Modeling infant segmentation of two morphologically diverse languages

Modeling infant segmentation of two morphologically diverse languages Georgia Rengina Loukatou1 Sabine Stoll 2 Damian Blasi2 Alejandrina Cristia1 (1) LSCP, Département d’études cognitives, ENS, EHESS, CNRS, PSL Research University, Paris, France (2) University of Zurich, Zurich, Switzerland georgia.loukatou@ens.fr RÉSUMÉ Les nourrissons doivent trouver des limites de mots dans le flux continu de la parole. De nombreuses études computationnelles étudient de tels mécanismes. Cependant, la majorité d’entre elles se sont concentrées sur l’anglais, une langue morphologiquement simple et qui rend la tâche de segmen- tation aisée. Les langues polysynthétiques - pour lesquelles chaque mot est composé de plusieurs morphèmes - peuvent présenter des difficultés supplémentaires lors de la segmentation. De plus, le mot est considéré comme la cible de la segmentation, mais il est possible que les nourrissons segmentent des morphèmes et non pas des mots. Notre étude se concentre sur deux langues ayant des structures morphologiques différentes, le chintang et le japonais. Trois algorithmes de segmentation conceptuellement variés sont évalués sur des représentations de mots et de morphèmes. L’évaluation de ces algorithmes nous mène à tirer plusieurs conclusions. Le modèle lexical est le plus performant, notamment lorsqu’on considère les morphèmes et non pas les mots. De plus, en faisant varier leur évaluation en fonction de la langue, le japonais nous apporte de meilleurs résultats. ABSTRACT A rich literature explores unsupervised segmentation algorithms infants could use to parse their input, mainly focusing on English, an analytic language where word, morpheme, and syllable boundaries often coincide. Synthetic languages, where words are multi-morphemic, may present unique diffi- culties for segmentation. Our study tests corpora of two languages selected to differ in the extent of complexity of their morphological structure, Chintang and Japanese. We use three conceptually diverse word segmentation algorithms and we evaluate them on both word- and morpheme-level representations. As predicted, results for the simpler Japanese are better than those for the more complex Chintang. However, the difference is small compared to the effect of the algorithm (with the lexical algorithm outperforming sub-lexical ones) and the level (scores were lower when evaluating on words versus morphemes). There are also important interactions between language, model, and evaluation level, which ought to be considered in future work. MOTS-CLÉS : variation interlinguistique, apprentissage statistique, segmentation des mots, acqui- sition du langage. KEYWORDS: cross-linguistic variation, statistical learning, word segmentation, language acquisi- tion. 1 Introduction Human infants are known to acquire a comprehension vocabulary of hundreds of words by two years of age (Hoff, 2013), and probably accumulate a protolexicon consisting solely of word forms, with no meaning attached, by the end of the first year (Ngon et al., 2013). Infants’ discovery of basic units in their input has been modeled as follows. A transcript of infant-directed speech is converted into phonological text, word boundaries are removed, and an algorithm is applied. Then, the word boundaries posited by the algorithm (and the resulting word tokens) are compared against the adult segmentation found in the original corpus. By and large, two classes of algorithms have been heavily studied …

Content truncated to 500 words