Kim, HyangHee
(Graduate Program in Speech-Language Pathology, Department & Research Institute of Rehabilitation Medicine, Yonsei University College of Medicine)
,
Choi, Ji-Myoung
(Interdisciplinary Graduate Program in Linguistics and Informatics, Institute of Language and Information Studies, Yonsei University)
,
Kim, Hansaem
(Interdisciplinary Graduate Program in Linguistics and Informatics, Institute of Language and Information Studies, Yonsei University)
,
Baek, Ginju
(Graduate Program in Speech-Language Pathology, Yonsei University)
,
Kim, Bo Seon
(Graduate Program in Speech-Language Pathology, Yonsei University)
,
Seo, Sang Kyu
(Department of Korean Language and Literature, Institute of Language and Information Studies, Yonsei University)
Spontaneous speech provides rich information defining the linguistic characteristics of individuals. As such, computational analysis of speech would enhance the efficiency involved in evaluating patients' speech. This study aims to provide a method to differentiate the persons with and without aphas...
Spontaneous speech provides rich information defining the linguistic characteristics of individuals. As such, computational analysis of speech would enhance the efficiency involved in evaluating patients' speech. This study aims to provide a method to differentiate the persons with and without aphasia based on language usage. Ten aphasic patients and their counterpart normal controls participated, and they were all tasked to describe a set of given words. Their utterances were linguistically processed and compared to each other. Computational analyses from PCA (Principle Component Analysis) to machine learning were conducted to select the relevant linguistic features, and consequently to classify the two groups based on the features selected. It was found that functional words, not content words, were the main differentiator of the two groups. The most viable discriminators were demonstratives, function words, sentence final endings, and postpositions. The machine learning classification model was found to be quite accurate (90%), and to impressively be stable. This study is noteworthy as it is the first attempt that uses computational analysis to characterize the word usage patterns in Korean aphasic patients, thereby discriminating from the normal group.
Spontaneous speech provides rich information defining the linguistic characteristics of individuals. As such, computational analysis of speech would enhance the efficiency involved in evaluating patients' speech. This study aims to provide a method to differentiate the persons with and without aphasia based on language usage. Ten aphasic patients and their counterpart normal controls participated, and they were all tasked to describe a set of given words. Their utterances were linguistically processed and compared to each other. Computational analyses from PCA (Principle Component Analysis) to machine learning were conducted to select the relevant linguistic features, and consequently to classify the two groups based on the features selected. It was found that functional words, not content words, were the main differentiator of the two groups. The most viable discriminators were demonstratives, function words, sentence final endings, and postpositions. The machine learning classification model was found to be quite accurate (90%), and to impressively be stable. This study is noteworthy as it is the first attempt that uses computational analysis to characterize the word usage patterns in Korean aphasic patients, thereby discriminating from the normal group.
* AI 자동 식별 결과로 적합하지 않은 문장이 있을 수 있으니, 이용에 유의하시기 바랍니다.
문제 정의
This pilot study is within the context of computational data-driven approach to detecting and differentiating persons with and without linguistic impairments. The goal of this study is thus to identify differential linguistic characteristics between aphasic patients and normal subjects through computational analysis of their Korean-specific morphological uses in spontaneous speech resulting from a word definition task. Morpho-syntactic information of spontaneous speech tells us how words are used and how they are related to each other.
The study is significant in that it has provided crucial directions for clinical research in evaluating the spontaneous utterances of aphasia. The computational linguistic analysis employed in this study has proved quite accurate in identifying aphasic patients by producing output of a linguistic description of the patient group.
[7] found different proportions of nouns, pronouns, adjectives and verbs between semantic dementia and normal control groups. These analyses could facilitate an accurate and objective comparison of language usages between the persons with and without aphasic symptoms.
This research is focused on the word usage patterns across the two groups, rather than on acoustic characteristics. Accordingly, the majority of linguistic features chosen as variables are part-of-speech tagged words, and some non-word phenomena to gauge utterance fluency such as repetitions and the use of filler interjections.
가설 설정
Thirdly, the more frequent use of demonstrative adjectives modifying nouns by the control group compared with the patient group can also be indicative of differences in two groups’ uses of morpho-syntactic and/or semantic functions. In this study, we trichotomized the types of demonstratives according to Korean linguistic features. A demonstrative adjective proceeds to modify a substantive indicating its specificity, whereas the use of a demonstrative pronoun conveys unspecificity of what is referred to by a specific noun.
제안 방법
From the two principal components that account for the majority of the variance, the most important linguistic features correlated with the components are extracted. After that, statistical tests were performed on every linguistic feature to identify the most relevant ones in distinguishing the two groups. In the process, it will be re-confirmed whether the features extracted in the PCA have explanatory power in differentiating the two groups.
Computational “data-driven” approaches such as in [7], [12], [13] have also been adopted to identify the characteristics of discourse performance of patients with language impairment. For these studies, spontaneous utterances were collected, processed, and comparatively analyzed using computational statistical and machine learning algorithms. For example, a PCA (Principal Component Analysis) algorithm was used to pick out the lexical and syntactic difference between patients with semantic dementia and normal controls [12].
The features identified from the statistical tests are likely to be critical elements in predicting (or classifying) whether or not an individual is going to be diagnosed as having aphasia based on her/his language use. For this prediction, a logistic regression analysis was carried out with the selected features as variables.
To validate the PCA results and the discriminatory features derived from dimension 2, statistical tests were conducted to find out which individual feature is statistically significant in distinguishing the two groups. Given the small number of samples (20 samples in total), permuted t-test and U-test, rather than ordinary t-test, were applied to all the 24 features excluding three frequency level features. To the three frequency level variables that are about frequency comparison, the Smirnov-Kolmogorov test was applied.
The least and the most number of words produced in each group were 165 (AP9) and 1070 (AP7), in the patient group, and 431 (NC2) and 2177 (NC8), in the normal control group. In order to set off the impact of the difference in text length between the groups and among the individuals, the values of the variables in the Table 2 were obtained by calculating relative frequency, which means the frequencies used in this study were adjusted against the varying amounts of speech not to bias the results.
In this study, of the 20 features of word usage patterns, demonstratives are subcategorised into 4 types, NP type (DEMON_NP), MM type (DEMON_MM), and VA type (DEMON_VA), and their combined all types (DEMON_ALL). These respectively signify demonstrative pronouns (e.
The second supplementary model, on the other hand, was constructed by removing one of the variables from the original model, which was the sentence final endings (EOMALS), a word class with the highest statistical significance and the highest correlation coefficient with group differentiation. It could be considered to be a predictable feature in that persons with aphasia tend to produce fragmentary sentences more frequently, and therefore, use more final endings.
This pilot study is within the context of computational data-driven approach to detecting and differentiating persons with and without linguistic impairments. The goal of this study is thus to identify differential linguistic characteristics between aphasic patients and normal subjects through computational analysis of their Korean-specific morphological uses in spontaneous speech resulting from a word definition task.
This study is noteworthy in that it is the first attempt to identify linguistic features, i.e. word usage patterns, to distinguish an aphasic patient from a non-aphasic subject, and to test how successfully the statistical and machine learning models based on the features can separate the two subject groups with and without aphasia.
To assess the stability of the model, two supplementary models were set up, one of which was constructed by incorporating two more variables into the initial model, and the other by removing a variable from the model. Into the first supplementary model to test the completeness of the original model, two variables were added that did not belong to word usage patterns but correlated with dimension 2 of the PCA model – namely, the frequency level of all words (FREQ.
To validate the PCA results and the discriminatory features derived from dimension 2, statistical tests were conducted to find out which individual feature is statistically significant in distinguishing the two groups. Given the small number of samples (20 samples in total), permuted t-test and U-test, rather than ordinary t-test, were applied to all the 24 features excluding three frequency level features.
Using the 27 linguistic features automatically extracted, computational analysis was conducted to identify the most salient features distinguishing the patient group from the normal control group. We first executed PCA to look at how the 20 participants are positioned in a two-dimensional space and how they group together.
대상 데이터
The study was carried out on a total of 20 participants: 10 individuals with aphasia (AP) (5 females, 5 males) aged between 19 and 79 years (M = 51.6; SD = 18.2) and 10 non-aphasic control subjects (NC) (5 females, 5 males) aged between 22 and 76 years (M = 50.5; SD = 17.7). The number of 20 participants may not be large enough to lead to a broad generalization yet, but it is large enough to get practical and clinical implications and thus suggest directions for computational approach to automatic diagnosis of language deficit problems such as aphasia as in previous pilot studies [1], [9], [15].
이론/모형
In a machine-learning classification task, a computer recognizes and learns patterns and their related categories from the data and predicts to which of the categories a new case belongs. In this study, the Bayesian logistic regression algorithm, a subtype of logistic regression, was chosen due to its resistance to data sparsity. The performance of the model was measured by prediction accuracy.
The classification model was built using the Bayesian logistic regression, whose classificatory performance is assessed in terms of accuracy. The fit was bootstrapped 100 times to guarantee the stability of the model performance.
For this study, concrete and abstract ten nouns [9] were provided to all participants: watermelon, pharmacy, electric fan, train, rabbit, jealousy, music, excursion, joke, friendship. The first five concrete nouns were chosen based on The Florida Semantic Battery [20] to represent semantic class, definitional class, animacy/inanimacy, and image. The abstract ones were chosen with their abstractness and clearness of the semantic boundaries and features taken into consideration.
All the words in the transcripts were assigned with their parts-of-speech tags, and some errors in parts-of-speech tagging were corrected semi-automatically in the post-edit stage. The grammatical tagging was conducted with a widely used Korean morphological analyzer, UTagger, based on the Sejong tagset. For detailed tagset, refer to the Appendix 1.
Given the small number of samples (20 samples in total), permuted t-test and U-test, rather than ordinary t-test, were applied to all the 24 features excluding three frequency level features. To the three frequency level variables that are about frequency comparison, the Smirnov-Kolmogorov test was applied. The threshold of 0.
성능/효과
5 shows the result of class predictions. In the model, one patient (AP7) was misclassified as the other class with a patient class probability of 21.9%, and one non-aphasic (NC2) was assigned to the aphasic group with a normal class probability of 29.8%. These (un-)successful predictions are forecast in Fig.
So, the removal of this predictable feature could provide a way to test the robustness of the original classification model. It turned out that the new model achieved an accuracy of 85%, with only three out of 20 persons misclassified. Two non-aphasic subjects (NC2 and NC5) and one aphasic patient (AP7) were misclassified.
The significance of the functional word classes as separating indicators between the two groups was demonstrated in the successful classification modelling. The classification model built on the features of function words, sentence final endings, demonstratives, and postpositions was able to identify the aphasic patients with a 90% success rate. The robustness of the model, i.
The main finding from the current computational analysis of Korean utterances in terms of word usage pattern is that the functional linguistic features, such as sentence final endings (EOMALS), postpositions (JOSAS) and demonstratives (DEMON_MM), are more crucial in distinguishing aphasic patients and healthy controls than the content-related words (e.g. nouns, verbs, adjectives). These strong classification features between the two groups are less related to ‘what’ a speaker says than to ‘how’ she/he says it.
The two predictors were not found to improve the overall performance of the initial model. The model achieved the same accuracy rate (90%) with the same two persons (AP7 and NC2) misclassified, though the mean probability of patient candidates being correctly classified as patients rose marginally from 78.1% to 84.6%, suggesting that the frequency levels of words as a variable could play a minor role in distinguishing the two groups.
후속연구
In the course of this work, we hope to enhance the level of understanding of the speech characteristics of Korean aphasic patients, and to lay the groundwork for automatic diagnosis of aphasic cases based on linguistic symptoms.
AI-Helper
※ AI-Helper는 부적절한 답변을 할 수 있습니다.