Guozheng Li, Peng Wang, Wenjun Ke. Findings of the Association for Computational Linguistics: EMNLP 2023. 2023.

The rapid proliferation of documents has made document intelligence increasingly critical across various industries. In recent years, Large Language Models (LLMs) have dramatically transformed the field of document intelligence, allowing for more advanced and accurate document processing solutions. Despite these advancements, most existing surveys have failed to focus on these breakthroughs, instead concentrating on traditional methods and earlier machine learning techniques. This survey seeks to fill that gap by offering an in-depth analysis of approximately 300 papers published between 2021 and mid-2025, thus providing a comprehensive overview of the impact of LLMs in document intelligence. The key topics explored include Retrieval-Augmented Generation (RAG), long-context processing, and fine-tuning LLMs for document comprehension. Furthermore, the survey highlights essential datasets, practical applications, current challenges, and future research directions, offering critical insights for both researchers and industry practitioners looking to advance the field.

Multimodal named entity recognition (MNER) is the task of recognizing named entities in multimodal context. Existing methods focus on utilizing co-attention mechanism to discover the relationships between multiple modalities. However, they still have two deficiencies: First, current methods fail to fuse the multimodal representations in a fine-grained way, which may bring noise of visual modalities. Second, current methods ignore bridging the semantic gap between heterogeneous modalities. To solve the above issues, we propose a novel MNER method with bottleneck fusion and contrastive learning (BFCL). Specifically, we first incorporate the transformer-based bottleneck fusion mechanism, subsequently, information between different modalities can only be exchanged through several bottleneck tokens, thus reducing the noise propagation. Then we propose two decoupled image-text contrastive losses to align the unimodal representations, making the representations of semantically similar modalities closer, while the representations of semantically different modalities farther away. Experimental results demonstrate that our method is competitive to the state-of-the-art models, and achieves 74.54% and 85.70% F1-scores on Twitter-2015 and Twitter-2017 datasets, respectively.

Knowledge distillation for knowledge graph embedding (KGE) aims to reduce the KGE model size to address the challenges of storage limitations and knowledge reasoning efficiency. However, current work still suffers from the performance drops when compressing a high-dimensional original KGE model to a low-dimensional distillation KGE model. Moreover, most work focuses on the reduction of inference time but ignores the time-consuming training process of distilling KGE models. In this paper, we propose IterDE, a novel knowledge distillation framework for KGEs. First, IterDE introduces an iterative distillation way and enables a KGE model to alternately be a student model and a teacher model during the iterative distillation process. Consequently, knowledge can be transferred in a smooth manner between high-dimensional teacher models and low-dimensional student models, while preserving good KGE performances. Furthermore, in order to optimize the training process, we consider that different optimization objects between hard label loss and soft label loss can affect the efficiency of training, and then we propose a soft-label weighting dynamic adjustment mechanism that can balance the inconsistency of optimization direction between hard and soft label loss by gradually increasing the weighting of soft label loss. Our experimental results demonstrate that IterDE achieves a new state-of-the-art distillation performance for KGEs compared to strong baselines on the link prediction task. Significantly, IterDE can reduce the training time by 50% on average. Finally, more exploratory experiments show that the soft-label weighting dynamic adjustment mechanism and more fine-grained iterations can improve distillation performance.

Knowledge graph embedding aims to embed entities and relations into low-dimensional vector spaces. Most existing methods only focus on triple facts in knowledge graphs. In addition, models based on translation or distance measurement cannot fully represent complex relations. As well-constructed prior knowledge, entity types can be employed to learn the representations of entities and relations. In this paper, we propose a novel knowledge graph embedding model named TransET, which takes advantage of entity types to learn more semantic features. More specifically, circle convolution based on the embeddings of entity and entity types is utilized to map head entity and tail entity to type-specific representations, then translation-based score function is used to learn the presentation triples. We evaluated our model on real-world datasets with two benchmark tasks of link prediction and triple classification. Experimental results demonstrate that it outperforms state-of-the-art models in most cases.