No description available.

No description available.

In this thesis we have present effective approaches for two classic Nature Language Processing tasks: Multi-label Text Classification(MLTC) and Sentiment Analysis(SA) based on two datasets. For MLTC, a robust deep learning approach based on convolution neural network(CNN) has been introduced. We have done this on almost one million records with a related label list consists of 20 labels. We have divided our data set into three parts, training set, validation set and test set. Our CNN based model achieved great result measured in F1 score. For SA, data set was more informative and well-structured compared with MLTC. A traditional word embedding method, Word2Vec was used for generating word vector of each text records. Following that, we employed several classic deep learning models such as Bi-LSTM, RCNN, Attention mechanism and CNN to extract sentiment features. In the next step, a classification frame was designed to graded. At last, the start-of-art language model, BERT which use transfer learning method was employed. In conclusion, we compared performance of RNN-based model, CNN-based model and pre-trained language model on classification task and discuss their applicability. / Thesis / Master of Science in Electrical and Computer Engineering (MSECE) / This theis purposed two deep learning solution to both multi-label classification problem and sentiment analysis problem.

An important means for disseminating information on social media platforms is by including URLs that point to external sources in user posts. In X, formally known as Twitter, we estimate that about 21% of the daily stream of English-language posts contain URLs. Given this prevalence, we assert that studying URLs in social media holds significant importance as they play a pivotal part in shaping the flow of information and influencing user behavior. Examining hyperlinked posts can help us gain valuable insights into online discourse and detect emerging trends. The first aspect of our analysis is the study of users' intentions behind including URLs in social media posts. We argue that gaining insights about the users' motivations for posting with URLs has multiple applications, including the appropriate treatment and processing of these posts in other tasks. Hence, we build a comprehensive taxonomy containing the various intentions behind sharing URLs on social media. In addition, we explore the labeling of intentions via the use of crowdsourcing. In addition to the intentions aspect of hyperlinked posts, we analyze their structure relative to the content of the web documents pointed to by the URLs. Hence, we define, and analyze the segmentation problem of hyperlinked posts and develop an effective algorithm to solve it. We show that our solution can benefit sentiment analysis on social media. In the final aspect of our analysis, we investigate the emergence of news outlets posing as local sources, known as "pink slime", and their spread on social media. We conduct a comprehensive study investigating hyperlinked posts featuring pink slime websites. Through our analysis of the patterns and origins of posts, we discover and extract syntactical features and utilize them for developing a classification approach to detect such posts. Our approach has achieved an accuracy rate of 92.5%. / Computer and Information Science

This dissertation focuses on the review understanding problem, which has gained attention from both industry and academia, and has found applications in many downstream tasks, such as recommendation, information retrieval and review summarization. In this dissertation, we aim to develop machine learning and natural language processing tools to understand and learn structured knowledge from unstructured reviews, which can be investigated in three research directions, including understanding review corpora, understanding review documents, and understanding review segments. For the corpus-level review understanding, we have focused on discovering knowledge from corpora that consist of short texts. Since they have limited contextual information, automatically learning topics from them remains a challenging problem. We propose a semantics-assisted non-negative matrix factorization model to deal with this problem. It effectively incorporates the word-context semantic correlations into the model, where the semantic relationships between the words and their contexts are learned from the skip-gram view of a corpus. We conduct extensive sets of experiments on several short text corpora to demonstrate the proposed model can discover meaningful and coherent topics. For document-level review understanding, we have focused on building interpretable and reliable models for the document-level multi-aspect sentiment analysis (DMSA) task, which can help us to not only recover missing aspect-level ratings and analyze sentiment of customers, but also detect aspect and opinion terms from reviews. We conduct three studies in this research direction. In the first study, we collect a new DMSA dataset in the healthcare domain and systematically investigate reviews in this dataset, including a comprehensive statistical analysis and topic modeling to discover aspects. We also propose a multi-task learning framework with self-attention networks to predict sentiment and ratings for given aspects. In the second study, we propose corpus-level and concept-based explanation methods to interpret attention-based deep learning models for text classification, including sentiment classification. The proposed corpus-level explanation approach aims to capture causal relationships between keywords and model predictions via learning importance of keywords for predicted labels across a training corpus based on attention weights. We also propose a concept-based explanation method that can automatically learn higher level concepts and their importance to model predictions. We apply these methods to the classification task and show that they are powerful in extracting semantically meaningful keywords and concepts, and explaining model predictions. In the third study, we propose an interpretable and uncertainty aware multi-task learning framework for DMSA, which can achieve competitive performance while also being able to interpret the predictions made. Based on the corpus-level explanation method, we propose an attention-driven keywords ranking method, which can automatically discover aspect terms and aspect-level opinion terms from a review corpus using the attention weights. In addition, we propose a lecture-audience strategy to estimate model uncertainty in the context of multi-task learning. For the segment-level review understanding, we have focused on the unsupervised aspect detection task, which aims to automatically extract interpretable aspects and identify aspect-specific segments from online reviews. The existing deep learning-based topic models suffer from several problems such as extracting noisy aspects and poorly mapping aspects discovered by models to the aspects of interest. To deal with these problems, we propose a self-supervised contrastive learning framework in order to learn better representations for aspects and review segments. We also introduce a high-resolution selective mapping method to efficiently assign aspects discovered by the model to the aspects of interest. In addition, we propose using a knowledge distillation technique to further improve the aspect detection performance. / Doctor of Philosophy / Nowadays, online reviews are playing an important role in our daily lives. They are also critical to the success of many e-commerce and local businesses because they can help people build trust in brands and businesses, provide insights into products and services, and improve consumers' confidence. As a large number of reviews accumulate every day, a central research problem is to build an artificial intelligence system that can understand and interact with these reviews, and further use them to offer customers better support and services. In order to tackle challenges in these applications, we first have to get an in-depth understanding of online reviews. In this dissertation, we focus on the review understanding problem and develop machine learning and natural language processing tools to understand reviews and learn structured knowledge from unstructured reviews. We have addressed the review understanding problem in three directions, including understanding a collection of reviews, understanding a single review, and understanding a piece of a review segment. In the first direction, we proposed a short-text topic modeling method to extract topics from review corpora that consist of primary complaints of consumers. In the second direction, we focused on building sentiment analysis models to predict the opinions of consumers from their reviews. Our deep learning models can provide good prediction accuracy as well as a human-understandable explanation for the prediction. In the third direction, we develop an aspect detection method to automatically extract sentences that mention certain features consumers are interested in, from reviews, which can help customers efficiently navigate through reviews and help businesses identify the advantages and disadvantages of their products.

Selecting the correct methodology to better understand how consumers perceive food products is a challenging task for the food industry and sensory researchers alike. Free comment tasks (FC) utilize the advantages of open-ended questions to generate intuitive comments from untrained consumers to help identify and describe sensory attributes of products. However, FC data is typically analyzed using text analysis done by hand and is very cumbersome to organize and interpret. There is a growing need and interest to add to the library of data analysis tools used to understand FC data and consumer acceptance studies. Sentiment analysis is an opinion mining tool commonly used in marketing and computer science that extracts the emotional valence of the author from an unstructured text in the form of a sentiment score. A few studies in sensory evaluation use lexicon-based sentiment analysis which has many drawbacks: it is time-consuming, requires a large amount of data and dictionaries need to be tailored for food. We used a deep learning sentiment analysis approach to analyze and predict consumer sentiment/acceptance. The research objectives of this study are 1) to explore quicker and automated methods of sentiment analysis to better understand and predict consumer acceptance, and 2) to examine the advantages and disadvantages of sentiment analysis as a data analysis tool in sensory evaluation. We avoided the pitfalls of creating a sentiment lexicon by using online beer reviews to train a word embedding model where all of the relevant words in the review are converted into vectors. We used the distance and similarity (clustering) of the vectors to determine taste/flavor attributes that correspond to negative and positive sentiment. Next, to validate and test our model we gathered FC data in a consumer acceptance study. Panelists (N=68) were presented with six beers, one at a time and were instructed to taste and smell before leaving comments. We performed sentiment analysis on the FC data, and we compared our deep learning sentiment analysis model with three other pre-existing sentiment analysis models: SentimentR, VADER, and Liu and Hu opinion lexicon. Our deep learning sentiment analysis model had the highest accuracy (69%) and precision rate (73%). Overall, our findings provide an early look into the advantages and disadvantages of sentiment analysis applied to FC data in sensory evaluation. / Master of Science in Life Sciences / It can be a challenging task for the food industry and sensory researchers to select the correct methodology to better understand how consumers perceive food products. One method is Free comment tasks (FC) which use open-ended questions to generate comments from consumers. However, FC data is typically analyzed using text analysis done by hand and is very cumbersome to organize and interpret. This thesis is interested in investigating the application of data analysis methods from computer science on FC data. Sentiment analysis is an opinion mining tool commonly used in marketing and computer science that finds the emotional tone of the author from a text in the form of a sentiment score. First, we created a deep learning sentiment analysis model which uses algorithms to find useful patterns in the text that indicate positive and negative sentiment with minimal human intervention. We were interested if there were any advantages in creating a model so we compared our model to three widely used sentiment analysis models: VADER, Liu Hu and Sentiment R. Next, to test our sentiment analysis model and the three others we gathered FC data in a consumer acceptance study. Panelists (N=68) were presented with six beers, one at a time and were instructed to taste and smell before leaving comments. The research objectives of this study are 1) to explore quicker and automated data analysis methods to understand FC data, and 2) to examine the advantages and disadvantages of data analysis tools from computer science in sensory evaluation. Our deep learning sentiment analysis model had the highest accuracy (69%) and precision rate (73%). Overall, our findings provide an early look into the advantages and disadvantages of sentiment analysis applied to FC data in sensory evaluation.

This paper has three chapters. In the first chapter, we develop a measure of housing sentiment for 24 cities in China by parsing through newspaper articles from 2006 to 2017.We find that the sentiment index has strong predictive power for future house prices even after controlling for past price changes and macroeconomic fundamentals. The index leads price movements by nearly 9 months, and it is highly correlated with other survey expectations measures that come with a significant time lag. In the second chapter, we show that short term house price movement is predictable by solely using newspaper and historical price change. In the last chapter, using the sentiment index constructed from newspaper, we got empirical results to show that some people are forward-looking when deciding default and a positive sentiment (anticipated house price appreciation) will lower the Z score of probability of default by 0.028. / Doctor of Philosophy / This paper has three chapters. In the first chapter, we develop a measure of housing sentiment for 24 cities in China by parsing through newspaper articles from 2006 to 2017. Two sentiment index were created using text mining method based on keywords matching and machine learning respectively.We find that the sentiment index has strong predictive power for future house prices even after controlling for past price changes and macroeconomic fundamentals. The index leads price movements by nearly 9 months, and it is highly correlated with other survey expectations measures that come with a significant time lag. In contrast, we find much weaker feedback coming from past prices to current sentiment. In the second chapter, we show that short term house price movement is predictable by solely using newspaper and historical price change. The accuracy of the prediction could be up to 0.96 for out of sample prediction. We first use a text mining method to transfer all the text information into numerical vector space, which is able to represent the extracted full information contained in a text. Then by adopting machine learning models of Neural networks, SVM, and random forest, we classified the newspaper into 1 (up) and 0 (down) group and constructed an index as the mean label accordingly. In the last chapter, by merging the Fannie Mae loan performance data with the sentiment index constructed from newspaper as well as the macro variables about local market, we got empirical results to show that some people are forward-looking when deciding default and a positive sentiment ( anticipated house price appreciation) will lower the Z score of probability of default by 0.028. We found that during the recession period, people access more information when they try to default, on top of the traditional econ conditions and historical house price, they also consider the future house price change. Moreover, borrowers with high income, high home value, and high FICO scores tend to pay more attention to future price change. However, for those who are less experienced in this game (first time home buyer), they only pay attention to the historical price change during the recession period.

Yes / The governments, nowadays, throughout the world are increasingly becoming dependent on public opinion regarding the framing and implementation of certain policies for the welfare of the general public. The role of social media is vital to this emerging trend. Traditionally, lack of public participation in various policy making decision used to be a major cause of concern particularly when formulating and evaluating such policies. However, the exponential rise in usage of social media platforms by general public has given the government a wider insight to overcome this long pending dilemma. Cloud-based e-governance is currently being realized due to IT infrastructure availability along with mindset changes of government advisors towards realizing the various policies in a best possible manner. This paper presents a pragmatic approach that combines the capabilities of both cloud computing and social media analytics towards efficient monitoring and controlling of governmental policies through public involvement. The proposed system has provided us some encouraging results, when tested for Goods and Services Tax (GST) implementation by Indian government and established that it can be successfully implemented for efficient policy making and implementation.

Yes / This paper investigates the impact of investor sentiment on the mean-variance relationship in 14 European stock markets. Applying three approaches to define investors’ neutrality and determine high and low sentiment periods, we find that individual investors’ increased presence and trading over high-sentiment periods would undermine the risk-return tradeoff. More importantly, we report that investors’ optimism (pessimism) is more determined by their normal sentiment state, represented by the all-period average sentiment level, rather than the neutrality value set in sentiment surveys.

Yes / This paper investigates the role of institutional investor sentiment in the mean–variance relation. We find market returns are negatively (positively) related to market’s conditional volatility over bullish (bearish) periods. The evidence indicates institutional investors to be sentiment traders as well.