Emotion Recognition is attracting the attention of the research community due to the multiple areas where it can be applied, such as in healthcare or in road safety systems. In this paper, we propose a multimodal emotion recognition system that relies on speech and facial information. For the speech-based modality, we evaluated several transfer-learning techniques, more specifically, embedding extraction and Fine-Tuning. The best accuracy results were achieved when we fine-tuned the CNN-14 of the PANNs framework, confirming that the training was more robust when it did not start from scratch and the tasks were similar. Regarding the facial emotion recognizers, we propose a framework that consists of a pre-trained Spatial Transformer Network on saliency maps and facial images followed by a bi-LSTM with an attention mechanism. The error analysis reported that the frame-based systems could present some problems when they were used directly to solve a video-based task despite the domain adaptation, which opens a new line of research to discover new ways to correct this mismatch and take advantage of the embedded knowledge of these pre-trained models. Finally, from the combination of these two modalities with a late fusion strategy, we achieved 80.08% accuracy on the RAVDESS dataset on a subject-wise 5-CV evaluation, classifying eight emotions. The results revealed that these modalities carry relevant information to detect users’ emotional state and their combination enables improvement of system performance.

Emotion recognition is attracting the attention of the research community due to its multiple applications in different fields, such as medicine or autonomous driving. In this paper, we proposed an automatic emotion recognizer system that consisted of a speech emotion recognizer (SER) and a facial emotion recognizer (FER). For the SER, we evaluated a pre-trained xlsr-Wav2Vec2.0 transformer using two transfer-learning techniques: embedding extraction and fine-tuning. The best accuracy results were achieved when we fine-tuned the whole model by appending a multilayer perceptron on top of it, confirming that the training was more robust when it did not start from scratch and the previous knowledge of the network was similar to the task to adapt. Regarding the facial emotion recognizer, we extracted the Action Units of the videos and compared the performance between employing static models against sequential models. Results showed that sequential models beat static models by a narrow difference. Error analysis reported that the visual systems could improve with a detector of high-emotional load frames, which opened a new line of research to discover new ways to learn from videos. Finally, combining these two modalities with a late fusion strategy, we achieved 86.70% accuracy on the RAVDESS dataset on a subject-wise 5-CV evaluation, classifying eight emotions. Results demonstrated that these modalities carried relevant information to detect users’ emotional state and their combination allowed to improve the final system performance.

Spatial Transformer Networks are considered a powerful algorithm to learn the main areas of an image, but still, they could be more efficient by receiving images with embedded expert knowledge. This paper aims to improve the performance of conventional Spatial Transformers when applied to Facial Expression Recognition. Based on the Spatial Transformers’ capacity of spatial manipulation within networks, we propose different extensions to these models where effective attentional regions are captured employing facial landmarks or facial visual saliency maps. This specific attentional information is then hardcoded to guide the Spatial Transformers to learn the spatial transformations that best fit the proposed regions for better recognition results. For this study, we use two datasets: AffectNet and FER-2013. For AffectNet, we achieve a 0.35% point absolute improvement relative to the traditional Spatial Transformer, whereas for FER-2013, our solution gets an increase of 1.49% when models are fine-tuned with the Affectnet pre-trained weights.

Intent recognition is a key component of any task-oriented conversational system. The intent recognizer can be used first to classify the user’s utterance into one of several predefined classes (intents) that help to understand the user’s current goal. Then, the most adequate response can be provided accordingly. Intent recognizers also often appear as a form of joint models for performing the natural language understanding and dialog management tasks together as a single process, thus simplifying the set of problems that a conversational system must solve. This happens to be especially true for frequently asked question (FAQ) conversational systems. In this work, we first present an exploratory analysis in which different deep learning (DL) models for intent detection and classification were evaluated. In particular, we experimentally compare and analyze conventional recurrent neural networks (RNN) and state-of-the-art transformer models. Our experiments confirmed that best performance is achieved by using transformers. Specifically, best performance was achieved by fine-tuning the so-called BETO model (a Spanish pretrained bidirectional encoder representations from transformers (BERT) model from the Universidad de Chile) in our intent detection task. Then, as the main contribution of the paper, we analyze the effect of inserting unseen domain words to extend the vocabulary of the model as part of the fine-tuning or domain-adaptation process. Particularly, a very simple word frequency cut-off strategy is experimentally shown to be a suitable method for driving the vocabulary learning decisions over unseen words. The results of our analysis show that the proposed method helps to effectively extend the original vocabulary of the pretrained models. We validated our approach with a selection of the corpus acquired with the Hispabot-Covid19 system obtaining satisfactory results.

Not every visual media production is equally retained in memory. Recent studies have shown that the elements of an image, as well as their mutual semantic dependencies, provide a strong clue as to whether a video clip will be recalled on a second viewing or not. We believe that short textual descriptions encapsulate most of these relationships among the elements of a video, and thus they represent a rich yet concise source of information to tackle the problem of media memorability prediction. In this paper, we deepen the study of short captions as a means to convey in natural language the visual semantics of a video. We propose to use vector embeddings from a pretrained SBERT topic detection model with no adaptation as input features to a linear regression model, showing that, from such a representation, simpler algorithms can outperform deep visual models. Our results suggest that text descriptions expressed in natural language might be effective in embodying the visual semantics required to model video memorability.