We present Mockingjay as a new speech representation learning approach, where bidirectional Transformer encoders are pre-trained on a large amount of unlabeled speech. Previous speech representation methods learn through conditioning on past frames and predicting information about future frames. Whereas Mockingjay is designed to predict the current frame through jointly conditioning on both past and future contexts. The Mockingjay representation improves performance for a wide range of downstream tasks, including phoneme classification, speaker recognition, and sentiment classification on spoken content, while outperforming other approaches. Mockingjay is empirically powerful and can be fine-tuned with downstream models, with only 2 epochs we further improve performance dramatically. In a low resource setting with only 0.1% of labeled data, we outperform the result of Mel-features that uses all 100% labeled data.
Several lexica for sentiment analysis have been developed and made available in the NLP community. While most of these come with word polarity annotations (e.g. positive/negative), attempts at building lexica for finer-grained emotion analysis (e.g. happiness, sadness) have recently attracted significant attention. Such lexica are often exploited as a building block in the process of developing learning models for which emotion recognition is needed, and/or used as baselines to which compare the performance of the models. In this work, we contribute two new resources to the community: a) an extension of an existing and widely used emotion lexicon for English; and b) a novel version of the lexicon targeting Italian. Furthermore, we show how simple techniques can be used, both in supervised and unsupervised experimental settings, to boost performances on datasets and tasks of varying degree of domain-specificity.
Pooling is an essential component of a wide variety of sentence representation and embedding models. This paper explores generalized pooling methods to enhance sentence embedding. We propose vector-based multi-head attention that includes the widely used max pooling, mean pooling, and scalar self-attention as special cases. The model benefits from properly designed penalization terms to reduce redundancy in multi-head attention. We evaluate the proposed model on three different tasks: natural language inference (NLI), author profiling, and sentiment classification. The experiments show that the proposed model achieves significant improvement over strong sentence-encoding-based methods, resulting in state-of-the-art performances on four datasets. The proposed approach can be easily implemented for more problems than we discuss in this paper.
A novel approach for unsupervised domain adaptation for neural networks is proposed that relies on metric-based regularization of the learning process. The metric-based regularization aims at domain-invariant latent feature representations by means of maximizing the similarity between domain-specific activation distributions. The proposed metric results from modifying an integral probability metric such that it becomes translation-invariant on a polynomial function space. The metric has an intuitive interpretation in the dual space as the sum of differences of higher order central moments of the corresponding activation distributions. Error minimization guarantees are proven for the continuous case. As demonstrated by an analysis of standard benchmark experiments for sentiment analysis, object recognition and digit recognition, the outlined approach is robust regarding parameter changes and achieves higher classification accuracies than comparable approaches.
We are interested in estimating individual labels given only coarse, aggregated signal over the data points. In our setting, we receive sets ("bags") of unlabeled instances with constraints on label proportions. We relax the unrealistic assumption of known label proportions, made in previous work; instead, we assume only to have upper and lower bounds, and constraints on bag differences. We motivate the problem, propose an intuitive formulation and algorithm, and apply our methods to real-world scenarios. Across several domains, we show how using only proportion constraints and no labeled examples, we can achieve surprisingly high accuracy. In particular, we demonstrate how to predict income level using rough stereotypes and how to perform sentiment analysis using very little information. We also apply our method to guide exploratory analysis, recovering geographical differences in twitter dialect.
We introduce openXBOW, an open-source toolkit for the generation of bag-of-words (BoW) representations from multimodal input. In the BoW principle, word histograms were first used as features in document classification, but the idea was and can easily be adapted to, e.g., acoustic or visual low-level descriptors, introducing a prior step of vector quantisation. The openXBOW toolkit supports arbitrary numeric input features and text input and concatenates computed subbags to a final bag. It provides a variety of extensions and options. To our knowledge, openXBOW is the first publicly available toolkit for the generation of crossmodal bags-of-words. The capabilities of the tool are exemplified in two sample scenarios: time-continuous speech-based emotion recognition and sentiment analysis in tweets where improved results over other feature representation forms were observed.
SentiWordNet is an important lexical resource supporting sentiment analysis in opinion mining applications. In this paper, we propose a novel approach to construct a Vietnamese SentiWordNet (VSWN). SentiWordNet is typically generated from WordNet in which each synset has numerical scores to indicate its opinion polarities. Many previous studies obtained these scores by applying a machine learning method to WordNet. However, Vietnamese WordNet is not available unfortunately by the time of this paper. Therefore, we propose a method to construct VSWN from a Vietnamese dictionary, not from WordNet. We show the effectiveness of the proposed method by generating a VSWN with 39,561 synsets automatically. The method is experimentally tested with 266 synsets with aspect of positivity and negativity. It attains a competitive result compared with English SentiWordNet that is 0.066 and 0.052 differences for positivity and negativity sets respectively.
We present a novel summarization framework for reviews of products and services by selecting informative and concise text segments from the reviews. Our method consists of two major steps. First, we identify five frequently occurring variable-length syntactic patterns and use them to extract candidate segments. Then we use the output of a joint generative sentiment topic model to filter out the non-informative segments. We verify the proposed method with quantitative and qualitative experiments. In a quantitative study, our approach outperforms previous methods in producing informative segments and summaries that capture aspects of products and services as expressed in the user-generated pros and cons lists. Our user study with ninety users resonates with this result: individual segments extracted and filtered by our method are rated as more useful by users compared to previous approaches by users.
Sentiment analysis that classifies data into positive or negative has been dominantly used to recognize emotional aspects of texts, despite the deficit of thorough examination of emotional meanings. Recently, corpora labeled with more than just valence are built to exceed this limit. However, most Korean emotion corpora are small in the number of instances and cover a limited range of emotions. We introduce KOTE dataset. KOTE contains 50k (250k cases) Korean online comments, each of which is manually labeled for 43 emotion labels or one special label (NO EMOTION) by crowdsourcing (Ps = 3,048). The emotion taxonomy of the 43 emotions is systematically established by cluster analysis of Korean emotion concepts expressed on word embedding space. After explaining how KOTE is developed, we also discuss the results of finetuning and analysis for social discrimination in the corpus.
The quality of artificially generated texts has considerably improved with the advent of transformers. The question of using these models to generate learning data for supervised learning tasks naturally arises. In this article, this question is explored under 3 aspects: (i) are artificial data an efficient complement? (ii) can they replace the original data when those are not available or cannot be distributed for confidentiality reasons? (iii) can they improve the explainability of classifiers? Different experiments are carried out on Web-related classification tasks -- namely sentiment analysis on product reviews and Fake News detection -- using artificially generated data by fine-tuned GPT-2 models. The results show that such artificial data can be used in a certain extend but require pre-processing to significantly improve performance. We show that bag-of-word approaches benefit the most from such data augmentation.