1 Introduction

1Since 2014, the “National Institute for Documentation, Innovation and Educational Research” (INDIRE) manages for the Ministry of Education (MIUR) the induction program of the Italian Newly Qualified Teachers (NQTs), i.e. the induction phase of teachers professional development that aims to support teachers in their transition from their initial teacher education into working life in schools. Experimented for the first time in 2014, it became effective starting in 2015 with the DM 850/2015.¹ The program involves all new hiring teachers from primary to secondary school for a total of 130,000 NQTs committed in the last 3 years. The underlying theoretical framework developed by INDIRE, MIUR and University of Macerata is based on the alternation of laboratorial and traditional classroom activities with documentation and reflection activities. The purpose is “to influence practices through a process that alternates between moments of immersion and distancing, which are actualised in When I teach and When I reconsider my teaching to think of what happened” (Magnoler et al., 2016). An on-line environment developed and managed by INDIRE² was set up to support teachers to reflect about and document their educational and professional activities (see Figure 1) during the induction program. All evidences of the instructional tasks (surveys, writing tasks, lesson plans, instructional materials, etc.) are collected in the e-portfolio and printed by the teachers for the final exam. An yearly monitoring of teachers activities is carried on by INDIRE to assess the effectiveness of the whole induction program, as well as of the single instructional tasks. It is aimed to modify, whenever needed, the program in order to improve stakeholders’ scaffolding to the newly qualified teachers and lastly teachers’ professional development.

Figure 1: The on-line environment collecting the e-portfolio of the newly qualified teachers

2In this paper, we report first results of an on-going study devoted to investigate the potentialities offered by Natural Language Processing methods and tools for the analysis of the NQTs e-portfolio. We consider in particular the documents written by the 26,526 teachers hired in the 2016/17 school year. Many protocols (or models) have been proposed to assess reflection in teachers writing, e.g. (Sparks-Langer et al., 1990; Hatton and Smith, 1995; Kember et al., 2008; Larrivee, 2008; Harland and Wondra, 2011). These models rely on features that suggest either different levels of reflection (means focused on the depth of reflection) or content of reflection (focused on the breadth of reflection), and usually they have found to mix features of both classes (depth and breadth) []. We rather focus here on the analysis of the form to study which are the main linguistic phenomena, distinguishing reflective from non reflective writings. Specifically, we devised a methodology devoted to investigate whether and to which extent a wide set of linguistic features automatically extracted from texts can be exploited to characterize NQTs’ reflective writings.

3Our contribution: i) we collect a corpus of reflective writings manually annotated by experts in the learning science domain and classified with respect to different types of reflectivity; ii) we detect a wide set of linguistically phenomena, characterizing the collected writings; iii) we report the first results of an automatic classification experiment to assess which features contribute more in the automatic prediction of reflexivity.

2 Defining reflection

4Within the teaching and teacher education domain, a very large amount of studies have been dedicated to conceptualization and analysis of teachers’ reflection and teachers’ reflective practice. Dewey (1933), Van Manen (1977), Schon (1984; Schon (1987; Schon (1991), Mezirow (1990) are among the main references. The attention on reflective thinking in the teachers’ education field has increased starting from the 80s as a reaction to the overlay technical view of teaching. Scholars have intensely studied reflection as a concept, detected more levels and types of reflection, how it works during and after professional teachers’ practice, its role and purpose in teachers’ professional development, and how it can be embedded in the curriculum of teachers preparation or professional development, and which techniques may be used to promote it (groups of discussion, readings, oral interview, action research projects, writing tasks, etc). In his seminal work “How we think”, Dewey provides the most shared definition of reflective thinking as applied in the educational field: reflection may be seen as an “active, persistent, and careful consideration of any belief or supposed form of knowledge in the light of the grounds that support it and the further conclusions to which tends”. Hence, reflection is a systematic process of thinking that happens only if related to actual experiences, and includes observation of conditions and references to different pieces of knowledge, (i.e. references to previous experiences, domain knowledge, common sense knowledge, etc.), in order to respond to a dilemma (Mezirow, 1990). Teachers’ educators have extensively employed writing tasks, such as writing structured or unstructured journals, portfolios, essays, blogs, open-ended questions to foster reflection both in pre-service and experienced teachers. Operational definitions of reflectivity proposed to develop schemes for assessing it are focused on identifying the presence of “reflective content” in teachers’ writing, or how deep the reflection is.

5Based on these premises, we are currently developing a reflection assessment schema suitable to describe properly the peculiarities of the Italian teachers’ reflective writings written in the framework of the 2016/17 induction program. The schema designed so far, reported in Table 1, was devised according to the following criteria: a writing is reflective if it i) makes direct references to experienced teaching activity, ii) involves several topics (content/pedagogical knowledge) and references to previous experiences, classroom management, learners needs, iii) includes premises analysis (theoretical, context-related, personal) iv) debates a problem (a dilemma), a doubt, v) has an output: it sums up what was learned, sketches future plans, gives a new insight and understanding for immediate or future actions.

3 The Corpus

6The corpus of NQTs reflective writings is part of the wider collection of documents written by the 26,526 teachers engaged in the 2016/17 INDIRE induction program. The whole corpus includes all texts written in two of the seven activities of the e-portfolio: Didactic Activity 1 and 2 (DA) for a total of 265,200 texts. During these two activities, teachers were supported by guiding questions designed by INDIRE experts to help them to understand the consistency of the planned and acted teaching activities. For DA 1 and 2 they wrote 5 short texts as answers to 5 different groups of questions. The first 4 groups provide guidance for teachers to write general reflections only on the design of their teaching activity; the fifth group is meant to guide NQTs towards an overall reflection on their whole teaching experience, i.e. both the design and the real teaching activity, also including classroom assessment techniques.

7We focused here on the answers to this latter group of questions that were devised in order to encourage teachers to reflect on the following issues: i) differences and similarities between the designed and achieved activities, ii) the most effective choices adopted, also including classroom assessment techniques, iii) how the activity could be improved, iv) the role played by the tutor and documentation practices. We considered in particular a subset of this group of answers that were annotated by 3 experts in the learning science domain according to the reflectivity annotation scheme described in Section 2 (see Table 2). The agreement between the three annotators was calculated using the Fleiss’ kappa test and we obtained a k=0.66, i.e. substantial agreement.

4 Linguistic Features and Reflectivity

8The annotated corpus was tagged by the part-of-speech tagger described in Dell’Orletta (2009) and dependency-parsed by the DeSR parser (Attardi et al., 2009). This allowed to extract a wide set of multilevel features, i.e. raw text, lexical, morpho-syntactic and syntactic, fully described by Dell’Orletta et al. (2013). They was used to reconstruct the linguistic profile of reflective writings and to carry out a first classification experiment aimed at predicting whether a text is reflective.

4.1 Distribution of Linguistic Features

9Table 3 shows a selection of the features that vary significantly i) between reflective and non-reflective answers (column Reflectivity) and ii) among the different types of reflectivity we considered (column Types of Reflectivity)³. The analysis of variance was computed in the first case using the Wilcoxon Rank-sum test for paired samples, while in the second case we used the Kruskal-Wallis test since we aimed to assess the different distribution of features in the 4 classes.

10In both cases, features from all levels of analysis resulted to be significant. If we consider the first ten most discriminative features, reflective writings resulted to be longer in terms of number of words and sentences, they are characterized by longer sentences and by a lower Type/Token Ratio; they contain an higher number of verbal heads and of embedded complement ‘chains’ (governed by a nominal head). Interestingly, they mostly contain linguistic phenomena typically related to syntactic complexity, for example they are characterized by i) an higher use of verbal modification (e.g. higher % of adverbs, of auxiliary and modal verbs), ii) more complex verbal predicate structures (e.g. higher average verbal arity, calculated as the number of instantiated dependency links sharing the same verbal head), iii) more extensive use of subordination (e.g. higher % of subordinate clauses also embedded in deep chains), iv) features related to a non canonical word order (e.g. higher % of pre-verbal objects and post-verbal subjects), v) longer dependency links and higher parse trees, two features related to sentence length. On the contrary, non reflective NQTs’ answers contain an higher level of lexical complexity: they have an higher Type/Token Ratio, a lower percentage of “Fundamental words”, i.e. very frequent words according to the classification proposed by DeMauro:2000 in the Basic Italian Vocabulary (BIV), and an higher percentage of “High usage words”.

11If we focus on the linguistic profile of the different types of reflective writings, we can observe that answers annotated as Reflection and Radical reflection are mostly characterized by features typically related to structural complexity. This is particular the case of Radical reflection answers that are longer in terms of number of sentences and words; they have more complex verbal predicates (e.g. an higher % of adverbs and of an implicit mood such as gerundive that can be more ambiguous with respect to the referential subject), more complex use of subordination (e.g. average length of ‘chains’ of embedded subordinate clauses), long distance constructions (length of dependency links), non canonical constructions (post-verbal subject). The higher % of demonstrative pronouns and determiners can be related to one of the most representative characteristic of reflection, i.e. the direct reference to real life. On the contrary, they contain a simpler use of lexicon, e.g. a lower Type/Token ratio and an higher percentage of “Fundamental words”.

4.2 Prediction of Reflectivity

12Table 4 reports the results of the automatic classification experiment we devised in order to predict whether a text is reflective. We built a classifier based on LIBLINEAR (Fan et al., 2008) as machine learning library trained using the LIBLINEAR L2-regularized L2-loss support vector classification function. We followed a 5-fold cross-validation process and relied on a training set of 370 answers balanced between the reflective and non reflective texts, since the under sampling technique has been proofed to improve classification performance on unbalanced datasets (Qazi and Raza, 2012). The performance was calculated in terms of F-score in the correct classification of non reflective (0 in the table) or of reflective (1) writings. We used different classification models: the Raw text one uses only raw text features, the Lexical one uses the distribution of the lexicon belonging to the Basic Italian Vocabulary and up to bi-grams of words, the Morpho-syntactic one uses the unigram of part-of-speech and verbal morphology features, the All features model uses all the considered features including the syntactic ones. A very competitive baseline was computed: it exploits the distribution of unigrams of words (Unigrams). As it can be seen, the model that uses all the considered features resulted to be the best one. On the contrary, the model relying on very simple types of features (raw text features) that capture how much teachers have written achieves the worst results. We also carried out a very preliminary experiment to classify the three different types of reflective writings but it produced unsatisfactory results due to the unbalanced distribution of answers in the reflective classes. As expected, a balanced experiment yielded very low accuracies since we used very few data.

5 Conclusions and current developments

13We reported first results of a on-going study devoted to reconstruct the linguistic profile of a corpus of reflective writings by Italian newly recruited teachers that we collected for the specific purpose of this paper. We are currently enlarging the corpus with new manually annotated data to improve the accuracy of the automatic classification of different types of reflectivity.