chunk_id
stringlengths
3
7
chunk
stringlengths
1
823
source_url
stringclasses
416 values
title
stringclasses
416 values
chunk_idx
int64
0
294
chunk_start_char
int64
0
139k
chunk_end_char
int64
303
139k
19_28
After choosing the parameters, models were re-trained on all the available data. We consider some of the most popular baseline models in the literature: (i) minority class baseline (due to the heavily imbalanced dataset), (ii) Support Vector Machines trained on TF-IDF bi-gram language model, (iii) and Support Vector ...
https://arxiv.org/abs/1905.00472
A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project
28
15,702
16,261
19_29
To identify sentiment expressed towards an SF, we use the pretrained model described in BIBREF44 , in which a multiplicative LSTM cell is trained at the character level on a corpus of 82 million Amazon reviews. The model representation is then fed to a logistic regression classifier to predict sentiment. This model (w...
https://arxiv.org/abs/1905.00472
A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project
29
16,261
16,786
19_30
In our approach, we first map the text associated with the SF annotation with a segment from the document and pass the full segment to the pretrained OpenAI model identify the sentiment polarity for that segment. To identify sentiment targeted towards an entity, we use the recently released Target-Based Sentiment Ana...
https://arxiv.org/abs/1905.00472
A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project
30
16,786
17,140
19_31
In TBSA, two stacked LSTM cells are trained to predict both sentiment and target boundary tags (e.g., predicting S-POS to indicate the start of the target towards which the author is expressing positive sentiment, I-POS and E-POS to indicate intermediate and end of the target). In our submission, since input text docu...
https://arxiv.org/abs/1905.00472
A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project
31
17,140
17,768
19_32
In this model we limit our focus on the task of correctly identifying those segments with sentiment towards a SF. That is, given a pair of SF and segment, we train models to identify if this segment contains any sentiment towards that SF. This allows us to expand our dataset from 123 documents into one with $\sum _d ...
https://arxiv.org/abs/1905.00472
A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project
32
17,768
18,258
19_33
Summary of the training dataset after augmentation is given in Table 3 . Given the highly skewed label distribution in the training data, a majority of the constructed pairs do not have any sentiment towards a SF. Hence, our resulting dataset has a highly imbalanced distribution which we address by training our model...
https://arxiv.org/abs/1905.00472
A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project
33
18,258
18,901
19_34
Owing to the nature of the problem domain, there is considerable variance in the source of the text documents and their structure. For example, tweets only have one segment per sample whereas news articles contain an average of $7.07\pm 4.96$ and $6.31\pm 4.93$ segments for English and Spanish documents respectively....
https://arxiv.org/abs/1905.00472
A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project
34
18,901
19,352
19_35
Table 4 presents a breakdown of the train set for each sentiment across domains, as is evident tweets form a sizeable group of the training set. Motivated by this, we train different models for tweets and non-tweet documents in order to capture the underlying differences between the data sources. Initial experiments ...
https://arxiv.org/abs/1905.00472
A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project
35
19,352
19,929
19_36
To address this, we decided to use a model which only produced results for tweets given that these only contain one segment, making the segment identification sub-task trivial. Results Model performance during train is presented in Table 5 . While all the models outperformed the baselines, not all of them did so wit...
https://arxiv.org/abs/1905.00472
A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project
36
19,929
20,320
19_37
The ones found to be significantly better than the baselines were models IIb (Domain-specific) and IIc (Twitter-only) (permutation test, $n = 10^5$ both $p < 0.05$ ). The difference in precision between model IIb and IIc points out to the former making the wrong predictions for news articles. These errors are most lik...
https://arxiv.org/abs/1905.00472
A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project
37
20,320
20,889
19_38
Table 6 present the official evaluation results for English and Spanish. Some information is missing since at the time of submission only partial score had been made public. As previously mentioned, the pre-trained state-of-the-art models (model I) were directly applied to the evaluation data without any adaptation. ...
https://arxiv.org/abs/1905.00472
A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project
38
20,889
21,555
19_39
There are multiple difficulties beyond the limited amount of data. For example, in order to provide decision-makers with actionable and usable information, it is not enough for the system to correctly classify sentiment or emotional state, it also ought to identify the source and target of the expressed sentiment. To ...
https://arxiv.org/abs/1905.00472
A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project
39
21,555
22,134
19_40
Thus, it is important that systems take into account the possible differences in the way sentiment is expressed in each one of these sources. In this work, we presented two approaches to the task of providing actionable and useful information. Our results show that state-of-the-art sentiment classifiers can be leverag...
https://arxiv.org/abs/1905.00472
A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project
40
22,134
22,739
19_41
TABLE I EXAMPLE OF A SEGMENT CONTAINING A SITUATION FRAME WITH SENTIMENT RELATED ANNOTATIONS (IN BOLD). TABLE II FREQUENCY STATISTICS FOR THE PROVIDED TRAINING DATA PER LANGUAGE: NUMBER OF DOCUMENTS, NUMBER OF ANNOTATED SITUATION FRAMES, NUMBER OF SENTIMENT INSTANCES, PERCENTAGE OF NEGATIVE POLARITY. TABLE IV TRAIN...
https://arxiv.org/abs/1905.00472
A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project
41
22,739
23,161
20_0
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations This article compares two multimodal resources that consist of diagrams which describe topics in elementary school natural sciences. Both resources contain the same diagrams and represent their struct...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
0
0
748
20_1
The results show that the identity of diagram elements can be learned from their layout features, while the expert annotations provide better representations of diagram types. Introduction Diagrams are a common feature of many everyday media from newspapers to school textbooks, and not surprisingly, different forms ...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
1
748
1,385
20_2
There is also a long-standing interest in processing and generating diagrams computationally BIBREF4, BIBREF5, BIBREF6, which is now resurfacing as advances emerging from deep learning for computer vision and natural language processing are brought to bear on diagrammatic representations BIBREF7, BIBREF8, BIBREF9. Fr...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
2
1,385
2,036
20_3
Making sense of how diagrams exploit the 2D layout space falls arguably within the domain of computer vision. On the other hand, diagrams also have a discourse structure, which uses the layout space to set up discourse relations between instances of natural language, various types of images, arrows and lines, thus for...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
3
2,036
2,676
20_4
Recent examples of work in this area include alikhanietal2019 and ottoetal2019, who model discourse relations between natural language and photographic images, drawing on linguistic theories of coherence and text–image relations, respectively. In most cases, however, predicting a single discourse relation covers only ...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
4
2,676
3,270
20_5
Unlike many other areas, the study of diagrammatic representations is particularly well-resourced, as several multimodal resources have been published recently to support research on computational processing of diagrams BIBREF10, BIBREF8, BIBREF11. This study compares two such resources, AI2D BIBREF10 and AI2D-RST BI...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
5
3,270
3,680
20_6
Whereas AI2D features crowd-sourced, non-expert annotations, AI2D-RST provides multiple layers of expert annotations, which are informed by state-of-the-art approaches to multimodal communication BIBREF12 and annotation BIBREF13, BIBREF14. This provides an interesting setting for comparison and evaluation, as non-exp...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
6
3,680
4,261
20_7
Whether expert annotations provide better representations of diagrammatic structures and thus justify their higher cost is one question that this study seeks to answer. Both AI2D and AI2D-RST represent the multimodal structure of diagrams using graphs. This enables learning their representations using graph neural ne...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
7
4,261
4,692
20_8
This article reports on two experiments that evaluate the capability of AI2D and AI2D-RST to represent the multimodal structure of diagrams using graphs, focusing particularly on spatial layout, the hierarchical organisation of diagram elements and their connections expressed using arrows and lines. Data This sectio...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
8
4,692
5,336
20_9
Figure FIGREF1 provides an overview of the two datasets, explains their relation to each other and provides an overview of the experiments reported in Section SECREF4 Data ::: Crowd-sourced Annotations from AI2D The Allen Institute for Artificial Intelligence Diagrams dataset (AI2D) contains 4903 English-language di...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
9
5,336
5,820
20_10
The dataset was originally developed to support research on diagram understanding and visual question answering BIBREF16, but has also been used to study the contextual interpretation of diagrammatic elements, such as arrows and lines BIBREF17. The AI2D annotation schema models four types of diagram elements: text, g...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
10
5,820
6,348
20_11
Each diagram is represented using a Diagram Parse Graph (DPG), whose nodes stand for diagram elements while the edges between the nodes carry information about their semantic relations. The annotation for AI2D, which includes layout segmentations for the diagram images, DPGs and a multiple choice question-answer set, ...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
11
6,348
6,914
20_12
To exemplify, many forms of multimodal discourse are assumed to possess a hierarchical structure, whose representation requires a tree graph. Diagrams, however, use arrows and lines to draw connections between elements that are not necessarily part of the same subtree, and for this reason representing connectivity req...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
12
6,914
7,377
20_13
Whether computational modelling of diagrammatic structures, or more generally, multimodal discourse parsing, benefits from pulling apart different types of multimodal structure remains an open question, which we pursued by developing an alternative annotation schema for AI2D, named AI2D-RST, which is introduced below....
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
13
7,377
7,927
20_14
The annotation schema, which draws on state-of-the-art theories of multimodal communication BIBREF12, adopts a stand-off approach to describing the diagrams. Hence the three annotation layers in AI2D-RST are represented using three different graphs, which use the same identifiers for nodes across all three graphs to a...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
14
7,927
8,524
20_15
These groups are organised into a hierarchy, which represents the organisation of content in the 2D layout space BIBREF13, BIBREF14. Connectivity: A cyclic graph representing connections between diagram elements or their groups, which are signalled using arrows or lines BIBREF20. Discourse structure: A tree graph re...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
15
8,524
8,968
20_16
The grouping graph, which is initially populated by diagram elements from the AI2D layout segmentation, provides a foundation for describing connectivity and discourse structure by adding nodes to the grouping graph that stand for groups of diagram elements, as shown in the upper part of Figure FIGREF1. In addition, ...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
16
8,968
9,604
20_17
hiippalaetal2019-ai2d show that the proposed annotation schema can be reliably applied to the data by measuring inter-annotator agreement between five annotators on random samples from the AI2D-RST corpus using Fleiss' $\kappa $ BIBREF23.
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
17
9,604
9,844
20_18
The results show high agreement on grouping ($N = 256, \kappa = 0.84$), diagram types ($N = 119, \kappa = 0.78$), connectivity ($N = 239, \kappa = 0.88$) and discourse structure ($N = 227, \kappa = 0.73$). It should be noted, however, that these measures may be affected by implicit knowledge that tends to develop amon...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
18
9,844
10,224
20_19
Graph-based Representations Both AI2D and AI2D-RST use graphs to represent the multimodal structure of diagrams. This section explicates how the graphs and their node and edge types differ across the two multimodal resources. Graph-based Representations ::: Nodes ::: Node Types AI2D and AI2D-RST share most node ty...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
19
10,224
10,822
20_20
In AI2D-RST, the image constant acts as the root node of the tree in the grouping graph. In addition to text, graphics, arrows and the image constant, AI2D-RST features two additional node types for groups and discourse relations, whereas AI2D includes an additional node for arrowheads. To summarise, AI2D contains fiv...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
20
10,822
11,314
20_21
Graph-based Representations ::: Nodes ::: Node Features The same features are used for both AI2D and AI2D-RST for nodes with layout information, namely text, graphics, arrows and arrowheads (in AI2D only).
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
21
11,314
11,522
20_22
The position, size and shape of each diagram element are described using the following features: (1) the centre point of the bounding box or polygon, divided by the height and width of the diagram image, (2) area, or the number of pixels within the polygon, divided by the total number of pixels in the image, and (3) t...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
22
11,522
12,044
20_23
Each dimension is set to zero for grouping nodes in AI2D-RST and image constant nodes in AI2D and AI2D-RST. Graph-based Representations ::: Nodes ::: Discourse Relations AI2D-RST models discourse relations using nodes, which have a 25-dimensional, one-hot encoded feature vector to represent the type of discourse rel...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
23
12,044
12,429
20_24
In AI2D, the discourse relations derived from engelhardt2002 are represented using a 10-dimensional one-hot encoded vector, which is associated with edges connecting diagram elements participating in the relation. Because the two resources draw on different theories and represent discourse relations differently, I use...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
24
12,429
13,062
20_25
Graph-based Representations ::: Edges Whereas AI2D encodes information about semantic relations using edges, in AI2D-RST the information carried by edges depends on the graph in question. The edges of the grouping graph do not have features, whereas the edges of the connectivity graph have a 3-dimensional, one-hot e...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
25
13,062
13,654
20_26
For the experiments reported in Section 4, self-loops are added to each node in the graph. A self-loop is an edge that originates in and terminates at the same node. Self-loops essentially add the graph's identity matrix to the adjacency matrix, which allow the graph neural networks to account for the node's own feat...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
26
13,654
14,240
20_27
Experiments ::: Graph Neural Networks I evaluated the following graph neural network architectures for both graph and node classification tasks: Graph Convolutional Network (GCN) BIBREF25 Simplifying Graph Convolution (SGC) BIBREF26, averaging incoming node features from up to 2 hops away Graph Attention Network ...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
27
14,240
14,750
20_28
0. For GCN, GAT and SAGE, each network consists of two of the aforementioned layers with a Rectified Linear Unit (ReLU) activation, followed by a dense layer and a final softmax function for predicting class membership probabilities. For SGC, the network consists of a single SGC layer without an activation function. Th...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
28
14,750
15,166
20_29
Experiments ::: Hyperparameters and Training I used the Tree of Parzen Estimators (TPE) algorithm BIBREF31 to tune model hyperparameters separately for each dataset, architecture and task using the implementation in the Tune BIBREF32 and hyperopt BIBREF33 libraries. For each dataset, architecture and task, I evaluat...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
29
15,166
15,673
20_30
Tables TABREF20 and TABREF21 give the hyperparameters and spaces searched for node and graph classification. Following shcuretal2018, I shuffled the training and validation splits for each run to prevent overfitting and used the same training procedure throughout. I used the Adam optimiser BIBREF34 for both hyperparam...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
30
15,673
16,282
20_31
These weights were passed to the loss function during hyperparameter search and training. After hyperparameter optimisation, I trained each model with the best hyperparameter combination for 20 runs, using 850 diagrams for training, 75 for validation and 75 for testing, shuffling the splits for each run while monitor...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
31
16,282
16,861
20_32
Experiments ::: Tasks ::: Node Classification The purpose of the node classification task is to evaluate how well algorithms learn to classify the parts of a diagram using the graph-based representations in AI2D and AI2D-RST and node features representing the position, size and shape of the element, as described in ...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
32
16,861
17,454
20_33
Furthermore, the node representations learned during this task can be used as node features for graph classification, as will be shown shortly below in Section SECREF26 Table TABREF25 presents a baseline for node classification from a dummy classifier, together with results for random forest and support vector machin...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
33
17,454
17,830
20_34
Both AI2D and AI2D-RST include five node types, of which four are the same: the difference is that whereas AI2D includes arrowheads, AI2D-RST includes nodes for groups of diagram elements, as outlined in Section SECREF9 The results seem to reflect the fact that image constants and grouping nodes have their features se...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
34
17,830
18,412
20_35
Table TABREF22 shows the results for node classification using various graph neural network architectures. Because the results are not entirely comparable due to different node types present in the two resources, it is more reasonable to compare architectures. SAGE, GCN and GAT clearly outperform SGC in classifying n...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
35
18,412
18,974
20_36
AI2D-RST, in turn, must learn representations from scratch for both image constants ($N = 1000$) and grouping nodes ($N = 7300$). Because SAGE learns useful node representations for both resources, as reflected in high performance for all metrics, I chose this architecture for extracting node features for graph class...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
36
18,974
19,473
20_37
Here the aim is to evaluate to what extent graph neural networks can learn about the generic structure of primary school science diagrams from the graph-based representations in AI2D and AI2D-RST. Correctly identifying what the diagram attempts to communicate and how carries implications for tasks such as visual quest...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
37
19,473
20,041
20_38
Whereas AI2D includes 17 classes that represent the semantic content of diagrams, as exemplified by categories such as `parts of the Earth', `volcano', and `food chains and webs', AI2D-RST classifies diagrams into abstract diagram types, such as cycles, networks, cross-sections and cut-outs. More specifically, AI2D-RS...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
38
20,041
20,560
20_39
The 11 fine-grained classes in AI2D-RST shown in Figure FIGREF8 are complemented by an additional class (`mixed'), which includes diagrams that combine multiple diagram types, whose inclusion avoids performing multi-label classification (see the example in Figure FIGREF28). The coarse classes, which are derived by gr...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
39
20,560
21,072
20_40
For this task, the node features consist of the representations learned during node classification in Section SECREF24 These representations are extracted by feeding the features representing node position, size and shape to the graph neural network, which in both cases uses the GraphSAGE architecture BIBREF28, and r...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
40
21,072
21,664
20_41
Table TABREF29 provides a baseline for graph classification from a dummy classifier, as well as results for random forest (RF) and support vector machine (SVM) classifiers trained on 850 and tested on 150 diagrams. The macro F1 scores show that the RF classifier with 100 decision trees offers competitive performance ...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
41
21,664
22,192
20_42
The results for graph classification using graph neural networks presented in Table TABREF27 show certain differences between AI2D and AI2D-RST. When classifying diagrams into the original semantic categories defined in AI2D ($N = 17$), the AI2D graphs significantly outperform AI2D-RST when using the GraphSAGE archit...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
42
22,192
22,633
20_43
This is not surprising as the AI2D graphs were tailored for the original classes, yet the AI2D-RST graphs seem to capture generic properties that help to classify diagrams into semantic categories nearly as accurately as AI2D graphs designed specifically for this purpose, although no semantic features apart from the l...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
43
22,633
23,203
20_44
This classification task obviously benefits AI2D-RST, whose classification schema was originally designed for abstract diagram types. This may also suggest that the AI2D graphs do not capture regularities that would support learning to generalise about diagram types. The situation is somewhat different for fine-graine...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
44
23,203
23,853
20_45
The edges introduced from the connectivity graph naturally increase the flow of information in the graph, but this does not seem to help learn distinctive features between diagram types. On the other hand, it should be noted that the nodes are not typed, that is, the model cannot distinguish between edges from the gro...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
45
23,853
24,432
20_46
The lack of visual features may also affect overall classification performance: certain fine-grained classes, which are also prominent in the data, such as 2D cross-sections and 3D cut-outs, may have similar graph-based representations. Extracting visual features from diagram images may help to discern between diagram...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
46
24,432
24,883
20_47
Discussion The results for AI2D-RST show that the grouping graph, which represents visual perceptual groups of diagram elements and their hierarchical organisation, provides a robust foundation for describing the spatial organisation of diagrammatic representations. This kind of generic schema can be expanded beyond...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
47
24,883
25,467
20_48
The promising results AI2D-RST suggest is that domain experts in multimodal communication should be involved in planning crowd-sourced annotation tasks right from the beginning. Segmentation, in particular, warrants attention as this phase defines the units of analysis: cut-outs and cross-sections, for instance, use ...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
48
25,467
26,041
20_49
Tasks related to grouping and connectivity annotation could be crowd-sourced relatively easily, whereas annotating diagram types and discourse relations may require multi-step procedures and assistance in the form of prompts, as yungetal2019 have recently shown for RST. Involving both expert and crowd-sourced annotat...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
49
26,041
26,688
20_50
Whereas the expert-annotated grouping graph in AI2D-RST has been pruned from isolated nodes, which ensures that features are propagated to neighbouring nodes, the crowd-sourced AI2D graphs contain both isolated nodes and subgraphs. To what extent these disconnections affect the performance for AI2D warrant a separate ...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
50
26,688
27,228
20_51
To begin with, a comparison of representations for discourse structures using the question-answering set accompanying AI2D would be particularly interesting, especially if both AI2D and AI2D-RST graphs were enriched with features from state of the art semantic representations for natural language and graphic elements....
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
51
27,228
27,800
20_52
The comparison involved two tasks, graph and node classification, using four different architectures for graph neural networks, which were compared to baselines from dummy, random forest and support vector machine classifiers. The results showed that graph neural networks can learn to accurately identify diagram elem...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
52
27,800
28,430
20_53
Improving accuracy may require additional features that capture visual properties of the diagrams, as these distinctions cannot be captured by graph-based representations and features focusing on layout. Overall, the results nevertheless suggest that simple layout features can provide a foundation for representing di...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
53
28,430
29,099
20_54
AI2D-RST provides alternative, expert-annotated stand-off descriptions for a subset of 1000 diagrams from the original AI2D dataset. The grouping layer in AI2D provides a foundation for further annotation layers by allowing references to groups of nodes. Figure 2: Fine-grained classes, their number and frequencies in...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
54
29,099
29,565
20_55
The arrows indicate choices: if the diagram designer chooses depiction, a further choice must be made between pictorial– diagrammatic and 2D/3D representations. The dashed lines indicate coarse groups of diagram types. Table 1: Hyperparameter ranges for graph classification Table 2: Hyperparameter ranges for node cl...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
55
29,565
30,018
20_56
The following abbreviations indicate the graph used: ‘AI2D’ for the original crowd-sourced graphs from AI2D, ‘G’ for the grouping graph and ‘G+C’ for the combination of grouping and connectivity graph from AI2D-RST. An asterisk indicates that the difference between AI2D and the best AI2D-RST graph is statistically sig...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
56
30,018
30,500
20_57
Figure 3: Diagram #4120 in AI2D combines two diagram types: a cross-section with a cycle (cf. Figure 2) Table 4: Baseline accuracy, macro F1 and weighted F1 scores for node classification from dummy (D), random forest (RF; 100 estimators) and support vector machine (SVM; C = 1.0) classifiers with balanced class weig...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
57
30,500
30,918
20_58
Each node is represented by a 4-dimensional vector. Table 5: Mean accuracy, macro F1 and weighted F1 scores for graph classification. The results are averaged over 20 runs. The following abbreviations indicate the graph used: ‘AI2D’ for the original crowd-sourced graphs from AI2D, ‘G’ for the grouping graph and ‘G+C’...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
58
30,918
31,308
20_59
* indicates that the difference between AI2D and the best AI2D-RST graph is statistically significant at p < 0.05 when comparing the results over 20 runs for the given metric using Mann–Whitney U test. + indicates the same for AI2D-RST grouping graph and the combination of grouping and connectivity graphs. The best re...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
59
31,308
31,696
20_60
Table 6: Baseline accuracy, macro F1 and weighted F1 scores for graph classification using dummy (D), random forest (RF; 100 estimators) and support vector machine (SVM; C = 1.0) classifiers with balanced class weights. The results are averaged over 20 runs. All models were implemented using scikit-learn 0.21.3. Each...
https://arxiv.org/abs/1912.02866
Classifying Diagrams and Their Parts using Graph Neural Networks: A Comparison of Crowd-Sourced and Expert Annotations
60
31,696
32,128
21_0
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions Detecting controversy in general web pages is a daunting task, but increasingly essential to efficiently moderate discussions and effectively filter problematic content. Unfortunately, controversies oc...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
0
0
755
21_1
We train neural networks that can capture semantic information from texts using weak signal data. By leveraging the semantic properties of word embeddings we robustly improve on existing controversy detection methods. To evaluate model stability over time and to unseen topics, we asses model performance under varying ...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
1
755
1,447
21_2
Controversial content can signal the need for moderation on social platforms, either to prevent conflict between users or limit the spread of misinformation. More generally, controversies provide insight into societies BIBREF0 . Often, the controversial content is outside the direct control of a platform on which it i...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
2
1,447
2,118
21_3
We take the task of controversy detection and evaluate robustness of different methodologies with respect to the varying nature of controversies. Prior work on detecting controversies has taken three kinds of approaches: 1) lexical approaches, which seek to detect controversies through signal terms, either through ba...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
3
2,118
2,630
21_4
Features such as mutual reverts BIBREF2 , user-provided flags BIBREF3 , interaction networks BIBREF4 or stance-distributions BIBREF5 have been used as platform-specific indicators of controversies. The downside of these approaches is the lack of generalizability due to their platform-specific nature. 3) matching model...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
4
2,630
3,155
21_5
Controversy detection is a difficult task because 1) controversies are latent, like ideology, meaning they are often not directly mentioned as controversial in text. 2) Controversies occur across a vast range of topics with varying topic-specific vocabularies. 3) Controversies change over time, with some topics and a...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
5
3,155
3,777
21_6
fulltext) features are likely to memorize the controversial topics in the training set rather than the `language of controversy'. Alleviating dependence on platform specific features and reducing sensitivity to an exact lexical representation is paramount to robust controversy detection. To this end, we focus only on ...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
6
3,777
4,405
21_7
In essence these embeddings are latent-vector representations that aim to capture the underlying meaning of words. Distances between such latent-vectors are taken to express semantic relatedness, despite having different surface forms. By using embeddings, neural architectures are also able to leverage features learne...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
7
4,405
5,009
21_8
We use two often applied network architectures adopting word embeddings, to classify controversy: Recurrent Neural Networks BIBREF10 and Convolutional Neural Networks BIBREF11 to answer the following research question. RQ: Can we increase robustness of controversy detection using neural methods? Currently, there is n...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
8
5,009
5,615
21_9
Methods A proven approach in modelling text with neural networks is to use Recurrent Neural Networks (RNNs) which enjoy weight sharing capabilities to model words irrespective of their sequence location. A specific type, the Hierarchical Attention Network (HAN) proposed by BIBREF10 makes use of attention to build do...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
9
5,615
6,279
21_10
Recently, Convolutional Neural Networks (CNNs) have enjoyed increasing success in text classification. One such network introduced by BIBREF11 looks at patterns in words within a window, such as "Scientology [...] brainwashes people". The occurrences of these patterns are then summarized to their 'strongest' observat...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
10
6,279
6,872
21_11
We explore the potential of RNNs and CNNs for controversy detection using both the HAN BIBREF10 and the CNN BIBREF11 model. Similar to BIBREF10 , each bi-directional GRU cell is set to a dimension of 50, resulting in a word/sentence representation of size 100 after concatenation. The word/sentence attention vectors s...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
11
6,872
7,349
21_12
Each model is trained using mini batches of size 64 and uses both dropout (0.5) and INLINEFORM0 regularization (1e-3) at the dense prediction layer. Both networks use pre-trained embeddings, trained on 100 billion words of a Google News corpus, which are further fine-tuned during training on the controversy dataset. T...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
12
7,349
7,838
21_13
For cross-temporal, cross-topic and cross-domain training & evaluation, we generate a new dataset based on Wikipedia crawl data. This dataset is gathered by using Wikipedia's `List of Contoversial articles' overview page of 2018 (time of writing) and 2009 (for comparison with baselines) . Using this as a `seed' set of...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
13
7,838
8,399
21_14
The snowball-sample approach includes general, non-Wikipedia, pages that are referred to from Wikipedia pages. The dataset thus extends beyond just the encyclopedia genre of texts. Labels are assumed to propagate: a page linked from a controversial issue is assumed to be controversial. The resulting dataset statistics...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
14
8,399
9,053
21_15
The test-train split depends on the task investigated and is listed in the results section for the respective task. To test for significant results, all models were evaluated using a bootstrap approach: by drawing 1000 samples with replacements INLINEFORM0 documents from the test set equal to the test-set size. The re...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
15
9,053
9,642
21_16
Together with an SVM baseline they act as controversy detection alternatives using only full text features, thus meeting the task-requirements of platform-independence. Note: the implementation of BIBREF7 additionally requires ranking methods to select a subset of the training data for each language model. A simplifie...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
16
9,642
10,252
21_17
Similarly, for completeness sake, we also include both the state-of-the-art matching model, the TILE-Clique model from BIBREF1 and the sentiment analysis baseline (using the state-of-the-art Polyglot library for python) from BIBREF6 in the comparison with previous work. Comparison of results with previous work Table...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
17
10,252
10,799
21_18
The TILE-Clique matching model outperforms all other models on Precision although this difference is not significant compared to the neural approaches. Similarly, the language model trained on the DBPedia dataset outperforms other models on Recall but shows no significant difference compared to the CNN model. Notably,...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
18
10,799
11,417
21_19
Robustness of the model across time Controversy is expected to change over time. Some issues become controversial, others cease to be so. To investigate robustness of controversy detection models with respect to changes over time, we evaluate model performance in two variants: trained and tested on 2018, or trained ...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
19
11,417
11,873
21_20
Within year, the hierarchical attention model (HAN) outperforms all other models on Recall, F1 and AUC, losing Precision to the CNN and SVM models. However, our main interest is the robustness when a model is trained on a different year (2009) than the test set (2018).
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
20
11,873
12,144
21_21
These between year experiments show a superior score for the HAN model compared to the non-neural models on Recall, and show significant improvements on F1 (p < 0.05) and AUC (p < 0.05), losing only to the SVM model on Precision (non significantly). In terms of robustness, we can also take the percentage change betwee...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
21
12,144
12,690
21_22
In Figure 1, we show the pooled results for the lexical and neural models to illustrate the overall increase in robustness by neural approaches. Interestingly, the SVM and HAN model show some unexpected improvement with regard to Precision when applied to unseen timeframes. For both models, this increase in Precision...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
22
12,690
13,276
21_23
Robustness of the model across topics To evaluate robustness towards unseen topics, 10-fold cross validation was used on the top ten largest topics present in the Wikipedia dataset in a leave-one-out fashion. The results are shown in table 4. In line with previous results, the language model scores best on Recall, b...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
23
13,276
13,804
21_24
Overall, when grouping together all neural and lexical results, the neural methods outperform the lexical models in Precision (p < 0.01), F1 (p < 0.05) and AUC (p < 0.01) with no significant difference found on the overall Recall scores. These results indicate that neural methods seem better able to generalize to unse...
https://arxiv.org/abs/1812.00382
Improved and Robust Controversy Detection in General Web Pages Using Semantic Approaches under Large Scale Conditions
24
13,804
14,306