Single-domain fine-grained sentiment analysis
Fine-grained sentiment analysis has attracted a great deal of attention recently due to its various applications and related challenging research topics. One of the main tasks is to extract both sentiment words and topic words from a sentence, where a sentiment word is an opinion expression and the...
محفوظ في:
المؤلف الرئيسي: | |
---|---|
مؤلفون آخرون: | |
التنسيق: | Final Year Project |
اللغة: | English |
منشور في: |
2019
|
الموضوعات: | |
الوصول للمادة أونلاين: | http://hdl.handle.net/10356/76908 |
الوسوم: |
إضافة وسم
لا توجد وسوم, كن أول من يضع وسما على هذه التسجيلة!
|
المؤسسة: | Nanyang Technological University |
اللغة: | English |
الملخص: | Fine-grained sentiment analysis has attracted a great deal of attention recently due to its various applications and related challenging research topics. One of the main tasks is to extract both sentiment words and topic words from a sentence, where a sentiment word is an opinion expression and the topic word is an aspect on which the opinion word is expressed. In this project, I use a joint model which integrates recursive neural networks and sequence labelling models for explicit aspect and opinion terms extraction. Based on the experiment result, I also design an interactive application which utilizes the joint model and is able to take user input and extract aspect and opinion expressions from the given text.
Word embedding is a popular natural language processing(NLP) method which aims at learning vector representations of words from documents. In this project, I use Yelp Challenge dataset for word embedding pre-training and SemEval Challenge 2014 dataset to evaluate my models. Previous studies have shown that a joint model of Recursive Neural Networks(RNN) and sequence labelling methods are promising for this task because it learns high-level discriminative features and dually propagates information between aspect and opinion terms (Wang, Pan, Dahlmeier, & Xiao, 2016). Hence, I conduct experiments using Recursive Neural Network integrated with different sequence labelling methods respectively: Conditional Random Fields(CRFs) and Bi-directional LSTM(Bi-LSTM). The experimental result verifies the robustness of the joint models. Based on the well-tuned sentiment analysis models, an interactive application is built using a Python web framework, Flask. Two types of UI are designed for the application: one takes text input, the other takes url input. Analysis results are shown as sentences where aspect and opinion terms are highlighted in different colors. |
---|