Recent studies have demonstrated significant improvements in selection tasks, and a considerable portion of this success is attributed to incorporating informative negative samples during training. While traditional methods for constructing hard negatives provide meaningful supervision, they depend on static samples that do not evolve during training, leading to sub-optimal performance. Dynamic hard negative sampling addresses this limitation by continuously adapting to the model’s changing state throughout training. However, the high computational demands of this method restrict its applicability to certain model architectures. To overcome these challenges, we introduce an efficient dynamic hard negative sampling (EDHNS). EDHNS enhances efficiency by pre-filtering easily discriminable negatives, thereby reducing the number of candidates the model needs to compute during training. Additionally, it excludes question-candidate pairs where the model already exhibits high confidence from loss computations, further reducing training time. These approaches maintain learning quality while minimizing computation and streamlining the training process. Extensive experiments on DSTC9, DSTC10, Ubuntu, and E-commerce benchmarks demonstrate that EDHNS significantly outperforms baseline models, proving its effectiveness in dialogue selection tasks.

Instruction tuning has emerged as a powerful technique, significantly boosting zero-shot performance on unseen tasks. While recent work has explored cross-lingual generalization by applying instruction tuning to multilingual models, previous studies have primarily focused on English, with a limited exploration of non-English tasks. For in-depth exploration of cross-lingual generalization in instruction tuning, we perform instruction tuning individually for two distinct language meta-datasets. Subsequently, we assess the performance on unseen tasks in the language different from the one used for training. To facilitate this investigation, we introduce a novel non-English meta-dataset named “KORANI” (Korean Natural Instruction), comprising 51 Korean benchmarks. Moreover, we design cross-lingual templates to mitigate discrepancies in language and instruction-format of the template between training and inference within the cross-lingual setting. Our experiments reveal consistent improvements through cross-lingual generalization in both English and Korean, outperforming baseline by average scores of 20.7% and 13.6%, respectively. Remarkably, these enhancements are comparable to those achieved by mono-lingual instruction tuning and even surpass them in some tasks. The result underscores the significance of relevant data acquisition across languages over linguistic congruence with unseen tasks during instruction tuning.

In task-oriented dialogue systems, intent classification is crucial for accurately understanding user queries and providing appropriate services. This study explores the use of intent descriptions with large language models for unseen domain intent classification. By examining the effects of description quality, quantity, and input length management, we identify practical guidelines for optimizing performance. Our experiments using FLAN-T5 3B demonstrate that 1) high-quality descriptions for both training and testing significantly improve accuracy, 2) diversity in training descriptions doesn’t greatly affect performance, and 3) off-the-shelf rankers selecting around ten intent options reduce input length without compromising performance. We emphasize that high-quality testing descriptions have a greater impact on accuracy than training descriptions. These findings provide practical guidelines for using intent descriptions with large language models to achieve effective and efficient intent classification in low-resource settings.

Despite the recent advances in dialogue state tracking (DST), the joint goal accuracy (JGA) of the existing methods on MultiWOZ 2.1 still remains merely 60%. In our preliminary error analysis, we find that beam search produces a pool of candidates that is likely to include the correct dialogue state. Motivated by this observation, we introduce a novel framework, called BREAK (Beam search and RE-rAnKing), that achieves outstanding performance on DST. BREAK performs DST in two stages: (i) generating k-best dialogue state candidates with beam search and (ii) re-ranking the candidates to select the correct dialogue state. This simple yet powerful framework shows state-of-the-art performance on all versions of MultiWOZ and M2M datasets. Most notably, we push the joint goal accuracy to 80-90% on MultiWOZ 2.1-2.4, which is an improvement of 23.6%, 26.3%, 21.7%, and 10.8% over the previous best-performing models, respectively. The data and code will be available at https://github.com/tony-won/DST-BREAK

Previous studies have constantly observed that a language model repeats itself, creating repetitions in an output sequence. To cope with the issue, stochastic decoding schemes have been the de facto approaches; the strategies add randomness in inference, hence avoiding the “self-loop”. However, the remedy comes at the cost of sacrificing output quality due to the randomness involved. In this work, we introduce a deterministic decoding scheme, local temperature beam search. This inference algorithm is an embarrassingly simple variant of beam search, yet it reduces repetition, whose level is superior to that of a sampling-based decoding algorithm, while maintaining the level of coherence as in beam search. Our idea is rooted in the concept of model calibration; we view a repetition as a casualty from overconfidence in a model. Therefore, our work mitigates the miscalibration present in the course of inference with a post-calibration approach applied in beam-specific manner. Our inference scheme is validated on text completion tasks, in which the repetition problem is seen most clearly, and is exhaustively compared with existing inference schemes.

The goal of DSTC11 track 5 is to build task-oriented dialogue systems that can effectively utilize external knowledge sources such as FAQs and reviews. This year’s challenge differs from previous ones as it includes subjective knowledge snippets and requires multiple snippets for a single turn. We propose a pipeline system for the challenge focusing on entity tracking, knowledge selection and response generation. Specifically, we devise a novel heuristic to ensemble the outputs from the rule-based method and neural model for entity tracking and knowledge selection. We also leverage metadata information in the knowledge source to handle fine-grained user queries. Our approach achieved the first place in objective evaluation and the third place in human evaluation of DSTC11 track 5.

Language Models (LMs) become outdated as the world changes; they often fail to perform tasks requiring recent factual information which was absent or different during training, a phenomenon called temporal misalignment. This is especially a challenging problem because the research community still lacks a coherent dataset for assessing the adaptability of LMs to frequently-updated knowledge corpus such as Wikipedia. To this end, we introduce TemporalWiki, a lifelong benchmark for ever-evolving LMs that utilizes the difference between consecutive snapshots of English Wikipedia and English Wikidata for training and evaluation, respectively. The benchmark hence allows researchers to periodically track an LM’s ability to retain previous knowledge and acquire updated/new knowledge at each point in time. We also find that training an LM on the diff data through continual learning methods achieves similar or better perplexity than on the entire snapshot in our benchmark with 12 times less computational cost, which verifies that factual knowledge in LMs can be safely updated with minimal training data via continual learning.

Retrieval-based dialogue systems display an outstanding performance when pre-trained language models are used, which includes bidirectional encoder representations from transformers (BERT). During the multi-turn response selection, BERT focuses on training the relationship between the context with multiple utterances and the response. However, this method of training is insufficient when considering the relations between each utterance in the context. This leads to a problem of not completely understanding the context flow that is required to select a response. To address this issue, we propose a new fine-grained post-training method that reflects the characteristics of the multi-turn dialogue. Specifically, the model learns the utterance level interactions by training every short context-response pair in a dialogue session. Furthermore, by using a new training objective, the utterance relevance classification, the model understands the semantic relevance and coherence between the dialogue utterances. Experimental results show that our model achieves new state-of-the-art with significant margins on three benchmark datasets. This suggests that the fine-grained post-training method is highly effective for the response selection task.