Decoding Weibo Sentiments: Unveiling Nuanced Emotions with

Bidirectional LSTM Analysis

Kaiwen Deng

Business School, Beijing Institute of Technology, Zhuhai, China

Keywords: Sentiment Analysis, Weibo Comments, Long Short-Term Memory, Pre-Trained Word Embeddings.

Abstract: This study highlights the critical role that deep learning plays in deciphering the complex emotional subtleties

prevalent in social media interactions via sentiment analysis of Weibo comments. The primary goal is to

deploy a bidirectional Long Short-Term Memory (LSTM) model that is intended to thoroughly analyze and

understand user attitudes, providing priceless information for improving marketing tactics and gauging public

opinion. Utilizing character-level segmentation and Tencent Artificial Intelligence (AI) Lab's pre-trained

word embeddings, the study advances sentiment analysis by enhancing sensitivity to contextual subtleties

within textual data. The study, which used a dataset of Weibo comments with sentiment annotations,

demonstrates the remarkable 96% accuracy with which the bidirectional LSTM model can categorize

sentiments. This result demonstrates how well the model captures complex emotional expressions,

outperforming both other deep learning methods and conventional machine learning techniques in sentiment

analysis tasks. The novel elements of the model's architecture, such character-level analysis and the intelligent

use of pre-trained embeddings, improve its classification accuracy and contextual comprehension. These

aspects represent significant advancements in sentiment analysis, with broad implications for both academic

research and practical applications in understanding social media discourse.

1 INTRODUCTION

The Weibo comment is the text content that users

reply and discuss the content on the Weibo platform,

covering rich emotions, attitudes, and views. The

significance of analyzing Weibo comments is that

being able to understand the emotional relationships,

attitudes, and views of users in social media. This is

of great significance for enterprises to formulate

accurate marketing strategies for public opinion

(Chen, 2022; Yuan, 2019), government

understanding of the people's conditions, and

academic research on social public opinion.

Therefore, this paper aims to use deep learning

technology, especially the text classification model

based on technologies based on long-term memory

(LSTM) to help reveal the motivation and factors

behind user behavior, and provide strong support for

decision-making in related fields (Li, 2019; Halawani,

2023). Scholars in the discipline of analysis of

sentiment have put forth a number of approaches to

handle text data and propel technological progress.

https://orcid.org/0009-0008-3657-6112

Deep learning-based techniques have advanced

substantially in the past few years in sentiment

analysis tasks. Among them, for text classification

tasks, recurrent neural network, or Recurrent Neural

Network (RNN), models like LSTM, or long-short-

term memory, are used extensively, particularly in

sentiment analysis (Khan, 2022; Wu, 2023). Through

these models, researchers are able to effectively

capture the semantic and contextual information in

written content, strengthening sentiment analysis's

precision and effectiveness (Yang, 2022).

Furthermore, sentiment analysis tasks have also been

handled by convolutional neural networks (CNNs).

Convolutional and pooling processes are used by

CNN models to extract features from text, which are

subsequently used for sentiment categorization.

Compared to traditional methods based on bag-of-

words models, CNNs can more effectively convey

local information through text, improving the

functionality of sentiment analysis (Kaur, 2023).

Moreover, sentiment analysis heavily relies on

traditional machine learning methods like Naive

Deng, K.

Decoding Weibo Sentiments: Unveiling Nuanced Emotions with Bidirectional LSTM Analysis.

DOI: 10.5220/0012923300004508

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Engineering Management, Information Technology and Intelligence (EMITI 2024), pages 221-226

ISBN: 978-989-758-713-9

221

Bayes and Support Vector Machines (SVM). These

methods typically rely on manually designed features

and statistical models for text classification. Although

these methods can achieve the goal of sentiment

analysis to some extent, they often perform poorly in

handling complex text data and struggle to capture

deep semantic information in text. In conclusion,

Deep learning-based techniques are gradually gaining

popularity in the past decade and become mainstream

in the field of sentiment analysis, especially LSTM

and CNN models have achieved significant results in

sentiment classification tasks. These methods can

better understand the semantic and contextual

information of text data, therefore raising sentiment

analysis's precision and effectiveness (Alsini, 2023;

Contreras, 2023).

The project's objective is to use LSTMs to

develop a sentiment analysis model for Weibo

comments. By breaking down remarks into individual

characters, it preprocesses the material and improves

comprehension of linguistic subtleties. Leveraging

pre-trained word embeddings from Tencent Artificial

Intelligence (AI) Lab enriches the model's

comprehension of context and sentiment. A

bidirectional LSTM architecture is employed to

capture both past and future context, improving

sentiment classification accuracy. The predictive

performance of the model is contrasted with different

deep learning architectures and conventional machine

learning techniques. The experiment demonstrates

the LSTM model's effectiveness in analyzing Weibo

comments' sentiment, underscoring its robust support

for social sentiment analysis and decision-making.

2 METHODOLOGIES

2.1 Dataset Description and

Preprocessing

The study utilizes a dataset comprising Weibo

comments, aiming to analyze sentiment through an

LSTM-based model. The dataset consists of text

comments which have been annotated for sentiment,

with the intention of facilitating a comprehensive

understanding of public sentiment on various topics

discussed on Weibo (ChineseNlpCorpus, 2018). The

dataset is a collection of Weibo comments, each

associated with a sentiment label indicating the

comment's overall sentiment (positive or negative).

These comments have been meticulously collected to

ensure a diverse representation of topics, linguistic

styles, and sentiments, providing a robust foundation

for analyzing sentiment nuances in social media text.

The preprocessing of the dataset is a critical step

to prepare the raw text data for the LSTM model. The

paper preprocessing pipeline involves several key

stages: Character Segmentation: Given the nature of

the Chinese language, which does not use spaces to

separate words, this paper adopts a character-level

segmentation approach. This method involves

breaking down each comment into individual

characters, thereby capturing the linguistic features

more effectively. Vocabulary Construction: A

vocabulary index is created from the segmented

dataset, with a maximum size set to 10,000 unique

tokens. Special tokens such as <UNK> for unknown

characters and <PAD> for padding are included to

handle out-of-vocabulary words and maintain

uniform comment lengths, respectively. Sequence

Padding and Truncation: To ensure uniform input

sizes for the LSTM model, comments are either

padded or truncated to a fixed length of 50 characters,

as defined by the pad_size parameter. This step

ensures that each input tensor to the model maintains

a consistent shape. Tokenization and Indexing: Each

character in a comment is replaced with its

corresponding index from the vocabulary, converting

the textual data into a numerical format that can be

processed by the model. Characters not found in the

vocabulary are replaced with the index for <UNK>.

Training, validation, and test sets are separated from

the pre-processed dataset with a distribution of sixty

percent, twenty percent, and twenty percent,

accordingly. By splitting the data this way, the model

can be trained on an important part of the data, refined

and confirmed on another portion, and then tested for

generalization on unseen data.

2.2 Proposed Approach

This study explores the sentiment orientation in

Weibo comments by leveraging a model based on

LSTM networks. LSTM, recognized for its capacity

to effectively process sequential data and model long-

term dependencies, emerges as an optimal tool for

addressing Natural Language Processing (NLP)

tasks, especially emotion analysis. The methodology

unfolds in several pivotal phases: data preprocessing,

involving character-level segmentation and sequence

normalization; model construction, which

incorporates pre-trained word embeddings to bolster

the understanding of textual sentiment; and a series of

training, fine-tuning, and evaluation steps, employing

diverse performance metrics such as F1 scores and

accuracy to ascertain the efficacy of the model.

Aimed at conducting a thorough analysis of sentiment

orientations in Weibo comments, the research

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence

222

Figure 1: The pipeline of this study

(Photo/Picture credit: Original).

employs deep learning technologies to delve into

public emotions. The approach not only seeks to

improve the precision of sentiment analysis but also

to introduce new angles and methods for examining

sentiment on social media platforms. The process is

shown in the Figure 1.

2.2.1 LSTM

The core of the proposed approach is an LSTM-based

model for sentiment analysis. LSTM networks, a

special kind of RNN, are made to teach sequence

prediction problems about order dependence. Unlike

traditional feedforward neural networks, LSTM can

handle complete data sequences in addition to

individual data points since it contains feedback

connections. This feature is notably helpful for

applications involving natural language processing,

because sentiment interpretation depends heavily on

word order and context. The sentiment analysis

model based on the LSTM forms the basis of the

suggested methodology. An advanced kind of RNN

called LSTM networks is developed specifically to

identify order dependency in sequence prediction

issues. Because LSTM incorporates feedback

connections, it can handle complete data sequences in

addition to individual data points, unlike traditional

feedforward neural networks. This characteristic is

especially beneficial for natural language processing

tasks where context and order of words play a crucial

role in understanding sentiment. The LSTM model's

defining feature is its ability to remember and utilize

past information through its internal state, commonly

referred to as the cell state, to make informed

predictions. This is particularly important in

sentiment analysis where the sentiment conveyed by

a sentence can be heavily dependent on the context

provided by preceding words or phrases. The use of

LSTM in sentiment analysis of Weibo comments

allows for a more nuanced understanding of user

sentiment, which can be leveraged for market

analysis, public opinion monitoring, or even for

sociological research. In this experiment, the

implementation process begins with data

preprocessing, involving character-level

segmentation, vocabulary construction, sequence

padding and truncation, and tokenization and

indexing. The LSTM model is then initialized with

pre-trained embeddings and configured with

hyperparameters such as hidden layer size quantity of

layers and rate of learning. Cross-entropy loss and the

Adam optimizer are used to train the pre-processed

data, and performance is measured using metrics like

accuracy, precision, recall, and F1 scores. There are

three main parts to the suggested model architecture:

the Embedding Layer, which uses word embeddings

that have already been trained to convert vocabulary

indices into vector representations; the LSTM Layer,

which uses bidirectional LSTM units to gather

information from both previous and subsequent

contexts, making it easier to learn long-term

dependencies from sequential data and the Fully

Connected Layer, which translates abstract features

learned by the LSTM into predictive outputs for

sentiment classification, typically through the

modelling of sentiment scores and conversion into a

probability distribution using the SoftMax function.

2.2.2 Model Configuration

A model architecture is crafted with a careful balance

between complexity and computational efficiency,

selecting 128 units for the size of the hidden layers. A

dual-layer LSTM structure is used to improve

sentiment context recognition and to further

comprehend text sequences. Furthermore, as part of a

regularization strategy to mitigate model overfitting,

a dropout rate of 0.5 is set, randomly ignoring a

fraction of the network's nodes during training.

Additionally, the batch size is determined to be 128,

based on a trade-off consideration between memory

demands and gradient estimation stability during the

model training process.

2.2.3 Loss Function

The Adam optimizer is the optimization function used

in this model. Adam, which stands for Adaptive

Moment Estimation, is based on the idea of

combining the advantages of Momentum and

RMSProp, two additional optimization techniques.

Adam records each weight in the neural network's

first moment vector (m) and second moment vector

(v). The first moment (the mean) and the second

moment (the uncentered variance) of the gradients are

estimated by the parameters m and v, respectively.

Decoding Weibo Sentiments: Unveiling Nuanced Emotions with Bidirectional LSTM Analysis

223

The squared gradient and the gradient's exponential

moving average are calculated mathematically. The

parameters updating guidelines with Adam are as

follows:

11 1

(1 )

tt t

mm g

ββ

−

=⋅ +− ⋅

(1)

21 2

(1 )

tt t

vv g

ββ

−

=+−⋅

(2)

(1 )

=⋅−

(3)

(1 )

=⋅−

(4)

()

tt t t

θθη ε

=−− +⋅

(5)

The gradient and the decay rates for the moment

estimates, which are typically set to 0.9 and 0.999,

respectively, indicate the parameters of the model at

time step t. The rate of learning is denoted by η, and

the tiny scalar, ε, is used to avoid dividing by zero,

usually around 10

ି଼

The advantage of Adam is its adaptive learning

rate capability, which allows for individual learning

rates for each parameter. This adaptive mechanism

often leads to faster convergence and has proven to be

effective in various contexts, particularly in complex

tasks such as sentiment analysis with large datasets

and models with a significant number of parameters.

2.3 Implementation Details

The LSTM-based sentiment analysis model was

implemented in the research using Python 3.9 and the

integrated programming environment PyCharm

2023.3.3. For development, a Microsoft Windows 11

Home Edition system with an RTX 3070 graphics

card and a 12th generation Intel(R) Core (TM) i7-

12700H processor operating at the frequencies of

2.30 megahertz frequencies was employed. Data

augmentation techniques were applied to enrich the

dataset and mitigate overfitting issues, ensuring a

robust model training process.

3 RESULTS AND DISCUSSION

In the conducted study, the analysis succinctly

evaluates the model's performance through three

crucial visual representations: an ascending training

accuracy curve indicating swift initial learning, a

descending loss graph signifying effective

optimization, and a confusion matrix that highlights

competent classification with an imbalance in false

negatives, suggesting potential areas for refinement.

Figure 2: The training accuracy of a model over epochs

(Photo/Picture credit: Original).

As depicted in Figure 2, the model's training

accuracy incrementally increases with the

progression of epochs. The accuracy exhibits a

precipitous climb from approximately 75% to over

85% within the initial epochs, signalling the model's

rapid learning from the training data. Subsequently,

the increment in accuracy decelerates, yet it continues

to demonstrate a slow and steady enhancement until

it plateaus around 96% at the 30-epoch mark. This

pattern indicates that after a period of swift learning,

the model begins to converge, and the stability of

accuracy suggests that it has reached its performance

potential given the current architecture and dataset.

Although the high accuracy denotes strong model

performance on the training set, the absence of

validation or test accuracy precludes a full assessment

of overfitting. Without appropriate regularization,

there is a potential for the model to overfit the training

data, diminishing its generalizability to new data.

Figure 3: The training loss of a model over epochs

(Photo/Picture credit: Original).

Figure 3 portrays the fluctuation in the model's

loss during the training process. The loss markedly

decreases in the initial epochs, which indicates that

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence

224

the model rapidly reduces error, effectively

optimizing towards a superior direction. In

subsequent epochs, the decline in loss decelerates,

suggesting that the model is entering a phase of fine-

tuning and plateaus after 30 epochs. This descending

and stabilization of loss reflect the model's approach

to optimal performance; concurrently, vigilance is

warranted against the risk of overfitting associated

with excessively low loss. Ideally, the decrease in loss

should coincide with the model's enhanced

understanding of data representation, yet an overly

complex model might learn noise rather than

underlying useful patterns.

Figure 4: The predictions of the model's confusion matrix

(Photo/Picture credit: Original).

The confusion matrix, shown in Figure 4, offers

an intuitive perspective of the model's predictive

performance. The matrix reveals that 11,193 negative

samples and 10,418 active samples were correctly

classified, signifying a strong predictive accuracy for

both classes. However, there were also 803 false

positives where negative samples were incorrectly

labeled as active, and 1,556 false negatives where

active samples were mistakenly labeled as negative.

The relative abundance of false negatives suggests a

propensity of the model to misclassify active samples

as negative, which may be attributable to dataset

imbalance or improper threshold settings in the

classifier. This imbalance in error distribution could

impact the model's utility in practical applications,

especially if accurate identification of one class is

paramount. Further refinement of the model may

necessitate adjustments in data preprocessing

strategies or further tuning of the model parameters.

In conclusion, the comprehensive experiments

conducted in this chapter have significantly

illuminated both the capabilities and areas of

improvement for the machine learning model under

scrutiny. Through detailed analyses encompassing

training accuracy, loss patterns, and confusion matrix

insights, the experiments have elucidated a trajectory

of rapid learning and convergence, while also

cautioning against the potential for overfitting due to

high training accuracy without corresponding

validation. Furthermore, the analysis of predictive

accuracy through the confusion matrix has

highlighted challenges related to class imbalance and

classification thresholds, underscoring the necessity

for ongoing model refinement. Collectively, these

findings not only validate the significance of the

experimental efforts undertaken but also pave the way

for future enhancements to optimize model

performance.

4 CONCLUSIONS

This study presents a groundbreaking approach to

sentiment analysis within the realm of Weibo

comments, harnessing the power of a bidirectional

LSTM model. By intricately combining character-

level segmentation and pre-trained word embeddings,

this innovative methodology delves deep into the

intricate emotional fabric woven within social media

discourse. Through meticulous experimentation, the

model exhibits remarkable proficiency, achieving a

commendable accuracy plateau of 96% in sentiment

classification, thus solidifying its status as a

formidable tool in the realm of sentiment analysis.

Looking forward, the scope of exploration

extends to the dynamic nature of sentiment within

Weibo comments, with a particular emphasis on

understanding how sentiments evolve over time in

response to various social, political, and cultural

stimuli. It is believed that a deeper analysis of

sentiment fluctuations holds the key to unveiling

invaluable insights into collective social behavior,

thereby informing decision-making processes across

diverse domains. This study demonstrates the

effectiveness of the bidirectional LSTM model and

represents a major advancement in the area of

sentiment evaluation in deciphering the nuanced

emotional undertones embedded within Weibo

comments. As the journey of exploration continues,

the commitment remains unwavering in further

refining the model's sensitivity and applicability,

thereby contributing meaningfully to the ongoing

discourse surrounding sentiment analysis in the

digital age.

Decoding Weibo Sentiments: Unveiling Nuanced Emotions with Bidirectional LSTM Analysis

225

REFERENCES

Alsini, R. 2023. Analysis of Real Time Twitter Sentiments

using Deep Learning Models. Journal of Applied Data

Sciences, vol. 4(4).

Chen, X. 2022. Research on Emotional Tendency Analysis

of Weibo Comments Based on Deep Neural Network.

Health Science and Technology.

ChineseNlpCorpus. 2018. Weibo sentiments.

https://github.com/SophonPlus/ChineseNlpCorpus/blo

b/master/datasets/weibo_senti_100k/intro.ipynb

Contreras, Hernández, S., Tzili Cruz, M. P., Espínola

Sánchez, J. M., & Pérez Tzili, A. 2023. Deep Learning

Model for COVID-19 Sentiment Analysis on Twitter.

New Generation Computing.

Halawani, H. T., Mashraqi, A., Badr, S. K., & Alkhalaf, S.

2023. Automated sentiment analysis in social media

using Harris Hawks optimisation and deep learning

techniques. Alexandria Engineering Journal.

Kaur, G., & Sharma, A. 2023. A deep learning-based model

using hybrid feature extraction approach for consumer

sentiment analysis. Journal of Big Data, 10.

Khan, L., Amjad, A., Afaq, K. M., & Chang, H.-T. 2022.

Deep Sentiment Analysis Using CNN-LSTM

Architecture of English and Roman Urdu Text Shared

in Social Media. Applied Sciences, vol. 12(5), p: 2694.

Li, J., & Dong, D. 2019. Analysis of Weibo Comments

Based on SVM and LDA Models. 2019 Chinese

Control And Decision Conference (CCDC).

Wu, L. 2023. Application of deep learning in social network

public opinion sentiment. International Journal of

Artificial Intelligence.

Yang, H. 2022. Network Public Opinion Risk Prediction

and Judgment Based on Deep Learning: A Model of

Text Sentiment Analysis. Computational Intelligence

and Neuroscience.

Yuan, Z., Yuwei, G., Feng, Z., & Reimei, W. 2019. A

Literature Review of Sentiment Analysis on Chinese

Social Media. Journal of Physics: Conference Series,

vol. 1314(1), p: 012140.

EMITI 2024 - International Conference on Engineering Management, Information Technology and Intelligence

226