Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

In this thesis, we examined the automation of linguistic features for the detection of aphasia through spontaneous speech. The data for our research consisted of automatically generated and manually corrected transcriptions of a picture description task (CAT-NL) from elderly neurotypical controls and aphasia patients. More specifically, we examined (1) three types of repetition patterns, (2) lexical diversity and (3) semantic paraphasia. The repetition patterns were extracted from the transcriptions by means of n-grams: we looked at bigrams and trigrams containing identical words and identical consecutive bigrams and trigrams. The lexical diversity was measured through the MATTR metric. The last feature, semantic paraphasia, was measured through the semantic similarity using SpaCy between the nouns extracted from the transcriptions and the key nouns from the picture. Additionally, we looked at the accuracy of SpaCy’s part-of-speech (POS) tagger for the aphasia group as well as the control group. To improve the performance of the tagger on colloquial speech, we trained the tagger on data containing discourse particles and interjections. Furthermore, we investigated the reliability of the automatically generated transcriptions from an automatic speech recognizer (ASR) by comparing the performance of the automated analysis of the linguistic features between the ASR transcriptions and the manually corrected transcriptions. The results showed us that all of the repetition types significantly distinguished between the control group and the aphasia group for manually corrected transcripts. The aphasia group had more transcriptions with higher frequencies of these repetition patterns. For the ASR transcriptions, this was only the case for the bigrams containing identical words. Likewise, there was a significant difference between both groups for the lexical diversity for the ASR transcriptions as well as the manually corrected transcripts. Regarding the semantic similarity scores for semantic paraphasia, there was a significant difference between both groups in both types of transcriptions. It should be noted that both groups revealed a high percentage of out-of-vocabulary words in the ASR as well as the manually corrected transcriptions, which likely affected the results. SpaCy’s POS tagger demonstrated low F1 scores in the case of proper nouns and symbols for both groups for the manually corrected transcripts. The trained POS tagger achieved an accuracy of approximately 16 percent, which may be due to the small dataset. Regarding the two types of transcripts, there was no significant difference found for the repetition patterns, the lexical diversity and the F1 scores of the POS tagger. This suggests that ASR transcriptions can be reliably used for the analysis of these linguistic features. However, a significant difference was found between both types of transcriptions for the semantic similarity scores. This is largely due to the higher percentage of OOV words in the ASR transcriptions. This research contributes to a more time-efficient and objective diagnosis of aphasia, which can help in the rehabilitation of aphasia patients.