10–minute chat sessions in their respective target languages, helping each other with pronunciation. Short sessions maintain focus, while helping each other creates stronger motivation.
Smart chat further empowers schools to upload their own curriculum, and combine it, where desired with Smart Chat’s conversation-oriented modules, making teachers and department heads stakeholders in the learning process. Smart Chat can be used either to practice conversational speaking or to focus on form implicitly. It further allows explorations of custom topics and modules, and provides multiple levels of difficulty, providing new lexical and grammatical structures, all while having a great chat.
- Eight levels of difficulty.
- Nine modules on relevant, useful topics, such as travel, activities, food, and social chat.
- Partners can access content in any module during chat through the “Explore” function.
- Pinyin on or off (when off, replaced by Traditional or Simplified Chinese characters)
- Auto-play of Audio clips on or off
- Supports both video and text messaging
Smart Chat content is entered and managed via a Content Management System (CMS). Sentences may be sequenced to build logical semantic relationships with other related sentences. This is done through establishing parent and/or child relationships and affinity ranking between sentences, creating a realistic branching content flow. Admin also controls which topics or modules content can appear in.
Smart Chat’s coding architecture further empowers learners to move through topics quickly or dive deeply into them. Semi-random content selection adds an element of unpredictability to each chat. Lastly, while there is no conscious focus on form, all conversation follows proper grammatical rules, which allows for gradual layering of later-acquired forms, as learners move to higher levels.
In short, Smart Chat delivers content that engages and motivates users and aligns with their priorities and preferences. It insures a good partner match, encourages partners to help each other, keeps each session short enough to maintain focus, and provides content that enhances engagement, as both learners explore a variety of real life topics together.
Smart Chat and relevant studies
Smart Chat is unique among video chat formats in its reliance on user selected and spoken pre-translated content. While this may be a novel method for video chat, various studies have found incidental acquisition in L2 while viewing L1 subtitles; e.g., Koolstra, C. & Beentjes, W. J. (1999). Studies have also shown how glossed words in L2 readings, containing analogous known L2 words, or L1 words or hints contribute to processing of the L2 target word, Plonsky and Ziegler (2016).
Lastly, a considerable body of evidence that shows that new technologies focused mainly on just speaking or listening have the ability to enhance SLA, whether synchronous or asynchronous (Payne and Whitney, 2002), (Hirotani, M. 2009).
Smart Chat vs Chat Bots
Real life conversation, without pressure to produce, and motivated by instrumental or integrative orientation produces the lowest affective filter (Krashen, Dulay & Burt, above). Chat bots currently provide a narrowly circumscribed series of questions or responses within specific topics. They can play the role of taxi drivers, hotel clerks, etc. and may be quite useful for practicing practical tasks, even to the point of generating some SLA.
It’s unlikely that any currently available language learning bots can provide enough real life input on its own to produce the long-term engagement and motivation. Until NLP drastically improves its ability to parse semantic content and create convincing human-like responses, its usefulness as an “intelligent tutor” will be limited. Bots may be a great place to start the journey of learning a new language, but we have doubts about their ability to single-handedly take anyone to the Promised Land.
Smart Chat aims to do just one thing well—provide enough real life, on-the-fly content to keep learners hearing and using their L2 with each other for as much time as is needed for them to achieve speaking proficiency. Our efforts are geared towards finding ways to create high levels of engagement and motivation, coupled with a large enough corpus of interesting, varied content to sustain months of conversation.
Best AI use cases for now and the future
We believe that the best use of AI for SLA, as it applies to beginning students, at least, revolves around two critical factors—a format that leverages students’ openness and interest in learning the language and maintains it over the long haul, coupled with enough comprehensible input to stimulate real proficiency, irrespective of how it’s delivered.
While the conditions for SLA are affected by many variables, including age, stress, individual aptitude, types of interactions with others, and content, the first two conditions must always be present for successful acquisition to occur. With that in mind let’s examine the best use cases for currently available AI-driven technologies.
In short, AI’s role in SLA within a gaming or interactive environment is reflected in the exploration, excitement, motivation and engagement it generates through its human-like responses to user behaviors, thereby enhancing its users’ focus on the game’s (L2) messaging in ways they don’t in the classroom or when doing homework or reviewing online materials. Its ability to generate strong instrumental orientation to extract meaning from “mission critical” L2 messaging (to score more points, etc.) is nearly unparalleled. “Serious” games, where properly configured and professionally rendered, can produce high engagement levels, just as commercial ones do.
We should be thinking of new ways to better harness AI’s ability to mimic human behavior, within narrowly circumscribed bounds, to create far more sophisticated and complex L2 interaction between game and player. Voice recognition can (and should) be built into new serious games, in which the system either responds to or prompts the learner to enter voice commands or other output at various points to open weapons caches or doors or drawers, or select red or green scarves, etc. or where “wizards” pop up at various points to ask questions. This approach avoids the huge investment in data science necessary to avoid errors in responding to even simple user queries that we still find in NLP (bots).
Bots and Intelligent tutors
Bot chat is effective where the system can implement a standardized protocol with a large enough corpus of relevant text to train on and negotiate meaning with users. If the subject is booking a hotel room, there’s a clear protocol and the system can accurately predict most responses. Bots are still error prone outside of their comfort zone. As bots become better at parsing meaning from conversational flow and context there’s little doubt that they will eventually be able to have more free-ranging and productive and perhaps engaging conversations with learners. That time has not yet come.
AI (intelligent tutors) are effective for activities that focus on form. Expert at applying rules and correcting errors through explicit correction, recasts, or model building, intelligent tutors can also vary exercise difficulty based on test proficiency. Perhaps it’s worth considering how such a bot can provide additional L2 input that adds communicative value to these explicit rule tasks, while correctly demonstrating the same rules.
Peer-to-peer “smart chat”
We think that any technology that brings L2 learners together with native (or fluent) speakers of that language and keeps them engaged over the long haul is worth developing. After all, this is the closest approximation of the method for L2 that has proven itself over thousands of years of human interaction. Smart Chat is one approach to achieving this goal. We hope to see other projects that explore this area and how best to utilize the power of AI to enhance outcomes in it.
Burston, J. (2015). Twenty years of MALL project implementation: A meta-analysis of learning outcomes. ReCALL, 27(1): 4–20.
Dulay, H. and Burt, M. (1974) Natural sequences in child second language acquisition. Language Learning 24: 37-53.
Ellis, R. (2005), Instructed Second Language Acquisition: A Literature Review. Wellington: Ministry of Education. p. 35
Hirotani, M., (2009). Synchronous Versus Asynchronous CMC and Transfer to Japanese Oral Performance, CALICO Journal, 26(2): 413-438.
Jia, J., (2009) An AI Framework to Teach English as a Foreign Language: CSIEC. AI Magazine, Summer 2009: 59–71.
Koolstra, C. & Beentjes, W. J. (1999). Children’s vocabulary acquisition in a foreign language through watching subtitled television programmes at home. Educational Technology Research & Development 47: 51-60.
Krashen, S. (1981) Second Language Acquisition and Second Language Learning. Oxford: Pergamon Press.
Krashen, S. (1982) Principles and practice in second language acquisition. S Krashen. Oxford
Payne, J. S. and Whitney, P. J. (2002) ‘Developing L2 oral proficiency through synchronous CMC: output, working memory, and interlanguage development’, CALICO Journal, 20(1): 7–32.
Plonsky, L., Ziegler, N. (2016). The CALL–SLA Interface: Insights From A Second Order Synthesis. Language Learning & Technology, 20(2): 17–37.
Reinders, H. (Ed.) (2012). Digital games in language learning and teaching. New York: Palgrave Macmillan: 95-114
Suh, SW Kim, NJ Kim (2010) – Journal of computer assisted learning, 26 (5): 370–378
Weizenbaum, J (1966) “ELIZA: A computer program for the study of natural language communication between man and machine,” Communication of the ACM, 9(1): 35-36.
Yudintseva, A.(2015) Game-Enhanced Second Language Vocabulary Acquisition Strategies: A Systematic Review. Open Journal of Social Sciences, 3: 101-109. http://dx.doi.org/10.4236/jss.2015.310015