Detecting Signs of Dementia Using Word Vector Representations

Abstract

Recent approaches to word vector representations, e.g., 'w2vec' and 'GloVe', have been shown to be powerful methods for capturing the semantics and syntax of words in a text. The approaches model the co-occurrences of words and recent successful applications on written text have shown how the vector representations and their interrelations represent the meaning or sentiment in the text. Most applications have targeted written language, however, in this paper, we investigate how these models port to the spoken language domain where the text is the result of (erroneous) automatic speech transcription. In particular, we are interested in the task of detecting signs of dementia in a person's spoken language. This is motivated by the fact that early signs of dementia are known to affect a person's ability to express meaning articulately for example when they engage in a conversation -something which is known to be cognitively very demanding. We analyse conversations designed to probe people's short and long-term memory and propose three different methods for how word vectors may be used in a classification setup. We show that it is possible to identify dementia from the output of a speech recogniser despite a high occurrence of recognition errors.

Automatic Hierarchical Attention Neural Network for Detecting AD

“…Our result on automatic transcripts is also state-of-art. Compared with the automatic transcripts based dementia detection accuracy (62.3%) on DementiaBank presented in [9], we got an F-score of 76.09%. In [28], an almost similar precision (79%) was achieved by selecting features extracted from audio and transcripts at the same time.…”

“…In dementia detection, word embedding has been proposed to be used for converting spoken transcripts into vectors for detecting cognitive decline [9]. Two recent techniques, 'w2vec' [15] and 'GloVe' [16], which consider the context in the text, were proposed and achieved a better performance compared with traditional methods like bag-of-words (BOW) [17].…”

“…The benefit of this operation is to capture the semantic information about the words in sentences. The efficiency of using word vectors for dementia detection has been demonstrated before [9]. In our model, wit with t ∈ [1, T ] and i ∈ [1, L] is used to represent the tth word in ith sentence.…”

“…For language models we followed the in-domain 3/4 grams with the KN/Turing smoothing. Note that we added an additional 64 hours worth of conversational data (see [9] for more details) to boost the acoustic model of the ASRs. Finally, we got the transcripts with a word error rate (WER) of 41.6% on Demen-tiaBank and 33.8% on IVA.…”

“…To evaluate the model, the use of both manual and automatic transcripts (generated by using an automatic speech recogniser (ASR)) is explored. Cookie Theft picture descriptions from the DementiaBank database [7], as well as from an in-house database collected using our Intelligent Virtual Agent (IVA; 'digital doctor') system [8,9] are used to test our model. To the best of our knowledge, this is the first use of a bidirectional hierarchical recurrent neural network combined with an attention mechanism (BHANN) for dementia detection.…”