Key information extraction (KIE) from visually rich documents, such as receipts and forms, involves a deep understanding of textual, visual, and layout feature information. Transformers fine-tuned for KIE achieve state-of-the-art performance but lack generality and portability across different domains. In contrast, vision large language models (VLLMs) offer higher flexibility and zero-shot capability but fall short with domain-specific layout relations unless performing a resource-demanding supervised fine-tuning. To reach the best compromise solution between lightweight models and VLLMs, we propose KIEPrompter, a cost-effective LLM-based KIE approach that leverages the predictions of lightweight models as external knowledge injected into VLLM prompts. By incorporating these auxiliary predictions, VLLMs are guided to attend relevant multimodal content without ad hoc training. The accuracy results achieved by KIEPrompter in three benchmark document collections are superior to those of VLLMs in both zero-shot and layout-sensitive scenarios. We compare various strategies for incorporating lightweight model predictions, ranging from coarse-grained predictions without explicit confidence scores to fine-grained per-element network logits. We also demonstrate that our approach is robust to the absence of specific classes in trained lightweight models, as the VLLMs' pre-training compensates for the limited generality of lightweight models.

Online hate speech detection has become an important issue due to the growth of online content, but resources in languages other than English are extremely limited. We introduce K-MHaS, a new multi-label dataset for hate speech detection that effectively handles Korean language patterns. The dataset consists of 109k utterances from news comments and provides a multi-label classification using 1 to 4 labels, and handles subjectivity and intersectionality. We evaluate strong baselines on K-MHaS. KR-BERT with a sub-character tokenizer outperforms others, recognizing decomposed characters in each hate speech class.

The Federal Reserve System (the Fed) plays a significant role in affecting monetary policy and financial conditions worldwide. Although it is important to analyse the Fed's communications to extract useful information, it is generally long-form and complex due to the ambiguous and esoteric nature of content. In this paper, we present FedNLP, an interpretable multi-component Natural Language Processing (NLP) system to decode Federal Reserve communications. This system is designed for end-users to explore how NLP techniques can assist their holistic understanding of the Fed's communications with NO coding. Behind the scenes, FedNLP uses multiple NLP models from traditional machine learning algorithms to deep neural network architectures in each downstream task. The demonstration shows multiple results at once including sentiment analysis, summary of the document, prediction of the Federal Funds Rate movement and visualization for interpreting the prediction model's result. Our application system and demonstration are available at https://fednlp.net.

Yu-Hsiang Lin, Chian-Yu Chen, Jean Lee, Zirui Li, Yuyan Zhang, Mengzhou Xia, Shruti Rijhwani, Junxian He, Zhisong Zhang, Xuezhe Ma, Antonios Anastasopoulos, Patrick Littell, Graham Neubig. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics. 2019.

360° videos and Head-Mounted Displays (HMDs) are getting increasingly popular. However, streaming 360° videos to HMDs is challenging. This is because only video content in viewers' Field-of-Views (FoVs) is rendered, and thus sending complete 360° videos wastes resources, including network bandwidth, storage space, and processing power. Optimizing the 360° video streaming to HMDs is, however, highly data and viewer dependent, and thus dictates real datasets. However, to our best knowledge, such datasets are not available in the literature. In this paper, we present our datasets of both content data (such as image saliency maps and motion maps derived from 360° videos) and sensor data (such as viewer head positions and orientations derived from HMD sensors). We put extra efforts to align the content and sensor data using the timestamps in the raw log files. The resulting datasets can be used by researchers, engineers, and hobbyists to either optimize existing 360° video streaming applications (like rate-distortion optimization) and novel applications (like crowd-driven camera movements). We believe that our dataset will stimulate more research activities along this exciting new research direction.

We study the problem of predicting the Field-of-Views (FoVs) of viewers watching 360° videos using commodity Head-Mounted Displays (HMDs). Existing solutions either use the viewer's current orientation to approximate the FoVs in the future, or extrapolate future FoVs using the historical orientations and dead-reckoning algorithms. In this paper, we develop fixation prediction networks that concurrently leverage sensor- and content-related features to predict the viewer fixation in the future, which is quite different from the solutions in the literature. The sensor-related features include HMD orientations, while the content-related features include image saliency maps and motion maps. We build a 360° video streaming testbed to HMDs, and recruit twenty-five viewers to watch ten 360° videos. We then train and validate two design alternatives of our proposed networks, which allows us to identify the better-performing design with the optimal parameter settings. Trace-driven simulation results show the merits of our proposed fixation prediction networks compared to the existing solutions, including: (i) lower consumed bandwidth, (ii) shorter initial buffering time, and (iii) short running time.