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 |
|---|---|---|---|---|---|---|
17_12 | As we apply supervised training, the objective is to maximize the likelihood of all sentence labels $\mathbf {y}_L=(y_L^1, \cdots , y_L^m)$ given the input document $D$ and model parameters $\theta $ :
$$\log p(\mathbf {y}_L |D; \theta ) = \sum \limits _{i=1}^{m} \log p(y_L^i |D; \theta )$$ (Eq. | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 12 | 7,093 | 7,394 |
17_13 | 5)
Although extractive methods yield naturally grammatical summaries and require relatively little linguistic analysis, the selected sentences make for long summaries containing much redundant information. For this reason, we also develop a model based on word extraction which seeks to find a subset of words in $D$ ... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 13 | 7,394 | 7,845 |
17_14 | Compared to sentence extraction which is a sequence labeling problem, this task occupies the middle ground between full abstractive summarization which can exhibit a wide range of rewrite operations and extractive summarization which exhibits none. We formulate word extraction as a language generation task with an out... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 14 | 7,845 | 8,216 |
17_15 | In our supervised setting, the training goal is to maximize the likelihood of the generated sentences, which can be further decomposed by enforcing conditional dependencies among their constituent words:
$$\hspace*{-5.69046pt}\log p(\mathbf {y}_s |D;
\theta )\hspace*{-2.84544pt}=\hspace*{-2.84544pt}\sum \limits _{i=... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 15 | 8,216 | 8,584 |
17_16 | D, w^{\prime }_1,\hspace*{-2.84544pt}\cdots \hspace*{-2.84544pt}, w^{\prime }_{i-1}; \theta )$$ (Eq. 7)
In the following section, we discuss the data elicitation methods which allow us to train neural networks based on the above defined objectives.
Training Data for Summarization
Data-driven neural summarization... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 16 | 8,584 | 9,032 |
17_17 | Until now such corpora have been limited to hundreds of examples (e.g., the DUC 2002 single document summarization corpus) and thus used mostly for testing BIBREF7 . To overcome the paucity of annotated data for training, we adopt a methodology similar to hermann2015teaching and create two large-scale datasets, one fo... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 17 | 9,032 | 9,559 |
17_18 | The highlights (created by news editors) are genuinely abstractive summaries and therefore not readily suited to supervised training. To create the training data for sentence extraction, we reverse approximated the gold standard label of each document sentence given the summary based on their semantic correspondence B... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 18 | 9,559 | 10,069 |
17_19 | The rules take into account the position of the sentence in the document, the unigram and bigram overlap between document sentences and highlights, the number of entities appearing in the highlight and in the document sentence. We adjusted the weights of the rules on 9,000 documents with manual sentence labels created... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 19 | 10,069 | 10,664 |
17_20 |
For the creation of the word extraction dataset, we examine the lexical overlap between the highlights and the news article. In cases where all highlight words (after stemming) come from the original document, the document-highlight pair constitutes a valid training example and is added to the word extraction dataset... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 20 | 10,664 | 11,330 |
17_21 | Following this procedure, we obtained a word extraction dataset containing 170K articles, again from the DailyMail.
Neural Summarization Model
The key components of our summarization model include a neural network-based hierarchical document reader and an attention-based hierarchical content extractor. The hierarchi... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 21 | 11,330 | 11,963 |
17_22 | Such a representation yields minimum information loss and is flexible allowing us to apply neural attention for selecting salient sentences and words within a larger context. In the following, we first describe the document reader, and then present the details of our sentence and word extractors.
Document Reader
The... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 22 | 11,963 | 12,628 |
17_23 | Next, we build representations for documents using a standard recurrent neural network (RNN) that recursively composes sentences. The CNN operates at the word level, leading to the acquisition of sentence-level representations that are then used as inputs to the RNN that acquires document-level representations, in a h... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 23 | 12,628 | 13,126 |
17_24 | Firstly, single-layer CNNs can be trained effectively (without any long-term dependencies in the model) and secondly, they have been successfully used for sentence-level classification tasks such as sentiment analysis BIBREF19 . Let $d$ denote the dimension of word embeddings, and $s$ a document sentence consisting of... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 24 | 13,126 | 13,587 |
17_25 | We apply a temporal narrow convolution between $\mathbf {W}$ and a kernel $\mathbf {K} \in \mathbb {R}^{c \times d}$ of width $c$ as follows:
$$\mathbf {f}^{i}_{j} = \tanh (\mathbf {W}_{j : j+c-1} \otimes \mathbf {K} + b)$$ (Eq. 12)
where $\otimes $ equates to the Hadamard Product followed by a sum over all elem... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 25 | 13,587 | 13,912 |
17_26 | $\mathbf {f}^i_j $ denotes the $j$ -th element of the $i$ -th feature map $\mathbf {f}^i$ and $b$ is the bias. We perform max pooling over time to obtain a single feature (the $i$ th feature) representing the sentence under the kernel $\mathbf {K}$ with width $c$ :
$$\mathbf {s}_{i, \mathbf {K}}= \max _j \mathbf {f}... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 26 | 13,912 | 14,245 |
17_27 | 13)
In practice, we use multiple feature maps to compute a list of features that match the dimensionality of a sentence under each kernel width. In addition, we apply multiple kernels with different widths to obtain a set of different sentence vectors. Finally, we sum these sentence vectors to obtain the final sente... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 27 | 14,245 | 14,844 |
17_28 | The sentence embeddings obtained under each kernel width are summed to get the final sentence representation (denoted by green).
At the document level, a recurrent neural network composes a sequence of sentence vectors into a document vector. Note that this is a somewhat simplistic attempt at capturing document organ... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 28 | 14,844 | 15,567 |
17_29 |
The RNN we used has a Long Short-Term Memory (LSTM) activation unit for ameliorating the vanishing gradient problem when training long sequences BIBREF20 . | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 29 | 15,567 | 15,724 |
17_30 | Given a document $d=(s_1,
\cdots , s_m)$ , the hidden state at time step $t$ , denoted by $\mathbf {h_t}$ , is updated as:
$$\begin{bmatrix}
\mathbf {i}_t\\ \mathbf {f}_t\\ \mathbf {o}_t\\ \mathbf {\hat{c}}_t
\end{bmatrix} =
\begin{bmatrix} \sigma \\ \sigma \\ \sigma \\ \tanh \end{bmatrix} \mathbf {W}\cdot \ | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 30 | 15,724 | 16,036 |
17_31 | begin{bmatrix} \mathbf {h}_{t-1}\\ \mathbf {s}_t
\end{bmatrix}$$ (Eq. 15)
$$ \mathbf {c}_t = \mathbf {f}_t \odot \mathbf {c}_{t-1} +
\mathbf {i}_t \odot \mathbf {\hat{c}}_t$$ (Eq. 16)
where $\mathbf {W}$ is a learnable weight matrix. | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 31 | 16,036 | 16,277 |
17_32 | Next, we discuss a special attention mechanism for extracting sentences and words given the recurrent document encoder just described, starting from the sentence extractor.
Sentence Extractor
In the standard neural sequence-to-sequence modeling paradigm BIBREF11 , an attention mechanism is used as an intermediate st... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 32 | 16,277 | 16,974 |
17_33 | The complete architecture for the document encoder and the sentence extractor is shown in Figure 2 . As can be seen, the next labeling decision is made with both the encoded document and the previously labeled sentences in mind. | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 33 | 16,974 | 17,203 |
17_34 | Given encoder hidden states $(h_1, \cdots , h_m)$ and extractor hidden states $(\bar{h}_1, \cdots , \bar{h}_m)$ at time step $t$ , the decoder attends the $t$ -th sentence by relating its current decoding state to the corresponding encoding state:
$$\bar{\mathbf {h}}_{t} = \text{LSTM} ( p_{t-1} \mathbf {s}_{t-1}, \m... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 34 | 17,203 | 17,554 |
17_35 | 20)
$$p(y_L(t)=1 | D ) = \sigma (\text{MLP} (\mathbf {\bar{h}}_t : \mathbf {h}_t) )$$ (Eq. 21)
where MLP is a multi-layer neural network with as input the concatenation of $\mathbf {\bar{h}}_t$ and $\mathbf {h}_t$ . | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 35 | 17,554 | 17,776 |
17_36 | $p_{t-1}$ represents the degree to which the extractor believes the previous sentence should be extracted and memorized ( $p_{t-1}$ =1 if the system is certain; 0 otherwise).
In practice, there is a discrepancy between training and testing such a model. During training we know the true label $p_{t-1}$ of the previous... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 36 | 17,776 | 18,328 |
17_37 | To mitigate this, we adopt a curriculum learning strategy BIBREF21 : at the beginning of training when $p_{t-1}$ cannot be predicted accurately, we set it to the true label of the previous sentence; as training goes on, we gradually shift its value to the predicted label $p(y_L(t-1)=1
| d )$ .
Word Extractor
Compare... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 37 | 18,328 | 18,847 |
17_38 | A small extension to the structure of the sequential labeling model makes it suitable for generation: instead of predicting a label for the next sentence at each time step, the model directly outputs the next word in the summary. | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 38 | 18,847 | 19,077 |
17_39 | The model uses a hierarchical attention architecture: at time step $t$ , the decoder softly attends each document sentence and subsequently attends each word in the document and computes the probability of the next word to be included in the summary $p(w^{\prime }_t = w_i|
d, w^{\prime }_1, \cdots , w^{\prime }_{t-1})... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 39 | 19,077 | 19,507 |
17_40 | bar{h}}_{t-1})\footnote {We empirically found that feeding
the previous sentence-level attention vector as additional
input to the LSTM would lead to small performance improvements.
This is not shown in the equation.}$$ (Eq. 25)
$$a_j^t = \mathbf {z}^\mathtt {T} \tanh (\mathbf {W}_e \mathbf {\bar{h}}_t + \mathbf {W... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 40 | 19,507 | 19,865 |
17_41 | 26)
In the above equations, $\mathbf {w}_i$ corresponds to the vector of the $i$ -th word in the input document, whereas $\mathbf {z}$ , $\mathbf {W}_e$ , $\mathbf {W}_r$ , $\mathbf {v}$ , $\mathbf {W}_{e^{\prime }}$ , and $\mathbf {W}_{r^{\prime }}$ are model weights. The model architecture is shown in Figure 3 .
... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 41 | 19,865 | 20,279 |
17_42 | In practice, it is not powerful enough to enforce grammaticality due to the lexical diversity and sparsity of the document highlights. A possible enhancement would be to pair the extractor with a neural language model, which can be pre-trained on a large amount of unlabeled documents and then jointly tuned with the ex... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 42 | 20,279 | 20,918 |
17_43 |
Experimental Setup
In this section we present our experimental setup for assessing the performance of our summarization models. We discuss the datasets used for training and evaluation, give implementation details, briefly introduce comparison models, and explain how system output was evaluated.
Results
Table 1 (u... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 43 | 20,918 | 21,445 |
17_44 | The table also includes results for the lead baseline, the logistic regression classifier (lreg), and three previously published systems (ilp, tgraph, and urank).
The nn-se outperforms the lead and lreg baselines with a significant margin, while performing slightly better than the ilp model. This is an encouraging re... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 44 | 21,445 | 22,084 |
17_45 | Overall, our sentence extraction model achieves performance comparable to the state of the art without sophisticated constraint optimization (ilp, tgraph) or sentence ranking mechanisms (urank). We visualize the sentence weights of the nn-se model in the top half of Figure 4 . As can be seen, the model is able to loca... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 45 | 22,084 | 22,546 |
17_46 | This is somewhat expected given that Rouge is $n$ -gram based and not very well suited to measuring summaries which contain a significant amount of paraphrasing and may deviate from the reference even though they express similar meaning. However, a meaningful comparison can be carried out between nn-we and nn-abs whic... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 46 | 22,546 | 23,131 |
17_47 | The extraction-based generation approach is more robust for proper nouns and rare words, which pose a serious problem to open vocabulary models. An example of the generated summaries for nn-we is shown at the lower half of Figure 4 .
Table 1 (lower half) shows system results on the 500 DailyMail news articles (test s... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 47 | 23,131 | 23,632 |
17_48 | This is due to the fact that the gold standard summaries (aka highlights) tend to be more laconic and as a result involve a substantial amount of paraphrasing. More experimental results on this dataset are provided in the appendix.
The results of our human evaluation study are shown in Table 2 . Specifically, we show... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 48 | 23,632 | 24,036 |
17_49 | Perhaps unsurprisingly, the human-written descriptions were considered best and ranked 1st 27% of the time, however closely followed by our nn-se model which was ranked 1st 22% of the time. The ilp system was mostly ranked in 2nd place (38% of the time). The rest of the systems occupied lower ranks. We further convert... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 49 | 24,036 | 24,466 |
17_50 | This allowed us to perform Analysis of Variance (ANOVA) which revealed a reliable effect of system type. Specifically, post-hoc Tukey tests showed that nn-se and ilp are significantly ( $p < 0.01$ ) better than lead, nn-we, and nn-abs but do not differ significantly from each other or the human goldstandard.
Conclusi... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 50 | 24,466 | 24,972 |
17_51 | Our models can be trained on large scale datasets and learn informativeness features based on continuous representations without recourse to linguistic annotations. Two important ideas behind our work are the creation of hierarchical neural structures that reflect the nature of the summarization task and generation by... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 51 | 24,972 | 25,546 |
17_52 | One way to improve the word-based model would be to take structural information into account during generation, e.g., by combining it with a tree-based algorithm BIBREF31 . It would also be interesting to apply the neural models presented here in a phrase-based setting similar to lebret2015phrase. A third direction wo... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 52 | 25,546 | 26,103 |
17_53 |
Acknowledgments
We would like to thank three anonymous reviewers and members of the ILCC at the School of Informatics for their valuable feedback. The support of the European Research Council under award number 681760 “Translating Multiple Modalities into Text” is gratefully acknowledged.
Appendix
In addition to t... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 53 | 26,103 | 26,537 |
17_54 | Since there is no established evaluation standard for this task, we experimented with three different ROUGE limits: 75 bytes, 275 bytes and full length.
Figure 3: Neural attention mechanism for word extraction.
Figure 2: A recurrent convolutional document reader with a neural sentence extractor.
Table 2: Rankings (... | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 54 | 26,537 | 27,084 |
17_55 | The top half shows the relative attention weights given by the sentence extraction model. Darkness indicates sentence importance. The lower half shows the summary generated by the word extraction. | https://arxiv.org/abs/1603.07252 | Neural Summarization by Extracting Sentences and Words | 55 | 27,084 | 27,281 |
18_0 | Improved Representation Learning for Predicting Commonsense Ontologies
Recent work in learning ontologies (hierarchical and partially-ordered structures) has leveraged the intrinsic geometry of spaces of learned representations to make predictions that automatically obey complex structural constraints. We explore two ... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 0 | 0 | 613 |
18_1 | Our second extension exploits the partial order structure of the training data to find long-distance triplet constraints among embeddings which are poorly enforced by the pairwise training procedure. We find that both incorporating free text and augmented training constraints improve over the original order-embedding ... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 1 | 613 | 1,294 |
18_2 | This background knowledge is crucial for solving many difficult, ambiguous natural language problems in coreference resolution and question answering, as well as the creation of other reasoning machines.
More than just curating a static collection of facts, we would like commonsense knowledge to be represented in a w... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 2 | 1,294 | 1,958 |
18_3 | In these knowledge graphs, nodes represent entities or terms $t$ , and hyperedges are relations $R$ between these entities or terms, with each fact in the knowledge graph represented as a triplet $<t_1, R, t_2>$ . Researchers have developed many models for knowledge representation and learning in this setting BIBREF4 ... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 3 | 1,958 | 2,497 |
18_4 |
While a knowledge graph completion model can represent relations such as Is-A and entailment, there is no mechanism to ensure that its predictions are internally consistent. For example, if we know that a dog is a mammal, and a pit bull is a dog, we would like the model to also predict that a pit bull is a mammal. Th... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 4 | 2,497 | 2,967 |
18_5 |
Recently, a thread of research on representation learning has aimed to create embedding spaces that automatically enforce consistency in these predictions using the intrinsic geometry of the embedding space BIBREF9 , BIBREF0 , BIBREF10 . In these models, the inferred embedding space creates a globally consistent stru... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 5 | 2,967 | 3,542 |
18_6 | While the original work included results on ontology prediction on WordNet, we focus exclusively on the model's application to commonsense knowledge, with its unique characteristics including complex ordering structure, compositional, multi-word entities, and the wealth of commonsense knowledge to be found in large-sc... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 6 | 3,542 | 4,072 |
18_7 | We find incorporating unstructured text brings accuracy from 92.0 to 93.0 on a commonsense dataset containing Is-A relations from ConceptNet and Microsoft Concept Graph (MCG), with larger relative gains from smaller amounts of labeled data.
The second extension uses the complex partial-order structure of real-world o... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 7 | 4,072 | 4,701 |
18_8 |
We find that order embeddings' ease of extension, both by incorporating non-ordered data, and additional training constraints derived from the structure of the problem, makes it a promising avenue for the development of further algorithms for automatic learning and jointly consistent prediction of ontologies.
Data
... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 8 | 4,701 | 5,304 |
18_9 | For our task, we use the hypernym relations only. ConceptNet is a KB of triples consisting of a left term $t_1$ , a relation $R$ , and a right term $t_2$ . The relations come from a fixed set of size 34. But unlike WordNet, terms in ConceptNet can be phrases. We focus on the Is-A relation in this work. MCG also consis... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 9 | 5,304 | 5,774 |
18_10 |
For experiments involving unstructured text, we use the WaCkypedia corpus BIBREF13 .
Models
We introduce two variants of order embeddings. The first incorporates non-hierarchical unstructured text data into the supervised ontology. The second improves the training procedure by adding additional examples representin... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 10 | 5,774 | 6,284 |
18_11 | The vector embeddings satisfy the following property with respect to the partial order: $
x \preceq y \text{ if and only if } \bigwedge _{i=1}^{N}x_{i}\ge y_i
$
where $x$ is the subcategory and $y$ is the supercategory. This means the general concept embedding should be smaller than the specific concept embedding in... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 11 | 6,284 | 6,699 |
18_12 | We can define a surrogate energy for this ordering function as $d(x, y) = \left\Vert \max (0,y-x) \right\Vert ^2$ . | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 12 | 6,699 | 6,816 |
18_13 | The learning objective for order embeddings becomes the following, where $m$ is a margin parameter, $x$ and $y$ are the hierarchically supervised pairs, and $x^{\prime }$ and $y^{\prime }$ are negatively sampled concepts: $
L_{\text{Order}} = \sum _{x,y}\max (0, m+d(x,y)-d(x^{\prime }, y^{\prime }))
$
Joint Text and... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 13 | 6,816 | 7,272 |
18_14 | A standard method for learning word representations is word2vec BIBREF14 , which predicts current word embeddings using a context of surrounding word embeddings. | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 14 | 7,272 | 7,434 |
18_15 | We incorporate a modification of the CBOW model in this work, which uses the average embedding from a window around the current word as a context vector $v_2$ to predict the current word vector $v_1$ : $
v_2 = \frac{1}{window}\sum _{k \in \lbrace -window/2,...,window/2\rbrace \setminus \lbrace t\rbrace }v_{t+k}
$
B... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 15 | 7,434 | 7,916 |
18_16 | and not dot product, $v^{\prime }_1$ and $v^{\prime }_2$ are the negative examples selected from the vocabulary during training: $
& d_\text{pos} = d(v_1,v_2) = \left\Vert v_1- v_2\right\Vert \\
& d_\text{neg} = d(v^{\prime }_1, v^{\prime }_2) = \left\Vert v^{\prime }_1- v^{\prime }_2\right\Vert \\
& L_{ | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 16 | 7,916 | 8,225 |
18_17 | \text{CBOW}}= \sum _{w_c,w_t}\max (0, m+d_\text{pos}-d_\text{neg})
$
Finally, after each gradient update, we map the embeddings back to the positive domain by applying the absolute value function. | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 17 | 8,225 | 8,423 |
18_18 | We propose jointly learning both the order- and text- embedding model with a simple weighted combination of the two objective functions: $
&L_{\text{Joint}} = \alpha _{1}L_{\text{Order}}+\alpha _{2}L_{\text{CBOW}}
$
We perform two sets of experiments on the combined ConceptNet and MCG Is-A relations, using different... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 18 | 8,423 | 8,781 |
18_19 | The first data set, called Data1, uses 119,159 training examples, 1,089 dev examples, and 1,089 test examples. The second dataset, Data2, evenly splits the data in 47,662 examples for each set.
Our baselines for this model are a standard order embedding model, and a bilinear classifier BIBREF6 trained to predict Is-A... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 19 | 8,781 | 9,226 |
18_20 |
We see in Table 2 that while adding extra text data helps all models, the best performance is consistently achieved by a combination of order embeddings and unstructured text.
Long-Range Join and Meet Constraints
Order embeddings map words to a partially-ordered space, which we can think of as a directed acyclic gr... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 20 | 9,226 | 9,648 |
18_21 | For example, if we have $<$ dog IsA mammal $>$ , $<$ mammal IsA animal $>$ , we can produce the training example $<$ dog IsA animal $>$ .
We observe that even more training examples can be created by treating our partial-order structure as a lattice. A lattice is a partial order equipped with two additional operation... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 21 | 9,648 | 10,149 |
18_22 | In our case, the vector join and meet would be the pointwise max and min of two embeddings.
We can add many additional training examples to our data by enforcing that the vector join and meet operations satisfy the joins and meets found in the training lattice/DAG. | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 22 | 10,149 | 10,416 |
18_23 | If $w_c$ and $w_p$ are the nearest common child and parent for a pair $w_1, w_2$ , the loss for join and meet learning can be written as the following: $
& d_c(w_1,w_2,w_c) = \left\Vert \max (0,w_1 \vee w_2-w_c) \right\Vert ^2 \\
& d_p(w_1,w_2,w_p) = \left\Vert \max (0,w_p - w_1 \wedge w_2) \right\Vert ^ | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 23 | 10,416 | 10,724 |
18_24 | 2 \\
& {\small L_\text{join} = \sum _{w_1,w_2,w_c}\max (0, m+d_c(w_1,w_2,w_c))}\\
& {\small L_\text{meet} = \sum _{w_1,w_2,w_p}\max (0, m+d_p(w_1,w_2,w_p))}\\
& L = L_\text{join} + L_\text{meet}
$
In this experiment, we use the same dataset as BIBREF0 , created by taking 40,00 edges | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 24 | 10,724 | 11,009 |
18_25 | from the 838,073-edge transitive closure of the WordNet hierarchy for the dev set, 4,000 for the test set, and training on the rest of the transitive closure. We additionally add the long-range join and meet constraints (3,028,302 and 4,006 respectively) between different concepts and see that the inclusion of this ad... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 25 | 11,009 | 11,441 |
18_26 |
Experiments
In both sets of experiments we train all models using the Adam optimizer BIBREF15 , using embeddings of dimension 50, with all hyperparameters tuned on a development set. When embedding multi-word phrases, we represent them as the average of the constituent word embeddings.
Conclusion and Future Work
I... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 26 | 11,441 | 12,056 |
18_27 | In future work we would like to explore embedding models for structured prediction that automatically incorporate additional forms of reasoning such as negation, joint learning of ontological and other commonsense relations, and the application of improved training methods to new models for ontology prediction such as... | https://arxiv.org/abs/1708.00549 | Improved Representation Learning for Predicting Commonsense Ontologies | 27 | 12,056 | 12,687 |
19_0 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project
During the course of a Humanitarian Assistance-Disaster Relief (HADR) crisis, that can happen anywhere in the world, real-time information is often posted online by the people in need of help which, in turn, can be used by diff... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 0 | 0 | 378 |
19_1 | Automated processing of such posts can considerably improve the effectiveness of such efforts; for example, understanding the aggregated emotion from affected populations in specific areas may help inform decision-makers on how to best allocate resources for an effective disaster response. However, these efforts may b... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 1 | 378 | 977 |
19_2 | In this work, we describe our submission for the 2019 Sentiment, Emotion and Cognitive state (SEC) pilot task of the LORELEI project. We describe a collection of sentiment analysis systems included in our submission along with the features extracted. Our fielded systems obtained the best results in both English and Sp... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 2 | 977 | 1,651 |
19_3 | This information contributes to the creation and dissemination of situational awareness BIBREF2 , BIBREF3 , BIBREF4 , BIBREF0 , and crisis response agencies such as government departments or public health-care NGOs can make use of these channels to gain insight into the situation as it unfolds BIBREF2 , BIBREF5 . Addi... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 3 | 1,651 | 2,131 |
19_4 | While many of these organizations recognize the value of the information found online—specially during the on-set of a crisis—they are in need of automatic tools that locate actionable and tactical information BIBREF7 , BIBREF0 .
Opinion mining and sentiment analysis techniques offer a viable way of addressing these ... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 4 | 2,131 | 2,788 |
19_5 | For example, identifying tweets labeled as “fear” might support responders on assessing mental health effects among the affected population BIBREF11 . Given the critical and global nature of the HADR events, tools must process information quickly, from a variety of sources and languages, making it easily accessible to... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 5 | 2,788 | 3,395 |
19_6 |
The LORELEI program provides a framework for developing and testing systems for real-time humanitarian crises response in the context of low-resource languages. The working scenario is as follows: a sudden state of danger requiring immediate action has been identified in a region which communicates in a low resource ... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 6 | 3,395 | 4,069 |
19_7 | The program's objective is the rapid deployment of systems that can process text or speech audio from a variety of sources, including newscasts, news articles, blogs and social media posts, all in the local language, and populate these Situation Frames. While the task of identifying Situation Frames is similar to exis... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 7 | 4,069 | 4,644 |
19_8 |
The Sentiment, Emotion, and Cognitive State (SEC) evaluation task was a recent addition to the LORELEI program introduced in 2019, which aims to leverage sentiment information from the incoming documents. This in turn may be used in identifying severity of the crisis in different geographic locations for efficient di... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 8 | 4,644 | 5,197 |
19_9 | To this end, our models are based on a combination of state-of-the-art sentiment classifiers and simple rule-based systems. We evaluate our systems as part of the NIST LoREHLT 2019 SEC pilot task.
Previous Work
Social media has received a lot of attention as a way to understand what people communicate during disaste... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 9 | 5,197 | 5,753 |
19_10 | To organize and make sense of the sentiment information found in social media, particularly those messages sent during the disaster, several works propose the use of machine learning models (e.g., Support Vector Machines, Naive Bayes, and Neural Networks) trained on a multitude of linguistic features. These features i... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 10 | 5,753 | 6,298 |
19_11 | Most of the work is centered around identifying messages expressing sentiment towards a particular situation as a way to distinguish crisis-related posts from irrelevant information BIBREF25 . Either in a binary fashion (positive vs. negative) (e.g., BIBREF25 ) or over fine-grained emotional classes (e.g., BIBREF16 ).... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 11 | 6,298 | 6,739 |
19_12 | This can be attributed to a more challenging task due to the nature of the domain since, for example, journalists will often refrain from using clearly positive or negative vocabulary when writing news articles BIBREF27 . However, certain aspects of these communication channels are still apt for sentiment analysis, su... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 12 | 6,739 | 7,271 |
19_13 | Most of which are done by manually constructing resources for a particular language (e.g., in tweets BIBREF30 , BIBREF31 , BIBREF32 and in disaster-related news coverage BIBREF33 ), or by applying cross-language text categorization to build language-specific models BIBREF31 , BIBREF34 .
In this work, we develop syste... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 13 | 7,271 | 7,837 |
19_14 | This makes our approach easily extendable to other languages, bypassing the scalability issues that arise from the need to manually construct lexica resources.
Problem Definition
This section describes the SEC task in the LORELEI program along with the dataset, evaluation conditions and metrics.
The Sentiment, Emot... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 14 | 7,837 | 8,414 |
19_15 | The source is defined as a person or a group of people expressing the sentiment, and can be either a PER/ORG/GPE (person, organization or geo political entity) construct in the frame, the author of the text document, or an entity not explicitly expressed in the document. The target toward which the sentiment is expres... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 15 | 8,414 | 8,888 |
19_16 | Situation Frames (SF) are similar in nature to those used in Natural Language Understanding (NLU) systems: in essence they are data structures that record information corresponding to a single incident at a single location BIBREF15 . A SF frame includes a situation Type taken from a fixed inventory of 11 categories (e... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 16 | 8,888 | 9,487 |
19_17 | A list of situation frames and documents serve as input for our sentiment analysis systems.
Data
Training data provided for the task included documents were collected from social media, SMS, news articles, and news wires. This consisted of 76 documents in English and 47 in Spanish. The data are relevant to the HADR ... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 17 | 9,487 | 10,001 |
19_18 | Sentiment annotations were done at a segment (sentence) level, and included Situation Frame, Polarity (positive / negative), Sentiment Score, Emotion, Source and Target. Sentiment labels were annotated between the values of -3 (very negative) and +3 (very positive) with 0.5 increments excluding 0. Additionally, the pr... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 18 | 10,001 | 10,556 |
19_19 |
Evaluation
Systems participating in the task were expected to produce outputs with sentiment polarity, emotion, sentiment source and target, and the supporting segment from the input document. This output is evaluated against a ground truth derived from two or more annotations. For the SEC pilot evaluation, a refere... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 19 | 10,556 | 11,255 |
19_20 | That is, consider two annotators in agreement even if their judgments vary on sentiment values or perceived emotions. Designate those annotations with agreement as “D” and those which were not agreed upon as “S”. When computing precision, recall and f measure, each of the sentiment annotations in D will count as two o... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 20 | 11,255 | 11,769 |
19_21 | The updated precision, recall and f-measure were defined as follows: $
\text{precision} &= \frac{2 * \text{Matches in D} + \text{Matches in S}}{2 * \text{Matches in D} + \text{Matches in S} + \text{Unmatched}}\\[10pt]
\text{recall} &= \frac{2 * \text{Matches in D} + \text{Matches in S}}{2|D| + |S|}\\[10pt]
\text{f1} &... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 21 | 11,769 | 12,100 |
19_22 | \text{precision} * \text{recall}}{(\text{precision} + \text{recall})}
$
Method
We approach the SEC task, particularly the polarity and emotion identification, as a classification problem. Our systems are based on English, and are extended to other languages via automatic machine translation (to English). In this se... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 22 | 12,100 | 12,649 |
19_23 | For the pilot evaluation, we translated all of the Spanish documents into English, and included them as additional training data. At this time we do not translate English to Spanish, but plan to explore this thread in future work.
Linguistic Features
We extract word unigrams and bigrams. These features were then tra... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 23 | 12,649 | 13,291 |
19_24 | In our feature set we include a 300-dimensional word2vec word representation trained on a large news corpus BIBREF36 . We obtain a representation for each segment by averaging the embedding of each word in the segment. We also experimented with the use of GloVe BIBREF37 , and Sent2Vec BIBREF38 , an extension of word2v... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 24 | 13,291 | 13,856 |
19_25 | We obtained word percentages across 192 lexical categories using Empath BIBREF39 , which extends popular tools such as the Linguistic Inquiry and Word Count (LIWC) BIBREF22 and General Inquirer (GI) BIBREF40 by adding a wider range of lexical categories. These categories include emotion classes such as surprise or dis... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 25 | 13,856 | 14,359 |
19_26 | For this work, we learn sentiment representations using a bilateral Long Short-Term Memory model BIBREF41 trained on the Stanford Sentiment Treebank BIBREF42 . This model was selected because it provided a good trade off between simplicity and performance on a fine-grained sentiment task, and has been shown to achieve... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 26 | 14,359 | 14,997 |
19_27 | These models were pre-trained on larger corpora and evaluated directly on the task without any further adaptation. In a second approach we explore a data augmentation technique based on a proposed simplification of the task. In this approach, traditional machine learning classifiers were trained to identify which segm... | https://arxiv.org/abs/1905.00472 | A system for the 2019 Sentiment, Emotion and Cognitive State Task of DARPAs LORELEI project | 27 | 14,997 | 15,702 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.