Measures of Lexical Diversity and Detection of Alzheimer’s Using

Speech

Muskan Kothari, Darshil Vipul Shah, Moulya T., Swasthi P. Rao and Jayashree R.

Department of Computer Science and Engineering, PES University, Bangalore, India

Keywords: Alzheimer’s Disease, Speech, Feature Extraction, Brunet’s Measure, Sichel’s Measure, Lexical Diversity,

MATTR, MTLD.

Abstract: Alzheimer's disease is the most common cause of dementia — a continuous decline in thinking, behavioral

and social skills that affects a person's ability to function independently. Another area of concern is the overlap

of symptoms with a similar disease of dementia - Frontotemporal Dementia(FTD). This paper aims to analyze

the difference in linguistic features between control and dementia groups with respect to lexical diversity

through measures like Brunet’s and Sichel’s measure, frequency rates of adverb, verb, and linguistic

deterioration through repetition, disfluency, incomplete sentences, hesitation and long pauses through dataset

obtained by DementiaBank. This is achieved through gauging the cognitive ability in speech, which is an

inexpensive and non-invasive mode of analysis, qualifying as a screening test. The subjects are given certain

description tasks such as the famous cookie theft picture, analyzed through conversations. The result displays

the difference in lexical diversity which is a significant marker.

1 INTRODUCTION

Alzheimer’s disease is a progressive neurologic

disorder that causes the brain to shrink (atrophy) and

brain cells to die. Researchers across the world are

constantly making efforts to find methods for the

detection and treatment of this disorder in an

effective, non-invasive and cost-efficient way.

Speech is one of the most effective, inexpensive and

non-invasive modes of testing.

AD affects one in ten adults over the age of 65

years in the United States (Alzheimer’s Association,

2015). Diagnosis is possibly more effective in the

early stages of dementia. In low and middle income

countries, diagnosis of AD frequently occurs several

years after the onset of the disease. This leads to a

treatment gap for early dementia sufferers

(Alzheimer’s Disease International, 2011). This gap

reduces the effectiveness of treatments, prolonging

the patients’ state of reduced independence.

Sometimes AD might be misclassified into what’s

known as Fronto Temporal Dementia, as the

symptoms are very common to Alzheimer’s Disease

and can jeopardize the appropriate diagnosis and

medication for a patient with cognitive impairment

since it is now considered to be as common as

Alzheimer’s in middle aged patients. AD is often

difficult to differentiate with FTD, especially in the

early stage. Currently, there are no disease-modifying

treatments for FTD. The acetylcholinesterase

inhibitors widely used in patients with AD could lead

to worsening of symptoms in those with FTD.

Therefore, accurate diagnosis from a

differentiating perspective of FTD and AD and the

reduction of misdiagnosis is of essential utility in

clinical trials. FTD is also a highly heritable group of

neurodegenerative disorders, with around 30% of

patients having a strong family history. Diagnosing

and confirming it early could be very helpful for the

descendants of the patient as well.

Realizing the necessity, scope and potential of this

area of research, the work described in this paper aims

to resolve some core issues related to Alzheimer’s

detection among patients, taking the first step in

classifying it precisely through linguistic features

extracted from transcribed files in CHAT (Codes for

the Human Analysis of Transcripts) protocol

(MacWhinney, 2000). It cannot be denied that

examining linguistic features is one of the best and

most inexpensive ways to detect Alzheimer’s, which

is why this research will be using them in the unique

model, inspired by the approaches we have explored.

806

Kothari, M., Shah, D., T., M., Rao, S. and R., J.

Measures of Lexical Diversity and Detection of Alzheimer’s Using Speech.

DOI: 10.5220/0011779000003393

In Proceedings of the 15th International Conference on Agents and Artiﬁcial Intelligence (ICAART 2023) - Volume 3, pages 806-812

ISBN: 978-989-758-623-1; ISSN: 2184-433X

 2023 by SCITEPRESS – Science and Technology Publications, Lda. Under CC license (CC BY-NC-ND 4.0)

2 RELATED WORK

The work done in (Renxuan Albert Li, 2020) mainly

focuses on Mild Cognitive Impairment (MCI) using

the Brain, Stress, Hypertension, and Aging Research

Program (B-SHARP) dataset. Three speech tasks

were given to the subjects and the recordings of 1-2

mins each were transcribed using Temi (Daniela

Beltrami, 2018), a tool that automatically transcribed

speech and linguistic features were further analyzed

with a helpful tool called ELIT (Jacob Devlin, 2019).

Three tasks involved speaking about picture

description, room environment and daily activity.

Task 2 out of 3 has highest accuracy which proved

spatial descriptions to be most useful.

The methodology proposed in (N. Wang, 2020) is

highly personalized. It analyzes the hidden linguistic

patterns of each subject separately using their own

linguistic biomarkers over a duration. The main

analysis done over here was a case study on President

Reagan’s speeches. Uses a lot of speech features such

as pronoun-noun ratio, word frequency ratio,

Honore’s measure, Brunet’s measure etc. Focuses on

trying to predict in an automated manner rather than

on trained data, and uses SVM for this approach, but

it’s observed that prediction using t-SNE is more

accurate than the automated SVM approach.

The aim of the research done in (Haulcy, 2021)

was to classify Alzheimer’s using ADRess Dataset.

The dataset consists of audio recordings along with

the transcripts, and metadata for non-AD and AD

patients. Feature sets were formed with LDA, and

with PCA, and training of classifiers on feature sets

to observe the effect of dimensionality reduction. One

main advantage of using linguistic features is the

usage of punctuation. The semantic and syntactic

information is used by the model. The classifiers used

are LDA, Decision Tree classifier, the k-nearest

neighbors classifier, SVM and RF classifier.

So far, most of the work done was in English and

no other language had been worked upon in detail.

But in (Zhiqiang Guo, 2020), AD was detected in

Mandarin. The dataset used here consists of

transcriptions of the cookie theft picture in Mandarin.

208 transcriptions were recorded equally for both

healthy and AD patients. The results of this

experiment show that the contrastive learning method

can achieve better accuracy than conventional CNN-

based and BERT-based methods. The output was

achieved by a model containing two pooling layers of

english and mandarin and two auto-encoders of both

the languages. The accuracy obtained here was

81.4%.

In (Chloé Pou-Prom, 2018), the researchers

leverage the multiview nature of DementiaBank, to

learn an embedding that captures different modes of

cognitive impairment. Generalized canonical

correlation analysis (GCCA) was applied to the

dataset and the benefits of using multiview

embeddings on identifying AD and predicting clinical

scores were demonstrated. The short-coming of the

research being that while GCCA allowed for an

arbitrary number of views, it learnt only linear

projections to the embedding space. In this case,

DGCCA can be used which makes use of neural

networks to learn non-linear mappings to the

embedding space.

Semantic Verbal Fluency tests were used in

(Felipe Paula, 2018) to detect certain clinical

conditions like dementia The SVF dataset of a 100

patients was classified into groups of 25 controls each

in classes like Amnestic Mild Cognitive Deficit

(aMCD), Multi-domain Mild Cognitive Deficit

(mMCD) and Alzheimer’s Disease (AD). The SVF

test uses a binary function called switch which

operates on a sequence of N words. Three heuristics

of the switch function were explored. These were the

Detection based on global mean, detection based on

local mean and hybrid detection.

An approach of using CNN and LSTM was seen

in (Flavio Di Palo, 2019). The purpose of CNN and

LSTM was to enable the learning of both implicitly

learned features and targeted features to perform

classification. A bi-directional LSTM was used

instead, and an attention mechanism was applied on

the hidden states of the LSTM. Class weights that

were added to the loss function in this approach took

the dataset imbalance into account.

Kathleen et al. in (Zhou, 2016) have devised ways

to differentiate and identify between having AD and

depression. To analyze further, textual and acoustic

features were extracted from the patient’s speech

data. A subset of the extracted features were selected

by using a correlation-based filter. A detailed analysis

of correlation between depression and dementia was

carried out by the authors. The selected features were

then fed in ML classifiers like SVM and Logistic

Regression (LR) models.

3 DATASET

From the review done in (Haulcy R, 2021), there are

various datasets available for the study of Dementia

in languages such as English, French, Greek,

Hungarian, Italian, Mandarin, Portuguese, Spanish,

Swedish and Turkish. While most of them are

Measures of Lexical Diversity and Detection of Alzheimer’s Using Speech

807

available upon request, the availability of the rest of

the datasets is undefined. In English, there are 3 major

datasets widely known to be available, namely

DementiaBank, Pitt Corpus and WRAP. All of these

datasets are available upon request.

For the purpose of this paper, the dataset chosen

was Pitt corpus, available in English under non-

protocol data where the media included audio files

obtained from DementiaBank. This is an open-source

repository of various corpora available on request. In

DementiaBank, you have corpora available in 5

languages namely English, German, Spanish,

Mandarin and Taiwanese, categorized under protocol

data, non-protocol data and PPA non-protocol data.

This corpus is maintained by Francois Boller and

James Becker as part of a larger protocol

administered at the University of Pittsburgh School of

Medicine.

The dataset includes audio as well as

downloadable transcripts which follow the CHAT

protocol. The dataset includes the conversation

between two participants playing two roles, one as the

investigator (INV) and the other as the participant

(PAR) who is the patient. The data includes responses

for both control and dementia groups where control

groups have elderly individuals and dementia groups

include patients with probable and possible

Alzheimer’s disease. The group also includes a few

patients from other dementia diseases. The

conversations between the two roles is transcribed for

4 language tasks -

1. Cookie Theft - includes participants describing

the cookie theft picture

2. Fluency - includes responses to the word fluency

task for the dementia group only.

3. Recall - includes responses to story recall tasks

for the dementia group only.

4. Sentence - includes responses to sentence

construction task for dementia group only.

The focus for this paper is only for the cookie

theft task since it includes both the groups. The reason

for choosing the DementiaBank dataset over other

available datasets in English is the fact that this

dataset is balanced. It also includes other

demographic information of the patients such as age,

sex, diagnosis and MMSE score.

MMSE stands for Mini-Mental State

Examination which is a set of 11 questions that a

doctor asks the patient to assess the cognitive

impairment. A total of 6 areas of mental abilities are

checked through this examination which includes

orientation to time and place, concentration, short-

term memory recall which can be reasoned for the

story recall task, language skills, visuospatial abilities

which can be reasoned for the cookie theft task and

finally, the ability to follow instructions. The

maximum obtainable score for MMSE is 30. A score

below 24 is usually indicative of possible cognitive

impairment.

A total of 548 files are used for further analysis

and research. 305 of the total files are from the

dementia group, and 243 files are from the control

group. To read the CHAT files in .cha format,

pylangacq was used, which is a library to read

conversational data represented in this format. It has

various methods which allows to obtain information

about the participants (in this case, it returns PAR and

INV), the metadata stored in transcribed files (which

usually start with the @ symbol), number of files,

number of words, and number of utterances filtered

by participants, through a reader object. It also gives

information about tokens in each file which returns an

object of tuples with 4 fields. Tokens give you word

based annotations, and the fields include the word

itself, the part-of-speech tag, morphological

information and the grammatical relation. The

grammatical relation is an object which tells the

relation between two words, including 3 attributes

which are the position of the dependent (the word

itself), position of the head, and the relation between

them.

The metadata transcribed in the files includes

information like the encoding (in this case UTF8),

language, participants, information about the

participants like language, corpus, age, sex, role,

group and education.The control files contain a total

of 3896 utterances and 33931 words while the

dementia files contain a total of 5585 utterances and

43471 words. A subset of the information obtained

from one of the.cha files of the dementia group is

detailed in table 1. The results of words, utterances,

tokens and meta-data along with the method used

from pylangacq is displayed.

4 METHODOLOGY

4.1 Data Preprocessing and

Preparation

The first step for preparing the data was to analyze the

different essential components that constitute the

CHAT files. From previous methods explored, the

utterances function posed to be very useful, along

with the tokens methods. The dataset preparation

started with extracting all the utterances by

participants in each file. This means that using the

utterances method, filtered by ‘PAR’, each file was

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

808

Table 1: Analysis of CHAT files.

Sentence

Information about participant conversations

Methods used Subhead

“He's

taking

jar.

that’s

all.”

.words()

["he's", 'taking', 'cookie', 'jar', '.',

"that's", 'all', '.']

.words(by_utte

rances=True)

[["he's", 'taking', 'cookie', 'jar',

'.'], ["that's", 'all', '.']]

.tokens()

Token(word='taking', pos='part',

mor='take-PRESP',

gra=Gra(dep=3, head=0,

rel='ROOT'))

.headers()

{'UTF8': '',

'PID': '11312/t-00002422-1',

'Languages': ['eng'],

'Participants': {'PAR': {'name':

'Participant',

'language': 'eng',

'corpus': 'Pitt',

'age': '56;',

'sex': 'male',

'group': 'ProbableAD',

'ses': '',

'role': 'Participant',

'education': '20',

'custom': ''},

'INV': {'name': 'Investigator',

'language': 'eng',

'corpus': 'Pitt',

'age': '',

'sex': '',

'group': '',

'ses': '',

'role': 'Investigator',

'education': '',

'custom': ''}},

'Media': '003-0, audio',

'G': 'Cookie'}

processed and the associated label was also prepared

for the group that the ‘PAR’ belonged to. Control

group was labeled 0 and the dementia group was

labeled 1. The other important feature extraction was

using POS tags. Parts of speech tagging have been

proving to be essential to extract and learn some of

the key features of speech. For patients with

Alzheimer’s, some of the POS tags are more frequent

than normal patients. Using spaCy, an open source

library highly suitable for tasks in Natural Language

Processing and written in Python and Cython,

deemed useful for POS tagging. Each utterance in

each file was passed to a function that added the POS

tag after the token in each row. Using this library,

extraction or preparation tasks become easier because

of the attributes that each token is embedded with.

The transcription files also included some of the

key transcription symbols to signify the manner of

speech or the verbal fluencies. Verbal utterances like

repetitions, retractions, pauses of both types - short

and long, incomplete words, incomplete sentences,

assimilations, various errors, hesitations and

disfluencies were captured through transcription

symbols, which is elaborated in table 2.

Table 2: Transcription Symbols.

Sl. No. Symbol Meaning

1 [/] Repetition

2 [//] Retraction

3 [..] Pause

4 [.] Short pause

5 [...] Long pause

6 [+sgram] Grammatical error

7 &uh/&um/&mm/&hm Hesitation

8 &w+ Disfluenc

These transcription symbols were replaced with

the expansions of what they represented. The concept

of regular expressions was used to identify these

symbols and each annotation was hereby replaced

with the direct meaning.

At the end, we had a dataframe consisting of the

label column, all utterances belonging to each file,

POS tagged column consisting of the token followed

by its POS tag after each, the expanded representation

of the annotation in each utterance, and a final column

without annotations to prevent skewing of POS tags.

4.2 Ratios and Measures

For the research pertaining to this paper, the linguistic

features are divided into POS features and lexical

diversity. For POS features, 3 values were computed,

which are pronoun-noun ratio, adverb frequency rate

and verb frequency rate. These measures are deemed

important from the correlation result obtained in (N.

Wang, 2020). Alzheimer’s patients seemingly use

more pronouns than nouns. The utterances of AD

patients are also rich in adverbs and verbs compared

to other POS tags. The results were consistent with

the observations except for a slight variation in verb

frequency rate. The P-N ratio obtained for AD

patients was 0.6923 and for normal patients was

Measures of Lexical Diversity and Detection of Alzheimer’s Using Speech

809

0.5181, which indicates that normal patients’ speech

included more nouns resulting in a P-N ratio less than

AD patients who used more pronouns than nouns. For

adverb frequency rate, the result obtained for AD

patients and normal patients was 48.95 and 60.08

respectively. Our implementation computed the

frequency by dividing the number of tokens by the

number of adverbs. Thus, a higher number of adverbs

per number of tokens would result in a lesser adverb

frequency rate according to our implementation. This

was consistent with the observation that AD patients

use more adverbs. For verb frequency rate, using the

same implementation as adverb frequency rate, the

result obtained for AD patients and normal patients

was 16.50 and 12.50 respectively. This implies that

less number of verbs were used per number of tokens

by AD patients compared to normal patients.

There are 4 measures computed for lexical

diversity. From the case study in (Zhou, 2016) on

President Reagan’s speech, it was established that AD

patients have a declined vocabulary richness in their

speech. Here’s where POS tags come to use once

again. It proves that the speech including the

vocabulary and the gaps can give a lot to infer. Three

popular measures for vocabulary richness are the

Honore’s statistic (HS), Brunet’s index (BM) and

Sichel measure (SICH).

It is important to know what hapax legomena and

hapax dislegomena mean. Hapax legomena are the

word types that occur once in a text while

dislegomena are those that occur twice in a text. By

logic, hapax legomena is usually the indicator of

lexical diversity. Honore’s statistic which is usually

denoted by R is based on the understanding that texts

with rich vocabulary have larger proportions of words

that are hapax legomena. But this measure is sensitive

to sample size. Both Honore’s and Sichel’s result in a

higher value when vocabulary is rich. In case of

Brunet’s (W), smaller the value, higher the

vocabulary richness and is also not sensitive to the

text length. The range of values is usually between 10

and 20. For the purpose of this study, Sichel’s and

Brunet’s measure was chosen, which balances the

results for lexical diversity since they are both

inversely proportional.

The other two measures used were MTLD

(Measure of Textual Lexical Diversity) and MATTR

(Moving Average TTR), based on TTR (Type-Token

Ratio) which is the number of different words in a

sample of text. MTLD tells the average number of

consecutive words that maintains a certain TTR

before dropping. MATTR is simple enough, in that it

calculates the TTR for a window of a certain size.

4.3 Equations and Measures

Brunet’s measure was implemented using (1)

𝑊 = 𝑁





(1)

where -a is a scaling constant, usually equals -0.172.

N denotes the length of text and V denotes the number

of different words. Sichel’s measure is as simple as

computing hapax dislegomena on the text.

4.4 Training Models

The training of models started with the preparation of

transcribed speeches of AD patients. As explained

earlier, the CHAT protocol and its meanings were

thoroughly analyzed and POS tagging was applied.

In addition to the POS tag and preprocessed

utterances, 4 measures of lexical diversity and 3 ratios

of linguistic features were included in the dataset. The

gaps in utterances are of equal importance to

differentiate a control patient from an AD patient. It

is observed that the speech of AD patients shows

higher occurrences of repetitions, retractions,

disfluency, long pauses, hesitation, grammatical

errors and incomplete sentences.

To conclude, all the features mentioned and

described thus far have been used to prepare and store

the dataset. From the research of existing work,

CNNs, SVMs and LSTMs give the best results. For

this research, a total of 7 models were trained on the

features scaled appropriately. The top three models to

give the highest accuracy were MultinomialNB, SVC

and Random Forest Classifier.

A comparison of the mean values obtained for

each feature in both groups were also compared and

the results were consistent with the existing work

except for verb frequency ratio which deviated from

the existing inferences. All the linguistic features and

gaps denoting retraction, repetition, disfluency,

hesitation and more showed a higher mean in values

for AD patients compared to the control group.

5 RESULTS AND DISCUSSION

The highest accuracy as seen in table 4 obtained was

88.92% by KNearestNeighbors classifier followed by

SVC and MultinomialNB. From the preparation and

analysis of all the measures, it was clear that the AD

patients have a degraded linguistic sense of speech

which is seen in poor lexical diversity, higher use of

adverbs and pronouns, less use of nouns and we have

also identified through verb frequency rate that

despite the observation in (Zhou, 2016), verbs are not

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

810

that frequent in AD patients. The results obtained

after computing the mean values for other measures

like MATTR, MTLD showed that MATTR and

MTLD for AD patients was less than normal patients

which is an indication of reduced lexical diversity in

the speech of AD patients, and the number of

occurrences of repetition, retraction, hesitations,

grammatical errors etc, were higher than normal

patients.

Reported in table 3 are the pair of values obtained

for lexical diversity measures and number of

occurrences in the utterances of AD and control

group.

Table 3: Comparison of Linguistic Measures.

Measure

Group

Control AD

MATTR 0.597633 0.566128

MTLD 34.004573 32.048859

Repetition 0.711934 1.780328

Retraction 1.300412 2.101639

Long pause 0.069959 0.098361

Disfluency 0.732510 1.655738

Hesitation 3.419753 3.603279

Grammatical error 1.234568 1.436066

Incom

lete sentence 0.172840 0.518033

It can be inferred that the values for repetition,

retraction, disfluency and incomplete sentence were

significantly higher for AD than control and could

pose as a useful measure for training the model and

detection purpose.

Finally, the most significant accuracies obtained

are tabulated below for the top 3 models. The test size

was set to 0.15 and a random state of 61 was applied.

Decision tree resulted in the lowest accuracy of

60.24%. The confusion matrix was plotted along with

the computation of F1 score, precision and recall for

each model of the 7 models.

Table 4: Results from Top 3 Models.

Model

Results

Accuracy Precision Recall F1 score

KNN 88.92 0.8592 0.8537 0.8563

SVC 84.33 0.8133 0.8128 0.8197

MultinomialNB 84.33 0.8164 0.8216 0.8225

6 CONCLUSION

This research highlights a significant marker in

analyzing speech of AD patients. From a medical

perspective, using speech is an inexpensive and a

non-invasive process which qualifies as screening

tests. Capable of quick and reliable results, the

inferences from this work include the degradation of

lexical diversity in the speech of AD patients, where

measures like Brunet’s and Sichel’s gave

differentiable mean values for the two control groups.

MATTR and MTLD are another pair of measures

where the mean values for AD patients were less than

the control group. In terms of utterances and manner

of speech, the top 4 significant markers were

repetition, retraction, disfluency and incomplete

sentences; the mean number of occurrences was ~ 78-

201% higher in AD group.

7 FUTURE WORK

This paper talks about the validation of existing

inferences with a deviation in verb frequency ratio

and also contributes by implementing 4 lexical

diversity ratios. There is some potential to include the

demographic information from the transcripts and

analyze the differences in the onset and changes in

cognitive impairments between male and female. To

contribute to the work described in this paper in future

in order to make it more complete, we want to

implement Conditional Random Fields (CRF) to

predict the relation between consecutive POS tags

and analyze useful inferences obtained, if any.

Another addition would be to train models like t-SNE

and hybrid CNN-LSTM, like in (

Sweta Karlekar,

2018)

on the prepared dataset.

Measures of Lexical Diversity and Detection of Alzheimer’s Using Speech

811

ACKNOWLEDGEMENTS

Expressing profound gratitude to Dr. Jayashree R for

encouraging and guiding us along the way and the

Dept. of Computer Science and Engineering at PES

University, for providing this opportunity to expand

our potential of impact, for conducting frequent

research and inculcating problem-solving disciplines.

This opportunity would not be possible without the

grant support in the research conducted by the

maintainers and researchers of DementiaBank and

Pitt Corpus. We are thankful to Carnegie Mellon

University, for facilitating resources and granting

access.

REFERENCES

MacWhinney B. 2000. The CHILDES Project: Tools for

analyzing talk, 3rd edition. Lawrence Erlbaum

Associates, Mahwah, New Jersey.

Renxuan Albert Li, Ihab Hajjar, Felicia Goldstein, and

Jinho D. Choi. 2020. Analysis of Hierarchical Multi-

Content Text Classification Model on B-SHARP

Dataset for Early Detection of Alzheimer’s Disease. In

Proceedings of the 1st Conference of the Asia-Pacific

Chapter of the Association for Computational

Linguistics and the 10th International Joint Conference

on Natural Language Processing, pages 358–365,

Suzhou, China. Association for Computational

Linguistics.

Daniela Beltrami, Gloria Gagliardi, Rema Rossini Favretti,

Enrico Ghidoni, Fabio Tamburini, and Laura Calzà.

2018. Speech Analysis by Natural Language

Processing Techniques: A Possible Tool for Very Early

Detection of Cognitive Decline? Frontiers in Aging

Neuroscience, 10(369):1–13.

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina

Toutanova. 2019. BERT: Pre-training of Deep

Bidirectional Transformers for Language

Understanding. In Proceedings of the Conference of the

North American Chapter of the Association for

Computational Linguistics: Human Language

Technologies, pages 4171–4186.

N. Wang, F. Luo, P. Vishal, K. Subbalakshmi, and R.

Chandramouli. 2020. “Personalized early stage

Alzheimer’s disease detection: a case study of president

Reagan’s speeches.” In Proceedings of the 19th

SIGBioMed Workshop on Biomedical Language

Processing, pages 133–139, Online. Association for

Computational Linguistics.

Haulcy R and Glass J (2021) Classifying Alzheimer’s

disease using audio and text-based representations of

speech. Front. Psychol. 11:624137. doi:

10.3389/fpsyg.2020.624137

Zhiqiang Guo, Zhaoci Liu, Zhenhua Ling, Shijin Wang,

Lingjing Jin, and Yunxia Li. 2020. Text Classification

by Contrastive Learning and Cross-lingual Data

Augmentation for Alzheimer’s Disease Detection. In

Proceedings of the 28th International Conference on

Computational Linguistics, pages 6161–6171,

Barcelona, Spain (Online). International Committee on

Computational Linguistics.

Chloé Pou-Prom and Frank Rudzicz. 2018. Learning

multiview embeddings for assessing dementia. In

Proceedings of the 2018 Conference on Empirical

Methods in Natural Language Processing, pages 2812–

2817, Brussels, Belgium. Association for

Computational Linguistics.

Felipe Paula, Rodrigo Wilkens, Marco Idiart, and Aline

Villavicencio. 2018. Similarity Measures for the

Detection of Clinical Conditions with Verbal Fluency

Tasks. In Proceedings of the 2018 Conference of the

North American Chapter of the Association for

Computational Linguistics: Human Language

Technologies, Volume 2 (Short Papers), pages 231–

235, New Orleans, Louisiana. Association for

Computational Linguistics.

Flavio Di Palo and Natalie Parde. 2019. Enriching Neural

Models with Targeted Features for Dementia

Detection. In Proceedings of the 57th Annual Meeting

of the Association for Computational Linguistics:

Student Research Workshop, pages 302–308, Florence,

Italy. Association for Computational Linguistics.

Zhou, Luke & Fraser, Kathleen & Rudzicz, Frank. (2016).

Speech Recognition in Alzheimer’s Disease and in its

Assessment. 1948-1952. 10.21437/Interspeech.2016-

1228.

Yamada Y, Shinkawa K, Kobayashi M, Nishimura M,

Nemoto M, Tsukada E, (2021), “Tablet-based

automatic assessment for early detection of

Alzheimer’s disease using speech responses to daily life

questions.” Front. Digit. Health 3:653904. doi:

10.3389/fdgth.2021.653904

Chen L, Dodge HH, Asgari M. Topic-Based Measures of

Conversation for Detecting Mild Cognitive

Impairment. Proc Conf Assoc Comput Linguist Meet.

2020 Jul;2020:63-67. PMID: 33642674; PMCID:

PMC7909094.

Sweta Karlekar, Tong Niu, and Mohit Bansal. 2018.

Detecting Linguistic Characteristics of Alzheimer’s

Dementia by Interpreting Neural Models. In

Proceedings of the 2018 Conference of the North

American Chapter of the Association for

Computational Linguistics: Human Language

Technologies, Volume 2 (Short Papers), pages 701–

707, New Orleans, Louisiana. Association for

Computational Linguistics.

ICAART 2023 - 15th International Conference on Agents and Artiﬁcial Intelligence

812