Diffusion models have achieved impressive generative performance across diverse domains such as image, video, and scientific data generation. However, fine-tuning these models for new tasks remains challenging due to their large scale, architectural diversity, and high sensitivity to hyperparameters—particularly learning rates. In this work, we propose Wasserstein-Aware Transfer (WAT), a principled and effective fine-tuning strategy grounded in diffusion trajectory analysis and optimal transport theory. Our key insight is that the distributional discrepancies between diffusion trajectories from different datasets decrease progressively over time and converge near the noise end. Based on this observation, we introduce a class-wise matching mechanism that minimizes the Wasserstein distance between class distributions of source and target datasets. This enables alignment at the class level without modifying the standard fine-tuning pipeline. To further enhance knowledge retention, we propose a novel sampling strategy that linearly combines class-conditional outputs from both pretrained and fine-tuned models. This method is simple yet effective, requiring negligible computational overhead while preserving domain-specific and generalizable knowledge. Extensive experiments across seven diverse benchmarks demonstrate that WAT reliably enhances generation quality under distribution shifts, outperforming competitive baselines. These results underscore its robustness and affirm the potential of optimal transport as a principled basis for knowledge transfer in diffusion models.

Generating a set of text is a common challenge for many NLP applications, for example, automatically providing multiple keyphrases for a document to facilitate user reading. Existing generative models use a sequential decoder that generates a single sequence successively, and the set generation problem is converted to sequence generation via concatenating multiple text into a long text sequence. However, the elements of a set are unordered, which makes this scheme suffer from biased or conflicting training signals. In this paper, we propose a branching decoder, which can generate a dynamic number of tokens at each time-step and branch multiple generation paths. In particular, paths are generated individually so that no order dependence is required. Moreover, multiple paths can be generated in parallel which greatly reduces the inference time. Experiments on several keyphrase generation datasets demonstrate that the branching decoder is more effective and efficient than the existing sequential decoder.

Reasoning capabilities are crucial for Large Language Models~(LLMs), yet a notable gap exists between English and non-English languages. To bridge this disparity, some works fine-tune LLMs to relearn reasoning capabilities in non-English languages, while others replace non-English inputs with an external model's outputs such as English translation text to circumvent the challenge of LLM understanding non-English. Unfortunately, these methods often underutilize the built-in skilled reasoning and useful language understanding capabilities of LLMs. In order to better utilize the minds of reasoning and language understanding in LLMs, we propose a new method, namely MergeMinds, which merges LLMs with the external language understanding capabilities from multilingual models to boost the multilingual reasoning performance. Furthermore, a two-step training scheme is introduced to first train to embeded the external capabilities into LLMs and then train the collaborative utilization of the external capabilities and the built-in capabilities in LLMs. Experiments on three multilingual reasoning datasets and a language understanding dataset demonstrate that MergeMinds consistently outperforms all baselines, especially in low-resource languages. Without updating the parameters of LLMs, the average accuracy improved by 6.7 and 8.0 across all languages and low-resource languages on the MGSM dataset, respectively.

Dataset search, or more specifically, ad hoc dataset retrieval which is a trending specialized IR task, has received increasing attention in both academia and industry. While methods and systems continue evolving, existing test collections for this task exhibit shortcomings, particularly suffering from lexical bias in pooling and limited to keyword-style queries for evaluation. To address these limitations, in this paper, we construct ACORDAR 2.0, a new test collection for this task which is also the largest to date. To reduce lexical bias in pooling, we adapt dense retrieval models to large structured data, using them to find an extended set of semantically relevant datasets to be annotated. To diversify query forms, we employ a large language model to rewrite keyword queries into high-quality question-style queries. We use the test collection to evaluate popular sparse and dense retrieval models to establish a baseline for future studies. The test collection and source code are publicly available.

Many questions should be answered by not a single answer but a set of multiple answers. This emerging Multi-Span Reading Comprehension (MSRC) task requires extracting multiple non-contiguous spans from a given context to answer a question. Existing methods extend conventional single-span models to predict the positions of the start and end tokens of answer spans, or predict the beginning-inside-outside tag of each token. Such token-centric paradigms can hardly capture dependencies among span-level answers which are critical to MSRC. In this paper, we propose SpanQualifier, a span-centric scheme where spans, as opposed to tokens, are directly represented and scored to qualify as answers. Explicit span representations enable their interaction which exploits their dependencies to enhance representations. Experiments on three MSRC datasets demonstrate the effectiveness of our span-centric scheme and show that SpanQualifier achieves state-of-the-art results.

Distance computation is a bottleneck that limits the performance of many applications based on knowledge graphs (KGs). One common approach to improving online distance computation is to offline precompute certain information to be stored in an index called distance oracle. However, its effectiveness remains under-studied in the setting where edges are methodologically weighted to capture the structure and semantics of edge types in a KG. To fill the gap, in this paper, we present the first evaluation of representative distance oracles on KGs with commonly used edge weighting schemes. Our negative results and empirical justifications provide insights and a motivation for future studies of this unique setting.

Zixian Huang, Ao Wu, Jiaying Zhou, Yu Gu, Yue Zhao, Gong Cheng. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022.

A prominent application of knowledge graph (KG) is document enrichment. Existing methods identify mentions of entities in a background KG and enrich documents with entity types and direct relations. We compute an entity relation subgraph (ERG) that can more expressively represent indirect relations among a set of mentioned entities. To find compact, representative, and relevant ERGs for effective enrichment, we propose an efficient best-first search algorithm to solve a new combinatorial optimization problem that achieves a trade-off between representativeness and compactness, and then we exploit ontological knowledge to rank ERGs by entity-based document-KG and intra-KG relevance. Extensive experiments and user studies show the promising performance of our approach.

Knowledge graphs (KGs) have been extensively used to annotate text, e.g., news articles, in order to enhance its comprehension by readers. This requires to map entities occurring in the news to the target entities of the KG and to extract a so-called relationship sub-graph (RSG) that spans these entities. RSG extraction is computationally demanding and cannot scale to large KGs. Existing approximation algorithms that focus on structurally compact RSGs are not satisfactory since they often return no answers. We address this problem and develop an efficient algorithm to find approximations that connect the most salient subset of the target entities. Moreover, we propose a context-aware method to rank RSGs by their relevance to the news and their semantic cohesion. In the demo we will present our approach and the attendees will be able to experience how our system MiCRon helps to make sense of news article by computing and presenting RSGs relevant to these articles.

Zixian Huang, Yulin Shen, Xiao Li, Yu’ang Wei, Gong Cheng, Lin Zhou, Xinyu Dai, Yuzhong Qu. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.