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 |
|---|---|---|---|---|---|---|
10_37 | Our results lead to the following conclusions:
Acknowledgements
Gözde Gül Şahin was a PhD student at Istanbul Technical University and a visiting research student at University of Edinburgh during this study. She was funded by Tübitak (The Scientific and Technological Research Council of Turkey) 2214-A scholarship d... | https://arxiv.org/abs/1805.11937 | Character-Level Models versus Morphology in Semantic Role Labeling | 37 | 19,556 | 20,031 |
10_38 | This work was supported by ERC H2020 Advanced Fellowship GA 742137 SEMANTAX and a Google Faculty award to Mark Steedman. We would like to thank Adam Lopez for fruitful discussions, guidance and support during the first author's visit.
Table 1: Sample outputs of different ρ functions
Table 2: Training data statistics... | https://arxiv.org/abs/1805.11937 | Character-Level Models versus Morphology in Semantic Role Labeling | 38 | 20,031 | 20,557 |
10_39 | Best F1 for each language is shown in bold. First row: results on test, Second row: results on development.
Figure 2: x axis: Number of morphological features; y axis: Targeted F1 scores
Table 4: Results of ensembling via averaging (Avg) and stack generalization (SG). IOB: Improvement Over Best of baseline models
F... | https://arxiv.org/abs/1805.11937 | Character-Level Models versus Morphology in Semantic Role Labeling | 39 | 20,557 | 21,037 |
10_40 | X axis: Number of sentences, Y axis: F1 score
Table 6: Effect of layer size on model performances. I: Improvement over model with one layer.
Table 5: F1 scores on out of domain data. Best scores are shown with bold.
Figure 5: F1 scores for best-char (best of the CLMs) and model with predicted (predictedmorph) and g... | https://arxiv.org/abs/1805.11937 | Character-Level Models versus Morphology in Semantic Role Labeling | 40 | 21,037 | 21,392 |
11_0 | Look, Read and Enrich - Learning from Scientific Figures and their Captions
Compared to natural images, understanding scientific figures is particularly hard for machines. However, there is a valuable source of information in scientific literature that until now has remained untapped: the correspondence between a figu... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 0 | 0 | 706 |
11_1 | We also show that transferring lexical and semantic knowledge from a knowledge graph significantly enriches the resulting features. Finally, we demonstrate the positive impact of such features in other tasks involving scientific text and figures, like multi-modal classification and machine comprehension for question a... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 1 | 706 | 1,475 |
11_2 | In the case of scientific figures, like charts, images and diagrams, these are usually accompanied by a text paragraph, a caption, that elaborates on the analysis otherwise visually represented.
In this paper, we make use of this observation and tap on the potential of learning from the enormous source of free superv... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 2 | 1,475 | 2,109 |
11_3 | To this purpose, we explore how multi-modal scientific knowledge can be learnt from the correspondence between figures and captions.
The main contributions of this paper are the following:
An unsupervised Figure-Caption Correspondence task (FCC) that jointly learns text and visual features useful to address a range ... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 3 | 2,109 | 2,805 |
11_4 |
A corpus of scientific figures and captions extracted from SN SciGraph and AI2 Semantic Scholar.
We present the FCC task in section SECREF3, including the network architecture, training protocol, and how adding pre-trained word and semantic embeddings can enrich the resulting text and visual features. In section SEC... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 4 | 2,805 | 3,302 |
11_5 | Then, we relate our work to the state of the art in image-sentence matching and evaluate our approach in two challenging transfer learning tasks: caption and figure classification and multi-modal machine comprehension. In section SECREF5 we perform a qualitative study that illustrates how the FCC task leads to detaile... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 5 | 3,302 | 3,939 |
11_6 | However, reasoning with other visual representations like scientific figures and diagrams has not received the same attention yet and entails additional challenges: Scientific figures are more abstract and symbolic, their captions tend to be significantly longer and use specialized lexicon, and the relation between a ... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 6 | 3,939 | 4,563 |
11_7 | Similar two-branch neural architectures focus on image-sentence BIBREF5, BIBREF6 and audio-video BIBREF7 matching. Others like BIBREF8 learn common embeddings from images and text. However, in such cases one or both networks are typically pre-trained.
Focused on geometry, BIBREF9 maximize the agreement between text a... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 7 | 4,563 | 5,054 |
11_8 | In BIBREF11, they parse diagram components and connectors as a Diagram Parse Graph (DPG), semantically interpret the DPG and use the model to answer diagram questions. While we rely on the correspondence between figures and captions, they train a specific classifier for each component and connector type and yet anothe... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 8 | 5,054 | 5,683 |
11_9 | They assume pre-trained entity representations exist in each individual modality, e.g. the visual features encoding the image of a ball, the word embeddings associated to the token "ball", and the KG embeddings related to the ball entity, which are then stitched together. In contrast, FCC co-trains text and visual fea... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 9 | 5,683 | 6,366 |
11_10 | The information captured in the caption explains the corresponding figure in natural language, providing guidance to identify the key features of the figure and vice versa. By seeing a figure and reading the textual description in its caption we ultimately aim to learn representations that capture e.g. what it means t... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 10 | 6,366 | 7,065 |
11_11 | Negative pairs are extracted from combinations of figures and any other randomly selected captions. The network is then made to learn text and visual features from scratch, without additional labelled data.
Figure-Caption Correspondence ::: FCC Architecture and Model
We propose a 2-branch neural architecture (figure... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 11 | 7,065 | 7,655 |
11_12 |
The vision subnetwork follows a VGG-style BIBREF13 design, with 3x3 convolutional filters, 2x2 max-pooling layers with stride 2 and no padding. It contains 4 blocks of conv+conv+pool layers, where inside each block the two convolutional layers have the same number of filters, while consecutive blocks have doubling nu... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 12 | 7,655 | 8,054 |
11_13 | The final layer produces a 512-D vector after 28x28 max-pooling. Each convolutional layer is followed by batch normalization BIBREF14 and ReLU layers. Based on BIBREF15, the language subnetwork has 3 convolutional blocks, each with 512 filters and a 5-element window size with ReLU activation. Each convolutional layer ... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 13 | 8,054 | 8,491 |
11_14 | The language subnetwork has a 300-D embeddings layer at the input, with a maximum sequence length of 1,000 tokens. The fusion subnetwork calculates the element-wise product of the 512-D visual and text feature vectors into a single vector $r$ to produce a 2-way classification output (correspond or not). It has two ful... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 14 | 8,491 | 8,943 |
11_15 | $\hat{y} = softmax(r) \in \mathbb {R}^{2}$. During training, we minimize the negative log probability of the correct choice.
This architecture enables the FCC task to learn visual and text features from scratch in a completely unsupervised manner, just by observing the correspondence of figures and captions. Next, we... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 15 | 8,943 | 9,475 |
11_16 | Adding pre-trained visual features is also possible and indeed we also evaluate its impact in the FCC task in section SECREF14.
Let $V$ be a vocabulary of words from a collection of documents $D$. Also, let $L$ be their lemmas, i.e. base forms without morphological or conjugational variations, and $C$ the concepts (o... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 16 | 9,475 | 9,842 |
11_17 | made, has one lemma $l_k$ (make) and may be linked to one or more concepts $c_k$ in $C$ (create or produce something).
For each word $w_k$, the FCC task learns a d-D embedding $\vec{w}_k$, which can be combined with pre-trained word ($\vec{w^{\prime }}_k$), lemma ($\vec{l}_k$) and concept ($\vec{c}_k$) embeddings to ... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 17 | 9,842 | 10,198 |
11_18 | If no pre-trained knowledge is transferred from an external source, then $\vec{t}_k=\vec{w}_k$. Note that we previously lemmatize and disambiguate $D$ against the KG in order to select the right pre-trained lemma and concept embeddings for each particular occurrence of $w_k$. | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 18 | 10,198 | 10,475 |
11_19 | Equation DISPLAY_FORM8 shows the different combinations of learnt and pre-trained embeddings we consider: (a) learnt word embeddings only, (b) learnt and pre-trained word embeddings and (c) learnt word embeddings and pre-trained semantic embeddings, including both lemmas and concepts, in line with our recent findings ... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 19 | 10,475 | 11,100 |
11_20 | Since some words may not have associated pre-trained word, lemma or concept embeddings, we pad these sequences with $\varnothing _W$, $\varnothing _L$ and $\varnothing _C$, which are never included in the vocabulary. The dimensionality of $\vec{t}_k$ is fixed to 300, i.e. the size of each sub-vector in configurations ... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 20 | 11,100 | 11,477 |
11_21 | In doing so, we aimed at limiting the number of trainable parameters and balance the contribution of each information source.
In its most basic form, i.e. configuration $(a)$, the FCC network has over 32M trainable parameters (28M in the language subnetwork, 4M in the vision subnetwork and 135K in the fusion subnetwo... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 21 | 11,477 | 11,938 |
11_22 | We used 10-fold cross validation, Adam optimization BIBREF18 with learning rate $10^{-4}$ and weight decay $10^{-5}$. The network was implemented in Keras and TensorFlow, with batch size 32. The number of positive and negative cases is balanced within the batches.
Figure-Caption Correspondence ::: Semantic Embeddings... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 22 | 11,938 | 12,334 |
11_23 | The latter extends the Swivel algorithm BIBREF20 to jointly learn word, lemma and concept embeddings on a corpus disambiguated against the KG, outperforming the previous state of the art in word and word-sense embeddings by co-training word, lemma and concept embeddings as opposed to training each individually. In con... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 23 | 12,334 | 12,897 |
11_24 | Following up with the work presented in BIBREF16, our experiments focus on Sensigrafo, the KG underlying Expert System's Cogito NLP proprietary platform. Similar to WordNet, on which Vecsigrafo has also been successfully trained, Sensigrafo is a general-purpose KG with lexical and semantic information that contains ov... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 24 | 12,897 | 13,369 |
11_25 | All the semantic (lemma and concept) embeddings produced with HolE or Vecsigrafo are 100-D.
Results and Discussion
In this section, first we evaluate the actual FCC task against two supervised baselines. Then, we situate our work in the more general image-sentence matching problem, showing empirical evidence of the ... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 25 | 13,369 | 13,802 |
11_26 | Next, we test the visual and text features learnt in the FCC task in two different transfer learning settings: classification of scientific figures and captions and multi-modal machine comprehension for question answering given a context of text, figures and images.
Results and Discussion ::: Datasets
We have used t... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 26 | 13,802 | 14,292 |
11_27 | From its 39M articles, we downloaded 3,3M PDFs (the rest were behind paywalls, did not have a link or it was broken) and extracted 12.5M figures and captions through PDFFigures2 BIBREF22. We randomly selected 500K papers to train the FCC task on their figures and captions and another 500K to train Vecsigrafo on the te... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 27 | 14,292 | 14,758 |
11_28 | Since SciGraph does not provide a link to the PDF of the publication, we selected the intersection with SemScholar, producing a smaller corpus of 80K papers (in addition to the 1M papers from SemScholar mentioned above) and 82K figures that we used for training certain FCC configurations and supervised baselines (sect... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 28 | 14,758 | 15,240 |
11_29 | Its complexity and scope make it a challenging textual and visual question answering dataset.
Wikipedia. We used the January 2018 English Wikipedia dataset as one of the corpora on which to train Vecsigrafo. As opposed to SciGraph or SemScholar, specific of the scientific domain, Wikipedia is a source of general-purp... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 29 | 15,240 | 15,810 |
11_30 | We also compare the performance of the FCC task against two supervised baselines, training them on a classification task against the SciGraph taxonomy. For such baselines we first train the vision and language networks independently and then combine them. The feature extraction parts of both networks are the same as d... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 30 | 15,810 | 16,455 |
11_31 | If it exceeds a threshold, which we heuristically fixed on 0.325, the result is positive. The supervised pre-training baseline freezes the weights of the feature extraction trunks from the two trained networks, assembles them in the FCC architecture as shown in section SECREF6, and trains the FCC task on the fully con... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 31 | 16,455 | 17,034 |
11_32 | $FCC_k$ denotes the corpus and word representation used to train the FCC task. Acc$_{vgg}$ shows the accuracy after replacing our visual branch with pre-trained VGG16 features learnt on ImageNet. This provides an estimate of how specific of the scientific domain scientific figures and therefore the resulting visual fe... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 32 | 17,034 | 17,642 |
11_33 |
We trained the FCC network on two different scientific corpora: SciGraph ($FCC_{1-5}$) and SemScholar ($FCC_{6-7}$). Both $FCC_1$ and $FCC_6$ learnt their own word representations without transfer of any pre-trained knowledge. Even in its most basic form our approach substantially improves over the supervised baselin... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 33 | 17,642 | 18,124 |
11_34 |
Adding pre-trained knowledge at the input layer of the language subnetwork provides an additional boost, particularly with lemma and concept embeddings from Vecsigrafo ($FCC_5$). Vecsigrafo clearly outperformed HolE ($FCC_3$), which was also beaten by pre-trained fastText BIBREF24 word embeddings ($FCC_2$) trained on... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 34 | 18,124 | 18,456 |
11_35 |
Since graph-based KG embedding approaches like HolE only generate embeddings of the artifacts explicitly contained in the KG, this may indicate that Sensigrafo, the KG used in this task, provides a partial coverage of the scientific domain, as could be expected since we are using an off-the-shelf version. Deeper insp... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 35 | 18,456 | 18,856 |
11_36 | On the other hand, Vecsigrafo, trained on the same KG, also captures lexical information from the text corpora it is trained on, Wikipedia or SemScholar, raising lemma coverage to 42% and 47%, respectively.
Although the size of Wikipedia is almost triple of our SemScholar corpus, training Vecsigrafo on the latter res... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 36 | 18,856 | 19,347 |
11_37 | Training FCC on SemScholar, much larger than SciGraph, further improves accuracy, as shown in $FCC_6$ and $FCC_7$.
Results and Discussion ::: Image-Sentence Matching
We put our FCC task in the context of the more general problem of image-sentence matching through a bidirectional retrieval task where images are sough... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 37 | 19,347 | 19,900 |
11_38 | The selected baselines (Embedding network, 2WayNet, VSE++ and DSVE-loc) report results obtained on the Flickr30K and COCO datasets, which we also include in table TABREF20. Performance is measured in recall at k ($Rk$), with k={1,5,10}. From the baselines, we successfully reproduced DSVE-loc, using the code made avail... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 38 | 19,900 | 20,283 |
11_39 |
We trained the FCC task on all the datasets, both in a totally unsupervised way and with pre-trained semantic embeddings (indicated with subscript $vec$), and executed the bidirectional retrieval task using the resulting text and visual features. We also experimented with pre-trained VGG16 visual features extracted f... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 39 | 20,283 | 20,802 |
11_40 |
We can see a marked division between the results obtained on natural images datasets (table TABREF20) and those focused on scientific figures (table TABREF21). In the former case, VSE++ and DSVE-loc clearly beat all the other approaches. In contrast, our model performs poorly on such datasets although results are ame... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 40 | 20,802 | 21,282 |
11_41 | While the recall of DSVE-loc drops dramatically in SciGraph, and even more in SemScholar, our approach shows the opposite behavior in both figure and caption retrieval. Using visual features enriched with pre-trained semantic embeddings from Vecsigrafo during training of the FCC task further improves recall in the bid... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 41 | 21,282 | 21,838 |
11_42 |
Unlike in Flickr30K and COCO, replacing the FCC visual features with pre-trained ones from ImageNet brings us little benefit in SciGraph and even less in SemScholar, where the combination of FCC and Vecsigrafo ("Oursvec") obtains the best results across the board. This and the extremely poor performance of the best i... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 42 | 21,838 | 22,329 |
11_43 | Indeed, the best results in figure-caption correspondence ("Oursvec" in SemScholar) are still far from the SoA in image-sentence matching (DSVE-loc in COCO).
Results and Discussion ::: Caption and Figure Classification
We evaluate the language and visual representations emerging from FCC in the context of two classi... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 43 | 22,329 | 22,803 |
11_44 | The latter is a particularly hard task due to the whimsical nature of the figures that appear in our corpus: figure and diagram layout is arbitrary; charts, e.g. bar and pie charts, are used to showcase data in any field from health to engineering; figures and natural images appear indistinctly, etc. Also, note that w... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 44 | 22,803 | 23,214 |
11_45 |
We pick the text and visual features that produced the best FCC results with and without pre-trained semantic embeddings (table TABREF15, $FCC_7$ and $FCC_6$, respectively) and use the language and vision subnetworks presented in section SECREF6 to train our classifiers on SciGraph in two different scenarios. First, ... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 45 | 23,214 | 23,728 |
11_46 | In both cases, we compare against a baseline using the same networks initialized with random weights, without FCC training. In doing so, through the first, non-trainable scenario, we seek to quantify the information contributed by the FCC features, while training from scratch on the target corpus should provide an upp... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 46 | 23,728 | 24,343 |
11_47 | In figure classification, we use learning rate $10^{-4}$, weight decay $10^{-5}$ and batch size 32.
The results in table TABREF23 show that our approach amply beats the baselines, including the upper bound (training from scratch on SciGraph). The delta is particularly noticeable in the non trainable case for both cap... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 47 | 24,343 | 24,786 |
11_48 | This includes both the random and VGG baselines and illustrates again the additional complexity of analyzing scientific figures compared to natural images, even if the latter is trained on a considerably larger corpus like ImageNet. Fine tuning the whole networks on SciGraph further improves accuracies. In this case, ... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 48 | 24,786 | 25,355 |
11_49 |
Results and Discussion ::: Textbook Question Answering (TQA) for Multi-Modal Machine Comprehension
We leverage the TQA dataset and the baselines in BIBREF23 to evaluate the features learnt by the FCC task in a multi-modal machine comprehension scenario. We study how our model, which was not originally trained for th... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 49 | 25,355 | 25,841 |
11_50 | We also study how pre-trained semantic embeddings impact in the TQA task: first, by enriching the visual features learnt in the FCC task as shown in section SECREF6 and then by using pre-trained semantic embeddings to enrich word representations in the TQA corpus.
We focus on multiple-choice questions, 73% of the dat... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 50 | 25,841 | 26,166 |
11_51 | Table TABREF24 shows the performance of our model against the results reported in BIBREF23 for five TQA baselines: random, BiDAF (focused on text machine comprehension), text only ($TQA_1$, based on MemoryNet), text+image ($TQA_2$, VQA), and text+diagrams ($TQA_3$, DSDP-NET). We successfully reproduced the $TQA_1$ and... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 51 | 26,166 | 26,532 |
11_52 | Then, we replaced the visual features in $TQA_2$ with those learnt by the FCC visual subnetwork both in a completely unsupervised way ($FCC_6$ in table TABREF15) and with pre-trained semantic embeddings ($FCC_7$), resulting in $TQA_4$ and $TQA_5$, respectively.
While $TQA_{1-5}$ used no pre-trained embeddings at all,... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 52 | 26,532 | 26,932 |
11_53 | Unlike FCC, where we used concatenation to combine pre-trained lemma and concept embeddings with the word embeddings learnt by the task, element-wise addition worked best in the case of TQA.
Following the recommendations in BIBREF23, we pre-processed the TQA corpus to i) consider knowledge from previous lessons in th... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 53 | 26,932 | 27,580 |
11_54 | This optimization allowed reducing the amount of text to consider for each question, improving the signal to noise ratio. Finally, we obtained the most relevant paragraphs for each question through tf-idf and trained the models using 10-fold cross validation, Adam, learning rate $10^{-2}$ and batch size 128. In text M... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 54 | 27,580 | 28,107 |
11_55 | Enhancing word representation with pre-trained semantic embeddings during training of the TQA task provides an additional boost that results in the highest accuracies for both text MC and diagram MC. These are significantly good results since, according to the TQA authors BIBREF23, most diagram questions in the TQA co... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 55 | 28,107 | 28,706 |
11_56 | The findings reported herein are qualitatively consistent for all the FCC variations in table TABREF15.
Vision features. The analysis was carried out on an unconstrained variety of charts, diagrams and natural images from SciGraph, without filtering by figure type or scientific field. To obtain a representative sampl... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 56 | 28,706 | 29,391 |
11_57 |
Figure FIGREF27 shows a selection of 6 visual features with the 4 figures that activate each feature more significantly and their activation heatmaps. Only figures are used as input, no text. As can be seen, the vision subnetwork has automatically learnt, without explicit supervision, to recognize different types of ... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 57 | 29,391 | 30,007 |
11_58 | Furthermore, as shown by the heatmaps, our model discriminates the key elements associated to the figures that most activate each feature: the actual whiskers, the blots, the borders of each image under comparison, the blots and their complementary bar charts, as well as the line plots and the correspondence between t... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 58 | 30,007 | 30,662 |
11_59 |
We also estimated a notion of semantic specificity based on the concepts of a KG. For each visual feature, we aggregated the captions of the figures that most activate it and used Cogito to disambiguate the Sensigrafo concepts that appear in them. Then, we estimated how important each concept is to each feature by ca... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 59 | 30,662 | 31,235 |
11_60 | This seems to indicate a correlation between activation and the semantic specificity of each visual feature. For example, the heatmaps of the figures related to the feature with the lowest tf-idf (left-most column) highlights a particular visual pattern, i.e. the whiskers, that may spread over many, possibly unrelated... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 60 | 31,235 | 31,729 |
11_61 | Others, like the feature illustrated by the figures in the fifth column, capture the semantics of a specific type of 2D charts relating two magnitudes x and y. Analyzing their captions with Cogito, we see that concepts like e.g. isochronal and exponential functions are mentioned. If we look at the second and four top-... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 61 | 31,729 | 32,256 |
11_62 | Similar to the visual case, we selected the features from the last block of the language subnetwork with the highest activation. For visualization purposes, we picked the figures corresponding to the captions in SciGraph that most activate such features (figure FIGREF28). No visual information is used.
Several distin... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 62 | 32,256 | 32,789 |
11_63 | Interestingly, it also seems to have learnt some type of is-a relations (western blot is a type of immunoblot). The second feature focuses on variations of the term radiograph, e.g. radiograph-y/s. The third feature specializes in text related to curve plots involving several statistic analysis, e.g. Real-time PCR, on... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 63 | 32,789 | 33,410 |
11_64 | The fourth feature extracts citations and models named after prominent scientists, e.g. Evans function (first and fourth figure), Manley (1992) (second), and Aliev-Panfilov model (third). The fifth feature extracts chromatography terminology, e.g. 3D surface plot, photomicrograph or color map and, finally, the right-m... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 64 | 33,410 | 33,936 |
11_65 | Figure FIGREF29 shows the activation heatmaps for two sample captions, calculated on the embeddings layer of the language subnetwork. The upper one corresponds to the fourth column left-right and third figure top-down in figure FIGREF28. Its caption reads: "The Aliev-Panfilov model with $\alpha =0.01$...The phase port... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 65 | 33,936 | 34,337 |
11_66 | Below, (first column, fourth figure in figure FIGREF28): "Relative protein levels of ubiquitin-protein conjugates in M. quadriceps...A representative immunoblot specific to ubiquitin...". Consistently with our analysis, activation focuses on the most relevant tokens for each text feature: "Aliev-Panfilov model" and "i... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 66 | 34,337 | 34,922 |
11_67 | In this paper, we provide empirical evidence of this and show that co-training text and visual features from a large corpus of scientific figures and their captions in a correspondence task (FCC) is an effective, flexible and elegant unsupervised means towards overcoming such complexity. We show how such features can ... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 67 | 34,922 | 35,625 |
11_68 | In the future, it will be interesting to further the study of the interplay between the semantic concepts explicitly represented in different KGs, contextualized embeddings e.g. from SciBERT BIBREF31, and the text and visual features learnt in the FCC task. We also plan to continue to charter the knowledge captured in... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 68 | 35,625 | 36,199 |
11_69 |
Figure 1: Proposed 2-branch architecture of the FCC task.
Table 1: FCC and supervised baselines results (% accuracy).
Table 2: Caption length: natural images vs scientific datasets.
Table 3: Bidirectional retrieval. FCC vs. image-sentence matching baselines (%recall@k). Natural images datasets.
Table 4: Bidirecti... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 69 | 36,199 | 36,702 |
11_70 | FCC vs. random, BiDAF, MemoryNet, VQA and DSDP-NET baselines.
Figure 2: Selected visual features and activation heatmaps. The top row labels the dominant pattern for each feature.
Figure 3: Selected text features. Top row labels the dominant pattern for each text feature.
Figure 4: Sample caption activation heatmap... | https://arxiv.org/abs/1909.09070 | Look, Read and Enrich - Learning from Scientific Figures and their Captions | 70 | 36,702 | 37,056 |
12_0 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity
In this paper we describe a deep learning system that has been designed and built for the WASSA 2017 Emotion Intensity Shared Task. We introduce a representation learning approach based on inner attention on top of an RNN. Results show tha... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 0 | 0 | 527 |
12_1 |
Introduction
Twitter is a huge micro-blogging service with more than 500 million tweets per day from different locations in the world and in different languages. This large, continuous, and dynamically updated content is considered a valuable resource for researchers. In particular, many of these messages contain em... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 1 | 527 | 1,207 |
12_2 | For example, dejected and wistful denote some amount of sadness, and are thus associated with sadness. On the other hand, some words are associated with affect even though they do not denote affect. For example, failure and death describe concepts that are usually accompanied by sadness and thus they denote some amoun... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 2 | 1,207 | 1,841 |
12_3 | To this end, the WASSA-2017 Shared Task on Emotion Intensity BIBREF0 represents the first task where systems have to automatically determine the intensity of emotions in tweets. Concretely, the objective is to given a tweet containing the emotion of joy, sadness, fear or anger, determine the intensity or degree of the... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 3 | 1,841 | 2,365 |
12_4 | Given the 140 character limit of tweets, it is also possible to find some phenomena such as the intensive usage of emoticons and of other special Twitter features, such as hashtags and usernames mentions —used to call or notify other users. In this paper we describe our system designed for the WASSA-2017 Shared Task o... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 4 | 2,365 | 2,830 |
12_5 | In particular, we use a Bi-LSTM model with intra-sentence attention on top of word embeddings to generate a tweet representation that is suitable for emotion intensity. Our results show that our proposed model offers interesting capabilities compared to approaches that do rely on external information sources.
Propose... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 5 | 2,830 | 3,456 |
12_6 | Since in the task the input-output alignment is explicit, they investigated how the alignment can be best utilized in encoder-decoder models concluding that the attention mechanisms are helpful.
EmoAtt is based on a bidirectional RNN that receives an embedded input sequence INLINEFORM0 and returns a list of hidden ve... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 6 | 3,456 | 3,837 |
12_7 | To improve the capabilities of the RNN to capture short-term temporal dependencies BIBREF4 , we define the following: DISPLAYFORM0
Where INLINEFORM0 can be regarded as a context window of ordered word embedding vectors around position INLINEFORM1 , with a total size of INLINEFORM2 . To further complement the context... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 7 | 3,837 | 4,347 |
12_8 | Finally, we combine our INLINEFORM5 augmented hidden states, compressing them into a single vector, using a global intra-sentence attentional component in a fashion similar to vinyalsgrammar2015. Formally, DISPLAYFORM0
Where INLINEFORM0 is the vector that compresses the input sentence INLINEFORM1 , focusing on the r... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 8 | 4,347 | 4,964 |
12_9 |
Experimental Setup
To test our model, we experiment using the training, validation and test datasets provided for the shared task BIBREF5 , which include tweets for four emotions: joy, sadness, fear, and anger. These were annotated using Best-Worst Scaling (BWS) to obtain very reliable scores BIBREF6 .
We experimen... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 9 | 4,964 | 5,371 |
12_10 | These are vectors trained on a dataset of 2B tweets, with a total vocabulary of 1.2 M. To pre-process the data, we used Twokenizer BIBREF8 , which basically provides a set of curated rules to split the tweets into tokens. We also use Tweeboparser BIBREF9 to get the POS-tags for each tweet.
Table TABREF3 summarizes th... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 10 | 5,371 | 5,924 |
12_11 | We also see there is an important vocabulary gap between the dataset and GloVe, with an average coverage of only 64.3 %. To tackle this issue, we used a set of binary features derived from POS tags to capture some of the semantics of the words that are not covered by the GloVe embeddings. We also include features for ... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 11 | 5,924 | 6,525 |
12_12 |
While the structure of our introduced model allows us to easily include more linguistic features that could potentially improve our predictive power, such as lexicons, since our focus is to study sentence representation for emotion intensity, we do not experiment adding any additional sources of information as input.... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 12 | 6,525 | 7,137 |
12_13 | Given this setting, we explored different hyper-parameter configurations, including context window sizes of 1, 3 and 5 as well as RNN hidden state sizes of 100, 200 and 300. We experimented with unidirectional and bidirectional versions of the RNNs.
To avoid over-fitting, we used dropout regularization, experimenting... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 13 | 7,137 | 7,583 |
12_14 | We experimented with different values for weight INLINEFORM2 , with a minimum value of 0.01 and a maximum of 0.2.
To evaluate our model, we wrapped the provided scripts for the shared task and calculated the Pearson correlation coefficient and the Spearman rank coefficient with the gold standard in the validation set... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 14 | 7,583 | 8,060 |
12_15 |
For training, we used mini-batch stochastic gradient descent with a batch size of 16 and padded sequences to a maximum size of 50 tokens, given the nature of the data. We used exponential decay of ratio INLINEFORM0 and early stopping on the validation when there was no improvement after 1000 steps. Our code is availa... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 15 | 8,060 | 8,531 |
12_16 | In general, as Table TABREF13 shows, our intra-sentence attention RNN was able to outperform the Weka baseline BIBREF5 on the development dataset by a solid margin. Moreover, the model manages to do so without any additional resources, except pre-trained word embeddings. These results are, however, reversed for the te... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 16 | 8,531 | 9,117 |
12_17 |
To validate the usefulness of our binary features, we performed an ablation experiment and trained our best models for each corpus without them. Table TABREF15 summarizes our results in terms of Pearson correlation on the development portion of the datasets. As seen, performance decreases in all cases, which shows th... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 17 | 9,117 | 9,764 |
12_18 |
On the other hand, our model also offers us very interesting insights on how the learning is performed, since we can inspect the attention weights that the neural network is assigning to each specific token when predicting the emotion intensity. By visualizing these weights we can have a clear notion about the parts ... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 18 | 9,764 | 10,395 |
12_19 | However, we also see some examples where the lack of semantic information about the input words, specially for hashtags or user mentions, makes the model unable to identify some of these the most salient words to predict emotion intensity. Several pre-processing techniques can be implemented to alleviate this problem,... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 19 | 10,395 | 10,993 |
12_20 | However on ly the first of these values was significant, with a p-value of INLINEFORM0 . Regarding the hidden size of the RNN, we could not find statistical difference across the tested sizes. Dropout also had inconsistent effects, but was generally useful.
Joy Dataset
In the joy dataset, our experiments showed us t... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 20 | 10,993 | 11,520 |
12_21 | Regarding the hidden size of the RNN, we observed that 100 hidden units offered better performance in our experiments, with an average absolute gain of 0.052 ( INLINEFORM2 ) over 50 hidden units. Compared to the models with 200 hidden units, the performance difference was statistically not significant. | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 21 | 11,520 | 11,824 |
12_22 |
Fear Dataset
On the fear dataset, again we observed that embeddings of size 50 provided the best results, offering average gains of 0.12 ( INLINEFORM0 ) and 0.11 ( INLINEFORM1 ) for sizes 25 and 100, respectively. | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 22 | 11,824 | 12,040 |
12_23 | When it comes to the size of the RNN hidden state, our experiments showed that using 100 hidden units offered the best results, with average absolute gains of 0.117 ( INLINEFORM2 ) and 0.108 ( INLINEFORM3 ) over sizes 50 and 200.
Sadness Dataset
Finally, on the sadness datasets again we experimentally observed that ... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 23 | 12,040 | 12,508 |
12_24 | Results were statistically equivalent for size 100. We also observed that using 50 or 100 hidden units for the RNN offered statistically equivalent results, while both of these offered better performance than when using a hidden size of 200.
Conclusions
In this paper we introduced an intra-sentence attention RNN for... | https://arxiv.org/abs/1708.05521 | EmoAtt at EmoInt-2017: Inner attention sentence embedding for Emotion Intensity | 24 | 12,508 | 12,926 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.