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 |
|---|---|---|---|---|---|---|
36_45 | Specifically, we count the friends and followers of each user, their verified status, and the number of tweets posted within six-month snapshots of their timelines, as in BIBREF11, BIBREF4, and BIBREF8.
Feature Engineering ::: Timeline Features
Here, we consider linguistic features, drawn from both the author and ta... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 45 | 25,822 | 26,333 |
36_46 |
Feature Engineering ::: Timeline Features ::: Message Behavior
To more clearly represent the social relationship between the author and target users, we consider the messages sent between them as follows:
Downward mention count: How many messages has the author sent to the target?
Upward mention count: How many me... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 46 | 26,333 | 26,840 |
36_47 | We compute the ratio $\frac{|M_a \cap M_t|}{|M_a \cup M_t|}$.
Multiset mention overlap: Let $\hat{M}_a$ be the multiset of all accounts mentioned by author $a$ (with repeats for each mention), and let $\hat{M}_t$ be the multiset of all accounts mentioned by target $t$. | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 47 | 26,840 | 27,111 |
36_48 | We measure $\frac{|\hat{M}_a \cap ^{*} \hat{M}_t|}{|\hat{M}_a \cup \hat{M}_t|}$ where $\cap ^{*}$ takes the multiplicity of each element to be the sum of the multiplicity from $\hat{M}_a $ and the multiplicity from $\hat{M}_b$
The direct mention count measures the history of repeated communication between the author ... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 48 | 27,111 | 27,446 |
36_49 | For harmful messages, downward overlap is higher ($D=0.178$) and upward overlap is lower ($D=0.374$) than for harmless messages, as shown in Figure FIGREF38. This means malicious authors tend to address the target repeatedly while the target responds with relatively few messages.
Mention overlap is a measure of socia... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 49 | 27,446 | 28,057 |
36_50 | We see in Figure FIGREF38 that repeated aggression is linked to slightly greater mention overlap ($D=0.07$, $p=0.07$), but the trend is significant only for multiset mention overlap ($D=0.08$, $p=0.03$).
Feature Engineering ::: Timeline Features ::: Timeline Similarity
Timeline similarity is used to indicate common ... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 50 | 28,057 | 28,589 |
36_51 | This can be used to distinguish cyberbullying from harmless banter between friends and associates. To compute this metric, we represent the author and target timelines as TF-IDF vectors $\vec{A}$ and $\vec{T}$. We then take the cosine similarity between the vectors as
A cosine similarity of 1 means that users' timeli... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 51 | 28,589 | 29,116 |
36_52 |
In Figure FIGREF44 (a), we see that the timelines were significantly less similar when the target was in a position of greater power ($D=0.294$). This is not surprising, since power can be derived from such differences between social groups. We do not observe the same dissimilarity when the author was more powerful (... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 52 | 29,116 | 29,582 |
36_53 |
Turning to Figure FIGREF44 (b), we see that aggressive messages were less likely to harbor harmful intent if they were sent between users with similar timelines ($D=0.285$). Aggressive banter between friends is generally harmless, so again, this confirms our intuitions.
Feature Engineering ::: Timeline Features ::: ... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 53 | 29,582 | 30,186 |
36_54 | We do this in two ways: through a simple ratio of new words, and through the use of language models.
To estimate historical language behavior, we count unigram and bigram frequencies from a 4-year snapshot of the author's timeline. Then, after removing all URLs, punctuation, stop words, mentions, and hashtags from th... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 54 | 30,186 | 30,721 |
36_55 | We can also build a language model from the bigram frequencies, using Kneser-Ney smoothing as implemented in NLTK BIBREF36. From the language model, we compute the surprise of the original message $m$ according to its cross-entropy, given by
where $m$ is composed of bigrams $b_1, b_2, \dots , b_N$, and $P(b_i)$ is th... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 55 | 30,721 | 31,099 |
36_56 |
We see in Figure FIGREF47 that harmfully intended messages have a greater density of new words ($D=0.06$). This is intuitive, since attacks may be staged around new topics of conversation. However, the cross entropy of these harmful messages is slightly lower than for harmless messages ($D=0.06$). This may be due to ... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 56 | 31,099 | 31,680 |
36_57 |
Feature Engineering ::: Thread Features ::: Visibility
To determine the public visibility of the author's post, we collect basic measurements from the interactions of other users in the thread. They are as follows.
Message count: Count the messages posted in the thread
Reply message count: Count the replies posted... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 57 | 31,680 | 32,257 |
36_58 |
Maximum author retweets: The largest number of retweets the author received on a message in the thread.
Feature Engineering ::: Thread Features ::: Aggression
To detect repeated aggression, we again employ the hate speech and offensive language classifier of BIBREF35. Each message is given a binary label according ... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 58 | 32,257 | 32,802 |
36_59 |
Aggressive message count: Count the messages in the thread classified as aggressive
Aggressive author message count: Count the author's messages that were classified as aggressive
Aggressive user count: Of the users who posted a reply in the thread after the author first commented, count how many had a message clas... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 59 | 32,802 | 33,460 |
36_60 | We averaged results using five-fold cross-validation, with 80% of the data allocated for training and 20% of the data allocated for testing at each iteration. To account for the class imbalance in the training data, we used the synthetic minority over-sampling technique (SMOTE) BIBREF39. We did not over-sample testing... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 60 | 33,460 | 34,031 |
36_61 | Note that because we do not include thread features in the User set, it can be used for cyberbullying prediction and early intervention. The Proposed set can be used for detection, sinct it is a collection of all newly proposed features, including thread features. The Combined adds these to the baseline text features.... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 61 | 34,031 | 34,602 |
36_62 | This was expected and the score aligns well with the success of other published results of Table TABREF8. Cyberbullying detection is more complex than simply identifying aggressive text, however. We find that these same baseline methods fail to reliably detect repetition, harmful intent, visibility among peers, and po... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 62 | 34,602 | 35,282 |
36_63 | When we combine all features, we see our $F_1$ scores beat baselines for each criterion. This demonstrates the effectiveness of our approach, using linguistic similarity and community measurements to encode social characteristics for cyberbullying classification.
Similar results were obtained by replacing our logisti... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 63 | 35,282 | 35,818 |
36_64 | We chose to highlight logistic regression because it can be more easily interpreted. As a result, we can identify the relative importance of our proposed features. The feature weights are also given in Appendix 2, Tables TABREF78-TABREF78. There we observe a trend. The aggressive language and repetition criteria are d... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 64 | 35,818 | 36,489 |
36_65 | Precision scores are generally much lower than recall scores across all models. To reduce our misclassification of false positives and better distinguish between joking or friendly banter and cyberbullying, it may be necessary to mine for additional social features. Overall, we should work to increase all $F_1$ scores... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 65 | 36,489 | 37,116 |
36_66 | This classifier has an F1 score of 0.90 BIBREF35. Even with this filter, only 0.7% of tweets were deemed by a majority of MTurk workers as cyberbullying (Table TABREF17). This extreme class imbalance can disadvantage a wide range of machine learning models. Moreover, the MTurk workers exhibited only moderate inter-ann... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 66 | 37,116 | 37,654 |
36_67 | For these reasons, we recognize that cyberbullying still has not been unambiguously defined. Moreover, their underlying constructs are difficult to identify. In this study, we did not train workers to recognize subtle cues for interpersonal popularity, nor the role of anonymity in creating a power imbalance.
Furtherm... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 67 | 37,654 | 38,312 |
36_68 | However, they are representative of the literature in both the social science and machine learning communities, and they can be used in weighted combinations to accommodate new definitions.
The main contribution of our paper is not that we solved the problem of cyberbullying detection. Instead, we have exposed the ch... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 68 | 38,312 | 38,995 |
36_69 | With cyberbullying being a systemic issue across social media platforms, we urge the development of a methodology for data sharing with researchers that provides adequate access to rich data to improve on the early detection of cyberbullying while also addressing the sensitive privacy issues that accompany such instan... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 69 | 38,995 | 39,630 |
36_70 | In order to more accurately represent the nature of cyberbullying, we decomposed this complex issue into five representative characteristics. Our classes distinguish cyberbullying from other related behaviors, such as isolated aggression or crude joking. To help annotators infer these distinctions, we provided them wi... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 70 | 39,630 | 40,270 |
36_71 | Unlike previous text-based or user-based features, our features measure the relationship between a message author and target. We show that these features improve the performance of standard text-based models. These results demonstrate the relevance of social-network and language-based measurements to account for the n... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 71 | 40,270 | 40,926 |
36_72 |
Acknowledgements
This material is based upon work supported by the Defense Advanced Research Projects Agency (DARPA) under Agreement No. HR0011890019, and by the National Science Foundation (NSF) under Grant No. 1659886 and Grant No. 1553579. | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 72 | 40,926 | 41,171 |
36_73 |
Appendix 1: Analysis of the Real-World Class Distribution for Cyberbullying Criteria
To understand the real-world class distribution for the cyberbullying criteria, we randomly selected 222 directed English tweets from an unbiased sample of drawn from the Twitter Decahose stream across the entire month of October 20... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 73 | 41,171 | 41,680 |
36_74 | Upon analysis, we found that the positive class balance was prohibitively small, especially for repetition, harmful intent, visibility among peers, and author power, which were all under 5%.
Appendix 2: Model Evaluation
For the sake of comparison, we provide precision, recall, and $F_1$ scores for five different mac... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 74 | 41,680 | 42,256 |
36_75 |
Table 1: Datasets built from different related definitions of cyberbullying. For each dataset, we report the size, positive class balance, inter-annotator agreement, and whether the study incorporated social context in the annotation process.
Table 2: State of the Art in Cyberbullying Detection. Here, results are re... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 75 | 42,256 | 42,809 |
36_76 | Although the thread in the middle contains repeated use of aggressive language, there is no harmful intent, visibility among peers, or power imbalance. Overlap measures. (right) Graphical representation of the neighborhood overlap measures of author a and target t.
Figure 2: Cumulative Distribution Functions for neig... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 76 | 42,809 | 43,511 |
36_77 | These measures are shown to be predictive of power imbalance and harmful intent.
Figure 5: Cumulative Distribution Functions for language models on relevant features. These measures are shown to be predictive of harmful intent. | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 77 | 43,511 | 43,740 |
36_78 |
Table 4: Feature Combinations
Table 6: Recall
Table 5: Precision
Table 7: F1 Scores
Table 8: Analysis of Unfiltered Decahose Data
Table 9: Random Forest Precision
Table 10: AdaBoost Precision
Table 11: MLP Precision
Table 19: Top Absolute Weights for Repetition Features
Table 20: Top Absolute Weights for Har... | https://arxiv.org/abs/2004.01820 | Aggressive, Repetitive, Intentional, Visible, and Imbalanced: Refining Representations for Cyberbullying Classification | 78 | 43,740 | 44,162 |
37_0 | Knowledge Amalgam: Generating Jokes and Quotes Together
Generating humor and quotes are very challenging problems in the field of computational linguistics and are often tackled separately. In this paper, we present a controlled Long Short-Term Memory (LSTM) architecture which is trained with categorical data like jok... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 0 | 0 | 551 |
37_1 | Importantly, we believe the neural net has more knowledge as it's trained on different datasets and hence will enable it to generate more creative jokes or quotes from the mixture of information. May the network generate a funny inspirational joke!
Introduction
Generating texts containing human sentiments is one of ... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 1 | 551 | 1,103 |
37_2 |
Detailed studies have been conducted to identify the structure and principles of humor using mathematical models and generate fixed formulaic jokes using unsupervised learning from big data BIBREF0 . Recurrent Neural Networks are popular for text generation tasks and they were used for humor generation BIBREF1 . The ... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 2 | 1,103 | 1,731 |
37_3 |
A controlled LSTM can be used to train a network in a supervised way on multiple categorical data like jokes, quotes, and tweets by augmenting the category tag to the input word at every time-step. This way the neural net can learn the difference in the semantics of a joke and quote and generate more creative content... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 3 | 1,731 | 2,270 |
37_4 | We also found that a network trained on the combined dataset generated fewer offensive jokes compared to the one trained on just the jokes (as the jokes scraped from the internet were offensive with high probability). This is the first time anyone has used controlled LSTM architecture to generate texts with different ... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 4 | 2,270 | 2,930 |
37_5 |
Finally, we trained our model with three categories of data namely: jokes, quotes, and tweets. We show that the network is able to generate texts belonging to the specific category when we pass the category as input along with the seed text.
Recurrent Neural Networks
RNNs have been historically used for text genera... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 5 | 2,930 | 3,449 |
37_6 | Character level RNN has been shown to generate syntactically coherent texts by BIBREF2 but they are not semantically meaningful since only character level information is available to the char-RNN. But encoder-decoder models BIBREF3 have been proved to be successful in machine translation BIBREF4 and image captioning t... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 6 | 3,449 | 3,931 |
37_7 |
A variant of RNN called Long Short-Term Memory (LSTM) introduced by BIBREF7 have been shown to perform better than conventional RNNs by BIBREF8 which overcomes the above-mentioned modeling limitations of RNNs. LSTMs use three gates to regulate the hidden state variable of LSTM which functions as the memory unit.
Hum... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 7 | 3,931 | 4,443 |
37_8 | In the former work, an LSTM model with attention is used to generate jokes from a dataset consisting of 7699 jokes written by a single author giving the corpus a homogeneity of style. The jokes data is mixed with news data and a deep recurrent neural network is trained with weighted-pick strategy above the output laye... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 8 | 4,443 | 4,838 |
37_9 | This is the only work to the best of our knowledge which tried to mix jokes and non-jokes during training in order to bring in more information but it didn't train the network with a category tag so we have no control over what the model generates which makes judging it more subjective. It also means the network is tr... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 9 | 4,838 | 5,434 |
37_10 | We combined multiple sources and de-duplicated them to arrive at a large corpus for training. The two sources for jokes are CrowdTruth and Subreddits. After cleaning, we ended up with 96910 jokes and a vocabulary size of 8922 words. The two sources for quotes are Quotables and the TheWebMiner. After cleaning, we ended... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 10 | 5,434 | 5,812 |
37_11 | We downloaded the scraped tweets from kaggle and ended up with 130250 tweets with a vocabulary size of 10805 words after cleaning. We constrained the vocabulary to about 10000 words in each case. Finally, we combined the jokes, quotes, and tweets along with their class labels (joke is 0, quote is 1, tweet is 2) into a... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 11 | 5,812 | 6,244 |
37_12 | Each sentence starts with a 'sos' tag and ends with a 'eos' tag to denote the start and end of sentences. The final datasets can be found on our github repository. When we train the controlled LSTM with the combined data, we use weighted sample strategy so that the three categories contribute equally to loss even thou... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 12 | 6,244 | 6,796 |
37_13 | We tokenized each of our training examples into words and punctuation.As shown in figure ,We had the choice of training a many-to-one or many-to-many recurrent neural network. We choose many-to-many model so that the LSTM has more output to learn from. At every time step of the recurrent neural network, we pass the on... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 13 | 6,796 | 7,296 |
37_14 | Figure FIGREF13 shows the final architecture we used for training a single deep learning neural network with jokes, quotes, and tweets together with category tag and were able to successfully generate text with corresponding category later by setting the category input.
Model Parameters
Our final model has the follo... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 14 | 7,296 | 7,794 |
37_15 |
Vocabulary size: 12614
GloVe embedding dim: 200
Input embedding dim: 512
First dense layer dim: 512
LSTM-1 hidden layer dim: 1024
LSTM-2 hidden layer dim: 512
Final Dense layer dim: 512
DropOut factor: 0.2
Regularization factor: 0.00001
Input sequence length: 13
Experiments
To show that controlled LSTMs wo... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 15 | 7,794 | 8,154 |
37_16 | The first experiment was to train the network with just jokes for a baseline model. Second, we trained the model with the words of the jokes in the correct order (tag = 0) and reverse order (tag = 1) to show that the LSTM will learn to generate jokes in forward and reverse order according to the input category tag as ... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 16 | 8,154 | 8,728 |
37_17 |
Just Jokes
The first experiment was training the model with just jokes.
Examples of generated jokes:
i had to use a new word. plagiarism!
i had to use a lot of money in the closet. i just got laid by a chick.
yo mama is so hairy that bigfoot is seeing her last night.
Forward and Reverse Jokes
To validate our c... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 17 | 8,728 | 9,209 |
37_18 | We train our LSTM model with the combined data with control bit set different for the two instances and later ask it to generate jokes in the required order by setting the control bit. We found that the LSTM was surprisingly good at learning the semantics of correct and reversed sentences.
Example reversed joke: eos ... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 18 | 9,209 | 9,707 |
37_19 | eos
Importantly, the sentence is semantically correct in the reverse order meaning the LSTM learns both the forward and reverse semantics and generates text in the required order depending on input category/control bit.
Jokes, Quotes, and Tweets
Finally, we trained the network with all three types: jokes, quotes, a... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 19 | 9,707 | 10,249 |
37_20 | Also, we found the jokes generated by this model to be far less offensive compared to the jokes generated by a network trained on just humor dataset proving the mixture of information hypothesis. Figure FIGREF40 , figure FIGREF44 and figure FIGREF45 show examples of sentences with different sentiments generated on dif... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 20 | 10,249 | 10,897 |
37_21 | If exploration factor is 0 then we will always pick the word with the highest probability else if it's 0.3 then 30% of the times we pick a word according to softmax output layer probabilities. We found that a low non-zero exploration factor gave the best results.
Evaluation
We evaluated the similarity of the generat... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 21 | 10,897 | 11,491 |
37_22 | The Phrase Overlap criterion introduced by BIBREF9 and extended by BIBREF10 is chosen because it gives more weight to multi token phrase overlaps as its rarer. DISPLAYFORM0 DISPLAYFORM1
The lower the INLINEFORM0 value the novel the generated text which is indicative of the fact that our model is able to generalize b... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 22 | 11,491 | 12,063 |
37_23 | We then average the maximums of the Phrase Overlap metric between the generated texts (excluding the seed text) and all remaining training examples in our corpus to arrive at our final aggregated similarity score.
While comparing sentences with high Phrase Overlap score we found that the metric is giving a high score... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 23 | 12,063 | 12,663 |
37_24 | or "Yo mama so fat..." so it is expected that our model will pick up these common phrases.
For example, the generated quote "i love the music that i love because i love it . i can't tell you how to live without it" and "i am always training because i love it" has a Phrase Overlap score of 0.66 because of the presence... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 24 | 12,663 | 13,131 |
37_25 | Overlap of bigrams, trigrams, and even 4-grams can be expected between texts with very different meaning but phrase overlap score heavily penalizes such matches.
We also used the K-gram-Jaccard score to compute the similarity measure. We use INLINEFORM0 to overcome the drawbacks of Phrase Overlap and compute the simi... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 25 | 13,131 | 13,693 |
37_26 | This is expected because with higher exploration value we are allowing the generator to pick a not so best word which leads to more variations in the text but with lesser coherence.
To evaluate the syntactic correctness of the generated sentences we have used the Link Grammar Parser for English language developed by ... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 26 | 13,693 | 14,329 |
37_27 | Since the parsing is dependent on capitalization and our training data has been normalized to all small letters, we add necessary capitalization before feeding sentences to the parser. Here are examples of linkage diagram produced by syntactic parsing of a generated sentence. The lines represents the links between the... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 27 | 14,329 | 14,921 |
37_28 | If the word in a sentence is out of vocabulary of the parser, it is followed by [?], followed by one of .n, .v, .a, or .e, depending on whether the word is being interpreted as a noun, verb, adjective, or adverb. If a word is enclosed in square brackets, this indicates that the parser was forced to delete this word in... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 28 | 14,921 | 15,340 |
37_29 | In the first phase, attempt is made to find "complete" linkage for a sentence without leaving out any word, in which all the words are linked together. If the parser cannot interpret the sentence, it begins to relax this constraint. The value INLINEFORM0 indicates that the parser is allowing INLINEFORM1 words to be ig... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 29 | 15,340 | 15,947 |
37_30 |
In the above example we can see, out of total possible 261 linkages 100 linkages were without any p.p (post processing) error. We can also see the parser identifies the parts of speech of the words and the syntactic parsing is correct.
To evaluate the overall syntactic accuracy of our corpus we consider total percen... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 30 | 15,947 | 16,434 |
37_31 | From 150 generated texts we obtained 251 sentences and processed them adequately with capitalization. We have used exploration factor 0.1 while generating the texts. The results are presented in table TABREF42 . The accuracy is INLINEFORM0 i.e INLINEFORM1 of sentences were entirely correct and INLINEFORM2 were almost ... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 31 | 16,434 | 16,939 |
37_32 |
To evaluate the quality of the generated jokes, quotes, or tweets we rely on human judgment as there is no proven system to measure the quality of content objectively.
Giving the different category tags as input for the same prefix, we were able to generate texts of different categories.
Conclusion
We have propose... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 32 | 16,939 | 17,579 |
37_33 | The mixture of information is able to generate more creative content. We gathered a large corpus of jokes, quotes, and tweets from multiple sources for training our model which can be used for further research. The produced texts were subjectively and objectively evaluated, they are found to be semantically and syntac... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 33 | 17,579 | 18,296 |
37_34 |
In future, we will build upon our idea of inferencing patterns and relations between different categories of text to generate more creative and quality content with high probability.
Figure 1: Sequence to Sequence
Figure 2: Controlled Stacked LSTM
Figure 3: Sequence to Sequence with Category Tag
Figure 4: Loss an... | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 34 | 18,296 | 18,782 |
37_35 |
Figure 8: Generating Joke, Quotes, and Tweets
Table 3: Syntactic Accuracy of generated sentences
Figure 6: Generating Joke, Quotes, and Tweets
Figure 7: Generating Joke, Quotes, and Tweets | https://arxiv.org/abs/1806.04387 | Knowledge Amalgam: Generating Jokes and Quotes Together | 35 | 18,782 | 18,975 |
38_0 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization
In this paper, we introduce an embedding model, named CapsE, exploring a capsule network to model relationship triples (subject, relation, object). Our CapsE represents each triple as a 3-column matrix where each column v... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 0 | 0 | 645 |
38_1 | The length of this vector is used to measure the plausibility score of the triple. Our proposed CapsE obtains better performance than previous state-of-the-art embedding models for knowledge graph completion on two benchmark datasets WN18RR and FB15k-237, and outperforms strong search personalization baselines on SEAR... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 1 | 645 | 1,236 |
38_2 | However, large knowledge graphs, even containing billions of triples, are still incomplete, i.e., missing a lot of valid triples BIBREF1 . Therefore, much research efforts have focused on the knowledge graph completion task which aims to predict missing triples in KGs, i.e., predicting whether a triple not in KGs is l... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 2 | 1,236 | 1,610 |
38_3 | To this end, many embedding models have been proposed to learn vector representations for entities (i.e., subject/head entity and object/tail entity) and relations in KGs, and obtained state-of-the-art results as summarized by BIBREF5 and BIBREF6 . These embedding models score triples (s, r, o), such that valid triple... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 3 | 1,610 | 2,158 |
38_4 |
Triple modeling is applied not only to the KG completion, but also for other tasks which can be formulated as a triple-based prediction problem. An example is in search personalization, one would aim to tailor search results to each specific user based on the user's personal interests and preferences BIBREF7 , BIBREF... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 4 | 2,158 | 2,524 |
38_5 | Here the triples can be formulated as (submitted query, user profile, returned document) and used to re-rank documents returned to a user given an input query, by employing an existing KG embedding method such as TransE BIBREF3 , as proposed by BIBREF12 . Previous studies have shown the effectiveness of modeling tripl... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 5 | 2,524 | 2,982 |
38_6 |
Conventional embedding models, such as TransE BIBREF3 , DISTMULT BIBREF13 and ComplEx BIBREF14 , use addition, subtraction or simple multiplication operators, thus only capture the linear relationships between entities. Recent research has raised interest in applying deep neural networks to triple-based prediction pr... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 6 | 2,982 | 3,647 |
38_7 | To the best of our knowledge, however, none of the existing models has a “deep” architecture for modeling the entries in a triple at the same dimension.
BIBREF16 introduced capsule networks (CapsNet) that employ capsules (i.e., each capsule is a group of neurons) to capture entities in images and then uses a routing... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 7 | 3,647 | 4,235 |
38_8 | Each capsule accounts for capturing variations of an object or object part in the image, which can be efficiently visualized. Our high-level hypothesis is that embedding entries at the same dimension of the triple also have these variations, although it is not straightforward to be visually examined.
To that end, we ... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 8 | 4,235 | 4,907 |
38_9 | In our CapsE, INLINEFORM0 , INLINEFORM1 and INLINEFORM2 are unique INLINEFORM3 -dimensional embeddings of INLINEFORM4 , INLINEFORM5 and INLINEFORM6 , respectively. The embedding triple [ INLINEFORM7 , INLINEFORM8 , INLINEFORM9 ] of (s, r, o) is fed to the convolution layer where multiple filters of the same INLINEFORM... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 9 | 4,907 | 5,429 |
38_10 | Thus, each capsule can encode many characteristics in the embedding triple to represent the entries at the corresponding dimension. These capsules are then routed to another capsule which outputs a continuous vector whose length is used as a score for the triple. Finally, this score is used to predict whether the trip... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 10 | 5,429 | 5,966 |
38_11 | To our best of knowledge, our work is the first consideration of exploring the capsule network to knowledge graph completion and search personalization.
INLINEFORM0 We evaluate our CapsE for knowledge graph completion on two benchmark datasets WN18RR BIBREF17 and FB15k-237 BIBREF18 . CapsE obtains the best mean rank... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 11 | 5,966 | 6,528 |
38_12 | We adapt our model to search personalization and evaluate on SEARCH17 BIBREF12 – a dataset of the web search query logs. Experimental results show that our CapsE achieves the new state-of-the-art results with significant improvements over strong baselines.
The proposed CapsE
Let INLINEFORM0 be a collection of valid ... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 12 | 6,528 | 7,076 |
38_13 |
We denote INLINEFORM0 , INLINEFORM1 and INLINEFORM2 as the INLINEFORM3 -dimensional embeddings of INLINEFORM4 , INLINEFORM5 and INLINEFORM6 , respectively. In our proposed CapsE, we follow BIBREF15 to view each embedding triple [ INLINEFORM7 , INLINEFORM8 , INLINEFORM9 ] as a matrix INLINEFORM10 , and denote INLINEFO... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 13 | 7,076 | 7,510 |
38_14 | This filter INLINEFORM15 is repeatedly operated over every row of INLINEFORM16 to generate a feature map INLINEFORM17 , in which INLINEFORM18 where INLINEFORM19 denotes a dot product, INLINEFORM20 is a bias term and INLINEFORM21 is a non-linear activation function such as ReLU. Our model uses multiple filters INLINEFO... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 14 | 7,510 | 7,860 |
38_15 | We denote INLINEFORM23 as the set of filters and INLINEFORM24 as the number of filters, thus we have INLINEFORM25 INLINEFORM26 -dimensional feature maps, for which each feature map can capture one single characteristic among entries at the same dimension.
We build our CapsE with two single capsule layers for a simpli... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 15 | 7,860 | 8,491 |
38_16 | These characteristics are generalized into one capsule in the second layer which produces a vector output whose length is used as the score for the triple.
The first capsule layer consists of INLINEFORM0 capsules, for which each capsule INLINEFORM1 has a vector output INLINEFORM2 . Vector outputs INLINEFORM3 are mult... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 16 | 8,491 | 8,970 |
38_17 | The capsule then performs the non-linear squashing function to produce a vector output INLINEFORM7 : DISPLAYFORM0
where INLINEFORM0 , and INLINEFORM1 are coupling coefficients determined by the routing process as presented in Algorithm SECREF2 . Because there is one capsule in the second layer, we make only one diff... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 17 | 8,970 | 9,471 |
38_18 |
[ht] 1.25
all capsule i INLINEFORM0 the first layer INLINEFORM1 0 INLINEFORM2 = 1, 2, ..., m INLINEFORM3 INLINEFORM4
INLINEFORM0
all capsule i INLINEFORM0 the first layer INLINEFORM1 The routing process is extended from BIBREF16 . | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 18 | 9,471 | 9,709 |
38_19 |
We illustrate our proposed model in Figure FIGREF1 where embedding size: INLINEFORM0 , the number of filters: INLINEFORM1 , the number of neurons within the capsules in the first layer is equal to INLINEFORM2 , and the number of neurons within the capsule in the second layer: INLINEFORM3 . The length of the vector ou... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 19 | 9,709 | 10,088 |
38_20 |
Formally, we define the score function INLINEFORM0 for the triple INLINEFORM1 as follows: DISPLAYFORM0
where the set of filters INLINEFORM0 is shared parameters in the convolution layer; INLINEFORM1 denotes a convolution operator; and INLINEFORM2 denotes a capsule network operator. We use the Adam optimizer BIBREF1... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 20 | 10,088 | 10,611 |
38_21 | INLINEFORM2 is generated by corrupting valid triples in INLINEFORM3 .
Knowledge graph completion evaluation
In the knowledge graph completion task BIBREF3 , the goal is to predict a missing entity given a relation and another entity, i.e, inferring a head entity INLINEFORM0 given INLINEFORM1 or inferring a tail ent... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 21 | 10,611 | 11,081 |
38_22 |
Experimental setup
Datasets: We use two recent benchmark datasets WN18RR BIBREF17 and FB15k-237 BIBREF18 . These two datasets are created to avoid reversible relation problems, thus the prediction task becomes more realistic and hence more challenging BIBREF18 . Table TABREF7 presents the statistics of WN18RR and FB... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 22 | 11,081 | 11,605 |
38_23 | We use the “Filtered” setting protocol BIBREF3 , i.e., not taking any corrupted triples that appear in the KG into accounts. We rank the valid test triple and corrupted triples in descending order of their scores. We employ evaluation metrics: mean rank (MR), mean reciprocal rank (MRR) and Hits@10 (i.e., the proportio... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 23 | 11,605 | 12,161 |
38_24 |
Training protocol: We use the common Bernoulli strategy BIBREF20 , BIBREF21 when sampling invalid triples. For WN18RR, BIBREF22 found a strong evidence to support the necessity of a WordNet-related semantic setup, in which they averaged pre-trained word embeddings for word surface forms within the WordNet to create s... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 24 | 12,161 | 12,744 |
38_25 |
We employ the TransE and ConvKB implementations provided by BIBREF24 and BIBREF15 . For ConvKB, we use a new process of training up to 100 epochs and monitor the Hits@10 score after every 10 training epochs to choose optimal hyper-parameters with the Adam initial learning rate in INLINEFORM0 and the number of filters... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 25 | 12,744 | 13,093 |
38_26 | We obtain the highest Hits@10 scores on the validation set when using N= 400 and the initial learning rate INLINEFORM3 on WN18RR; and N= 100 and the initial learning rate INLINEFORM4 on FB15k-237.
Like in ConvKB, we use the same pre-trained entity and relation embeddings produced by TransE to initialize entity and re... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 26 | 13,093 | 13,490 |
38_27 | We set the batch size to 128, the number of neurons within the capsule in the second capsule layer to 10 ( INLINEFORM1 ), and the number of iterations in the routing algorithm INLINEFORM2 in INLINEFORM3 . We run CapsE up to 50 epochs and monitor the Hits@10 score after each 10 training epochs to choose optimal hyper-p... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 27 | 13,490 | 13,820 |
38_28 | The highest Hits@10 scores for our CapsE on the validation set are obtained when using INLINEFORM4 , INLINEFORM5 and the initial learning rate at INLINEFORM6 on WN18RR; and INLINEFORM7 , INLINEFORM8 and the initial learning rate at INLINEFORM9 on FB15k-237.
Dataset: We use the SEARCH17 dataset BIBREF12 of query logs ... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 28 | 13,820 | 14,198 |
38_29 | A log entity consists of a user identifier, a query, top-10 ranked documents returned by the search engine and clicked documents along with the user's dwell time. BIBREF12 constructed short-term (session-based) user profiles and used the profiles to personalize the returned results. They then employed the SAT criteria... | https://arxiv.org/abs/1808.04122 | A Capsule Network-based Embedding Model for Knowledge Graph Completion and Search Personalization | 29 | 14,198 | 14,739 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.