L1 Personalized Lexical Complexity Prediction (LCP) Dataset for German-Language Learners

Psycholinguistics | Second Language Acquisition (SLA) | Ontology | Semantic Web Technologies | Statistical Modelling

This project conducted an annotation task to create an LCP dataset for German-language learners. It aims to assess how word difficulty varies among individuals from language backgrounds that share roots and commonalities with the target language and those that do not. The dataset can be used to train a contextualized LCP model with personalized features.

Description of data

Unstructured continuous text from various domains was selected, with specific target words chosen from different levels of German texts. The dataset aims to collect features that can be analyzed for correlation with respect to the complexity of target words. For this purpose, various learner-specific (depending on the learner performing the annotation) and language-specific (based on the target language) features were accounted for. A complete list of all the different learn-specific and language-specific features are listed below.

Learner-specific features

Feature	Description
speaker_id	Individual speaker id
german_level	The current German of the participant in the CEFR scale
seen_before	Previous exposure to the word
country_of_origin	Country from which speaker comes from
native_language	Native language of speaker
background	Professional background of the participant
years_in_germany	Number of years the participant has lived in Germany

Language-specific features

Feature	Description
domain	Dataset from which the text is from
msy_category	Morphosyntactic category of the target word
en_cognate	Cognate value depending on whether it’s an English cognate or not
es_cognate	If there’s a Spanish cognate or not
es_ff	If there’s a false friend equivalent of the target word in Spanish
semantic_sim	Semantic similarity of the target word with the other words in the corpus
length	Length of the presented text in number of words
dependency_depth	The syntactic dependency relation length of the presented target word
word_frequency	Using wordfreq, determining the frequency of the target word
child_lex_frequency	Binary probability of the target word existing in the ChildLex frequency dataset (Schroeder et al. (2015)¹)

Generating word difficulty scores

To generate complexity scores, for different target words for each participant, the GermaNet ontology was used to identify semantic distances and similarity of perceived meanings of different target words.

References