1.6 Structure of the thesis
2.1.2 More than 30 years of ICALL
In their book Errors and Intelligence in Computer-Assisted Language Learning: Parsers and Pedagogues, Heift and Schulze (2007: p. 55–56) identify 119 NLP projects in CALL during the period between 1982 and 2004. Heift and Schulze record a total of 70 projects between the mid-1980s and the mid-1990s, and a total of 40 between the second half of the 1990s and the first of the 2000s.
Over the years NLP has been applied to language learning by developing (Heift and Schulze, 2007: Ch. 2, Schulze, 2010: p. 70–78):2
• So-called writer aid tools, which can help improve the quality of the learner’s written production even though they are not designed to learn a language. Among these, some of them concentrate on the correction of FL learner errors in non-restricted domains and others in the correction of errors in limited domains. Most of them are specialised in the target audience. In this group we find among others Gamon et al. (2009)’s research on the correction of errors made by learners of English, or ICICLE (Michaud and McCoy, 2006), a system that supports the learning of written English to signers of American Sign Language as a first language. Rimrott and Heift (2008)’s research is also interesting since it analyses how learner-specific tools perform compared to tools developed for native speakers (of German).
• Systems concentrating on the teaching and the learning of specialised grammat- ical phenomena, such as the use of adjectival endings, the use of morphology and syntax in noun phrase elements, word order in sentences, the use of (clitic, zero) pronouns, and so on. In this group, we find VERBCON (Bailin, 1990) and TDTDT (Pijls et al., 1987), which focus on the appropriate usage of verbs with respect to a selection of linguistic phenomena; SWIM (Zock, 1992), focusing on the use of clitics in French; and ALICE (Cerri, 1989) focusing on the use of temporal constructions in Italian, French and English.
In this group we can also include the only three systems that are still used in instruction settings today, ROBO-SENSEI, E-Tutor and TAGARELA (see Section 2.1.4) which are designed to be used in particular phases of a task-based approach to language learning with the goal to reinforce certain formal aspects of the learning experience (Schulze, 2010: p. 76–79).
• Systems focusing on the teaching of specific communicative competences to lan- guage learners. Among them we find a system to chat with the computer about one’s family, another about buying food in the market, or role-play activities to play spies or private investigators. In this group Schulze (2010: p. 70–73) includes FAMILIA, a system to chat about one’s family that pays attention to
2This list excludes applications of NLP for automatic scoring of learner essays, as well as tools
for the automatic annotation of learner corpora because strictly speaking they are not applications for learners to learn a language.
particular verb complement combinations (Weizenbaum, 1976); Spion (Sanders and Sanders, 1995) and Herr Komissar (DeSmedt, 1995), two systems that re- spectively use the spies and the private investigator domains to engage learners in a game-like conversation; the work by Menzel and Schr¨oder (1999), where learners state utterances related to a graphical market scenario; and FLUENT- 1 (Hamburger and Hashim, 1992) and FLUENT-2 (Schoelles and Hamburger, 1996), a graphical system in which learners could move objects in a particular micro-world (a bathroom) per request.
• Reading support tools, such as dictionaries or morphological analysers hyper- linked to reading texts, or links from the reading text to concordancers as a means to learn more about the usage of selected words. In this group we find GLOSSER RuG (Nerbonne et al., 1998; Roosmaa and Pr´osz´eky, 1998) and ELDIT (Knapp, 2004), two slightly different tools that assist learners in reading activities and vocabulary acquisition; and QucikAssist (Wood, 2009), a tool that allows learners to obtain linguistic and encyclopaedic information related to words in a text by clicking on them.
According to Heift and Schulze (Heift and Schulze, 2007: Ch. 2, Schulze, 2010: p. 70–72), ICALL systems with smaller coverage and less ambitious goals are the ones that commonly went beyond the prototype and reached the language learner. Examples of such systems are Spion (Sanders and Sanders, 1995), Herr Komissar (DeSmedt, 1995), GLOSSER RuG (Nerbonne et al., 1998; Roosmaa and Pr´osz´eky, 1998), ELDIT (Knapp, 2004), ROBO-SENSEI (Nagata, 2010), ETutor (Heift, 2010b) and TAGARELA (Amaral, 2007; Ziai, 2009; Amaral et al., 2011; Amaral and Meur- ers, 2011). More ambitious projects have yielded interesting results, but they usually end up not being used by learners: Two interesting examples are Textana (Schulze and Hamel, 1998; Schulze, 1998, 1999, 2001, 2003) and Freetext (L’Haire and Faltin, 2003; Granger, 2003; L’Haire, 2004).
Core research issues and influences from other disciplines
Over these 30 years of ICALL, different kinds of problems were approached, and different solutions attempted or adopted. The three issues most frequently tackled over the years are (i) the analysis of learner language, (ii) the appropriate strategies for the provision of feedback, and (iii) the adaptation of feedback to learners with different learning profiles and styles. We focus on the analysis of learner language in Chapter 3, where we introduce the key issues in Natural Language Processing for the purposes of this thesis.
As for the other two topics, Feedback and Student Modelling, we introduce the aspects that were most significant in ICALL according to Heift and Schulze (2007: Chs. 3 and 4). As for Feedback, it is generally understood as corrective feedback, and the main challenges in ICALL are to make feedback clear, comprehensible, as profitable as possible, and, of course, pedagogically grounded (Heift and Schulze, 2007: pp. 115–116). Heift and Schulze argue that (I)CALL research addressing the topic of feedback benefits from considering the general points of view of human- computer interaction (HCI), learning theories, second language acquisition theories,
and formal grammar (2007: p. 116). As we will see in Section 4.5, several studies have analysed how language learning is affected by the number of feedback messages, the wording, the inclusion of graphical highlighting, the grouping or filtering of corrective feedback depending on the relevance and the nature of the errors, the steps in which learners access different levels of feedback and the corresponding cognitive load – see also (Garrett, 1987; Pujol`a, 2001; Nagata, 1993, 1995, 1996).
As for Student Modelling, ICALL research is influenced by the research in user and student modelling in Artificial Intelligence. This influence contributed to a better understanding of how user information can be stored, what information is required in order for the student model to communicate with other modules in the system, and how the characterisation of the student as a learner can be updated over time according to his or her progress. ICALL practice has brought up topics discussed in student modelling such as the criteria to balance the weight of learner performance according to the learner’s developmental stage, the goals of the activity, the frequency of a particular error, and so on. Student Modelling falls out the scope of this thesis and will not be addressed in the following chapters. Further readings on this topic: Matthews (1992), Bull et al. (1995), Heift and Schulze (2007), Amaral and Meurers (2008).
Sustained research and development
An important aspect of ICALL research is the sustained use and development of ICALL systems. CALL systems, in general, are systems that are, should be, per- manently improved and adapted to the teacher and learner needs. In fact, those systems that present a sustained use and progress over the years are the ones that managed to be successfully integrated in real-world instruction settings – and they deserve particular attention (Levy, 1997: p. 13–14, Heift and Schulze, 2007: p. 9).
As for ICALL systems, researchers (Schulze, 2010: p.76–77, Amaral and Meur- ers, 2011: p. 3) agree that, there are only three systems that have been used for a sustained period of time in real-world instructions settings: ROBO-SENSEI, ETutor and TAGARELA. The reasons for this are fundamentally related to the difficulty of combining research and development (including use with learners), and the com- plexity of putting and keeping together cross-disciplinary teams (Antoniadis et al., 2004; Heift and Schulze, 2007; Schulze, 2008; Amaral and Meurers, 2011). Thus, re- phrasing our intial reference to Heift and Schulze (2007: p. 9), we claim that ICALL is a field of investigation that demands multidisciplinary research and sustained de- velopment.