Search

Text classification is a fundamental task in many applications such as topic labeling, sentiment analysis, and spam detection. The text syntactic relationship and word sequence are important and useful for text classification. How to model and incorporate them to improve performance is one key challenge. Inspired by human behavior in understanding text. In this paper, we combine the syntactic relationship, sequence structure, and semantics for text representation, and propose an attention-enhanced capsule network-based text classification model. Specifically, we use graph convolutional neural networks to encode syntactic dependency trees, build multi-head attention to encode dependencies relationship in text sequence, merge with semantic information by capsule network at last. Extensive experiments on five datasets demonstrate that our approach can effectively improve the performance of text classification compared with state-of-the-art methods. The result also shows capsule network, graph convolutional neural network, and multi-headed attention has integration effects on text classification tasks.

Nowadays, social media experience an increase in hostility, which leads to many people suffering from online abusive behavior and harassment. We introduce a new publicly available annotated dataset for abusive language detection in short texts. The dataset includes comments from YouTube, along with contextual information: replies, video, video title, and the original description. The comments in the dataset are labeled as abusive or not and are classified by topic: politics, religion, and other. In particular, we discuss our refined annotation guidelines for such classification. We report a number of strong baselines on this dataset for the tasks of abusive language detection and topic classification, using a number of classifiers and text representations. We show that taking into account the conversational context, namely, replies, greatly improves the classification results as compared with using only linguistic features of the comments. We also study how the classification accuracy depends on the topic of the comment.

Multitask learning has led to significant advances in Natural Language Processing, including the decaNLP benchmark where question answering is used to frame 10 natural language understanding tasks in a single model. In this work we show how models trained to solve decaNLP fail with simple paraphrasing of the question. We contribute a crowd-sourced corpus of paraphrased questions (PQ-decaNLP), annotated with paraphrase phenomena. This enables analysis of how transformations such as swapping the class labels and changing the sentence modality lead to a large performance degradation. Training both MQAN and the newer T5 model using PQ-decaNLP improves their robustness and for some tasks improves the performance on the original questions, demonstrating the benefits of a model which is more robust to paraphrasing. Additionally, we explore how paraphrasing knowledge is transferred between tasks, with the aim of exploiting the multitask property to improve the robustness of the models. We explore the addition of paraphrase detection and paraphrase generation tasks, and find that while both models are able to learn these new tasks, knowledge about paraphrasing does not transfer to other decaNLP tasks.

In Information Retrieval (IR), Data Mining (DM), and Machine Learning (ML), similarity measures have been widely used for text clustering and classification. The similarity measure is the cornerstone upon which the performance of most DM and ML algorithms is completely dependent. Thus, till now, the endeavor in literature for an effective and efficient similarity measure is still immature. Some recently-proposed similarity measures were effective, but have a complex design and suffer from inefficiencies. This work, therefore, develops an effective and efficient similarity measure of a simplistic design for text-based applications. The measure developed in this work is driven by Boolean logic algebra basics (BLAB-SM), which aims at effectively reaching the desired accuracy at the fastest run time as compared to the recently developed state-of-the-art measures. Using the term frequency–inverse document frequency (TF-IDF) schema, the K-nearest neighbor (KNN), and the K-means clustering algorithm, a comprehensive evaluation is presented. The evaluation has been experimentally performed for BLAB-SM against seven similarity measures on two most-popular datasets, Reuters-21 and Web-KB. The experimental results illustrate that BLAB-SM is not only more efficient but also significantly more effective than state-of-the-art similarity measures on both classification and clustering tasks.

Sentiment analysis plays a key role in companies, especially stores, and increasing the accuracy in determining customers’ opinions about products assists to maintain their competitive conditions. We intend to analyze the users’ opinions on the website of the most immense online store in Iran; Digikala. However, the Persian language is unstructured which makes the pre-processing stage very difficult and it is the main problem of sentiment analysis in Persian. What exacerbates this problem is the lack of available libraries for Persian pre-processing, while most libraries focus on English. To tackle this, approximately 3 million reviews were gathered in Persian from the Digikala website using web-mining techniques, and the fastText method was used to create a word embedding. It was assumed that this would dramatically cut down on the need for text pre-processing through the skip-gram method considering the position of the words in the sentence and the words’ relations to each other. Another word embedding has been created using the TF-IDF in parallel with fastText to compare their performance. In addition, the results of the Convolutional Neural Network (CNN), BiLSTM, Logistic Regression, and Naïve Bayes models have been compared. As a significant result, we obtained 0.996 AUC and 0.956 F-score using fastText and CNN. In this article, not only has it been demonstrated to what extent it is possible to be independent of pre-processing but also the accuracy obtained is better than other researches done in Persian. Avoiding complex text preprocessing is also important for other languages since most text preprocessing algorithms have been developed for English and cannot be used for other languages. The created word embedding due to its high accuracy and independence of pre-processing has other applications in Persian besides sentiment analysis.

Science across all disciplines has become increasingly data-driven, leading to additional needs with respect to software for collecting, processing and analysing data. Thus, transparency about software used as part of the scientific process is crucial to understand provenance of individual research data and insights, is a prerequisite for reproducibility and can enable macro-analysis of the evolution of scientific methods over time. However, missing rigor in software citation practices renders the automated detection and disambiguation of software mentions a challenging problem. In this work, we provide a large-scale analysis of software usage and citation practices facilitated through an unprecedented knowledge graph of software mentions and affiliated metadata generated through supervised information extraction models trained on a unique gold standard corpus and applied to more than 3 million scientific articles. Our information extraction approach distinguishes different types of software and mentions, disambiguates mentions and outperforms the state-of-the-art significantly, leading to the most comprehensive corpus of 11.8 M software mentions that are described through a knowledge graph consisting of more than 300 M triples. Our analysis provides insights into the evolution of software usage and citation patterns across various fields, ranks of journals, and impact of publications. Whereas, to the best of our knowledge, this is the most comprehensive analysis of software use and citation at the time, all data and models are shared publicly to facilitate further research into scientific use and citation of software.

Background Fine-grained sentiment analysis is used to interpret consumers’ sentiments, from their written comments, towards specific entities on specific aspects. Previous researchers have introduced three main tasks in this field (ABSA, TABSA, MEABSA), covering all kinds of social media data (e.g., review specific, questions and answers, and community-based). In this paper, we identify and address two common challenges encountered in these three tasks, including the low-resource problem and the sentiment polarity bias. Methods We propose a unified model called PEA by integrating data augmentation methodology with the pre-trained language model, which is suitable for all the ABSA, TABSA and MEABSA tasks. Two data augmentation methods, which are entity replacement and dual noise injection, are introduced to solve both challenges at the same time. An ensemble method is also introduced to incorporate the results of the basic RNN-based and BERT-based models. Results PEA shows significant improvements on all three fine-grained sentiment analysis tasks when compared with state-of-the-art models. It also achieves comparable results with what the baseline models obtain while using only 20% of their training data, which demonstrates its extraordinary performance under extreme low-resource conditions.

The real-time availability of the Internet has engaged millions of users around the world. The usage of regional languages is being preferred for effective and ease of communication that is causing multilingual data on social networks and news channels. People share ideas, opinions, and events that are happening globally i.e., sports, inflation, protest, explosion, and sexual assault, etc. in regional (local) languages on social media. Extraction and classification of events from multilingual data have become bottlenecks because of resource lacking. In this research paper, we presented the event classification task for the Urdu language text existing on social media and the news channels by using machine learning classifiers. The dataset contains more than 0.1 million (102,962) labeled instances of twelve (12) different types of events. The title, its length, and the last four words of a sentence are used as features to classify the events. The Term Frequency-Inverse Document Frequency (tf-idf) showed the best results as a feature vector to evaluate the performance of the six popular machine learning classifiers. Random Forest (RF) and K-Nearest Neighbor (KNN) are among the classifiers that out-performed among other classifiers by achieving 98.00% and 99.00% accuracy, respectively. The novelty lies in the fact that the features aforementioned are not applied, up to the best of our knowledge, in the event extraction of the text written in the Urdu language.

Deep learning belongs to the field of artificial intelligence, where machines perform tasks that typically require some kind of human intelligence. Similar to the basic structure of a brain, a deep learning algorithm consists of an artificial neural network, which resembles the biological brain structure. Mimicking the learning process of humans with their senses, deep learning networks are fed with (sensory) data, like texts, images, videos or sounds. These networks outperform the state-of-the-art methods in different tasks and, because of this, the whole field saw an exponential growth during the last years. This growth resulted in way over 10,000 publications per year in the last years. For example, the search engine PubMed alone, which covers only a sub-set of all publications in the medical field, provides already over 11,000 results in Q3 2020 for the search term 'deep learning', and around 90% of these results are from the last three years. Consequently, a complete overview over the field of deep learning is already impossible to obtain and, in the near future, it will potentially become difficult to obtain an overview over a subfield. However, there are several review articles about deep learning, which are focused on specific scientific fields or applications, for example deep learning advances in computer vision or in specific tasks like object detection. With these surveys as a foundation, the aim of this contribution is to provide a first high-level, categorized meta-survey of selected reviews on deep learning across different scientific disciplines. The categories (computer vision, language processing, medical informatics and additional works) have been chosen according to the underlying data sources (image, language, medical, mixed). In addition, we review the common architectures, methods, pros, cons, evaluations, challenges and future directions for every sub-category., Comment: 83 pages, 22 figures, 9 tables, 100 references

Bengali is a low-resource language that lacks tools and resources for various natural language processing (NLP) tasks, such as sentiment analysis or profanity identification. In Bengali, only the translated versions of English sentiment lexicons are available. Moreover, no dictionary exists for detecting profanity in Bengali social media text. This study introduces a Bengali sentiment lexicon, BengSentiLex, and a Bengali swear lexicon, BengSwearLex. For creating BengSentiLex, a cross-lingual methodology is proposed that utilizes a machine translation system, a review corpus, two English sentiment lexicons, pointwise mutual information (PMI), and supervised machine learning (ML) classifiers in various stages. A semi-automatic methodology is presented to develop BengSwearLex that leverages an obscene corpus, word embedding, and part-of-speech (POS) taggers. The performance of BengSentiLex compared with the translated English lexicons in three evaluation datasets. BengSentiLex achieves 5%–50% improvement over the translated lexicons. For identifying profanity, BengSwearLex achieves documentlevel coverage of around 85% in an document-level in the evaluation dataset. The experimental results imply that BengSentiLex and BengSwearLex are effective resources for classifying sentiment and identifying profanity in Bengali social media content, respectively.

Melody and lyrics, reflecting two unique human cognitive abilities, are usually combined in music to convey emotions. Although psychologists and computer scientists have made considerable progress in revealing the association between musical structure and the perceived emotions of music, the features of lyrics are relatively less discussed. Using linguistic inquiry and word count (LIWC) technology to extract lyric features in 2,372 Chinese songs, this study investigated the effects of LIWC-based lyric features on the perceived arousal and valence of music. First, correlation analysis shows that, for example, the perceived arousal of music was positively correlated with the total number of lyric words and the mean number of words per sentence and was negatively correlated with the proportion of words related to the past and insight. The perceived valence of music was negatively correlated with the proportion of negative emotion words. Second, we used audio and lyric features as inputs to construct music emotion recognition (MER) models. The performance of random forest regressions reveals that, for the recognition models of perceived valence, adding lyric features can significantly improve the prediction effect of the model using audio features only; for the recognition models of perceived arousal, lyric features are almost useless. Finally, by calculating the feature importance to interpret the MER models, we observed that the audio features played a decisive role in the recognition models of both perceived arousal and perceived valence. Unlike the uselessness of the lyric features in the arousal recognition model, several lyric features, such as the usage frequency of words related to sadness, positive emotions, and tentativeness, played important roles in the valence recognition model.

Framing is a process of emphasizing a certain aspect of an issue over the others, nudging readers or listeners towards different positions on the issue even without making a biased argument. {Here, we propose FrameAxis, a method for characterizing documents by identifying the most relevant semantic axes ("microframes") that are overrepresented in the text using word embedding. Our unsupervised approach can be readily applied to large datasets because it does not require manual annotations. It can also provide nuanced insights by considering a rich set of semantic axes. FrameAxis is designed to quantitatively tease out two important dimensions of how microframes are used in the text. \textit{Microframe bias} captures how biased the text is on a certain microframe, and \textit{microframe intensity} shows how actively a certain microframe is used. Together, they offer a detailed characterization of the text. We demonstrate that microframes with the highest bias and intensity well align with sentiment, topic, and partisan spectrum by applying FrameAxis to multiple datasets from restaurant reviews to political news.} The existing domain knowledge can be incorporated into FrameAxis {by using custom microframes and by using FrameAxis as an iterative exploratory analysis instrument.} Additionally, we propose methods for explaining the results of FrameAxis at the level of individual words and documents. Our method may accelerate scalable and sophisticated computational analyses of framing across disciplines., Comment: 24 pages; published in PeerJ CS

The satisfaction of employees is very important for any organization to make sufficient progress in production and to achieve its goals. Organizations try to keep their employees satisfied by making their policies according to employees’ demands which help to create a good environment for the collective. For this reason, it is beneficial for organizations to perform staff satisfaction surveys to be analyzed, allowing them to gauge the levels of satisfaction among employees. Sentiment analysis is an approach that can assist in this regard as it categorizes sentiments of reviews into positive and negative results. In this study, we perform experiments for the world’s big six companies and classify their employees’ reviews based on their sentiments. For this, we proposed an approach using lexicon-based and machine learning based techniques. Firstly, we extracted the sentiments of employees from text reviews and labeled the dataset as positive and negative using TextBlob. Then we proposed a hybrid/voting model named Regression Vector-Stochastic Gradient Descent Classifier (RV-SGDC) for sentiment classification. RV-SGDC is a combination of logistic regression, support vector machines, and stochastic gradient descent. We combined these models under a majority voting criteria. We also used other machine learning models in the performance comparison of RV-SGDC. Further, three feature extraction techniques: term frequency-inverse document frequency (TF-IDF), bag of words, and global vectors are used to train learning models. We evaluated the performance of all models in terms of accuracy, precision, recall, and F1 score. The results revealed that RV-SGDC outperforms with a 0.97 accuracy score using the TF-IDF feature due to its hybrid architecture.

BackgroundRumor detection is a popular research topic in natural language processing and data mining. Since the outbreak of COVID-19, related rumors have been widely posted and spread on online social media, which have seriously affected people’s daily lives, national economy, social stability, etc. It is both theoretically and practically essential to detect and refute COVID-19 rumors fast and effectively. As COVID-19 was an emergent event that was outbreaking drastically, the related rumor instances were very scarce and distinct at its early stage. This makes the detection task a typical few-shot learning problem. However, traditional rumor detection techniques focused on detecting existed events with enough training instances, so that they fail to detect emergent events such as COVID-19. Therefore, developing a new few-shot rumor detection framework has become critical and emergent to prevent outbreaking rumors at early stages.MethodsThis article focuses on few-shot rumor detection, especially for detecting COVID-19 rumors from Sina Weibo with only a minimal number of labeled instances. We contribute a Sina Weibo COVID-19 rumor dataset for few-shot rumor detection and propose a few-shot learning-based multi-modality fusion model for few-shot rumor detection. A full microblog consists of the source post and corresponding comments, which are considered as two modalities and fused with the meta-learning methods.ResultsExperiments of few-shot rumor detection on the collected Weibo dataset and the PHEME public dataset have shown significant improvement and generality of the proposed model.

Small sample learning aims to learn information about object categories from a single or a few training samples. This learning style is crucial for deep learning methods based on large amounts of data. The deep learning method can solve small sample learning through the idea of meta-learning “how to learn by using previous experience.” Therefore, this paper takes image classification as the research object to study how meta-learning quickly learns from a small number of sample images. The main contents are as follows: After considering the distribution difference of data sets on the generalization performance of measurement learning and the advantages of optimizing the initial characterization method, this paper adds the model-independent meta-learning algorithm and designs a multi-scale meta-relational network. First, the idea of META-SGD is adopted, and the inner learning rate is taken as the learning vector and model parameter to learn together. Secondly, in the meta-training process, the model-independent meta-learning algorithm is used to find the optimal parameters of the model. The inner gradient iteration is canceled in the process of meta-validation and meta-test. The experimental results show that the multi-scale meta-relational network makes the learned measurement have stronger generalization ability, which further improves the classification accuracy on the benchmark set and avoids the need for fine-tuning of the model-independent meta-learning algorithm.

Platforms that feature user-generated content (social media, online forums, newspaper comment sections etc.) have to detect and filter offensive speech within large, fast-changing datasets. While many automatic methods have been proposed and achieve good accuracies, most of these focus on the English language, and are hard to apply directly to languages in which few labeled datasets exist. Recent work has therefore investigated the use ofcross-lingual transfer learningto solve this problem, training a model in a well-resourced language and transferring to a less-resourced target language; but performance has so far been significantly less impressive. In this paper, we investigate the reasons for this performance drop, via a systematic comparison of pre-trained models and intermediate training regimes on five different languages. We show that using a better pre-trained language model results in a large gain in overall performance and in zero-shot transfer, and that intermediate training on other languages is effective when little target-language data is available. We then use multiple analyses of classifier confidence and language model vocabulary to shed light on exactly where these gains come from and gain insight into the sources of the most typical mistakes.

Hate speech is one type of harmful online content which directly attacks or promotes hate towards a group or an individual member based on their actual or perceived aspects of identity, such as ethnicity, religion, and sexual orientation. With online hate speech on the rise, its automatic detection as a natural language processing task is gaining increasing interest. However, it is only recently that it has been shown that existing models generalise poorly to unseen data. This survey paper attempts to summarise how generalisable existing hate speech detection models are and the reasons why hate speech models struggle to generalise, sums up existing attempts at addressing the main obstacles, and then proposes directions of future research to improve generalisation in hate speech detection.

Semantic Text Similarity (STS) has several and important applications in the field of Natural Language Processing (NLP). The Aim of this study is to investigate the effect of stemming on text similarity for Arabic language at sentence level. Several Arabic light and heavy stemmers as well as lemmatization algorithms are used in this study, with a total of 10 algorithms. Standard training and testing data sets are used from SemEval-2017 international workshop for Task 1, Track 1 Arabic (ar–ar). Different features are selected to study the effect of stemming on text similarity based on different similarity measures. Traditional machine learning algorithms are used such as Support Vector Machines (SVM), Stochastic Gradient Descent (SGD) and Naïve Bayesian (NB). Compared to the original text, using the stemmed and lemmatized documents in experiments achieve enhanced Pearson correlation results. The best results attained when using Arabic light Stemmer (ARLSTem) and Farasa light stemmers, Farasa and Qalsadi Lemmatizers and Tashaphyne heavy stemmer. The best enhancement was about 7.34% in Pearson correlation. In general, stemming considerably improves the performance of sentence text similarly for Arabic language. However, some stemmers make results worse than those for original text; they are Khoja heavy stemmer and AlKhalil light stemmer.

Today, increased attention is drawn towards network representation learning, a technique that maps nodes of a network into vectors of a low-dimensional embedding space. A network embedding constructed this way aims to preserve nodes similarity and other specific network properties. Embedding vectors can later be used for downstream machine learning problems, such as node classification, link prediction and network visualization. Naturally, some networks have text information associated with them. For instance, in a citation network, each node is a scientific paper associated with its abstract or title; in a social network, all users may be viewed as nodes of a network and posts of each user as textual attributes. In this work, we explore how combining existing methods of text and network embeddings can increase accuracy for downstream tasks and propose modifications to popular architectures to better capture textual information in network embedding and fusion frameworks.

The cross-modal retrieval (CMR) has attracted much attention in the research community due to flexible and comprehensive retrieval. The core challenge in CMR is the heterogeneity gap, which is generated due to different statistical properties of multi-modal data. The most common solution to bridge the heterogeneity gap is representation learning, which generates a common sub-space. In this work, we propose a framework called “Improvement of Deep Cross-Modal Retrieval (IDCMR)”, which generates real-valued representation. The IDCMR preserves both intra-modal and inter-modal similarity. The intra-modal similarity is preserved by selecting an appropriate training model for text and image modality. The inter-modal similarity is preserved by reducing modality-invariance loss. The mean average precision (mAP) is used as a performance measure in the CMR system. Extensive experiments are performed, and results show that IDCMR outperforms over state-of-the-art methods by a margin 4% and 2% relatively with mAP in the text to image and image to text retrieval tasks on MSCOCO and Xmedia dataset respectively.

The popularity of the internet, smartphones, and social networks has contributed to the proliferation of misleading information like fake news and fake reviews on news blogs, online newspapers, and e-commerce applications. Fake news has a worldwide impact and potential to change political scenarios, deceive people into increasing product sales, defaming politicians or celebrities, and misguiding visitors to stop visiting a place or country. Therefore, it is vital to find automatic methods to detect fake news online. In several past studies, the focus was the English language, but the resource-poor languages have been completely ignored because of the scarcity of labeled corpus. In this study, we investigate this issue in the Urdu language. Our contribution is threefold. First, we design an annotated corpus of Urdu news articles for the fake news detection tasks. Second, we explore three individual machine learning models to detect fake news. Third, we use five ensemble learning methods to ensemble the base-predictors’ predictions to improve the fake news detection system’s overall performance. Our experiment results on two Urdu news corpora show the superiority of ensemble models over individual machine learning models. Three performance metrics balanced accuracy, the area under the curve, and mean absolute error used to find that Ensemble Selection and Vote models outperform the other machine learning and ensemble learning models.

Document representation with outlier tokens exacerbates the classification performance due to the uncertain orientation of such tokens. Most existing document representation methods in different languages including Nepali mostly ignore the strategies to filter them out from documents before learning their representations. In this article, we propose a novel document representation method based on a supervised codebook to represent the Nepali documents, where our codebook contains only semantic tokens without outliers. Our codebook is domain-specific as it is based on tokens in a given corpus that have higher similarities with the class labels in the corpus. Our method adopts a simple yet prominent representation method for each word, called probability-based word embedding. To show the efficacy of our method, we evaluate its performance in the document classification task using Support Vector Machine and validate against widely used document representation methods such as Bag of Words, Latent Dirichlet allocation, Long Short-Term Memory, Word2Vec, Bidirectional Encoder Representations from Transformers and so on, using four Nepali text datasets (we denote them shortly as A1, A2, A3 and A4). The experimental results show that our method produces state-of-the-art classification performance (77.46% accuracy on A1, 67.53% accuracy on A2, 80.54% accuracy on A3 and 89.58% accuracy on A4) compared to the widely used existing document representation methods. It yields the best classification accuracy on three datasets (A1, A2 and A3) and a comparable accuracy on the fourth dataset (A4). Furthermore, we introduce the largest Nepali document dataset (A4), called NepaliLinguistic dataset, to the linguistic community.

Online review platforms are a popular way for users to post reviews by expressing their opinions towards a product or service, as well as they are valuable for other users and companies to find out the overall opinions of customers. These reviews tend to be accompanied by a rating, where the star rating has become the most common approach for users to give their feedback in a quantitative way, generally as a likert scale of 1-5 stars. In other social media platforms like Facebook or Twitter, an automated review rating prediction system can be useful to determine the rating that a user would have given to the product or service. Existing work on review rating prediction focuses on specific domains, such as restaurants or hotels. This, however, ignores the fact that some review domains which are less frequently rated, such as dentists, lack sufficient data to build a reliable prediction model. In this paper, we experiment on 12 datasets pertaining to 12 different review domains of varying level of popularity to assess the performance of predictions across different domains. We introduce a model that leverages aspect phrase embeddings extracted from the reviews, which enables the development of both in-domain and cross-domain review rating prediction systems. Our experiments show that both of our review rating prediction systems outperform all other baselines. The cross-domain review rating prediction system is particularly significant for the least popular review domains, where leveraging training data from other domains leads to remarkable improvements in performance. The in-domain review rating prediction system is instead more suitable for popular review domains, provided that a model built from training data pertaining to the target domain is more suitable when this data is abundant., Comment: 16 pages, 1 figure

Word vector representations are a crucial part of natural language processing (NLP) and human computer interaction. In this paper, we propose a novel word vector representation, Confusion2Vec, motivated from the human speech production and perception that encodes representational ambiguity. Humans employ both acoustic similarity cues and contextual cues to decode information and we focus on a model that incorporates both sources of information. The representational ambiguity of acoustics, which manifests itself in word confusions, is often resolved by both humans and machines through contextual cues. A range of representational ambiguities can emerge in various domains further to acoustic perception, such as morphological transformations, word segmentation, paraphrasing for NLP tasks like machine translation, etc. In this work, we present a case study in application to automatic speech recognition (ASR) task, where the word representational ambiguities/confusions are related to acoustic similarity. We present several techniques to train an acoustic perceptual similarity representation ambiguity. We term this Confusion2Vec and learn on unsupervised-generated data from ASR confusion networks or lattice-like structures. Appropriate evaluations for the Confusion2Vec are formulated for gauging acoustic similarity in addition to semantic–syntactic and word similarity evaluations. The Confusion2Vec is able to model word confusions efficiently, without compromising on the semantic-syntactic word relations, thus effectively enriching the word vector space with extra task relevant ambiguity information. We provide an intuitive exploration of the two-dimensional Confusion2Vec space using principal component analysis of the embedding and relate to semantic relationships, syntactic relationships, and acoustic relationships. We show through this that the new space preserves the semantic/syntactic relationships while robustly encoding acoustic similarities. The potential of the new vector representation and its ability in the utilization of uncertainty information associated with the lattice is demonstrated through small examples relating to the task of ASR error correction.

The article presents a discriminative approach to complement the unsupervised probabilistic nature of topic modelling. The framework transforms the probabilities of the topics per document into class-dependent deep learning models that extract highly discriminatory features suitable for classification. The framework is then used for sentiment analysis with minimum feature engineering. The approach transforms the sentiment analysis problem from the word/document domain to the topics domain making it more robust to noise and incorporating complex contextual information that are not represented otherwise. A stacked denoising autoencoder (SDA) is then used to model the complex relationship among the topics per sentiment with minimum assumptions. To achieve this, a distinct topic model and SDA per sentiment polarity is built with an additional decision layer for classification. The framework is tested on a comprehensive collection of benchmark datasets that vary in sample size, class bias and classification task. A significant improvement to the state of the art is achieved without the need for a sentiment lexica or over-engineered features. A further analysis is carried out to explain the observed improvement in accuracy.

Appropriate embedding transformation of sentences can aid in downstream tasks such as NLP and emotion and behavior analysis. Such efforts evolved from word vectors which were trained in an unsupervised manner using large-scale corpora. Recent research, however, has shown that sentence embeddings trained using in-domain data or supervised techniques, often through multitask learning, perform better than unsupervised ones. Representations have also been shown to be applicable in multiple tasks, especially when training incorporates multiple information sources. In this work we aspire to combine the simplicity of using abundant unsupervised data with transfer learning by introducing an online multitask objective. We present a multitask paradigm for unsupervised learning of sentence embeddings which simultaneously addresses domain adaption. We show that embeddings generated through this process increase performance in subsequent domain-relevant tasks. We evaluate on the affective tasks of emotion recognition and behavior analysis and compare our results with state-of-the-art general-purpose supervised sentence embeddings. Our unsupervised sentence embeddings outperform the alternative universal embeddings in both identifying behaviors within couples therapy and in emotion recognition.