Editor's note: NTU researcher Soujanya Poria et al. proposed a combined model based on a pre-trained convolutional neural network to extract emotion, mood, and personality characteristics for satire detection. NTHU PhD student Elvis Saravia summarized the main ideas of the thesis concisely.
Overview
This paper uses a combined model based on Convolutional Neural Network (CNN) to process sarcasm detection for NLP tasks. Satire detection is very important in fields such as sentiment detection and sentiment analysis, because this expression will reverse the polarity of the sentence.
example
One can think that satire is used to sarcasm or taunt. For example, "Is it you or I should take medicine", "I work 40 hours a week to be so poor." (There are more examples on examples.yourdictionary.com.)
challenge
An important part of understanding and detecting irony is to understand the facts about the incident. This allows us to detect the contrast between objective polarity (usually negative) and the author's ironic features (usually positive).
Consider the following example, "I love the pain of breaking up". It is difficult to extract knowledge to detect whether there is irony in it. "I love its suffering" in the example provides knowledge of the emotion expressed by the author (positive in this example), while "break up" describes an opposite emotion (negative).
Other challenges in satirical sentences include referring to multiple events, as well as extracting large amounts of facts, common sense, reference analysis, and logical reasoning. The author of the paper relies on CNN to automatically learn features from the satire corpus.
contribution
Apply deep learning to sarcasm detection
Satire detection using user profile, emotion, and emotional characteristics
Automatically extract features using pre-trained models
model
Sentiment shifting is common in communication involving satire. Therefore, the author of the paper first trained an emotion model based on CNN to learn emotion-specific feature extraction. The model learns local features at the low level, and then converts to global features at the high level. The author found that satirical expressions are relevant to users—some users post more satirical content than others.
The framework proposed by the author integrates characteristics based on user personality, emotional characteristics, and emotional characteristics. Each group of features is learned through an independent model and becomes a pre-trained model for extracting ironic-related features from the data set.
CNN framework
CNN can effectively model local features to learn more global features. In essence, this is in the learning context. Sentences are represented by word vectors (embedding) (based on Google's word2vec vector). A non-static representation is used, so the parameters of the word vector are learned during the training phase. Next, apply max pooling on the feature map to generate features. Then comes the softmax layer and the fully connected layer to output the final prediction. (See below)
In order to obtain other features-sentiment (S), sentiment (E), personality (P)-CNN models are pre-trained, and these pre-trained models are used to extract features from the satire dataset. A different training data set is used to train each model. (Refer to the paper for more details)
Two classifiers were tested-a CNN classifier (CNN) and an SVM classifier (CNN-SVM, using features extracted by CNN as input).
In addition, a baseline classifier (B) was trained—only the CNN model was used, without combining other models (emotions, emotions, etc.).
test
The data is a balanced and unbalanced satirical tweet data set, taken from the work of Ptacek et al. in 2014 and The Sarcasm Detector. Removed username, URL, # mark, and used NLTK Twitter Tokenizer. (Refer to the paper for more details)
The following table shows the performance of CNN and CNN-SVM classifiers. We can observe that the models that combine ironic features, emotional features, emotional features, and personality features (especially CNN-SVM) outperform other models.
B = baseline, S = emotion, E = emotion, P = personality. All experiments used five-fold cross-validation
The following table compares with the current state-of-the-art model (first row) and another well-known satire detection model (second row). Similarly, the model proposed in the paper outperformed other models.
D3 => D1 means training on data set 3 and testing on data set 1
The paper tested the generalization ability of the model. The main finding is that if the data sets are very different in nature, it will significantly affect the results. (See the figure below based on the data set visualized by PCA). For example, training on data set 1 and then testing on data set 3, the F1 score of the model is 33.05%.
in conclusion
In general, the authors of the paper found that satire is highly subject-based and highly contextual. Therefore, sentiment and other contextual cues help to detect irony from the text. Use pre-trained emotion, mood, and personality models to capture contextual information from the text.
Manually constructed features (for example, n-grams), although helpful for irony detection to some extent, will produce very sparse feature vector representations. Therefore, word embedding is used as the input feature.
360 Laptop
360 laptop sometimes is also called as Yoga Laptop , cause usually has touch screen features. Therefore you can see other names at market, like 360 flip laptop, 360 Touch Screen Laptop,360 degree rotating laptop, etc. What `s the 360 laptop price? Comparing with intel yoga laptop. Usually price is similar, but could be much cheaper if clients can accept tablet 2 In 1 Laptop with keyboard. Except yoga type, the most competitive model for Hope project or business project is that 14 inch celeron n4020 4GB 64GB Student Laptop or 15.6 inch intel celeron business laptop or Gaming Laptop. There fore, just share the basic parameters, like size, processor, memory, storage, battery, application scenarios, SSD or SSD plus HDD, two enter buttons or one is also ok, if special requirements, oem service, etc. Then can provide the most suitable solution in 1 to 2 working days. Will try our best to support you.
To make client start business more easier and extend marker much quickly, issue that only 100pcs can mark client`s logo on laptop, Mini PC , All In One PC, etc. Also can deal by insurance order to first cooperation.
360 Laptop,360 Laptop Price,360 Flip Laptop,360 Touch Screen Laptop,360 Degree Rotating Laptop
Henan Shuyi Electronics Co., Ltd. , https://www.shuyitablet.com