Sentiment analysis of the Hamas-Israel war on YouTube comments using deep learning Scientific Reports

The key aspect of sentiment analysis is to analyze a body of text for understanding the opinion expressed by it. Typically, we quantify this sentiment with a positive or negative value, called polarity. The overall sentiment is often inferred as positive, neutral or negative from the sign of the polarity score. So far, I have shown how a simple unsupervised model can perform very well on a sentiment analysis task. As I promised in the introduction, now I will show how this model will provide additional valuable information that supervised models are not providing. Namely, I will show that this model can give us an understanding of the sentiment complexity of the text.

The results indicate that there is no statistically significant correlation between sentiment scores and market returns next day. However, there is weak positive correlation between negative sentiment at day t and the volatility of the next day. R-value of 0.24 and p-value below 0.05 indicate that the two variables (negative sentiment and volatility) move in tandem. For instance, if the negative sentiment at a given day t increases, the volatility of the market would also increase the next day.

One of the main challenges in traditional manual text analysis is the inconsistency in interpretations resulting from the abundance of information and individual emotional and cognitive biases. Human misinterpretation and subjective interpretation often lead to errors in data analysis (Keikhosrokiani and Asl, 2022; Keikhosrokiani and Pourya Asl, 2023; Ying et al., 2022). To address this issue, hybrid methods that combine manual annotation with computational strategies have been proposed to ensure accurate interpretations are made. However, it is important to acknowledge that computational methods have limitations due to the inherent variability of sociality.

These advancements have provided richer, more nuanced semantic insights that significantly enhance sentiment analysis. However, despite these advancements, challenges arise when dealing with the complex syntactic relationships inherent in language-connections between aspect terms, opinion expressions, and sentiment polarities42,43,44. To bridge this gap, Tree hierarchy models like Tree LSTM and Graph Convolutional Networks (GCN) have emerged, integrating syntactic tree structures into their learning frameworks45,46. This incorporation has led to a more granular analysis that combines semantic depth with syntactic precision, allowing for a more accurate sentiment interpretation in complex sentence constructions.

As presented in Table 5, after regularization, the accuracy of the model was improved, and the result shows that there is minimal difference observed among training, validation, and test accuracy. This further shows that the problem of over-fitting is solved as compared to the previous result achieved before regularization. Convolutional layers extract features from different parts of the text and the pooling layer reduces the number of features in the input. Then features obtained from the pooling layer are passed to the Bidirectional-LSTM to extract contextual information. Finally, the last states of the BiLSTM are concatenated and passed into the Sigmoid activation function, which squashes the final value in the range between 0 and 1. 2 that Bi-LSTM can learn in both directions and integrate the pieces of knowledge to make a prediction.

Rule-based systems are simple and easy to program but require fine-tuning and maintenance. For example, “I’m SO happy I had to wait an hour to be seated” may be classified as positive, when it’s negative due to the sarcastic context. Sentiment analysis allows businesses to get into the minds of their customers. The startup’s virtual assistant engages with customers over multiple channels and devices as well as handles various languages.

It has been used in various NLP applications and is known for its ability to capture semantic relationships. In information retrieval systems, word embeddings can enable more accurate matching of user queries with relevant documents, which improves the effectiveness of search engines and recommendation systems. You can track sentiment over time, prevent crises from escalating by prioritizing mentions with negative sentiment, compare sentiment with competitors and analyze reactions to campaigns. One of the tool’s features is tagging the sentiment in posts as ‘negative, ‘question’ or ‘order’ so brands can sort through conversations, and plan and prioritize their responses. Buffer offers easy-to-use social media management tools that help with publishing, analyzing performance and engagement.

Why sentiment analysis is necessary

Interestingly, the BERT-chunk model performed approximately the same as the BERT-truncated one. This is in line with the idea that most of the relevant information of a news article is contained at its beginning or that online readers focus mainly on the headline and the lead67. Sentiment analysis tools show the organization what it needs to watch for in customer text, including interactions or social media. Patterns of speech emerge in individual customers over time, and surface within like-minded groups — such as online consumer forums where people gather to discuss products or services. Sentiment analysis has been widely used by several types of industries for the last decades. Not only it can produce helpful insights, but also save time and energy by leveraging the power of machine learning rather than manually gathering and analyzing the information from a bunch of data.

To tackle these issues, natural language models are utilizing advanced machine learning (ML) to better understand unstructured voice and text data. This article provides an overview of the top global natural language processing trends in 2023. They range from virtual agents and sentiment analysis to semantic search and reinforcement learning. It can be observed that our proposed approach leverages binary label relations, which is a general mechanism for knowledge conveyance, to enable gradual learning.

Figure 11 is very revealing, in that it confirms the results of the sample analysed with Lingmotif 2. Positive emotions substantially decrease between pre-covid- and covid expansión (69.98–61.34%), while negative ones increase (30.02–38.66%). Although declining positive and increasing negative trends are also identified in Economist, the differences are not as strong (57.42–55.59% for positive, 42.58–44.41% for negative). This suggests that the approach of the English periodical to news reporting is more stable than its Spanish counterpart. As the Figures 5–7 show, pre-covid expansión has 64% positive sentences (257 positive sentences), against 36% (or 145) negative ones (rating ‘fairly positive’ overall), TSI being ‘very intense’ (TSI average of 74).

To reduce the model’s vulnerability to over-fitting, the researcher added one Dense layer (Hidden layers) with 64 neurons and the activation function ReLU. Then added a dropout layer to the Convolutional layer before feeding it into the pooling layer, then added a dense layer. After the dense layer, the researcher also added another dropout layer, which was then fed into the fully connected layer. Dropout was discovered to be incredibly essential since it allows the model to avoid over-fitting by dropping neurons at a random point. The batch size was increased from 64 to 100, and the epoch number was decreased from 10 to 9. Change is made based on manual tunning and the experimental result is presented in Table 5.

Each day, we are challenged with texts containing a wide range of insults and harsh language. Automatic intelligent software that detects flames or other offensive words would be beneficial and could save users time and effort. semantic analysis of text These works defy language conventions by being written in a spoken style, which makes them casual. Because of the expanding volume of data and regular users, the NLP has recently focused on understanding social media content2.

The integrated model achieved an enhanced accuracy on the three datasets used for performance evaluation. Moreover, a hybrid dataset corpus was used to study Arabic SA using a hybrid architecture of one CNN layer, two LSTM layers ChatGPT and an SVM classifier45. Stacked LSTM layers produced feature representations more appropriate for class discrimination. The results highlighted that the model realized the highest performance on the largest considered dataset.

Study 1

Talkwalker is a leading social listening platform that provides businesses with actionable social media insights via real-time listening and advanced analytics. This platform goes beyond monitoring social media mentions to offer a robust set of tools for understanding brand sentiment, identifying trends, and engaging with target audiences. Its AI-powered sentiment analysis tool helps users find negative comments or detect basic forms of sarcasm, so they can react to relevant posts immediately. IBM Watson NLU recently announced the general availability of a new single-label text classification capability. This new feature extends language support and enhances training data customization, suited for building a custom sentiment classifier.

It analyzes text to reveal the type of sentiment, emotion, data category, and the relation between words based on the semantic role of the keywords used in the text. According to IBM, semantic analysis has saved 50% of the company’s time on the information ChatGPT App gathering process. Today, with the rise of deep learning, embedding layers have become a standard component of neural network architectures for NLP tasks. Embeddings are now used not only for words but also for entities, phrases and other linguistic units.

This model passes benchmarks by a large margin and earns 76% of global F1 score on coarse-grained classification, 51% for fine-grained classification, and 73% for implicit and explicit classification. An embedding is a learned text representation in which words with related meanings are represented similarly. The most significant benefit of embedding is that they improve generalization performance particularly if you don’t have a lot of training data. It is a Stanford-developed unsupervised learning system for producing word embedding from a corpus’s global phrase co-occurrence matrix. The essential objective behind the GloVe embedding is to use statistics to derive the link or semantic relationship between the words.

Typically, sentiment analysis for text data can be computed on several levels, including on an individual sentence level, paragraph level, or the entire document as a whole. Often, sentiment is computed on the document as a whole or some aggregations are done after computing the sentiment for individual sentences. We usually start with a corpus of text documents and follow standard processes of text wrangling and pre-processing, parsing and basic exploratory data analysis. Based on the initial insights, we usually represent the text using relevant feature engineering techniques. Depending on the problem at hand, we either focus on building predictive supervised models or unsupervised models, which usually focus more on pattern mining and grouping.

In addition to natural language processing, DL were employed in computer vision, handwriting recognition, speech recognition, object detection, cancer detection, biological image classification, face recognition, stock market analysis, and many others13. Finally, this section contains the baseline results generated using many deep learning algorithms such as CNN-1D, LSTM,GRU, Bi-GRU, Bi-LSTM and our proposed model based on mBERT model. According to the results presented in Table 9, deep learning models outperforms machine learning and rule-based approach.

Sentiment Analysis with AFINN Lexicon

By calculating the cosine similarity between the construct of interest and the input text, CCR provides a measure of association (i.e., CCR loading). Recent research indicates that CCR outperforms other theory-driven text analysis methods, such as word-counting and word embeddings55, making it well-suited for our theoretically driven investigation. This research addresses gaps from previous works through a comprehensive experimental study. The researcher studied the impacts of datasets preparation, word embedding, and deep learning models, with a focus on the problem of sentiment analysis. Four deep learning models CNN, Bi-LSTM, GRU, and CNN-Bi-LSTM for Amharic sentiment analysis were compared, the experiment result showed that combining CNN with Bi-LSTM generated a model that outperformed the others.

The startup applies AI techniques based on proprietary algorithms and reinforcement learning to receive feedback from the front web and optimize NLP techniques. AyGLOO’s solution finds applications in customer lifetime value (CLV) optimization, digital marketing, and customer segmentation, among others. NLP Cloud is a French startup that creates advanced multilingual AI models for text understanding and generation. They feature custom models, customization with GPT-J, follow HIPPA, GDPR, and CCPA compliance, and support many languages. Besides, these language models are able to perform summarization, entity extraction, paraphrasing, and classification.

During Period 3 (2001–2010), China joined the World Trade Organization (WTO) in 2001 and won the bid for hosting the 2008 Olympic Games. With its increasing integration into the world and rapid economic progress, China overtook Japan as the world’s second-largest economy in 2010. Moreover, as a result of the September 11, 2001 terrorist attacks, the US government viewed international terrorism as its primary threat, thereby providing China with a strategic opportunity to develop its own military power (He, 2016). Over Period 4 (2011–2020), China began to get involved in remarkedly more clashes with other countries, particularly with the US, as its economic and political influence continued to grow. Here, please note that sentiment analysis is distinct from appraisal analysis (Martin and White, 2005) and prosody analysis (Sinclair, 1991, 2004). Whereas semantic prosody focuses on a pragmatic unit of meaning and consists of extensive searches and analyses of the unit in context, appraisal analysis emphasizes a pragmatic unit of meaning and involves extensive searches and analyses of the unit.

• The TSS calculates the polarity of each sentence, taking into account both the number and the position of sentiment-related items.
• Random Forest is an ensemble learning that parallel builds multiple random decision trees, and the prediction is based on the most voted by the trees.
• It is also called “convergence” by Laviosa (2002) to suggest “the relatively higher level of homogeneity of translated texts”.
• By understanding how your audience feels and reacts to your brand, you can improve customer engagement and direct interaction.
• In addition, The ability of Bi-LSTM to encapsulate bi-directional context was investigated in Arabic SA in49.

Another approach involves leveraging machine learning techniques to train sentiment analysis models on substantial quantities of data from the target language. This method capitalizes on large-scale data availability to create robust and effective sentiment analysis models. By training models directly on target language data, the need for translation is obviated, enabling more efficient sentiment analysis, especially in scenarios where translation feasibility or practicality is a concern. Traditional machine learning methods such as support vector machine (SVM), Adaptive Boosting (AdaBoost), Decision Trees, etc. have been used for NLP downstream tasks.

Environmental and sustainability issues were present in both periods, although not as dominant as other topics. Based on the frequent words from the Expansión newspaper corpus during the years 2018 and 2019, it seems that the articles cover a wide range of topics. For H2, we used a frequency list with a relative degree of co-occurrence frequency (DOCF) from Sketch Engine, as it allowed us to compare the relative frequency of different topics in each newspaper corpus and identify differences between the two periods. We then compared the relative frequency of topics related to critical financial matters and the global health crisis in each newspaper corpus in the first and second periods, respectively.

The data that support the findings of this study are available from the corresponding author upon reasonable request. A comprehensive search was conducted in multiple scientific databases for articles written in English and published between January 2012 and December 2021. The databases include PubMed, Scopus, Web of Science, DBLP computer science bibliography, IEEE Xplore, and ACM Digital Library. Guofeng Wang and Yilin Liu contributed to research design, methodology, data collection, analysis, and writing and editing.

It has a visual interface that helps users annotate, train, and deploy language models with minimal machine learning expertise. Its dashboard consists of a search bar, which allows users to browse resources, services, and documents. You can foun additiona information about ai customer service and artificial intelligence and NLP. Additionally, a sidebar lets you create new language resources and navigate through its home page, services, SQL database, and more. Secondly, since the analysis of textual entailment involves a comparison between English and Chinese texts, multilingual semantic resources are needed. In the current study, the reference knowledge base for the textual entailment analysis in this study is WordNet (Miller, 1995) and its multilingual counterpart Open Multilingual WordNet (OMW).

In practice, SLSA is highly valuable in the scenarios where comments are represented by concise and isolated sentences with arbitrary topics, requiring a holistic analysis of sentiment at the sentence level. In another application, social media platforms (e.g., Twitter and Facebook) usually analyze people’s comments and posts by SLSA to gain insights into public opinion and social trends. Deep learning applies a variety of architectures capable of learning features that are internally detected during the training process.

The update and reset gates are two crucial gates of GRU that decide what information should be passed to the output27. And T.B.L.; methodology, M.S; S.R.; K.S.; sofware, M.S.; validation, V.E.S.; S.N. And T.B.L.; formal analysis, V.E.S. and M.S.; investigation, S.N.; writing—original draf preparation, V.E.S.; S.R. This step gradually labels the instances with increasing hardness in a workload.

• Incorporating the best NLP software into your workflows will help you maximize several NLP capabilities, including automation, data extraction, and sentiment analysis.
• Also, we examine and compare five frequently used topic modeling methods, as applied to short textual social data, to show their benefits practically in detecting important topics.
• Moreover, the Oslo Accords in 1993–95 aimed for a settlement between Israel and Hamas.
• In my testing, longer prompts can result in ChatGPT losing the request and, instead, offering a summary or analysis.

In this paper, we focus on how to supervise feature extraction by DNNs and leverage them for improved gradual learning on the task of SLSA. Most recently, the research on SLSA has experienced a considerable shift towards large pre-trained Language models (e.g., BERT, RoBERTa and XLNet)4,5,27,28. Some researchers investigated how to integrate the traditional language features (e.g., part-of-speech, syntax dependency tree and knowledge-base) into pre-trained models for improved performance27,29,30. Other researchers focused on how to design new networks for sentiment analysis based on the standard transformer structure28,31. Typically, they fed the outputs of the BERT model to a new network, reloading the parameters of the original pre-trained model to a new network.

By examining these hypotheses and premises, we aim to provide a comprehensive understanding of the role of sentiment and emotion in financial journalism across languages and time periods studied. There definitely seems to be more positive articles across the news categories here as compared to our previous model. However, still looks like technology has the most negative articles and world, the most positive articles similar to our previous analysis. Let’s now do a comparative analysis and see if we still get similar articles in the most positive and negative categories for world news.

The result of perception is quantum cognitive state represented by vector in the qubit Hilbert space. Complex-valued structure of the quantum state space extends the standard vector-based approach to semantics, allowing to account for subjective dimension of human perception in which the result is constrained, but not fully predetermined by input information. In the case of two distinctions, the perception model generates a two-qubit state, entanglement of which quantifies semantic connection between the corresponding words. This two-distinction perception case is realized in the algorithm for detection and measurement of semantic connectivity between pairs of words. The developed approach to cognitive modeling unifies neurophysiological, linguistic, and psychological descriptions in a mathematical and conceptual structure of quantum theory, extending horizons of machine intelligence.

The original English sentence is split into two Chinese sentences through divide translation. Sentence 1 contains a two-layered hierarchical nestification structure while Sentence 2 contains a three-layered hierarchical nestification structure. Additionally, the number of adverbials (ADV) in CT is significantly bigger than that in ES while the number of manners (MNR) in CT is significantly smaller. For a more detailed view of the differences in syntactic subsumption between CT and ES, the current study analyzed the features of several important semantic roles. The word-by-word expansion of the uncut danmaku corpus is mainly applied to the recognition of neologisms of three or more characters. Taking the neologism “蚌埠住了” as an example, after the binary neologism “蚌埠” is counted, the mutual information between “蚌埠” and “住” is calculated by shifting to the right and finally expanding to “蚌埠住了”.

The data source of this study was the official social media pages affiliated with Prime Minister Dr. Abiy Ahmed, Fana Broadcasting Corporation (FBC), the Ezema political party’s official Facebook page, and the Prosperity Party’s official Facebook account. Analyzing Amharic political sentiment poses unique challenges due to the diversity and length of content in social media comments. The Amharic language encompasses a rich vocabulary and intricate grammatical structures that can vary across regions and contexts. This linguistic complexity complicates sentiment analysis, necessitating context-aware approaches. Moreover, social media comments are often lengthy and contextually nuanced, making it challenging to accurately capture the intended sentiment5.

After the semantic roles in each corpus are labelled, textual entailment analysis is then conducted based on the labelling results. For verbs, the analysis is mainly focused on their semantic subsumption since they are the roots of argument structures. For other semantic roles like locations and manners, the entailment analysis is mainly focused on their role in creating syntactic subsumption. In addition to a comprehensive analysis that includes all semantic roles, this study also focuses on several important roles to delve into the semantic discrepancies across the three text types.

Evidence for simplification in information structure is also found in the form of fewer syntactic nestifications, illustrated mainly by a shorter role length of patients (A1) and ranges (A2). Based on these divergences, it is safe to conclude that CT do show a syntactic-semantic characteristic significantly distinct from ES. This section mainly focuses on the discussion of S-universals and presents the results of the comparison between ES and CT. With all the data collected, several statistical tests were conducted on all the indices to explore whether CT exhibit significant semantic differences from ES. Then, a detailed inspection of specific semantic roles was conducted to discuss specific semantic divergences between the two text types. An interesting observation from the results is the trade-off between precision and recall in several models.

The 58,458 sentences with the sentiment and emotion categories are prepared for sentiment classification and emotion detection. The flow of data preparation for sentiment and emotion classification is shown in Fig. The data description of the data prepared for text classification to classify sentiment is tabulated in Table 12. Furthermore, while rule-based detection methods facilitate the identification of sentences containing sexual harassment words, they do not guarantee that these sentences conceptually convey instances of sexual harassment. Henceforth manual interpretation remains essential for accurately determining which sentences involve actual instances of sexual harassment. The distribution of sentences based on different types of sexual harassment and types of sexual offenses can be observed in Fig.