Automatic soccer commentary generation aims to bridge the gap between raw visual content and professional, tactical commentary.However, existing datasets tend to produce incomplete commentary that lacks semantic richness and fails to convey the full visual information present in standard video clips.To address these limitations, we propose two manually curated datasets: SN-Short, which enhances scene-level semantic descriptions, and SN-Long, which captures event continuity for context-aware commentary.Based on these, we design a commentary augmentation pipeline that transforms incomplete annotations into MatchText, a semantically complete and structurally standardized dataset.Leveraging this supervision, we introduce MatchAware, a generation model that incorporates contextual cues from previous events to produce coherent commentary aligned with the visual flow of the game.Experimental results show that proposed approach significantly outperforms existing baselines on the constructed datasets.~45: 03 ~42: 14 SN-Caption (Mkhallati et al.2023): Dwight Gayle (Crystal Palace) launches a cross from the corner, but David Ospina is alert to thwart the effort.SN-Short (ours): Dwight Gayle (Crystal Palace) launches a cross from the corner, but David Ospina is alert to thwart the effort.The cross was aimed at the far post, but the keeper stood firm and cleared the danger.SN-Long (ours): ・・・・・・ SN-Caption (Mkhallati et al.2023): Goal! Olivier Giroud (Arsenal) fires the rebound inside the right post after the ball breaks to him in the box.The score is 0:2.SN-Short (ours): Goal! Olivier Giroud (Arsenal) fires the rebound inside the right post after the ball breaks to him in the box.The score is 0:2.The away fans erupt in joy, celebrating the crucial goal just before halftime.SN-Long (ours): SN-Short + Both sides have been creating scoring opportunities, with the keepers making key saves at both ends.But it'

Knowledge graphs (KGs) have been shown to be effective in improving recommendation quality by introducing rich item properties as auxiliary information. The success of current KG-based recommender systems (RSs) lies in the capability of modeling high-quality item representations. This is achieved by identifying significant properties for items and exploring the intrinsic correlation between items on the KG. However, since current KG-based works only focus on learning user implicit knowledge from KGs through items, the limited user-item interaction behavior is still an obstacle to learning high-quality user representations. Furthermore, irrelevant connections in the KG may lead to erroneous messaging during the process of high-order graph feature learning of users and items. This could subsequently result in the inaccurate recommendation of items to users. To overcome above limitations, we propose a Review-enhanced Contrastive Learning on KGs (RCLKG) model for high-quality recommendation. We first construct a review-enhanced KG by exploring user explicit preferences in reviews with the extracted review entities. Then, we design a review-aware self-augmentation mechanism that seamlessly integrates explicit review knowledge with item-aligned KGs to discard irrelevant neighbor nodes of users and items. Furthermore, we develop a global-level graph aggregation schema with a refined constraint on the merged denoising KG to further optimize the denoising KG generation by considering high-order connections with less erroneous messaging. Finally, experimental results on the rating prediction and the click-through rate prediction (CTR) tasks with three real-word datasets demonstrate the superiority of our proposed RCLKG model in comparison with the state-of-the-art baselines. • A novel review-enhanced contrastive learning model on KGs called RCLKG is proposed. • Review knowledge is injected into the knowledge graph as explicit knowledge of users. • Review-enhanced self-augmentation mechanism filters out irrelevant nodes in the KG. • Global-level graph encoder on merged denoising KG refines denoising KG generation. • Extensive experiments demonstrate the superiority of RCLKG.

This paper proposes a method to develop a machine translation (MT) system from Myanmar Sign Language (MSL) to Myanmar Written Language (MWL) and vice versa for the deaf community. Translation of MSL is a difficult task since only a small amount of a parallel corpus between MSL and MWL is available. To address the challenge for MT of the low-resource language, transfer learning is applied. An MT model is trained first for a high-resource language pair, American Sign Language (ASL) and English, then it is used as an initial model to train an MT model between MSL and MWL. The mT5 model is used as a base MT model in this transfer learning. Additionally, a self-training technique is applied to generate synthetic translation pairs of MSL and MWL from a large monolingual MWL corpus. Furthermore, since the segmentation of a sentence is required as preprocessing of MT for the Myanmar language, several segmentation schemes are empirically compared. Results of experiments show that both transfer learning and self-training can enhance the performance of the translation between MSL and MWL compared with a baseline model fine-tuned from a small MSL-MWL parallel corpus only.

Word embeddings or word vectors have become fundamental in language processing techniques, especially deep learning approaches. Although many languages have compound words (e.g., “robot arm” and “maple leaf”), such words have not received much attention from researchers. Most research on compound word embeddings considered only two-word compounds; there has been little detailed analysis on the learning representations of arbitrary-length compound words. This paper discusses the necessity for learning-based approaches for estimating the distributed representations of compound words instead of a simple average of the representations of constituents. An evaluation of two downstream tasks confirms the effectiveness of compositional models in encoding useful information into vector spaces. The experimental results suggest that complex architectures such as long short-term memory, gated recurrent units, and transformers learn better representations for long entities, whereas simpler models such as recurrent neural networks are more applicable for downstream tasks where there are only short compounds (two or three words in length), as in the noun compound interpretation task.

A knowledge graph becomes a central data hub in the enterprise and the research communities. Nevertheless, the development of knowledge graphs is challenging due to the insufficient functionalities of knowledge graph management platforms. In this paper, we develop a knowledge graph management platform (UWKGM). This platform enables users to integrate arbitrary functionalities as RESTful API services in order to facilitate the knowledge graph development process. In the demonstration, we highlight the main features of UWKGM and its use cases on knowledge graph management tasks.

In the Thai language, tone information is necessary for Thai speech recognition systems. Previous studies show that many acoustic cues are attributed to shapes of tones. Nevertheless, most Thai tone classification studies mainly adopted F 0 values and their derivatives without considering other acoustic features. In this article, other acoustic features for Thai tone classification are investigated. In the experiment, energy values and spectral information represented by three spectral-based features including the LPC-based feature, PLP-based feature, and MFCC-based feature are applied to the HCRF-based Thai tone classification, which was reported as the best approach for Thai tone classification. The energy values provide an error rate reduction of 22.40% in the isolated word scenario, while there are slight improvements in the continuous speech scenario. On the contrary, spectral-based features greatly contribute to Thai tone classification in the continuous-speech scenario, whereas spectral-based features slightly degrade performances in the isolated-word scenario. The best achievement in the continuous-speech scenario is obtained from the PLP-based feature, which yields an error rate reduction of 13.90%. Therefore, findings in this article are that energy values and spectral-based features, especially the PLP-based feature, are the main contributors to the improvement of the performances of Thai tone classification in the isolated-word scenario and the continuous-speech scenario, respectively.