A novel method for finding linear mappings among word embeddings for several languages, taking as pivot a shared, universal embedding space, is proposed in this paper. Previous approaches learn translation matrices between two specific languages, but this method learn translation matrices between a given language and a shared, universal space. The system was first trained on bilingual, and later on multilingual corpora as well. In the first case two different training data were applied; Dinu’s English-Italian benchmark data, and English-Italian translation pairs extracted from the PanLex database. In the second case only the PanLex database was used. The system performs on English-Italian languages with the best setting significantly better than the baseline system of Mikolov et al. [1], and it provides a comparable performance with the more sophisticated systems of Faruqui and Dyer [2] and Dinu et al. [3]. Exploiting the richness of the PanLex database, the proposed method makes it possible to learn linear mappings among an arbitrary number of languages.

Online speech enhancement is of great interest, because of its diverse areas of application. The Demucs architecture has been recently shown to achieve a high level of performance, even with small audio segments, while requiring a relatively small amount of computational resources. However, an issue that it bears, as well as most of state-of-the-art speech enhancement techniques, is that of target speech selection: it is assumed that only one speech source is present. However, in real-life scenarios, more than one speech source may be present, in it is uncertain which speech source within the fed mixture should be enhanced. Thus, it is of interest to complement speech enhancement techniques (such as Demucs-based) with a target selection scheme, so as to ensure that only the source of interest is enhanced. In this work, two target selection strategies are explored: 1) an embedding-based strategy, using a codified sample of the target speech, and 2) a location-based strategy, using a beamforming-based pre-filter to select the target that is in front of a 2-microphone array. It is shown that while both strategies improve the performance of the Demucs-based technique when one or more speech interferences are present, they both have their pros and cons. Specifically, while the beamforming-based strategy achieves overall a better performance compared to the embedding-based strategy, it is sensitive against the location variation of the target speech source which the embedding-based strategy does not suffer from.

Due to the influence of background conditions, lighting conditions, occlusion issues and the image resolution, how to extract robust person features is one of the difficulties in ReID research. Vision in Transformers (ViT) has achieved significant results in the field of computer vision. However, the existing problems still limit its application in ReID due to slow extraction of person features and difficulty in utilizing local features of people. To solve the mentioned problems, we utilize Pyramid Vision Transformer (PVT) as the backbone of feature extraction and propose a PVT-based ReID method in conjunction with other studies. Firstly, some improvements suitable for ReID are used on the PVT backbone, and we establish a basic model by using powerful methods verified on CNN-based ReID. Secondly, in an effort to further promote the robustness of the person features extracted by the PVT backbone, two new modules are designed. (1) The local feature clustering (LFC) is recommend to enhance the robustness of person features by calculating the distance between local features and global feature to select the most discrete local features and clustering them. (2) The side information embeddings (SIE) are used to encode non-visual information and send it into the network for training to reduce its impact on person features. Finally, the experiments show that PVTReID has achieved excellent results in ReID datasets and are 20% faster on average than CNN-based ReID methods.

The proliferation of deep learning in natural language processing (NLP) has led to the development and release of innovative technologies capable of understanding and generating human language with remarkable proficiency. Atinuke, a Transformer-based neural network, optimises performance across various language tasks by utilising a unique configuration. The architecture interweaves layers for processing sequential data with attention mechanisms to draw meaningful affinities between inputs and outputs. Due to the configuration of its topology and hyperparameter tuning, it can emulate human-like language by extracting features and learning complex mappings. Atinuke is modular, extensible, and integrates seamlessly with existing machine learning pipelines. Advanced matrix operations like softmax, embeddings, and multi-head attention enable nuanced handling of textual, acoustic, and visual signals. By unifying modern deep learning techniques with software design principles and mathematical theory, the system achieves state-of-the-art results on natural language tasks whilst remaining interpretable and robust.