| { | |
| "paper_id": "2020", | |
| "header": { | |
| "generated_with": "S2ORC 1.0.0", | |
| "date_generated": "2023-01-19T02:12:21.752841Z" | |
| }, | |
| "title": "Analysis of Hierarchical Multi-Content Text Classification Model on B-SHARP Dataset for Early Detection of Alzheimer's Disease", | |
| "authors": [ | |
| { | |
| "first": "Renxuan", | |
| "middle": [ | |
| "A" | |
| ], | |
| "last": "Li", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Emory University Atlanta", | |
| "location": { | |
| "region": "GA", | |
| "country": "USA" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Ihab", | |
| "middle": [], | |
| "last": "Hajjar", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Neurology Emory University Atlanta", | |
| "location": { | |
| "region": "GA", | |
| "country": "USA" | |
| } | |
| }, | |
| "email": "[email protected]" | |
| }, | |
| { | |
| "first": "Felicia", | |
| "middle": [], | |
| "last": "Goldstein", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Neurology Emory University Atlanta", | |
| "location": { | |
| "region": "GA", | |
| "country": "USA" | |
| } | |
| }, | |
| "email": "" | |
| }, | |
| { | |
| "first": "Jinho", | |
| "middle": [ | |
| "D" | |
| ], | |
| "last": "Choi", | |
| "suffix": "", | |
| "affiliation": { | |
| "laboratory": "", | |
| "institution": "Emory University Atlanta", | |
| "location": { | |
| "region": "GA", | |
| "country": "USA" | |
| } | |
| }, | |
| "email": "" | |
| } | |
| ], | |
| "year": "", | |
| "venue": null, | |
| "identifiers": {}, | |
| "abstract": "This paper presents a new dataset, B-SHARP, that can be used to develop NLP models for the detection of Mild Cognitive Impairment (MCI) known as an early sign of Alzheimer's disease. Our dataset contains 1-2 min speech segments from 326 human subjects for 3 topics, (1) daily activity, (2) room environment, and (3) picture description, and their transcripts so that a total of 650 speech segments are collected. Given the B-SHARP dataset, several hierarchical text classification models are developed that jointly learn combinatory features across all 3 topics. The best performance of 74.1% is achieved by an ensemble model that adapts 3 types of transformer encoders. To the best of our knowledge, this is the first work that builds deep learningbased text classification models on multiple contents for the detection of MCI.", | |
| "pdf_parse": { | |
| "paper_id": "2020", | |
| "_pdf_hash": "", | |
| "abstract": [ | |
| { | |
| "text": "This paper presents a new dataset, B-SHARP, that can be used to develop NLP models for the detection of Mild Cognitive Impairment (MCI) known as an early sign of Alzheimer's disease. Our dataset contains 1-2 min speech segments from 326 human subjects for 3 topics, (1) daily activity, (2) room environment, and (3) picture description, and their transcripts so that a total of 650 speech segments are collected. Given the B-SHARP dataset, several hierarchical text classification models are developed that jointly learn combinatory features across all 3 topics. The best performance of 74.1% is achieved by an ensemble model that adapts 3 types of transformer encoders. To the best of our knowledge, this is the first work that builds deep learningbased text classification models on multiple contents for the detection of MCI.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Abstract", | |
| "sec_num": null | |
| } | |
| ], | |
| "body_text": [ | |
| { | |
| "text": "Alzheimer's Disease (AD) is a progressive neurodegenerative disorder that is associated with memory loss and declines in major brain functions including semantic and pragmatic levels of language processing (Vestal et al., 2006; Ferris and Farlow, 2013) . Traditional cognitive assessments such as positron emission tomography or cerebrospinal fluid analysis are expensive and time-consuming (Fyffe et al., 2011) . This may cause delay in treating AD, known to be irreversible and incurable (Korczyn, 2012) , and put an increasing pressure on public health, especially for seniors whose life expectancy is rapidly growing yet are more likely to develop AD. Thus, it is crucial to find a more intelligent way of detecting AD in the earliest stage possible (Karr et al., 2018) .", | |
| "cite_spans": [ | |
| { | |
| "start": 206, | |
| "end": 227, | |
| "text": "(Vestal et al., 2006;", | |
| "ref_id": "BIBREF20" | |
| }, | |
| { | |
| "start": 228, | |
| "end": 252, | |
| "text": "Ferris and Farlow, 2013)", | |
| "ref_id": "BIBREF5" | |
| }, | |
| { | |
| "start": 391, | |
| "end": 411, | |
| "text": "(Fyffe et al., 2011)", | |
| "ref_id": "BIBREF8" | |
| }, | |
| { | |
| "start": 490, | |
| "end": 505, | |
| "text": "(Korczyn, 2012)", | |
| "ref_id": "BIBREF13" | |
| }, | |
| { | |
| "start": 754, | |
| "end": 773, | |
| "text": "(Karr et al., 2018)", | |
| "ref_id": "BIBREF12" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Mild Cognitive Impairment (MCI) is considered the first phase that patients start having biomarker evidence of brain changes that can eventually lead to AD (Albert et al., 2011) . MCI involves subtle language changes from impairment in reasoning that may not be noticeable to people other than friends and relatives. Because of this, the detection of MCI is a much more challenging task than detecting dementia (Suzman and Beard, 2011) . Recent studies in NLP have shown that it is possible to detect early stages of AD by analyzing patients' language patterns; however, most previous works have focused on the detection of dementia instead and researches tackling the detection of MCI have been based on relatively small datasets (Section 2).", | |
| "cite_spans": [ | |
| { | |
| "start": 153, | |
| "end": 177, | |
| "text": "AD (Albert et al., 2011)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 411, | |
| "end": 435, | |
| "text": "(Suzman and Beard, 2011)", | |
| "ref_id": "BIBREF19" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "This paper presents a new dataset that comprises three types of speech segments from both normal controls and MCI patients (Section 3). Then, a hierarchical text classification model is proposed, which jointly learns features from all three types of speech segments to determine whether or not each subject has MCI (Section 4). Individual and ensemble models using three types of transformer encoders are evaluated on our dataset and show that different transformer encoders reveal strengths in distinct types of speeches (Section 5). We believe that this work takes the initiative of deep learningbased NLP for detecting MCI that will be broadly beneficial to global public health.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Introduction", | |
| "sec_num": "1" | |
| }, | |
| { | |
| "text": "Only few studies have tackled the detection of MCI using NLP. 1 Asgari et al. (2017) conducted interviews with (27 C , 14 M ), and developed SVM and random forest models on their transcribed speeches. Beltrami et al. (2018) conducted three speech tasks with (48 C , 32 M , 16 D ), and analyzed phonetic and linguistic features of their speeches and transcripts. Fraser et al. (2019) conducted 3 language tasks with (29 C , 26 M ), and built a cascade model to learn multimodal features such as audio, text, eye-tracking. Gosztolya et al. (2019) ing sessions with (25 C , 25 M , 25 A ), and trained a SVM model using acoustic and linguistic features. All of the previous works were based on fewer than 100 subjects using traditional linguistic features to develop NLP models, compared to our work that is based on 326 subjects and 650 recordings using the latest transformer-based deep neural models. The task of dementia detection has been more explored by the NLP community. Becker et al. (1994) presented the DementiaBank, that consists of 552 audio recordings describing the picture called \"The Boston Cookie Theft\" from 99 normal controls and 194 dementia patients, that have been used by the following works. Orimaye et al. 2016presented deep-deep neural network language models using higher-order n-grams and skip-grams. Pou-Prom and Rudzicz (2018) leveraged linguistic features and multiview embeddings by applying generalized canonical correlation analysis. Karleka et al. (2018) proposed a model based on convolutional and recurrent neural networks and gave interpretations of this model to explain linguistic characteristics for detecting dementia. Our work is distinguished as:", | |
| "cite_spans": [ | |
| { | |
| "start": 62, | |
| "end": 63, | |
| "text": "1", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 64, | |
| "end": 84, | |
| "text": "Asgari et al. (2017)", | |
| "ref_id": "BIBREF1" | |
| }, | |
| { | |
| "start": 201, | |
| "end": 223, | |
| "text": "Beltrami et al. (2018)", | |
| "ref_id": "BIBREF3" | |
| }, | |
| { | |
| "start": 362, | |
| "end": 382, | |
| "text": "Fraser et al. (2019)", | |
| "ref_id": "BIBREF6" | |
| }, | |
| { | |
| "start": 521, | |
| "end": 544, | |
| "text": "Gosztolya et al. (2019)", | |
| "ref_id": "BIBREF9" | |
| }, | |
| { | |
| "start": 976, | |
| "end": 996, | |
| "text": "Becker et al. (1994)", | |
| "ref_id": "BIBREF2" | |
| }, | |
| { | |
| "start": 1466, | |
| "end": 1487, | |
| "text": "Karleka et al. (2018)", | |
| "ref_id": "BIBREF11" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related Work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "\u2022 We tackle the detection of MCI, not dementia,", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related Work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "\u2022 Our documents are multi-contents compared to single-content documents in the DementiaBank.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Related Work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "\u2022 Our approach is based on the latest contextualized embeddings compared to the distributional embeddings adapted by the previous works. Table 2 shows the statistics of the control and the MCI groups in B-SHARP. Note that when subjects make multiple visits, there is a year gap in between so that subjects generally do not remember so much from their previous visits. Thus, speeches from the same subject are not necessarily more similar than ones from the other subjects. In fact, most speeches across subjects, regardless of their groups, are very similar when they are transcribed since all subjects follow the same speech protocol in Section 3.2. 3", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 137, | |
| "end": 144, | |
| "text": "Table 2", | |
| "ref_id": "TABREF3" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Related Work", | |
| "sec_num": "2" | |
| }, | |
| { | |
| "text": "A speech task protocol has been conducted to collect recordings of both control and MCI subjects who are asked to speak about Q 1 : daily activity, Q 2 : room environment, and Q 3 : picture description for 1-2 minutes each. All subjects are provided with the same instructions in A.2, and visual abilities of the subjects are confirmed before recording. To reduce potential variance, the subjects are guided to follow similar activities before Q 1 , located to similar room settings before Q 2 , and shown the same picture in Fig 2, \"The Circus Procession\", for Q 3 . The collected voice recordings are automatically transcribed by the online tool called Temi. 4 Table 1 shows linguistic features about our dataset analyzed by the open-source NLP toolkit, ELIT. 5 Transcripts from the control group depict significantly higher numbers of tokens, nouns, and complex structures while transcripts from the MCI group gives significantly more discourse elements, implying that the control subjects are more expressive while the MCI subjects include more disfluency in their speeches.", | |
| "cite_spans": [ | |
| { | |
| "start": 661, | |
| "end": 662, | |
| "text": "4", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 762, | |
| "end": 763, | |
| "text": "5", | |
| "ref_id": null | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 526, | |
| "end": 532, | |
| "text": "Fig 2,", | |
| "ref_id": "FIGREF2" | |
| }, | |
| { | |
| "start": 663, | |
| "end": 670, | |
| "text": "Table 1", | |
| "ref_id": "TABREF1" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Speech Task Protocol", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "Transformer1 (T1) w 11 w 12 \u22ef w 1n [CLS1] w 21 w 22 \u22ef w 2n [CLS2] w 31 w 32 \u22ef w", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Speech Task Protocol", | |
| "sec_num": "3.2" | |
| }, | |
| { | |
| "text": "Although transformer encoders have recently established the state-of-the-art results on most document classification tasks, they have a limit on the input size. As in Table 1 , the average number of tokens in our input documents well-exceeds 512 when combining transcripts from all three tasks, which is the max-number of tokens that the pretrained models of these transformers allow in general. This makes it difficult to simply join all transcripts together and feed into a transformer encoder. Thus, this section presents a hierarchical transformer to overcome the challenge of long documents while jointly training transcript contents from all three tasks (Figure 1 ).", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 167, | |
| "end": 174, | |
| "text": "Table 1", | |
| "ref_id": "TABREF1" | |
| }, | |
| { | |
| "start": 660, | |
| "end": 669, | |
| "text": "(Figure 1", | |
| "ref_id": "FIGREF0" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Hierarchical Transformer", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "Let W i = {w i1 , . . . , w in } be a transcript, where w ij represents the j'th token in the transcript produced by the i'th task Q i (in our case, i = {1, 2, 3}). W i is prepended by the special token [CLS i ] that is used to learn the transcript embedding, and fed into the transformer T i . The transformer then generates E i = {c i , e i1 , . . . , e in }, where c i and e ij are the embeddings for [CLS i ] and w ij , respectively. c i \u2208 R d is used to make two types of predictions.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Hierarchical Transformer", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "First, c i is fed into a multilayer perceptron layer, MLP i , that generates the output vector o i \u2208 R 2 to predict whether or not the subject has MCI based on the transcript from Q i alone. Second, the transcript embeddings from all three tasks are concatenated such that c e = c 1 \u2295 c 2 \u2295 c 3 \u2208 R 3d , which gets fed into another MLP e to generate the output vector o e \u2208 R 2 , and makes the binary decision based on the transcripts from all three tasks, Q 1 , Q 2 and Q 3 .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Hierarchical Transformer", | |
| "sec_num": "4" | |
| }, | |
| { | |
| "text": "There are 650 recordings in our dataset (Table 2) , that is rather small to divide into train, development, and test sets. Thus, 5-fold cross-validation (CV) is used to evaluate the performance of our models. Table 5 shows the distributions of the five CV sets for our experiments, where the transcript of each recording is treated as an independent document. Notice that the distributions are calculated based on analysis of the last MLP layer instead of simple majority vote on individual models.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 40, | |
| "end": 49, | |
| "text": "(Table 2)", | |
| "ref_id": "TABREF3" | |
| }, | |
| { | |
| "start": 209, | |
| "end": 216, | |
| "text": "Table 5", | |
| "ref_id": "TABREF8" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Experiments", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "It is worth mentioning that all recordings from the same subject given multiple visits are assigned to the same CV set; thus, there is no overlap in terms of subjects across these CV sets. This allows us to avoid potential inflation in accuracy due to unique language patterns used by individual subjects.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experiments", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "BERT RoBERTa ALBERT Q 1 Q 2 Q 3 Q 1 Q 2 Q 3 Q 1 Q 2", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experiments", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "Q 3 ACC 67.6 (\u00b10.4) 69.0 (\u00b11.2) 67.7 (\u00b10.7) 69.0 (\u00b11.5) 69.9 (\u00b10.2) 65.2 (\u00b10.3) 67.6 (\u00b11.5) 69.5 (\u00b10.3) 66.6 (\u00b11.3) SEN 48.9 (\u00b11.8) 57.1 (\u00b12.5) 41.5 (\u00b13.6) 44.3 (\u00b14.5) 55.3 (\u00b11.2) 37.1 (\u00b13.7) 45.9 (\u00b11.9) 52.2 (\u00b10.6) 37.4 (\u00b13.3) SPE 80.4 (\u00b11.2) 77.3 (\u00b12.8) 85.2 (\u00b13.0) 85.8 (\u00b12.1) 79.7 (\u00b10.7) 84.5 (\u00b13.0) 82.6 (\u00b13.7) 81.4 (\u00b10.3) 86.8 (\u00b13.3) Table 3 , respectively. B e +R e uses transcript embeddings from both Bert e and RoBERTa e (so the total of 6 embeddings), A e +R e uses transcript embeddings from both ALBERT e and RoBERTa e (6 embeddings), and B e +A e +R e uses transcript embeddings from all three models (9 embeddings).", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 340, | |
| "end": 347, | |
| "text": "Table 3", | |
| "ref_id": "TABREF5" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Experiments", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "Three transformer encoders are used, BERT (Devlin et al., 2019) , RoBERTa (Liu et al., 2020) , and ALBERT (Lan et al., 2019) for our experiments. Every model is trained 3 times and its average performance with the standard deviation are reported. 6 ", | |
| "cite_spans": [ | |
| { | |
| "start": 42, | |
| "end": 63, | |
| "text": "(Devlin et al., 2019)", | |
| "ref_id": "BIBREF4" | |
| }, | |
| { | |
| "start": 74, | |
| "end": 92, | |
| "text": "(Liu et al., 2020)", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 106, | |
| "end": 124, | |
| "text": "(Lan et al., 2019)", | |
| "ref_id": "BIBREF14" | |
| } | |
| ], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Experiments", | |
| "sec_num": "5" | |
| }, | |
| { | |
| "text": "Individual models are built by training transcripts from each task separately using MLP i in Section 4. Table 3 shows the performance of the 3 transformer models on the individual tasks. The performance on Q 2 shows the highest accuracy for all three models, achieving 69.9% with RoBERTa, implying that the room environment task of Q 2 , involving many spatial descriptions, are the most effective to distinguish the MCI group. The highest sensitivity of 57.1% is achieved by BERT on Q 2 , and the highest specificity of 86.8% is achieved by ALBERT on Q 3 . Such a low sensitivity and a high specificity imply that it is easier to recognize the normal controls but not the MCI patients given the short speeches.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 104, | |
| "end": 111, | |
| "text": "Table 3", | |
| "ref_id": "TABREF5" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Performance of Individual Models", | |
| "sec_num": "5.1" | |
| }, | |
| { | |
| "text": "Ensemble models are developed by jointly training multiple transcript embeddings from the individual models using MLP e in Section 4. Table 4 shows the 6 Details about the experimental settings are provided in A.1.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 134, | |
| "end": 141, | |
| "text": "Table 4", | |
| "ref_id": "TABREF6" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Performance of Ensemble Models", | |
| "sec_num": "5.2" | |
| }, | |
| { | |
| "text": "model performance of the ensemble models. Additionally, results from a model that takes transcripts from the 3 tasks as one input document and trains a convolutional neural network (CNN) are provided for comparison to Karleka et al. (2018) . 7 R e shows 1.7% improvement on accuracy over the RoBERTa model in Table 3 although its sensitivity is worse. Table 6 shows the voting distributions of each task combination; given the samples correctly predicted by RoBERTa e , we count how often the individual models are correct for those samples by comparing the weights in MLP e and estimate the percentages. The combination of (Q 1 , Q 3 ) shows the highest percentage of 30%, meaning that 30% of the corrected predicted samples are voted by both Q 1 and Q 3 . Table 6 : Voting distributions of each task combination for RoBERTa e . Q i : % of only the Q i model is correct, Q i,i,j : % of all Q i , Q i , and Q j models are correct.", | |
| "cite_spans": [ | |
| { | |
| "start": 218, | |
| "end": 239, | |
| "text": "Karleka et al. (2018)", | |
| "ref_id": "BIBREF11" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 309, | |
| "end": 316, | |
| "text": "Table 3", | |
| "ref_id": "TABREF5" | |
| }, | |
| { | |
| "start": 352, | |
| "end": 359, | |
| "text": "Table 6", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 758, | |
| "end": 765, | |
| "text": "Table 6", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Performance of Ensemble Models", | |
| "sec_num": "5.2" | |
| }, | |
| { | |
| "text": "A similar analysis is done for B e +R e +A e although displaying the distributions is quite infeasible since it involves 2 9 -1 combinations. Among the samples correctly predicted by B e +R e +A e , 86% are derived from majority votes; in other words, at least 5 out of 9 individual models agree with the predictions. Votes from 6 and 5 models are the largest groups, showing 35% and 28%, respectively. Only 0.21% are agreed by all 9 models. No case of votes from 3 or less models is found, implying that no individual model dominates the final decision of B e +R e +A e .", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Performance of Ensemble Models", | |
| "sec_num": "5.2" | |
| }, | |
| { | |
| "text": "This paper presents the B-SHARP dataset, that is the largest dataset for the task of MCI detection feasible to develop robust deep neural models. Our best ensemble model using hierarchical transformer gives the accuracy of 74% to distinguish MCI patients from normal controls that is very promising. We will also explore models to make a longevity analysis per patient with this dataset. 8", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Conclusion", | |
| "sec_num": "6" | |
| }, | |
| { | |
| "text": "A.1 Experimental Settings Table 7 shows the configuration of the transformer encoders in Section 5. The base pre-trained models are used for all encoders. Individual Models For training the BERT and RoBERTa models, the batch size of 5, the learning rate of 5 \u2022 10 \u22126 , and the gradient clip of norm 0.5 are used with the Adam optimizer. A dropout rate of 0.15 is applied to all layers. For the ALBERT model, the batch size of 8 is used. All three models are trained for 30 epochs.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 26, | |
| "end": 33, | |
| "text": "Table 7", | |
| "ref_id": "TABREF10" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "A Appendix", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Ensemble Models For training the two model ensembles, B e +R e and A e +R e , the batch size of 72 and the learning rate of 5 \u2022 10 \u22125 are used with the Adam optimizer for 200 epochs. A dropout rate of 0.25 is also applied. For training the B e +A e +R e model, the dropout rate is set to 0.3. Table 8 describes the instructions provided to the subjects for the three speech tasks in Section 3.2.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 293, | |
| "end": 300, | |
| "text": "Table 8", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "A Appendix", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "I would like you to describe to me everything we did from the moment we met today until now. Please try to recall as many details as possible in the order the events actually happened where we met, what we did, what we saw, where we went, and what you felt or thought during each of these events.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Q1", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "I would like you to describe everything that you see in this room.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "Q2", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "I am going to show you a picture and ask you to describe what you see in as much detail as possible. You can describe the activities, characters, and colors of things you see in this picture. Table 8 : Instructions of the 3 speech tasks, Q 1 , Q 2 , Q 3 , provided to the subjects. Figure 2 illustrates the image of the picture called \"The Circus Procession\" for the picture description task, Q 3 , copyrighted by the McLoughlin Brothers as part of the Juvenile Collection. ", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 192, | |
| "end": 199, | |
| "text": "Table 8", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 282, | |
| "end": 290, | |
| "text": "Figure 2", | |
| "ref_id": "FIGREF2" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Q3", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "DementiaBank is the largest public dataset for dementia detection that comprises recordings for 4 language tasks, picture description, verbal fluency, story recall, and sentence construction, from a large longitudinal study (Becker et al., 1994) . Subjects in this study are divided into two groups, normal controls and dementia patients. Among the four tasks, data from only the picture description task can be used for classification since the other tasks give data of dementia patients only. 9 The design of this task is similar to Q 3 in B-SHARP (Section 3.2); each subject is shown \"The Boston Cookie Theft\" picture in Figure 3 to describe for 1-2 minutes. Table 10 shows the statistics of the DementiaBank in comparison to Table 2 in Section 3. Subjects in this study made up to 5 visits compared to 3 in B-SHARP although the number of subjects in each visit is larger in B-SHARP. B-SHARP has \u2248100", | |
| "cite_spans": [ | |
| { | |
| "start": 224, | |
| "end": 245, | |
| "text": "(Becker et al., 1994)", | |
| "ref_id": "BIBREF2" | |
| } | |
| ], | |
| "ref_spans": [ | |
| { | |
| "start": 624, | |
| "end": 632, | |
| "text": "Figure 3", | |
| "ref_id": "FIGREF3" | |
| }, | |
| { | |
| "start": 662, | |
| "end": 670, | |
| "text": "Table 10", | |
| "ref_id": "TABREF1" | |
| }, | |
| { | |
| "start": 729, | |
| "end": 736, | |
| "text": "Table 2", | |
| "ref_id": "TABREF3" | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "A.3 B-SHARP Compared to DementiaBank", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "#C: the number of normal controls, #M/D/A: the number of MCI / Dementia / AD patients.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "A.3 compares B-SHARP with the DementiaBank in details.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "Temi: https://www.temi.com 5 ELIT: https://github.com/elitcloud/elit", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "We also experiemented with LSTM-RNN and CNN-LSTM models as suggested byKarleka et al. (2018); however, the CNN model gave the highest accuracy on our dataset.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| }, | |
| { | |
| "text": "The verbal fluency task gives 1 audio recording of a normal control, that is still not enough to train classification models.", | |
| "cite_spans": [], | |
| "ref_spans": [], | |
| "eq_spans": [], | |
| "section": "", | |
| "sec_num": null | |
| } | |
| ], | |
| "back_matter": [ | |
| { | |
| "text": "SentencesNouns Verbs Conjuncts Complex Discourse Control124.0 (\u00b159.7) 12.6 (\u00b15.1) 23.7 (\u00b111.8) 27.1 (\u00b111.9) 2.8 (\u00b12.8) 1.6 (\u00b11.6) 1.5 (\u00b11.6) Dementia 114.3 (\u00b161.3) 12.1 (\u00b16.4) 18.7 (\u00b110.4) 23.9 (\u00b112.9) 2.4 (\u00b12.4) 1.4 (\u00b11.4) 2.8 (\u00b12.9) p 0.0625 0.3204 < 0.0001 0.0029 0.0715 0.1184 < 0.0001 Table 9 shows the statistics of linguistic features in comparison to Table 1 in Section 3. The same tools, Temi and ELIT, are used to measure them. Unlike B-SHARP, the control group in the Demen-tiaBank does not reveal a significantly greater number of tokens than the dementia group. The document size in the DementiaBank is 4.9 times smaller than B-SHARP on average. In both datasets, the noun and discourse counts are significantly different between the control and the other groups. It is interesting that a significant difference is found in verbs whereas it is not the case for complex structures in the DementiaBank, which is opposite in B-SHARP. This may imply that the verb usage deteriorates as it progresses from MCI to dementia, but more thorough research is needed for further verification.", | |
| "cite_spans": [], | |
| "ref_spans": [ | |
| { | |
| "start": 290, | |
| "end": 297, | |
| "text": "Table 9", | |
| "ref_id": null | |
| }, | |
| { | |
| "start": 359, | |
| "end": 366, | |
| "text": "Table 1", | |
| "ref_id": null | |
| } | |
| ], | |
| "eq_spans": [], | |
| "section": "Tokens", | |
| "sec_num": null | |
| } | |
| ], | |
| "bib_entries": { | |
| "BIBREF0": { | |
| "ref_id": "b0", | |
| "title": "The Diagnosis of Mild Cognitive Impairment Due to Alzheimer's Disease: Recommendations From the National Institute on Aging-Alzheimer's Association Workgroups on Diagnostic Guidelines for Alzheimer's Disease", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Marilyn", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Albert", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "T", | |
| "middle": [], | |
| "last": "Steven", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Dennis", | |
| "middle": [], | |
| "last": "Dekosky", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Bruno", | |
| "middle": [], | |
| "last": "Dickson", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Dubois", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Howard", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Nick", | |
| "middle": [ | |
| "C" | |
| ], | |
| "last": "Feldman", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Anthony", | |
| "middle": [], | |
| "last": "Fox", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Gamst", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "David", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Holtzman", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "William", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Jagust", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "C", | |
| "middle": [], | |
| "last": "Ronald", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Petersen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J", | |
| "middle": [], | |
| "last": "Peter", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Maria", | |
| "middle": [ | |
| "C" | |
| ], | |
| "last": "Snyder", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Bill", | |
| "middle": [], | |
| "last": "Carrillo", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Creighton H", | |
| "middle": [], | |
| "last": "Thies", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Phelps", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "Alzheimer's Dementia", | |
| "volume": "7", | |
| "issue": "3", | |
| "pages": "270--279", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Marilyn S Albert, Steven T DeKosky, Dennis Dickson, Bruno Dubois, Howard H Feldman, Nick C Fox, Anthony Gamst, David M Holtzman, William J Ja- gust, Ronald C Petersen, Peter J Snyder, Maria C Carrillo, Bill Thies, and Creighton H Phelps. 2011. The Diagnosis of Mild Cognitive Impairment Due to Alzheimer's Disease: Recommendations From the National Institute on Aging-Alzheimer's As- sociation Workgroups on Diagnostic Guidelines for Alzheimer's Disease. Alzheimer's Dementia, 7(3):270-279.", | |
| "links": null | |
| }, | |
| "BIBREF1": { | |
| "ref_id": "b1", | |
| "title": "Predicting Mild Cognitive Impairment From Spontaneous Spoken Utterances", | |
| "authors": [ | |
| { | |
| "first": "Meysam", | |
| "middle": [], | |
| "last": "Asgari", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jeffrey", | |
| "middle": [], | |
| "last": "Kaye", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Hiroko", | |
| "middle": [], | |
| "last": "Dodge", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2017, | |
| "venue": "", | |
| "volume": "3", | |
| "issue": "", | |
| "pages": "219--228", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Meysam Asgari, Jeffrey Kaye, and Hiroko Dodge. 2017. Predicting Mild Cognitive Impairment From Spontaneous Spoken Utterances. Alzheimer's De- mentia, 3(2):219-228.", | |
| "links": null | |
| }, | |
| "BIBREF2": { | |
| "ref_id": "b2", | |
| "title": "The natural history of alzheimer's disease: description of study cohort and accuracy of diagnosis", | |
| "authors": [ | |
| { | |
| "first": "James", | |
| "middle": [ | |
| "T" | |
| ], | |
| "last": "Becker", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Francois", | |
| "middle": [], | |
| "last": "Boller", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Oscar", | |
| "middle": [ | |
| "L" | |
| ], | |
| "last": "Lopez", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Judith", | |
| "middle": [], | |
| "last": "Saxton", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Karen", | |
| "middle": [ | |
| "L" | |
| ], | |
| "last": "Mcgonigle", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1994, | |
| "venue": "Archives of Neurology", | |
| "volume": "51", | |
| "issue": "6", | |
| "pages": "585--594", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "James T. Becker, Francois Boller, Oscar L. Lopez, Ju- dith Saxton, and Karen L. McGonigle. 1994. The natural history of alzheimer's disease: description of study cohort and accuracy of diagnosis. Archives of Neurology, 51(6):585-594.", | |
| "links": null | |
| }, | |
| "BIBREF3": { | |
| "ref_id": "b3", | |
| "title": "Speech Analysis by Natural Language Processing Techniques: A Possible Tool for Very Early Detection of Cognitive Decline? Frontiers in Aging Neuroscience", | |
| "authors": [ | |
| { | |
| "first": "Daniela", | |
| "middle": [], | |
| "last": "Beltrami", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Gloria", | |
| "middle": [], | |
| "last": "Gagliardi", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Rema", | |
| "middle": [ | |
| "Rossini" | |
| ], | |
| "last": "Favretti", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Enrico", | |
| "middle": [], | |
| "last": "Ghidoni", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Fabio", | |
| "middle": [], | |
| "last": "Tamburini", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Laura", | |
| "middle": [], | |
| "last": "Calz\u00e0", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "", | |
| "volume": "10", | |
| "issue": "", | |
| "pages": "1--13", | |
| "other_ids": { | |
| "DOI": [ | |
| "https://www.frontiersin.org/articles/10.3389/fnagi.2018.00369/full" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Daniela Beltrami, Gloria Gagliardi, Rema Rossini Favretti, Enrico Ghidoni, Fabio Tamburini, and Laura Calz\u00e0. 2018. Speech Analysis by Natural Lan- guage Processing Techniques: A Possible Tool for Very Early Detection of Cognitive Decline? Fron- tiers in Aging Neuroscience, 10(369):1-13.", | |
| "links": null | |
| }, | |
| "BIBREF4": { | |
| "ref_id": "b4", | |
| "title": "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding", | |
| "authors": [ | |
| { | |
| "first": "Jacob", | |
| "middle": [], | |
| "last": "Devlin", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ming-Wei", | |
| "middle": [], | |
| "last": "Chang", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Kenton", | |
| "middle": [], | |
| "last": "Lee", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Kristina", | |
| "middle": [], | |
| "last": "Toutanova", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "4171--4186", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. BERT: Pre-training of Deep Bidirectional Transformers for Language Un- derstanding. In Proceedings of the Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Tech- nologies, pages 4171-4186.", | |
| "links": null | |
| }, | |
| "BIBREF5": { | |
| "ref_id": "b5", | |
| "title": "Language Impairment in Alzheimer's Disease and Benefits of Acetylcholinesterase Inhibitors", | |
| "authors": [ | |
| { | |
| "first": "H", | |
| "middle": [], | |
| "last": "Steven", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Martin", | |
| "middle": [], | |
| "last": "Ferris", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Farlow", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2013, | |
| "venue": "Clinical Interventions in Aging", | |
| "volume": "8", | |
| "issue": "", | |
| "pages": "1007--1014", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Steven H Ferris and Martin Farlow. 2013. Language Impairment in Alzheimer's Disease and Benefits of Acetylcholinesterase Inhibitors. Clinical Interven- tions in Aging, 8:1007-1014.", | |
| "links": null | |
| }, | |
| "BIBREF6": { | |
| "ref_id": "b6", | |
| "title": "Predicting MCI Status From Multimodal Language Data Using Cascaded Classifiers", | |
| "authors": [ | |
| { | |
| "first": "Kathleen", | |
| "middle": [ | |
| "C" | |
| ], | |
| "last": "Fraser", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Kristina", | |
| "middle": [ | |
| "Lundholm" | |
| ], | |
| "last": "Fors", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Marie", | |
| "middle": [], | |
| "last": "Eckerstr\u00f6m", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "Frontiers in Aging Neuroscience", | |
| "volume": "11", | |
| "issue": "205", | |
| "pages": "1--18", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Kathleen C. Fraser, Kristina Lundholm Fors, Marie Eckerstr\u00f6m, Fredrik \u00d6hman, and Dimitrios Kokki- nakis. 2019. Predicting MCI Status From Multi- modal Language Data Using Cascaded Classifiers. Frontiers in Aging Neuroscience, 11(205):1-18.", | |
| "links": null | |
| }, | |
| "BIBREF7": { | |
| "ref_id": "b7", | |
| "title": "All our resources including source codes and models are available at http://anonymous", | |
| "authors": [], | |
| "year": null, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "All our resources including source codes and models are available at http://anonymous.", | |
| "links": null | |
| }, | |
| "BIBREF8": { | |
| "ref_id": "b8", | |
| "title": "Explaining Differences in Episodic Memory Performance among Older African Americans and Whites: The Roles of Factors Related to Cognitive Reserve and Test Bias", | |
| "authors": [ | |
| { | |
| "first": "Denise", | |
| "middle": [ | |
| "C" | |
| ], | |
| "last": "Fyffe", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Shubhabrata", | |
| "middle": [], | |
| "last": "Mukherjee", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Lisa", | |
| "middle": [ | |
| "L" | |
| ], | |
| "last": "Barnes", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jennifer", | |
| "middle": [ | |
| "J" | |
| ], | |
| "last": "Manly", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "David", | |
| "middle": [ | |
| "A" | |
| ], | |
| "last": "Bennett", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Paul", | |
| "middle": [ | |
| "K" | |
| ], | |
| "last": "Crane", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "Journal of the International Neuropsychological Society", | |
| "volume": "17", | |
| "issue": "4", | |
| "pages": "625--638", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Denise C. Fyffe, Shubhabrata Mukherjee, Lisa L. Barnes, Jennifer J. Manly, David A. Bennett, and Paul K. Crane. 2011. Explaining Differences in Episodic Memory Performance among Older African Americans and Whites: The Roles of Fac- tors Related to Cognitive Reserve and Test Bias. Journal of the International Neuropsychological So- ciety, 17(4):625-638.", | |
| "links": null | |
| }, | |
| "BIBREF9": { | |
| "ref_id": "b9", | |
| "title": "Identifying Mild Cognitive Impairment and mild Alzheimer's disease based on spontaneous speech using ASR and linguistic features", | |
| "authors": [ | |
| { | |
| "first": "G\u00e1bor", | |
| "middle": [], | |
| "last": "Gosztolya", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Veronika", | |
| "middle": [], | |
| "last": "Vinczea", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "L\u00e1szl\u00f3", | |
| "middle": [], | |
| "last": "T\u00f3th", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Magdolna", | |
| "middle": [], | |
| "last": "P\u00e1k\u00e1skid", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "J\u00e1nos", | |
| "middle": [], | |
| "last": "K\u00e1lmand", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Ildik\u00f3", | |
| "middle": [], | |
| "last": "Hoffmann", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2019, | |
| "venue": "Computer Speech & Language", | |
| "volume": "53", | |
| "issue": "", | |
| "pages": "181--197", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "G\u00e1bor Gosztolya, Veronika Vinczea, L\u00e1szl\u00f3 T\u00f3th, Mag- dolna P\u00e1k\u00e1skid, J\u00e1nos K\u00e1lmand, and Ildik\u00f3 Hoff- mann. 2019. Identifying Mild Cognitive Impairment and mild Alzheimer's disease based on spontaneous speech using ASR and linguistic features. Computer Speech & Language, 53:181-197.", | |
| "links": null | |
| }, | |
| "BIBREF11": { | |
| "ref_id": "b11", | |
| "title": "Detecting linguistic characteristics of alzheimer's dementia by interpreting neural models", | |
| "authors": [ | |
| { | |
| "first": "Sweta", | |
| "middle": [], | |
| "last": "Karleka", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Tong", | |
| "middle": [], | |
| "last": "Niu", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Mohit", | |
| "middle": [], | |
| "last": "Bansal", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "701--707", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.18653/v1/N18-2110" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Sweta Karleka, Tong Niu, and Mohit Bansal. 2018. De- tecting linguistic characteristics of alzheimer's de- mentia by interpreting neural models. In Proceed- ings of the 2018 Conference of the North Ameri- can Chapter of the Association for Computational Linguistics: Human Language Technologies, Vol- ume 2 (Short Papers), pages 701-707, New Orleans, Louisiana. Association for Computational Linguis- tics.", | |
| "links": null | |
| }, | |
| "BIBREF12": { | |
| "ref_id": "b12", | |
| "title": "When Does Cognitive Decline Begin? A Systematic Review of Change Point Studies on Accelerated Decline in Cognitive and Neurological Outcomes Preceding Mild Cognitive Impairment, Dementia, and Death", | |
| "authors": [ | |
| { | |
| "first": "Justin", | |
| "middle": [ | |
| "E" | |
| ], | |
| "last": "Karr", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Raquel", | |
| "middle": [ | |
| "B" | |
| ], | |
| "last": "Graham", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "M", | |
| "middle": [], | |
| "last": "Scott", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Graciela", | |
| "middle": [], | |
| "last": "Hofer", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Muniz-Terrera", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Psychology and Aging", | |
| "volume": "33", | |
| "issue": "2", | |
| "pages": "195--218", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Justin E Karr, Raquel B Graham, Scott M Hofer, and Graciela Muniz-Terrera. 2018. When Does Cog- nitive Decline Begin? A Systematic Review of Change Point Studies on Accelerated Decline in Cognitive and Neurological Outcomes Preceding Mild Cognitive Impairment, Dementia, and Death. Psychology and Aging, 33(2):195-218.", | |
| "links": null | |
| }, | |
| "BIBREF13": { | |
| "ref_id": "b13", | |
| "title": "Why Have We Failed to Cure Alzheimer's Disease?", | |
| "authors": [ | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Amos D Korczyn", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2012, | |
| "venue": "Journal of Alzheimer's Disease", | |
| "volume": "29", | |
| "issue": "2", | |
| "pages": "275--282", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Amos D Korczyn. 2012. Why Have We Failed to Cure Alzheimer's Disease? Journal of Alzheimer's Dis- ease, 29(2):275-282.", | |
| "links": null | |
| }, | |
| "BIBREF14": { | |
| "ref_id": "b14", | |
| "title": "ALBERT: A Lite BERT for Self-supervised Learning of Language Representations", | |
| "authors": [ | |
| { | |
| "first": "Zhenzhong", | |
| "middle": [], | |
| "last": "Lan", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Mingda", | |
| "middle": [], | |
| "last": "Chen", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Sebastian", | |
| "middle": [], | |
| "last": "Goodman", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Kevin", | |
| "middle": [], | |
| "last": "Gimpel", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Piyush", | |
| "middle": [], | |
| "last": "Sharma", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Radu", | |
| "middle": [], | |
| "last": "Soricut", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 1909, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, and Radu Soricut. 2019. ALBERT: A Lite BERT for Self-supervised Learning of Language Representations. arXiv, 11942(1909).", | |
| "links": null | |
| }, | |
| "BIBREF16": { | |
| "ref_id": "b16", | |
| "title": "The Montreal Cognitive Assessment, MoCA: a brief screening tool for mild cognitive impairment", | |
| "authors": [ | |
| { | |
| "first": "S", | |
| "middle": [], | |
| "last": "Ziad", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Natalie", | |
| "middle": [ | |
| "A" | |
| ], | |
| "last": "Nasreddine", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Val\u00e9rie", | |
| "middle": [], | |
| "last": "Phillips", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Simon", | |
| "middle": [], | |
| "last": "B\u00e9dirian", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Victor", | |
| "middle": [], | |
| "last": "Charbonneau", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Isabelle", | |
| "middle": [], | |
| "last": "Whitehead", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Jeffrey", | |
| "middle": [ | |
| "L" | |
| ], | |
| "last": "Collin", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Howard", | |
| "middle": [], | |
| "last": "Cummings", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Chertkow", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2005, | |
| "venue": "Journal of the American Geriatrics Society", | |
| "volume": "53", | |
| "issue": "4", | |
| "pages": "695--699", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Ziad S. Nasreddine, Natalie A. Phillips, Val\u00e9rie B\u00e9dirian, Simon Charbonneau, Victor Whitehead, Isabelle Collin, Jeffrey L. Cummings, and Howard Chertkow. 2005. The Montreal Cognitive Assess- ment, MoCA: a brief screening tool for mild cogni- tive impairment. Journal of the American Geriatrics Society, 53(4):695-699.", | |
| "links": null | |
| }, | |
| "BIBREF17": { | |
| "ref_id": "b17", | |
| "title": "Deep-Deep Neural Network Language Models for Predicting Mild Cognitive Impairment", | |
| "authors": [ | |
| { | |
| "first": "Jojo", | |
| "middle": [], | |
| "last": "Sylvester Olubolu Orimaye", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Sze-Meng", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Judyanne Sharmini Gilbert", | |
| "middle": [], | |
| "last": "Wong", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "", | |
| "middle": [], | |
| "last": "Fernandez", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2016, | |
| "venue": "Proceedings of the IJCAI Workshop on Advances in Bioinformatics and Artificial Intelligence", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "14--20", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Sylvester Olubolu Orimaye, Jojo Sze-Meng Wong, and Judyanne Sharmini Gilbert Fernandez. 2016. Deep- Deep Neural Network Language Models for Predict- ing Mild Cognitive Impairment. In Proceedings of the IJCAI Workshop on Advances in Bioinformatics and Artificial Intelligence, pages 14-20.", | |
| "links": null | |
| }, | |
| "BIBREF18": { | |
| "ref_id": "b18", | |
| "title": "Learning multiview embeddings for assessing dementia", | |
| "authors": [ | |
| { | |
| "first": "Chlo\u00e9", | |
| "middle": [], | |
| "last": "Pou", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "-", | |
| "middle": [], | |
| "last": "Prom", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Frank", | |
| "middle": [], | |
| "last": "Rudzicz", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2018, | |
| "venue": "Proceedings of the Conference on Empirical Methods in Natural Language Processing", | |
| "volume": "18", | |
| "issue": "", | |
| "pages": "2812--2817", | |
| "other_ids": { | |
| "DOI": [ | |
| "10.18653/v1/D18-1304" | |
| ] | |
| }, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Chlo\u00e9 Pou-Prom and Frank Rudzicz. 2018. Learning multiview embeddings for assessing dementia. In Proceedings of the Conference on Empirical Meth- ods in Natural Language Processing, EMNLP'18, pages 2812-2817, Brussels, Belgium. Association for Computational Linguistics.", | |
| "links": null | |
| }, | |
| "BIBREF19": { | |
| "ref_id": "b19", | |
| "title": "Global health and aging", | |
| "authors": [ | |
| { | |
| "first": "Richard", | |
| "middle": [], | |
| "last": "Suzman", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "John", | |
| "middle": [], | |
| "last": "Beard", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2011, | |
| "venue": "", | |
| "volume": "", | |
| "issue": "", | |
| "pages": "", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Richard Suzman and John Beard. 2011. Global health and aging.", | |
| "links": null | |
| }, | |
| "BIBREF20": { | |
| "ref_id": "b20", | |
| "title": "Efficacy of Language Assessment in Alzheimer's Disease: Comparing In-Person Examination and Telemedicine", | |
| "authors": [ | |
| { | |
| "first": "Lindsey", | |
| "middle": [], | |
| "last": "Vestal", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Laura", | |
| "middle": [], | |
| "last": "Smith-Olinde", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Gretchen", | |
| "middle": [], | |
| "last": "Hicks", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "Terri", | |
| "middle": [], | |
| "last": "Hutton", | |
| "suffix": "" | |
| }, | |
| { | |
| "first": "John", | |
| "middle": [], | |
| "last": "Hart", | |
| "suffix": "" | |
| } | |
| ], | |
| "year": 2006, | |
| "venue": "Clinical Interventions in Aging", | |
| "volume": "1", | |
| "issue": "4", | |
| "pages": "467--471", | |
| "other_ids": {}, | |
| "num": null, | |
| "urls": [], | |
| "raw_text": "Lindsey Vestal, Laura Smith-Olinde, Gretchen Hicks, Terri Hutton, and John Hart Jr. 2006. Ef- ficacy of Language Assessment in Alzheimer's Disease: Comparing In-Person Examination and Telemedicine. Clinical Interventions in Aging, 1(4):467-471.", | |
| "links": null | |
| } | |
| }, | |
| "ref_entries": { | |
| "FIGREF0": { | |
| "uris": null, | |
| "text": "Overview of hierarchical transformer to combine content features from the three types of speech tasks.", | |
| "type_str": "figure", | |
| "num": null | |
| }, | |
| "FIGREF2": { | |
| "uris": null, | |
| "text": "The picture of \"The Circus Procession\" used in the B-SHARP dataset.", | |
| "type_str": "figure", | |
| "num": null | |
| }, | |
| "FIGREF3": { | |
| "uris": null, | |
| "text": "The picture of \"The Boston Cookie Theft\" used in the DementiaBank.", | |
| "type_str": "figure", | |
| "num": null | |
| }, | |
| "TABREF0": { | |
| "type_str": "table", | |
| "html": null, | |
| "num": null, | |
| "text": "conducted question answer-", | |
| "content": "<table><tr><td/><td/><td>Tokens</td><td>Sentences</td><td>Nouns</td><td>Verbs</td><td>Conjuncts</td><td>Complex</td><td>Discourse</td></tr><tr><td>Q 1</td><td colspan=\"2\">Control 186.6 (\u00b160.4) MCI 175.6 (\u00b154.5)</td><td>10.4 (\u00b14.5) 9.8 (\u00b14.1)</td><td>28.1 (\u00b19.6) 23.7 (\u00b18.3)</td><td>30.4 (\u00b111.5) 29.3 (\u00b110.4)</td><td>8.5 (\u00b14.5) 8.5 (\u00b14.2)</td><td>2.3 (\u00b11.7) 2.0 (\u00b11.6)</td><td>8.1 (\u00b15.4) 9.2 (\u00b16.0)</td></tr><tr><td>Q 2</td><td colspan=\"2\">Control 191.5 (\u00b111.8) MCI 178.6 (\u00b111.7)</td><td>11.7 (\u00b14.7) 11.6 (\u00b14.7)</td><td colspan=\"2\">41.1 (\u00b113.3) 24.3 (\u00b111.2) 36.7 (\u00b112.1) 23.2 (\u00b110.6)</td><td>6.6 (\u00b14.5) 6.4 (\u00b14.4)</td><td>3.6 (\u00b12.7) 2.9 (\u00b12.3)</td><td>7.1 (\u00b14.8) 8.4 (\u00b15.3)</td></tr><tr><td>Q 3</td><td colspan=\"2\">Control 193.4 (\u00b163.4) MCI 187.8 (\u00b163.4)</td><td>12.6 (\u00b15.4) 12.7 (\u00b15.1)</td><td colspan=\"2\">39.5 (\u00b113.5) 28.4 (\u00b110.1) 36.2 (\u00b113.2) 27.7 (\u00b110.9)</td><td>8.0 (\u00b14.8) 7.2 (\u00b14.2)</td><td>3.3 (\u00b12.1) 2.6 (\u00b12.0)</td><td>6.1 (\u00b15.5) 7.3 (\u00b15.5)</td></tr><tr><td/><td colspan=\"8\">Control 578.1 (\u00b1149.8) 34.5 (\u00b110.7) 110.5 (\u00b127.9) 84.2 (\u00b125.4) 23.5 (\u00b110.1) 9.3 (\u00b14.5) 21.4 (\u00b113.0)</td></tr><tr><td>All</td><td>MCI</td><td colspan=\"2\">548.7 (\u00b1140.6) 34.0 (\u00b110.5)</td><td colspan=\"3\">98.1 (\u00b126.1) 81.2 (\u00b124.1) 22.5 (\u00b19.7)</td><td colspan=\"2\">7.7 (\u00b14.2) 25.3 (\u00b115.0)</td></tr><tr><td/><td>p</td><td>0.0110</td><td>0.5541</td><td>< 0.0001</td><td>0.1277</td><td>0.2046</td><td>< 0.0001</td><td>0.0006</td></tr></table>" | |
| }, | |
| "TABREF1": { | |
| "type_str": "table", | |
| "html": null, | |
| "num": null, | |
| "text": "Average counts and their standard deviations of linguistic features per transcript in the B-SHARP dataset.", | |
| "content": "<table><tr><td>Complex: occurrences of complex structures (e.g., relative clauses, non-finite clauses), Discourse: occurrences of</td></tr><tr><td>discourse elements (e.g., interjections, disfluency).</td></tr></table>" | |
| }, | |
| "TABREF3": { | |
| "type_str": "table", | |
| "html": null, | |
| "num": null, | |
| "text": "", | |
| "content": "<table><tr><td>: Statistics of control (C) and MCI (M) groups.</td></tr><tr><td>Sbj: # of subjects, 2nd/3rd: # of subjects who made the</td></tr><tr><td>2nd/3rd visits, Rec: # of voice recordings, MoCA/BNT:</td></tr><tr><td>average scores and stdevs from MoCA/BNT. Note that</td></tr><tr><td>subjects with the 2nd/3rd visits take one/two additional</td></tr><tr><td>recordings; thus, Rec = Sbj + 1\u2022(2nd) + 2\u2022(3rd).</td></tr></table>" | |
| }, | |
| "TABREF5": { | |
| "type_str": "table", | |
| "html": null, | |
| "num": null, | |
| "text": "Model performance on the individual tasks. ACC: accuracy, SEN: sensitivity, SPE: specificity.", | |
| "content": "<table><tr><td/><td>CNN</td><td>BERTe</td><td>RoBERTae ALBERTe</td><td>Be + Re</td><td>Ae + Re</td><td>Be + Ae + Re</td></tr><tr><td>ACC</td><td>69.5 (\u00b10.2)</td><td colspan=\"2\">69.9 (\u00b11.1) 71.6 (\u00b11.5) 69.7 (\u00b12.9)</td><td colspan=\"2\">72.2 (\u00b10.7) 71.5 (\u00b11.9)</td><td>74.1 (\u00b10.3)</td></tr><tr><td>SEN</td><td>49.2 (\u00b10.8)</td><td colspan=\"2\">57.6 (\u00b13.4) 48.5 (\u00b16.1) 46.2 (\u00b18.3)</td><td colspan=\"2\">56.5 (\u00b12.5) 51.7 (\u00b11.3)</td><td>60.9 (\u00b15.2)</td></tr><tr><td>SPE</td><td>83.5 (\u00b10.9)</td><td colspan=\"2\">77.4 (\u00b14.8) 87.5 (\u00b11.8) 85.4 (\u00b10.5)</td><td colspan=\"2\">83.1 (\u00b10.9) 86.7 (\u00b13.4)</td><td>84.0 (\u00b12.4)</td></tr></table>" | |
| }, | |
| "TABREF6": { | |
| "type_str": "table", | |
| "html": null, | |
| "num": null, | |
| "text": "Performance of ensemble models. Bert e /RoBERTa e /ALBERT e use transcript embeddings from all 3 tasks trained by the BERT/RoBERTa/ALBERT models in", | |
| "content": "<table/>" | |
| }, | |
| "TABREF8": { | |
| "type_str": "table", | |
| "html": null, | |
| "num": null, | |
| "text": "Statistics of the CV sets for our experiments. Rec/Sbj: # of recordings/subjects, C/M: in control/MCI group. CV i : the i'th set. ALL:", | |
| "content": "<table/>" | |
| }, | |
| "TABREF10": { | |
| "type_str": "table", | |
| "html": null, | |
| "num": null, | |
| "text": "", | |
| "content": "<table><tr><td>: Configurations of the BERT, RoBERTa, and</td></tr><tr><td>ALBERT encoders for our experiments. L: # of layers,</td></tr><tr><td>AH: # of attended heads, IC: # of input cells, HC: # of</td></tr><tr><td>hidden cells, P: # of parameters.</td></tr></table>" | |
| } | |
| } | |
| } | |
| } |