Deep Learning for Chinese Language Sentiment Extraction and Analysis

Text sentiment analysis technology is a hot topic recently. As a short text, there is a feature of using complex sentences to express the author's true views and complex emotional tendencies in micro-blogs. In current researches on sentiment classification based on machine learning, few of them focus on complex sentences. This paper proposed a sentiment analysis method based on SVM and complex phrasing classifier, and made a full analysis of structural features of Chinese conditional sentences, transition sentences and multiple negative sentences, which were taken as text features. A variety of different combinations of features were chosen, including emotional words, speech, negative words, the degree of adverbs and punctuation, etc., to optimize the results of sentiment analysis through multiple sets of experiments. The experiments show that when we choose the combinations of features of emotional words, part of speech and complex sentence patterns, this method improved the accuracy of sentiment classification compared to the common method.

Topic-related Chinese message sentiment analysis

Sentiment analysis of Chinese micro-blog based on sentiment dictionary has become a challenging research subject in the field of artificial intelligence. However, due to insufficient sentiment words, Chinese micro-blog sentiment analysis is difficult to process high accuracy. Aimed at this issue, we propose a method for constructing multiple sentiment dictionaries, which mainly constructs original sentiment dictionary, emoji dictionary, and other related dictionaries. Among them, we have innovatively constructed a Chinese micro-blog new word sentiment dictionary. Multiple sentiment dictionaries increase the coverage of sentiment words. At the same time, we further analyze semantic rule sets between Chinese micro-blog texts and take the inter-sentence analysis rules and sentence pattern analysis rules into the sentiment analysis of Chinese micro-blog, which further improves the accuracy of Chinese micro-blog sentiment analysis. Finally, based on the method of multiple sentiment dictionaries and semantic rule sets, we propose an algorithm for Chinese micro-blog sentiment calculation from complex sentences to clauses and then from clauses to words, and finally combined with the emoji. This algorithm can accurately classify Chinese micro-blog into positive Chinese micro-blog, negative Chinese micro-blog, and neutral Chinese micro-blog. The experimental results show that this method has greatly improved the sentiment analysis of Chinese micro-blog.

In recent years, there has been significant progress in Chinese text sentiment analysis research. However, few studies have investigated the differences between languages, the effectiveness of domain knowledge, and the requirements of downstream tasks. Considering the uniqueness of Chinese text and the practical needs of sentiment analysis, this study aims to address these gaps. To achieve this, we propose a method that deeply integrates the knowledge vector obtained from the emotional knowledge triplets using the TransE model with feature vectors from models like BiGRU and attention mechanisms. We also introduce radical features and emotional part of speech features based on the characteristics of characters and words. In addition, we propose a collaborative approach that integrates characters, words, radicals, and multi-granularity semantic features such as part of speech. Our approach, as evidenced by the Douban Film Review dataset and the NLPECC dataset, proficiently leverages emotional insights alongside nuanced linguistic elements, significantly bolstering the accuracy of sentiment detection in Chinese texts. The method achieved F1-score of 89.23% and 84.84%, respectively, underscoring its efficacy in the realm of Chinese sentiment analysis.

Aiming at the problem that normal vector representation method cannot fully represent the emotional semantic information contained in the text when dealing with Chinese text sentiment analysis task, a multi-granularity convolutional capsule neural network model is constructed. The input of sentiment analysis model is vector representation of the texts trained through the language model. As there exists the problem that a single language model is not enough to abstract text features, a multi-granularity text vector representation method based on BERT is proposed. The text vector representations with different granularities are input into the improved multi-channel convolutional capsule neural network. The capsule layer can associate the low-level and high-level text features, and it will extract information selectively through the dynamic routing algorithm to construct emotional and semantic features of the whole text. Multiple comparative experiments confirm that the method proposed in this paper will efficiently improve the accuracy of Chinese sentiment analysis.

Public health surveillance has gained more importance recently due the global COVID-19 pandemic. It is important to track public opinions and positions on social media automatically, so that this information can be used to improve public health. Sentiment analysis and stance detection are two social media analysis methods that can be applied to health-related social media posts for this purpose. In this chapter, the authors perform sentiment analysis and stance detection in Turkish tweets about COVID-19 vaccination. A sentiment- and stance-annotated Turkish tweet dataset about COVID-19 vaccination is created. Different machine learning approaches (SVM and Random Forest) are applied on this dataset, and the results are compared. Widespread COVID-19 vaccination is claimed to be useful in order to cope with this pandemic. Therefore, results of automatic sentiment and stance analysis on Twitter posts on COVID-19 vaccination can help public health professionals during their decision-making processes.

This study highlights the advantages of sentiment analysis using algorithms in understanding public opinion, especially in the context of the increasing complexity of digital content. To investigate and present the latest developments in sentiment analysis, this study uses the Systematic Literature Review (SLR) method to identify and evaluate previously developed sentiment analysis methods. The research steps involve identifying significant sentiment analysis methods, assessing advantages and disadvantages, and critically reviewing recent advances. By applying algorithms in this process, it is expected to be able to analyze and describe the development of sentiment analysis research comprehensively. Through the application of the SLR method, this study is expected to provide in-depth insights into trends, challenges, and opportunities for future research in sentiment analysis, create a better understanding of effective sentiment analysis methods, and detail the expected results that can be expected in the development of sentiment analysis.

With the wide popularity of personal terminals, people prefer social media to share their lives, which provides a rich source for sentiment analysis methods. However, challenges still exist in small-sample sentiment analysis methods. A sentiment analysis method for Small Samples based on Image Caotion and BERT is proposed. Specifically, the model takes a pre-trained language model as the image description decoder and uses a cross-modal attention mechanism to eliminate the effects of misaligned regions. This can further increase the interaction from image to text. Then, the generated descriptions are coupled with the original text in the dataset. The BERT model is used to extract word vectors and output sentiment analysis results. The COCO dataset is used to train the model for image Captioning, and the MVSA dataset is used for training and evaluation of sentiment analysis. The experiment creates Less Sample Segmentation by randomly selecting samples from the dataset. Accuracy and F1 value are used to compare with baseline models to evaluate the model performance. The results show that the Image Captioning-BERT model has a certain performance improvement in sentiment analysis of image-text pairs with small samples.

People have more channels to express their opinions and feelings about events, products, and celebrities because of the development of social networks. They are becoming rich data sources, gaining attention for many practical applications and in the field of research. Sentiment analysis (SA) is one of the most common uses of this data source. Of the currently available SA datasets, most are only suitable for use in SA corresponding to a specific level, such as document, sentence, or aspect levels. This renders it difficult to develop practical systems that require a combination of sentiment analyzes at all three levels. Additionally, the previous datasets included opinions on only a single domain, although many people often mention multiple domains when expressing their views. This study introduces a new dataset called multi-level and multi-domain (M2SA) for SA. Each sample in M2SA contains a short text with at least two sentences and two aspects with different domains and sentiment polarities. The release of the M2SA dataset will contribute to the promotion of research in the field of SA, primarily by promoting the development and improvement of methods for multi-level SA or multi-aspect, multi-domain SA. The M2SA dataset was tested using state-of-the-art SA methods and was compared with other standard datasets. The results demonstrate that the M2SA dataset is better than the previous datasets in supporting to improve of the performance of SA methods.

Chinese short text sentiment analysis can help understand society’s views on various hot topics. Many existing sentiment analysis methods are based on sentiment dictionaries. Still, sentiment dictionaries are easily affected by subjective factors. They require a lot of time to build as well as maintenance to prevent obsolescence. For the aim of extracting rich information within texts more effectively, we propose a Channel Attention TextCNN with Feature Word Extraction model (CAT-FWE). The feature word extraction module helps us choose words that affect the sentiment of reviews. Then, these words are integrated with multi-level semantic information to enhance the information of sentences. In addition, the channel attention textCNN module that is a promotion of traditional TextCNN tends to pay more attention to those meaningful features. It eliminates the impacts of features that do not make any sense effectively. We apply our CAT-FWE model to both fine-grained classification and binary classification tasks for Chinese short texts. Experiment results show that it can improve the performance of emotion recognition.

This thesis examines the issue of ill-formed text (microtext) and shows its impact on sentiment analysis. The source of microtext is Twitter, WhatsApp, Facebook and other social media platforms. Words or phrases not in their standard language format like “lol” (laugh out loud), “c u 2nite” (see you tonight) and “plz” (please) are called out of vocabulary (OOV) terms. On the other hand, words or phrases in their standard forms, for example, “tonight”, “love”, or “talk to you later”, are called In Vocabulary (IV) terms. The motivation behind microtext normalization is its growing presence on both online and offline platforms. Microtext might appear to be an ill-formed text, which requires find-and-replace methods. Traditionally, in a Natural Language Processing (NLP) context, microtexts such as abbreviations and length shortening or extending are normalized, but phonemic and graphemic variation is often ignored. However, the features of microtext can be traced back to social, cognitive and geographic influences. To understand social media text, it is essential to develop a microtext normalization module for traditional and modern NLP models. The first section of this thesis focuses on combining unsupervised learning methods in 2 different chapters, 4 and 5. The first method includes lexicon creation to transform the most frequent microtexts and emoticons on social media and applies them to sentiment analysis. This method shows an accuracy increase of 4% when microtext normalization is applied before sentiment analysis. This method is extended with an application to a pre-trained chatbot to improve the chatbot’s understanding of microtext (social media slangs). This method shows a mean BLEU score of 0.8 for the SMS and Tweet dataset. The second method incorporates a phonetic based approach to transform any microtext into its phonetic equivalent before 1normalizing it. This method shows an accuracy improvement of 6% for sentiment analysis on SenticNet. The second part of the unsupervised learning is in Chapter 5, introducing the IPA-based method for microtext normalization and sentiment analysis. Epitran was used to transform a word to its International Phonetic Alphabet (IPA) equivalent. The International Phonetic Alphabet (IPA) is an alphabetic system of phonetic notation based primarily on the Latin script. There is a significant improvement in microtext normalization task and sentiment analysis over Soundex. The results also show that there is very little redundancy when words are transformed to phonetics in IPA. In addition, the accuracy of polarity detection using this method has an improvement over baseline by 5%. Chapter 6 discusses a simple sequence-to-sequence based Deep Learning method. As the language is dynamic, it is not easy to maintain the lexicon. Deep Learning alleviates the static lexicon with a probabilistic model to transform microtext into its standard form. The findings suggest that a semiotic class of microtext was easy to transform, whereas the phonetic class of microtext was not handled easily. The proposed model improves the pre-trained sentiment analysis model by 4%. We also introduce a corpus for English microtext normalization. The corpus contains microtext-containing sentences along with their correctly spelt sentences and their correct polarity. This dataset can tackle two NLP problems, i.e., microtext normalization and sentiment analysis. This corpus consists of 38% positive, 39.9% negative and 22.1% neutral. A pre-trained model shows an accuracy of 46.4 % on OOV text for sentiment analysis, showing considerable room for improvement. In summary, this thesis introduces several approaches to microtext normalization analysis and focuses on sentiment analysis application. Microtext cannot be thought of as a one-dimensional problem to NLP. It encapsulates many factors like regional (geographical), ethnic (national and racial), and social (class, age, gender, socioeconomic status and education), which cannot be ignored given the growing importance of social media in our daily life.

Aspect-based sentiment analysis as fine-grained sentiment analysis problem, mainly aims to clarify the sentiment of different aspect. Previous fine-grained sentiment analysis methods usually used Recurrent Neural Networks as feature extractors, and achieved good results in sentiment analysis. However, they were limited to the sequential processing of sequence by RNN. Its training and operation efficiency are relatively low and time consuming. Aiming at this problem, combining the multi-head attention mechanism in Transformer, fusion model of convolution neural network and hierarchical attention coding network is proposed, and convolution neural network is used as feature encoder to improve the efficiency of model training. At the same time, multi-head attention is used to calculate the degree of relevance between aspect words and context, thus capturing emotional elements corresponding to entity words. Experiments on the SemEval2014 dataset have improved the training and inference efficiency of aspect-based sentiment analysis.

An essential function of natural language processing is sentiment analysis. Which holds substantial significance in understanding public opinion across diverse domains. However, while sentiment analysis methodologies abound in English, there exists a notable scarcity of research addressing sentiment analysis in languages like Hindi. In response, the above paper provides a pioneering aspect to Hindi sentiment analysis through the development of a hybrid deep learning-machine learning model integrated with a metaheuristic optimization algorithm. By amalgamating the strengths for normal machine learning (ML) techniques and deep learning (DL), this model endeavours to boost accuracy and robustness in sentiment classification tasks specific to Hindi text. Furthermore, the inclusion of a metaheuristic optimization algorithm aims to optimize crucial model parameters, thereby improving convergence speed and overall performance. The proposed approach is motivated by the need for more comprehensive sentiment analysis techniques tailored for multilingual social media data, particularly in languages like Hindi, which are prevalent on various online platforms. Through empirical evaluation and comparative analysis, this paper demonstrates the efficacy and potential applications of the proposed hybrid model in real-world sentiment analysis scenarios. This research contributes to bridging the gap in sentiment analysis research for non-English languages and lays the foundation for further advancements in multilingual sentiment analysis methodologies.

<sec> <title>BACKGROUND</title> The COVID-19 pandemic caused many changes in primary health care systems in Europe. The fast adoption of telemedicine, the shift of health care resources to COVID-19–related tasks, and the tendency of patients to cancel their nonurgent appointments are some examples of these changes. Patient satisfaction is an important outcome of health care services, and the changes caused by COVID-19 in the system may have affected it. Google Maps reviews provide an important channel for patients to communicate about their experiences regarding the primary health care system. </sec> <sec> <title>OBJECTIVE</title> Drawing from research on social media data analytics and text mining, this study set out to investigate the changes in public sentiment regarding primary health care in Finland and Andalusia (Spain) before, during, and after the COVID-19 pandemic. </sec> <sec> <title>METHODS</title> We collected 55,043 Google Maps reviews from primary health care locations in Finland and Andalusia from January 1, 2013, to May 15, 2024. There are 604 primary health care locations in Finland and 1016 in Andalusia. The total number of Google Maps reviews collected was 12,247 for Finland and 42,796 for Andalusia. First, lexicon-based sentiment analysis using the open-source software AFINN was conducted for the Finnish- and Spanish-language datasets. Thereafter, transformer-based deep learning models for sentiment analysis were applied for both languages. The numeric user ratings and the results of the sentiment analysis were then analyzed. In addition, we conducted a word frequency analysis of the reviews. </sec> <sec> <title>RESULTS</title> There were important changes in the ratings and sentiment in the data for Andalusia. The ratings shifted from median 4 (IQR 3) before the COVID-19 pandemic to median 1 (IQR 2) during and median 1 (IQR 3) after the COVID-19 pandemic, on a scale from 1 to 5. The median of the sentiment values of the review texts shifted from neutral before the COVID-19 pandemic to –2 (IQR 0.055) or –1 (IQR 1) during and after the COVID-19 pandemic, depending on which sentiment analysis method was used. Interestingly, changes in numeric ratings and sentiment of the review texts in Finland were only minor, and the median values were the same during all 3 periods. Lexical analysis revealed changes in word frequencies across the periods, reflecting shifts in primary health care experiences during the pandemic, especially among the Spanish-language reviews. </sec> <sec> <title>CONCLUSIONS</title> The change toward a more negative public discussion on primary health care in Andalusia during the COVID-19 pandemic was considerable. This can be observed both in the numeric user ratings and in the sentiment analysis of the review texts. However, the data for Finland show that the public discourse stayed mostly neutral or slightly positive. The findings have implications on the quality management procedures in primary health care and on the use of user-generated content as an additional information source. </sec>

We present a case study as part of a work-in-progress project about multimodal sentiment analysis on historic German plays, taking Emilia Galotti by G. E. Lessing as our initial use case. We analyze the textual version and an audio version (audio book). We focus on ready-to-use sentiment analysis methods: For the textual component, we implement a naive lexicon-based approach and another approach that enhances the lexicon by means of several NLP methods. For the audio analysis, we use the free version of the Vokaturi tool. We compare the results of all approaches and evaluate them against the annotations of a human expert, which serves as a gold standard. For our use case, we can show that audio and text sentiment analysis behave very differently: textual sentiment analysis tends to predict sentiment as rather negative and audio sentiment as rather positive. Compared to the gold standard, the textual sentiment analysis achieves accuracies of 56% while the accuracy for audio sentiment analysis is only 32%. We discuss possible reasons for these mediocre results and give an outlook on further steps we want to pursue in the context of multimodal sentiment analysis on historic plays.