Dataset Information

Using Morphological Data in Language Modeling for Serbian Large Vocabulary Speech Recognition.

ABSTRACT: Serbian is in a group of highly inflective and morphologically rich languages that use a lot of different word suffixes to express different grammatical, syntactic, or semantic features. This kind of behaviour usually produces a lot of recognition errors, especially in large vocabulary systems-even when, due to good acoustical matching, the correct lemma is predicted by the automatic speech recognition system, often a wrong word ending occurs, which is nevertheless counted as an error. This effect is larger for contexts not present in the language model training corpus. In this manuscript, an approach which takes into account different morphological categories of words for language modeling is examined, and the benefits in terms of word error rates and perplexities are presented. These categories include word type, word case, grammatical number, and gender, and they were all assigned to words in the system vocabulary, where applicable. These additional word features helped to produce significant improvements in relation to the baseline system, both for n-gram-based and neural network-based language models. The proposed system can help overcome a lot of tedious errors in a large vocabulary system, for example, for dictation, both for Serbian and for other languages with similar characteristics.

SUBMITTER: Pakoci E

PROVIDER: S-EPMC6421827 | biostudies-literature | 2019

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Using Morphological Data in Language Modeling for Serbian Large Vocabulary Speech Recognition.

Pakoci Edvin E Popović Branislav B Pekar Darko D

Computational intelligence and neuroscience 20190303

Serbian is in a group of highly inflective and morphologically rich languages that use a lot of different word suffixes to express different grammatical, syntactic, or semantic features. This kind of behaviour usually produces a lot of recognition errors, especially in large vocabulary systems-even when, due to good acoustical matching, the correct lemma is predicted by the automatic speech recognition system, often a wrong word ending occurs, which is nevertheless counted as an error. This effe ...[more]

PMID: 30944554

Similar Datasets

Project description:This study investigated Chinese English-as-a-foreign-language (EFL) learners' use of vocabulary learning strategies (VLSs) and its relationship with vocabulary knowledge (VK), especially in relation to proficiency, gender, and discipline. Structural equation models were established following exploratory factor analysis (EFA) and confirmatory factor analysis (CFA) procedures, and mediation analyses and multiple-group analyses, as well as analyses of variance, were conducted. Four hundred nineteen sophomores' strategy use frequency, Vocabulary Size Test (VST) scores (indicative of breadth of VK), Word Associates Test (WAT) scores (indicative of depth of VK), College English Test Band-4 scores, and gender and discipline categories were used as data. Proficiency significantly predicted Attention and Guessing positively but was a negative predictor of Socializing (asking others for help). Girls liked making notes while using dictionaries (DictNote) and Socializing, and students of arts also took more notes. Attention and Guessing significantly predicted VST and WAT positively, but Socializing significantly predicted the breadth and depth of VK negatively, and DictNote, Association, and Repetition had no significant relationship. The predictive power of Attention, Guessing, and Socializing, however, was achieved mainly, or for an important part, via the mediating or indirect effects of proficiency. Gender moderated the predictive power of Attention, Socializing, and DictNote over VST, greater for male students, whereas discipline moderated the relationship between Guessing and WAT, stronger for arts students. The findings are related to strategy features, gender characteristics, disciplinary influence, the EFL context and culture, and effective learning. This study reveals the complex relationships among use of VLSs, VK, and learner variables. Attention is called for to third-party factors in understanding VLSs-VK relationships. Given the important mediating effects of proficiency, it is proposed that vocabulary learning be strategically integrated into the accumulative process of English learning.

Project description:OBJECTIVES:To describe the epidemiology and parent-child concordance of hearing, speech reception, vocabulary and language in Australian parent-child dyads at child age 11 to 12 years. DESIGN:Population-based cross-sectional study (Child Health CheckPoint) nested within the Longitudinal Study of Australian Children. SETTING:Assessment centres in seven Australian cities and eight regional towns or home visits around Australia, February 2015 to March 2016. PARTICIPANTS:Of all participating CheckPoint families (n=1874), 1516 children (50% female) and 1520 parents (87% mothers, mean age 43.8 years) undertook at least one of four measurements of hearing and language. OUTCOME MEASURES:Hearing threshold (better ear mean of 1, 2 and 4 kHz) from pure-tone audiometry, speech reception threshold, receptive vocabulary, expressive and receptive languages using a sentence repetition task. Parent-child concordance was examined using Pearson's correlation coefficients and adjusted linear regression models. Survey weights and methods accounted for Longitudinal Study of Australian Children's complex sampling and stratification. RESULTS:Children had a similar speech reception threshold to parents (children mean -14.3, SD 2.4; parents -14.9, SD 3.2 dB) but better hearing acuity (children 8.3, SD 6.3; parents 13.4, SD 7.0 decibels hearing level). Standardised sentence repetition scores were similar (children 9.8, SD 2.9; parents 9.1, SD 3.3) but, as expected, parents had superior receptive vocabularies. Parent-child correlations were higher for the cognitively-based language measures (vocabulary 0.31, 95% CI 0.26 to 0.36; sentence repetition 0.29, 95% CI 0.24 to 0.34) than the auditory measures (hearing 0.18, 95% CI 0.13 to 0.23; speech reception threshold 0.18, 95% CI 0.13 to 0.22). Mother-child and father-child concordances were similar for all measures. CONCLUSIONS:We provide population reference values for multiple measures spanning auditory and verbal communication systems in children and mid-life adults. Concordance values aligned with previous twin studies and offspring studies in adults, in keeping with polygenic heritability that is modest for audition but around 60% for language by late childhood.

Dataset Information

Using Morphological Data in Language Modeling for Serbian Large Vocabulary Speech Recognition.

Publications

Using Morphological Data in Language Modeling for Serbian Large Vocabulary Speech Recognition.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets