Kyosuke Nishida

Ryota Tanaka, Kyosuke Nishida, Kosuke Nishida et al.

Visual question answering on document images that contain textual, visual, and layout information, called document VQA, has received much attention recently. Although many datasets have been proposed for developing document VQA systems, most of the existing datasets focus on understanding the content relationships within a single image and not across multiple images. In this study, we propose a new multi-image document VQA dataset, SlideVQA, containing 2.6k+ slide decks composed of 52k+ slide images and 14.5k questions about a slide deck. SlideVQA requires complex reasoning, including single-hop, multi-hop, and numerical reasoning, and also provides annotated arithmetic expressions of numerical answers for enhancing the ability of numerical reasoning. Moreover, we developed a new end-to-end document VQA model that treats evidence selection and question answering in a unified sequence-to-sequence format. Experiments on SlideVQA show that our model outperformed existing state-of-the-art QA models, but that it still has a large gap behind human performance. We believe that our dataset will facilitate research on document VQA.

Kosuke Nishida, Naoki Yoshinaga, Kyosuke Nishida

Although named entity recognition (NER) helps us to extract domain-specific entities from text (e.g., artists in the music domain), it is costly to create a large amount of training data or a structured knowledge base to perform accurate NER in the target domain. Here, we propose self-adaptive NER, which retrieves external knowledge from unstructured text to learn the usages of entities that have not been learned well. To retrieve useful knowledge for NER, we design an effective two-stage model that retrieves unstructured knowledge using uncertain entities as queries. Our model predicts the entities in the input and then finds those of which the prediction is not confident. Then, it retrieves knowledge by using these uncertain entities as queries and concatenates the retrieved text to the original input to revise the prediction. Experiments on CrossNER datasets demonstrated that our model outperforms strong baselines by 2.35 points in F1 metric.

Kosuke Nishida, Kyosuke Nishida, Shuichi Nishioka

Humans can obtain the knowledge of novel visual concepts from language descriptions, and we thus use the few-shot image classification task to investigate whether a machine learning model can have this capability. Our proposed model, LIDE (Learning from Image and DEscription), has a text decoder to generate the descriptions and a text encoder to obtain the text representations of machine- or user-generated descriptions. We confirmed that LIDE with machine-generated descriptions outperformed baseline models. Moreover, the performance was improved further with high-quality user-generated descriptions. The generated descriptions can be viewed as the explanations of the model's predictions, and we observed that such explanations were consistent with prediction results. We also investigated why the language description improved the few-shot image classification performance by comparing the image representations and the text representations in the feature spaces.

Tsuyoshi Baba, Kosuke Nishida, Kyosuke Nishida

Automatic image editing has great demands because of its numerous applications, and the use of natural language instructions is essential to achieving flexible and intuitive editing as the user imagines. A pioneering work in text-driven image editing, StyleCLIP, finds an edit direction in the CLIP space and then edits the image by mapping the direction to the StyleGAN space. At the same time, it is difficult to tune appropriate inputs other than the original image and text instructions for image editing. In this study, we propose a method to construct the edit direction adaptively in the StyleGAN and CLIP spaces with SVM. Our model represents the edit direction as a normal vector in the CLIP space obtained by training a SVM to classify positive and negative images. The images are retrieved from a large-scale image corpus, originally used for pre-training StyleGAN, according to the CLIP similarity between the images and the text instruction. We confirmed that our model performed as well as the StyleCLIP baseline, whereas it allows simple inputs without increasing the computational time.

Sora Kadotani, Kosuke Nishida, Kyosuke Nishida

Large language models (LLMs) can generate fluent responses, but sometimes hallucinate facts. In this paper, we investigate whether LLMs can detect their own hallucinations. We formulate hallucination detection as a classification task of a sentence. We propose a framework for estimating LLMs' capability of hallucination detection and a classification method using Chain-of-Thought (CoT) to extract knowledge from their parameters. The experimental results indicated that GPT-$3.5$ Turbo with CoT detected $58.2\%$ of its own hallucinations. We concluded that LLMs with CoT can detect hallucinations if sufficient knowledge is contained in their parameters.

Kazutoshi Shinoda, Kosuke Nishida, Kyosuke Nishida

Reward models (RMs) play a central role in aligning large language models (LLMs) with human preferences. However, RMs are often sensitive to spurious features such as response length. Existing inference-time approaches for mitigating these biases typically focus exclusively on response length, resulting in performance trade-offs. In this paper, we propose causally motivated intervention for mitigating multiple types of biases in RMs at inference time. Our method first identifies neurons whose activations are strongly correlated with predefined bias attributes, and applies neuron-level intervention that suppresses these signals. We evaluate our method on RM benchmarks and observe reductions in sensitivity to spurious features across diverse bias types, without inducing performance trade-offs. Moreover, when used for preference annotation, small RMs (2B and 7B) with our method, which edits less than 2% of all the neurons in RMs, enable LLMs to improve alignment, achieving performance comparable to that of a state-of-the-art 70B RM on AlpacaEval and MT-Bench. Further analysis reveals that bias signals are primarily encoded by neurons in early layers, shedding light on the internal mechanisms of bias exploitation in RMs.

Ryota Tanaka, Taichi Iki, Taku Hasegawa et al.

We aim to develop a retrieval-augmented generation (RAG) framework that answers questions over a corpus of visually-rich documents presented in mixed modalities (e.g., charts, tables) and diverse formats (e.g., PDF, PPTX). In this paper, we introduce a new RAG framework, VDocRAG, which can directly understand varied documents and modalities in a unified image format to prevent missing information that occurs by parsing documents to obtain text. To improve the performance, we propose novel self-supervised pre-training tasks that adapt large vision-language models for retrieval by compressing visual information into dense token representations while aligning them with textual content in documents. Furthermore, we introduce OpenDocVQA, the first unified collection of open-domain document visual question answering datasets, encompassing diverse document types and formats. OpenDocVQA provides a comprehensive resource for training and evaluating retrieval and question answering models on visually-rich documents in an open-domain setting. Experiments show that VDocRAG substantially outperforms conventional text-based RAG and has strong generalization capability, highlighting the potential of an effective RAG paradigm for real-world documents.

Kazutoshi Shinoda, Nobukatsu Hojo, Kyosuke Nishida et al.

Existing Theory of Mind (ToM) benchmarks diverge from real-world scenarios in three aspects: 1) they assess a limited range of mental states such as beliefs, 2) false beliefs are not comprehensively explored, and 3) the diverse personality traits of characters are overlooked. To address these challenges, we introduce ToMATO, a new ToM benchmark formulated as multiple-choice QA over conversations. ToMATO is generated via LLM-LLM conversations featuring information asymmetry. By employing a prompting method that requires role-playing LLMs to verbalize their thoughts before each utterance, we capture both first- and second-order mental states across five categories: belief, intention, desire, emotion, and knowledge. These verbalized thoughts serve as answers to questions designed to assess the mental states of characters within conversations. Furthermore, the information asymmetry introduced by hiding thoughts from others induces the generation of false beliefs about various mental states. Assigning distinct personality traits to LLMs further diversifies both utterances and thoughts. ToMATO consists of 5.4k questions, 753 conversations, and 15 personality trait patterns. Our analysis shows that this dataset construction approach frequently generates false beliefs due to the information asymmetry between role-playing LLMs, and effectively reflects diverse personalities. We evaluate nine LLMs on ToMATO and find that even GPT-4o mini lags behind human performance, especially in understanding false beliefs, and lacks robustness to various personality traits.

Haruka Kawasaki, Ryota Tanaka, Kyosuke Nishida

Visual document understanding (VDU) is a challenging task for large vision language models (LVLMs), requiring the integration of visual perception, text recognition, and reasoning over structured layouts. Although recent LVLMs have shown progress on VDU benchmarks, their performance is typically evaluated based on generated responses, which may not necessarily reflect whether the model has actually captured the required information internally. In this paper, we investigate how information required to solve VDU tasks is represented across different layers of LLMs within LVLMs using linear probing. Our study reveals that (1) there is a clear gap between internal representations and generated responses, and (2) information required to solve the task is often encoded more linearly from intermediate layers than from the final layer. Motivated by these findings, we explore fine-tuning strategies that target intermediate layers. Experiments show that fine-tuning intermediate layers improves both linear probing accuracy and response accuracy while narrowing the gap.

Yui Oka, Taku Hasegawa, Kyosuke Nishida et al.

In the realm of large-scale language models, a significant challenge arises when extrapolating sequences beyond the maximum allowable length. This is because the model's position embedding mechanisms are limited to positions encountered during training, thus preventing effective representation of positions in longer sequences. We analyzed conventional position encoding methods for long contexts and found the following characteristics. (1) When the representation dimension is regarded as the time axis, Rotary Position Embedding (RoPE) can be interpreted as a restricted wavelet transform using Haar-like wavelets. However, because it uses only a fixed scale parameter, it does not fully exploit the advantages of wavelet transforms, which capture the fine movements of non-stationary signals using multiple scales (window sizes). This limitation could explain why RoPE performs poorly in extrapolation. (2) Previous research as well as our own analysis indicates that Attention with Linear Biases (ALiBi) functions similarly to windowed attention, using windows of varying sizes. However, it has limitations in capturing deep dependencies because it restricts the receptive field of the model. From these insights, we propose a new position representation method that captures multiple scales (i.e., window sizes) by leveraging wavelet transforms without limiting the model's attention field. Experimental results show that this new method improves the performance of the model in both short and long contexts. In particular, our method allows extrapolation of position information without limiting the model's attention field.

Daiki Chijiwa, Taku Hasegawa, Kyosuke Nishida et al.

Tokenization -- the process of decomposing a given text into a sequence of subwords called tokens -- is one of the key components in the development of language models. Particularly, auto-regressive language models generate texts token by token, i.e., by predicting the next-token distribution given the previous ones, and thus tokenization directly affects their efficiency in text generation. Since each language model has their own vocabulary as a set of possible tokens, they struggle to cooperate with each other at the level of next-token distributions such as model ensemble. In this paper, we establish a theoretical framework of lossless vocabulary reduction, which efficiently converts a given auto-regressive language model into the one with an arbitrarily small vocabulary without any loss in accuracy. As an application, we demonstrate that language models with different tokenization can cooperate with each other efficiently through their maximal common vocabulary.

Daiki Chijiwa, Taku Hasegawa, Kyosuke Nishida et al.

While foundation models have been exploited for various expert tasks through fine-tuning, any foundation model will become outdated due to its old knowledge or limited capability. Thus the underlying foundation model should be eventually replaced by new ones, which leads to repeated cost of fine-tuning these new models. Existing work addresses this problem by inference-time tuning, i.e., modifying the output probabilities from the new foundation model with the outputs from the old foundation model and its fine-tuned model, which involves an additional overhead in inference by the latter two models. In this paper, we propose a new fine-tuning principle, Portable Reward Tuning (PRT), that reduces the inference overhead by its nature, based on the reformulation of fine-tuning as the reward maximization. Specifically, instead of fine-tuning parameters of the foundation models, PRT trains the reward model explicitly through the same loss function as in fine-tuning. During inference, the reward model can be used with any foundation model (with the same set of vocabularies or labels) through the formulation of reward maximization. Experimental results, covering both vision and language models, demonstrate that the PRT-trained model can achieve comparable accuracy to the existing work of inference-time tuning, with less inference cost.

Ryota Tanaka, Taichi Iki, Kyosuke Nishida et al.

We study the problem of completing various visual document understanding (VDU) tasks, e.g., question answering and information extraction, on real-world documents through human-written instructions. To this end, we propose InstructDoc, the first large-scale collection of 30 publicly available VDU datasets, each with diverse instructions in a unified format, which covers a wide range of 12 tasks and includes open document types/formats. Furthermore, to enhance the generalization performance on VDU tasks, we design a new instruction-based document reading and understanding model, InstructDr, that connects document images, image encoders, and large language models (LLMs) through a trainable bridging module. Experiments demonstrate that InstructDr can effectively adapt to new VDU datasets, tasks, and domains via given instructions and outperforms existing multimodal LLMs and ChatGPT without specific training.

Kosuke Nishida, Kyosuke Nishida, Itsumi Saito et al.

Multi-hop QA with annotated supporting facts, which is the task of reading comprehension (RC) considering the interpretability of the answer, has been extensively studied. In this study, we define an interpretable reading comprehension (IRC) model as a pipeline model with the capability of predicting unanswerable queries. The IRC model justifies the answer prediction by establishing consistency between the predicted supporting facts and the actual rationale for interpretability. The IRC model detects unanswerable questions, instead of outputting the answer forcibly based on the insufficient information, to ensure the reliability of the answer. We also propose an end-to-end training method for the pipeline RC model. To evaluate the interpretability and the reliability, we conducted the experiments considering unanswerability in a multi-hop question for a given passage. We show that our end-to-end trainable pipeline model outperformed a non-interpretable model on our modified HotpotQA dataset. Experimental results also show that the IRC model achieves comparable results to the previous non-interpretable models in spite of the trade-off between prediction performance and interpretability.

Kosuke Nishida, Kyosuke Nishida, Sen Yoshida

Pre-trained language models (PTLMs) acquire domain-independent linguistic knowledge through pre-training with massive textual resources. Additional pre-training is effective in adapting PTLMs to domains that are not well covered by the pre-training corpora. Here, we focus on the static word embeddings of PTLMs for domain adaptation to teach PTLMs domain-specific meanings of words. We propose a novel fine-tuning process: task-adaptive pre-training with word embedding regularization (TAPTER). TAPTER runs additional pre-training by making the static word embeddings of a PTLM close to the word embeddings obtained in the target domain with fastText. TAPTER requires no additional corpus except for the training data of the downstream task. We confirmed that TAPTER improves the performance of the standard fine-tuning and the task-adaptive pre-training on BioASQ (question answering in the biomedical domain) and on SQuAD (the Wikipedia domain) when their pre-training corpora were not dominated by in-domain data.

Ryota Tanaka, Kyosuke Nishida, Sen Yoshida

Recent studies on machine reading comprehension have focused on text-level understanding but have not yet reached the level of human understanding of the visual layout and content of real-world documents. In this study, we introduce a new visual machine reading comprehension dataset, named VisualMRC, wherein given a question and a document image, a machine reads and comprehends texts in the image to answer the question in natural language. Compared with existing visual question answering (VQA) datasets that contain texts in images, VisualMRC focuses more on developing natural language understanding and generation abilities. It contains 30,000+ pairs of a question and an abstractive answer for 10,000+ document images sourced from multiple domains of webpages. We also introduce a new model that extends existing sequence-to-sequence models, pre-trained with large-scale text corpora, to take into account the visual layout and content of documents. Experiments with VisualMRC show that this model outperformed the base sequence-to-sequence models and a state-of-the-art VQA model. However, its performance is still below that of humans on most automatic evaluation metrics. The dataset will facilitate research aimed at connecting vision and language understanding.

Yuma Koizumi, Ryo Masumura, Kyosuke Nishida et al.

One of the problems with automated audio captioning (AAC) is the indeterminacy in word selection corresponding to the audio event/scene. Since one acoustic event/scene can be described with several words, it results in a combinatorial explosion of possible captions and difficulty in training. To solve this problem, we propose a Transformer-based audio-captioning model with keyword estimation called TRACKE. It simultaneously solves the word-selection indeterminacy problem with the main task of AAC while executing the sub-task of acoustic event detection/acoustic scene classification (i.e., keyword estimation). TRACKE estimates keywords, which comprise a word set corresponding to audio events/scenes in the input audio, and generates the caption while referring to the estimated keywords to reduce word-selection indeterminacy. Experimental results on a public AAC dataset indicate that TRACKE achieved state-of-the-art performance and successfully estimated both the caption and its keywords.

Itsumi Saito, Kyosuke Nishida, Kosuke Nishida et al.

Pre-trained sequence-to-sequence (seq-to-seq) models have significantly improved the accuracy of several language generation tasks, including abstractive summarization. Although the fluency of abstractive summarization has been greatly improved by fine-tuning these models, it is not clear whether they can also identify the important parts of the source text to be included in the summary. In this study, we investigated the effectiveness of combining saliency models that identify the important parts of the source text with the pre-trained seq-to-seq models through extensive experiments. We also proposed a new combination model consisting of a saliency model that extracts a token sequence from a source text and a seq-to-seq model that takes the sequence as an additional input text. Experimental results showed that most of the combination models outperformed a simple fine-tuned seq-to-seq model on both the CNN/DM and XSum datasets even if the seq-to-seq model is pre-trained on large-scale corpora. Moreover, for the CNN/DM dataset, the proposed combination model exceeded the previous best-performed model by 1.33 points on ROUGE-L.

Itsumi Saito, Kyosuke Nishida, Kosuke Nishida et al.

We propose a new length-controllable abstractive summarization model. Recent state-of-the-art abstractive summarization models based on encoder-decoder models generate only one summary per source text. However, controllable summarization, especially of the length, is an important aspect for practical applications. Previous studies on length-controllable abstractive summarization incorporate length embeddings in the decoder module for controlling the summary length. Although the length embeddings can control where to stop decoding, they do not decide which information should be included in the summary within the length constraint. Unlike the previous models, our length-controllable abstractive summarization model incorporates a word-level extractive module in the encoder-decoder model instead of length embeddings. Our model generates a summary in two steps. First, our word-level extractor extracts a sequence of important words (we call it the "prototype text") from the source text according to the word-level importance scores and the length constraint. Second, the prototype text is used as additional input to the encoder-decoder model, which generates a summary by jointly encoding and copying words from both the prototype text and source text. Since the prototype text is a guide to both the content and length of the summary, our model can generate an informative and length-controlled summary. Experiments with the CNN/Daily Mail dataset and the NEWSROOM dataset show that our model outperformed previous models in length-controlled settings.

Kosuke Nishida, Kyosuke Nishida, Itsumi Saito et al.

This study tackles unsupervised domain adaptation of reading comprehension (UDARC). Reading comprehension (RC) is a task to learn the capability for question answering with textual sources. State-of-the-art models on RC still do not have general linguistic intelligence; i.e., their accuracy worsens for out-domain datasets that are not used in the training. We hypothesize that this discrepancy is caused by a lack of the language modeling (LM) capability for the out-domain. The UDARC task allows models to use supervised RC training data in the source domain and only unlabeled passages in the target domain. To solve the UDARC problem, we provide two domain adaptation models. The first one learns the out-domain LM and in-domain RC task sequentially. The second one is the proposed model that uses a multi-task learning approach of LM and RC. The models can retain both the RC capability acquired from the supervised data in the source domain and the LM capability from the unlabeled data in the target domain. We evaluated the models on UDARC with five datasets in different domains. The models outperformed the model without domain adaptation. In particular, the proposed model yielded an improvement of 4.3/4.2 points in EM/F1 in an unseen biomedical domain.

Yasuhito Ohsugi, Itsumi Saito, Kyosuke Nishida et al.

Conversational machine comprehension (CMC) requires understanding the context of multi-turn dialogue. Using BERT, a pre-training language model, has been successful for single-turn machine comprehension, while modeling multiple turns of question answering with BERT has not been established because BERT has a limit on the number and the length of input sequences. In this paper, we propose a simple but effective method with BERT for CMC. Our method uses BERT to encode a paragraph independently conditioned with each question and each answer in a multi-turn context. Then, the method predicts an answer on the basis of the paragraph representations encoded with BERT. The experiments with representative CMC datasets, QuAC and CoQA, show that our method outperformed recently published methods (+0.8 F1 on QuAC and +2.1 F1 on CoQA). In addition, we conducted a detailed analysis of the effects of the number and types of dialogue history on the accuracy of CMC, and we found that the gold answer history, which may not be given in an actual conversation, contributed to the model performance most on both datasets.

Kosuke Nishida, Kyosuke Nishida, Masaaki Nagata et al.

Question answering (QA) using textual sources for purposes such as reading comprehension (RC) has attracted much attention. This study focuses on the task of explainable multi-hop QA, which requires the system to return the answer with evidence sentences by reasoning and gathering disjoint pieces of the reference texts. It proposes the Query Focused Extractor (QFE) model for evidence extraction and uses multi-task learning with the QA model. QFE is inspired by extractive summarization models; compared with the existing method, which extracts each evidence sentence independently, it sequentially extracts evidence sentences by using an RNN with an attention mechanism on the question sentence. It enables QFE to consider the dependency among the evidence sentences and cover important information in the question sentence. Experimental results show that QFE with a simple RC baseline model achieves a state-of-the-art evidence extraction score on HotpotQA. Although designed for RC, it also achieves a state-of-the-art evidence extraction score on FEVER, which is a recognizing textual entailment task on a large textual database.

Jun Suzuki, Yoshihiko Suhara, Hiroyuki Toda et al.

Knowledge discovery from GPS trajectory data is an important topic in several scientific areas, including data mining, human behavior analysis, and user modeling. This paper proposes a task that assigns personalized visited-POIs. Its goal is to estimate fine-grained and pre-defined locations (i.e., points of interest (POI)) that are actually visited by users and assign visited-location information to the corresponding span of their (personal) GPS trajectories. We also introduce a novel algorithm to solve this assignment task. First, we exhaustively extract stay-points as candidates for significant locations using a variant of a conventional stay-point extraction method. Then we select significant locations and simultaneously assign visited-POIs to them by considering various aspects, which we formulate in integer linear programming. Experimental results conducted on an actual user dataset show that our method achieves higher accuracy in the visited-POI assignment task than the various cascaded procedures of conventional methods.

Kyosuke Nishida, Itsumi Saito, Kosuke Nishida et al.

This study tackles generative reading comprehension (RC), which consists of answering questions based on textual evidence and natural language generation (NLG). We propose a multi-style abstractive summarization model for question answering, called Masque. The proposed model has two key characteristics. First, unlike most studies on RC that have focused on extracting an answer span from the provided passages, our model instead focuses on generating a summary from the question and multiple passages. This serves to cover various answer styles required for real-world applications. Second, whereas previous studies built a specific model for each answer style because of the difficulty of acquiring one general model, our approach learns multi-style answers within a model to improve the NLG capability for all styles involved. This also enables our model to give an answer in the target style. Experiments show that our model achieves state-of-the-art performance on the Q&A task and the Q&A + NLG task of MS MARCO 2.1 and the summary task of NarrativeQA. We observe that the transfer of the style-independent NLG capability to the target style is the key to its success.

Kyosuke Nishida, Itsumi Saito, Atsushi Otsuka et al.

This study considers the task of machine reading at scale (MRS) wherein, given a question, a system first performs the information retrieval (IR) task of finding relevant passages in a knowledge source and then carries out the reading comprehension (RC) task of extracting an answer span from the passages. Previous MRS studies, in which the IR component was trained without considering answer spans, struggled to accurately find a small number of relevant passages from a large set of passages. In this paper, we propose a simple and effective approach that incorporates the IR and RC tasks by using supervised multi-task learning in order that the IR component can be trained by considering answer spans. Experimental results on the standard benchmark, answering SQuAD questions using the full Wikipedia as the knowledge source, showed that our model achieved state-of-the-art performance. Moreover, we thoroughly evaluated the individual contributions of our model components with our new Japanese dataset and SQuAD. The results showed significant improvements in the IR task and provided a new perspective on IR for RC: it is effective to teach which part of the passage answers the question rather than to give only a relevance score to the whole passage.