Aspect-based Sentiment Analysis

Aspect-Based Sentiment Analysis (ABSA) is a task with many applications today in analyzing different aspects of comments and reviews about products, services, etc. An illustrative example is shown in Figure 1, with the input sentence: "The price was too high, but the cab was amazing.". In this sentence, there are two aspects being evaluated: "price" and "cab". The aspect "price" is evaluated as "too high", thus it is "negative", and the aspect "cab" is [evaluated/described; 'described' is slighty better because amazing describes, 'evaluated' implies a formal process] as "amazing", therefore it is "positive".

Based on defining the output values of the model, we can have several smaller subtasks as follows:

1. Aspect Term Extraction (ATE) or Aspect-Based Term Extraction: Extract the aspects that are evaluated in the comment.

2. Aspect Term Sentiment Classification (ATSC): Given the input sentence and the aspects evaluated in the comment, predict the sentiment of the comment towards those aspects. (Added "towards those aspects" for clarity).

3. Aspect Sentiment Pair Extraction (ASPE): Given the input sentence, extract the aspects and predict the sentiment associated with each aspect. (Added "associated with each aspect" for better clarity).

4. Aspect-Oriented Opinion Extraction (AOOE), also known as Targeted Opinion Word Extraction (TOWE), or Opinion Target Extraction (OTE): Given the input sentence and an aspect, predict the opinion text (span) expressing the sentiment towards that aspect. (This task has multiple common names, so listing them is important.)

5. Aspect-Opinion Pair Extraction (AOPE): Based on the input, extract aspect and opinion.

6. Aspect-Opinion-Sentiment Triplet Extraction (AOSPE or AOSTE): Given the input sentence, extract information about the aspect, the opinion text (span) expressing the sentiment towards that aspect, and the sentiment of the aspect in the comment.

   

In this section, we will develop a model to address the ASPE problem. This will be based on a two-step approach:

• Step 1: Predict the words in the text that represent aspects; this is essentially the ATE (Aspect Term Extraction) task.

• Step 2: Given the input text and the positions of the aspect terms, predict the sentiment for each aspect; this corresponds to the ATSC (Aspect Term Sentiment Classification) task.

  

The experiments are conducted on the SemEval-2014 Task 4: Aspect Based Sentiment Analysis dataset, which has undergone basic preprocessing, including: punctuation removal, normalization, and whitespace tokenization.

Setup

1. Clone the repository:

   git clone <repository_url>
   cd Aspect-based-sentiment-analysis

2. Get checkpoint (in ouput):

• Aspect Term Extraction (ATE):

  https://www.kaggle.com/code/tsunmm/ate-restaurants

• Aspect Term Sentiment Classification (ATSC)

  https://www.kaggle.com/code/tsunmm/atsc-restaurant

Project Structure

• checkpoints/: Contains configuration files.
• notebooks/: Holds raw and processed datasets.

Aspect-based-sentiment-analysis/
├── checkpoints/
│   ├── abte-restaurants-distilbert-base-uncased  # get in ATE checkpoints
|   ├── ate-restaurants-distilbert-base-uncased  # get in ATSC checkpoints
├── notebooks/
|   ├── ASPE_inferences.ipynb # Jupyter notebooks for experimentation 
├── tutorial.pdf  # Step by step description coding project
├── .gitignore  # Git ignore file for unnecessary files
└── README.md  # Project description and instructions

Deployment

Code deployment