This paper presents a comprehensive study that investigates memorization in large language models (LLMs) from multiple perspectives. Experiments are conducted with the Pythia and LLM-jp model suites, both of which offer LLMs with over 10B parameters and full access to their pre-training corpora. Our findings include: (1) memorization is more likely to occur with larger model sizes, longer prompt lengths, and frequent texts, which aligns with findings in previous studies; (2) memorization is less likely to occur for texts not trained during the latter stages of training, even if they frequently appear in the training corpus; (3) the standard methodology for judging memorization can yield false positives, and texts that are infrequent yet flagged as memorized typically result from causes other than true memorization.
This paper introduces a new task, abstractive multi-video captioning, which focuses on abstracting multiple videos with natural language. Unlike conventional video captioning tasks generating a specific caption for a video, our task generates an abstract caption of the shared content in a video group containing multiple videos. To address our task, models must learn to understand each video in detail and have strong abstraction abilities to find commonalities among videos. We construct a benchmark dataset for abstractive multi-video captioning named AbstrActs. AbstrActs contains 13.5k video groups and corresponding abstract captions. As abstractive multi-video captioning models, we explore two approaches: end-to-end and cascade. For evaluation, we proposed a new metric, CocoA, which can evaluate the model performance based on the abstractness of the generated captions. In experiments, we report the impact of the way of combining multiple video features, the overall model architecture, and the number of input videos.
Large language models (LLMs) have achieved impressive performance on various reasoning tasks. To further improve the performance, we propose MultiTool-CoT, a novel framework that leverages chain-of-thought (CoT) prompting to incorporate multiple external tools, such as a calculator and a knowledge retriever, during the reasoning process. We apply MultiTool-CoT to the Task 2 dataset of NumGLUE, which requires both numerical reasoning and domain-specific knowledge. The experiments show that our method significantly outperforms strong baselines and achieves state-of-the-art performance.
We present KWJA, a high-performance unified Japanese text analyzer based on foundation models.KWJA supports a wide range of tasks, including typo correction, word segmentation, word normalization, morphological analysis, named entity recognition, linguistic feature tagging, dependency parsing, PAS analysis, bridging reference resolution, coreference resolution, and discourse relation analysis, making it the most versatile among existing Japanese text analyzers.KWJA solves these tasks in a multi-task manner but still achieves competitive or better performance compared to existing analyzers specialized for each task.KWJA is publicly available under the MIT license at https://github.com/ku-nlp/kwja.
Currently, most knowledge-grounded dialogue response generation models focus on reflecting given external knowledge. However, even when conveying external knowledge, humans integrate their own knowledge, experiences, and opinions with external knowledge to make their utterances engaging. In this study, we analyze such human behavior by annotating the utterances in an existing knowledge-grounded dialogue corpus. Each entity in the corpus is annotated with its information source, either derived from external knowledge (database-derived) or the speaker’s own knowledge, experiences, and opinions (speaker-derived). Our analysis shows that the presence of speaker-derived information in the utterance improves dialogue engagingness. We also confirm that responses generated by an existing model, which is trained to reflect the given knowledge, cannot include speaker-derived information in responses as often as humans do.
We propose a method to learn contextualized and generalized sentence representations using contrastive self-supervised learning. In the proposed method, a model is given a text consisting of multiple sentences. One sentence is randomly selected as a target sentence. The model is trained to maximize the similarity between the representation of the target sentence with its context and that of the masked target sentence with the same context. Simultaneously, the model minimizes the similarity between the latter representation and the representation of a random sentence with the same context. We apply our method to discourse relation analysis in English and Japanese and show that it outperforms strong baseline methods based on BERT, XLNet, and RoBERTa.
The global pandemic of COVID-19 has made the public pay close attention to related news, covering various domains, such as sanitation, treatment, and effects on education. Meanwhile, the COVID-19 condition is very different among the countries (e.g., policies and development of the epidemic), and thus citizens would be interested in news in foreign countries. We build a system for worldwide COVID-19 information aggregation containing reliable articles from 10 regions in 7 languages sorted by topics. Our reliable COVID-19 related website dataset collected through crowdsourcing ensures the quality of the articles. A neural machine translation module translates articles in other languages into Japanese and English. A BERT-based topic-classifier trained on our article-topic pair dataset helps users find their interested information efficiently by putting articles into different categories.
Typical event sequences are an important class of commonsense knowledge. Formalizing the task as the generation of a next event conditioned on a current event, previous work in event prediction employs sequence-to-sequence (seq2seq) models. However, what can happen after a given event is usually diverse, a fact that can hardly be captured by deterministic models. In this paper, we propose to incorporate a conditional variational autoencoder (CVAE) into seq2seq for its ability to represent diverse next events as a probabilistic distribution. We further extend the CVAE-based seq2seq with a reconstruction mechanism to prevent the model from concentrating on highly typical events. To facilitate fair and systematic evaluation of the diversity-aware models, we also extend existing evaluation datasets by tying each current event to multiple next events. Experiments show that the CVAE-based models drastically outperform deterministic models in terms of precision and that the reconstruction mechanism improves the recall of CVAE-based models without sacrificing precision.
Considerable effort has been devoted to building commonsense knowledge bases. However, they are not available in many languages because the construction of KBs is expensive. To bridge the gap between languages, this paper addresses the problem of projecting the knowledge in English, a resource-rich language, into other languages, where the main challenge lies in projection ambiguity. This ambiguity is partially solved by machine translation and target-side knowledge base completion, but neither of them is adequately reliable by itself. We show their combination can project English commonsense knowledge into Japanese and Chinese with high precision. Our method also achieves a top-10 accuracy of 90% on the crowdsourced English–Japanese benchmark. Furthermore, we use our method to obtain 18,747 facts of accurate Japanese commonsense within a very short period.