Development and Verification of Mandarin Recognition Test Materials

• Main Authors: Kuen-Shian Tsai, 蔡昆憲
• Other Authors: Shuenn-Tsong Young
• Format: Others
• Language: en_US
• Published: 2009
• Online Access: http://ndltd.ncl.edu.tw/handle/t874j9

博士 === 國立陽明大學 === 醫學工程研究所 === 97 === Word recognition tests have been widely used for hearing diagnosis and aural rehabilitation. However, some problems existing in previous word recognition tests may affect the reliability and validity of tests, including: (a) the time consuming and difficulty of constructing word lists manually led to a small number of word lists and even the phonemic balance of word lists might not be achieved; (b) 25-item word lists are usually constructed by arbitrarily dividing 50-item word lists into two parts, and hence the phonemic distributions of these half lists might be inconsistent to affect their interlist equivalence; (c) test items in word lists are not homogeneous, and hence the recognition difficulty of these half lists might be inconsistent to affect their interlist equivalence; (d) the lack of standardized-recorded speech recognition materials in Taiwan. Therefore, the purpose of this study was to design a Mandarin monosyllable recognition test (MMRT) for clinical and research applications, as well as to develop a set of approaches to improve the reliability and validity of word recognition tests. The study first investigated how to convert the design of word lists into an optimization problem, and proposed a genetic algorithm to solve this problem. Our genetic algorithm can construct word lists automatically according to the desired number of each phoneme. Different population sizes were used to evaluate the convergence and efficiency of the genetic algorithm applying to the construction of phonemically balanced word lists in this study. The result showed that population size of 100 is a better setting to balance the convergence and efficiency. Under this condition, we can obtain 45 phonemically balanced word lists in 100 runs, and spend 76 seconds to obtain a phonemically balanced word list on average. By the results, the genetic algorithm performed an efficient, robust, and low-complexity search of the problem space and can be easily modified to adapt to the word-list construction of other languages. For the improvement of the reliability and validity in word recognition tests, in addition to satisfying the major design criteria of familiarity, phonemic balance, and interlist equivalence, this study emphasized the homogeneity of test items in word lists. In the development of MMRT, to achieve the goal of high subject familiarity with the test material, we selected the 700 most frequently occurring monosyllables to be the test material. The homogeneity of the test material was achieved by evaluating five psychometric characteristics of these 700 monosyllables to obtain 348 homogeneous monosyllables with similar psychometric functions for constructing the word lists. The phonemic balance of the 50-item word lists was achieved by deriving the desired numbers of initials, finals, and tones in these lists according to their occurrence frequencies in 4733 monosyllabic words. The phonemic balance of the 25-item word lists was achieved by equally dividing the desired numbers of initials, finals, and tones in the 50-item word lists into two parts, called half-A and half-B lists. Three half-A lists and three half-B lists were constructed by the genetic algorithm from the 348 homogeneous monosyllables, and they could be paired to form nine 50-item word lists. Accordingly, all MMRT word lists are familiar, homogeneous, and phonemically balanced. The statistic results indicated that the six 25-item word lists and nine 50-item word lists exhibited interlist equivalence with respect to their psychometric functions and five psychometric characteristics; moreover, their interitem and mean intersubject variability are lower than those of previously reported word lists. This thesis then evaluated the clinical reliability of MMRT word lists, using 30 listeners with sensorineural hearing loss, presented at their most comfortable level in quiet. The statistic results also indicated the test-retest, interlist, and split-half reliability of the MMRT word lists are better than previously reported word lists. In conclusion, our genetic algorithm exhibits excellent convergence and efficiency in the design of word lists. In addition, the study results also support our hypothesis that controlling the homogeneity of test items and making 25-item and 50-item word lists are all phonemically balanced can reduce the variability of test items and hence improve the reliability of word recognition test. The validity and diagnosis sensitivity of the MMRT should be examined in further studies since the MMRT word lists have good reliability. Some further studies and applications are recommended in the end of this thesis.