farinamhz
commented
1 year ago Main problem
The paperâs main problem is classifying the sentiment of the customerâs reviews at the document level. Most of the recent works have been working on the short text. However, they think reviews are more documents and long rather than short texts. First, they aim to find latent target-opinion distribution behind the documents, and then they go through fine-grained prior knowledge from the users in the classification process. However, this is important that contrary to what we assume, by latent target-opinion, they mean something like the review below in which we have the aspect and just the opinion is not clear. For example, in the expression âSwear Iâm not gonna take the onion soup twiceâ, a customer implicitly views negative to an objective âonion soupâ.
Existing work
At first, most of the studies were based on traditional machine learning methods like a work which was making document feature representation and then applying an SVM classifier. However, later on, researchers went through more complex neural networks like Tang et al. in 2015, who used Convolutional Neural Network (CNN) as a document classifier with the word-level user as the input and preference matrix feature as the output. Most recently, approaches are paying attention to user information as background information and getting better results like what Chen et al. did in 2016, who used user data via attention mechanisms into the hierarchical sequential model for generating sentiment polarity. Even after that, researchers tried to use hierarchical networks and outperformed the previous studies. However, there is a significant point that they did not consider, which is the latent target and opinions behind the documents. In this regard, this paper proposed a method that will be explained in the next part. Most of the research works were mainly in short text, but there are a few who worked on long documents also, and this paper is going to focus on the long ones.
Inputs
- A collection of reviews that are long documents
Outputs
- Sentiments for the reviews
Example
Review = âAt the end of the day traveling we choose a near hotel, right next to the metro, a wonderful location. So going into the room before dinner starts, I gotta let you guys know that the service here is poor enough. Later we go to the canteen room. Well, the food is great, I love the baked French onion soup! Btw the duck is amazing and the ribs are flavorful. All is good except for the terrible environment, too many flies around, plus the noises outside. Maybe itâs just me, but still worth a try at this hotel.â
Overall rating: 3-star
Latent target-opinion: Food: 4-star Service: 1-star Location: 5-star Environment: 1-star
For conciseness, they only exemplify four representative targets of the hotel.
Proposed Method
In this study, they first capture the latent target-opinion distribution behind the documents and then incorporate the prior knowledge into the classification process. They have used a variational autoencoder where a review r and its sentences s_r were encoded to two sets of latent variables that follow Gaussian distribution: a) Coarse-grain review-level z_r, b) Fine-grain sentence-level z_sr, Followed by Transformer-based hierarchical classification for the task of reviewâs sentiment classification. They suppose that a review is formed by a bag of aspect and opinion words that are explicitly mentioned while filtering out other words using opinion lexicons.
Experimental Setup
Dataset The proposed model has been evaluated on three datasets:
- IMDB (Movie reviews)
- Yelp2013 and Yelp2014 (Hotel and Restaurant domain)
Each dataset has its own training, development, and test sets.
Evaluation and Metrics
They use the metrics below:
Standard Accuracy (Acc.):
Acc. = T/NRMSE:
RMSE = \sqrt{\sum_{k=1}^N(g_k - p_k)/N}T = Predicted sentiment labels which are the same as ground-truth. N = The total numbers of review document g_k = ground-truth p_k = predicted sentiment label. The smaller the RMSE score is, the stronger the method is. RMSE depicts the error divergences between the gold and the predicted labels for the multilabel classification tasks with more detail.
Baselines
- TF+SVM: Presenting advanced text features (word/character ngrams, sentiment lexicon features, cluster features, etc.) with an SVM classifier. The features are also concatenated with UPF for enhancement.
- SSWE+SVM: Sentiment-specific word embeddings with max/min/average pooling for yielding document representation with an SVM classifier.
- Para.Vector: The Distributed Memory Model of paragraph vector for the document sentiment classification.
- RNTN+RNN: Recursive Neural Tensor Network (RNTN) to obtain sentence representations fed into the Recurrent Neural Network (RNN) for sentiment classification.
- JMARS: Collaborative filtering algorithm, integrating with users and aspects information for the document-level sentiment classification.
- UPNN: A preference representation with a CNN network, which is concatenated with the user and product representations
- NSC: A hierarchical model based on LSTM for encoding review documents for sentiment classification.
- LUPDR: Compose temporal relations of reviews for better learning user and product embeddings with RNN for the document-level sentiment classification.
- CMA: Cascading multi-way attention model to form multiple user and product representations for the document-level sentiment classification.
- HUAPA: Consists of two individual hierarchical networks for better representing the document, achieving state-of-the-art performance for the document-level sentiment classification.
Results
Their experimental results on the three datasets showed that their model outperformed strong baselines and achieved state-of-the-art performance. They also proved that their model could recognize the latent fine-grained target and opinion information, which will bring more accurate document classification.
Code
There is no available code for this work.
Presentation
https://www.youtube.com/watch?v=hR0ZnW_qPM0
Criticism
The proposed unsupervised method extracts explicit aspects and neglects their latent occurrences, as I mentioned earlier by explaining the meaning of latent in this work. However, we aim to propose a method to detect latent aspects when they are hidden with no surface form.
hosseinfani
Latent Target-Opinion as Prior for Document-Level Sentiment Classification: A Variational Approach from Fine-Grained Perspective
I chose this paper because they stated in the subject that they have focused on latent aspect-opinion pairs. However, as I understood and will explain in summary, they have a different explanation for the word "latent" in their work which helped me to get the idea of bolding the meaning and examples of latent in our work while we write the documents.