 |
|||||||||||
|
|
|||||||||||
|
Feature Articles: R&D and Commercialization of NTT’s Large Language Model “tsuzumi” Vol. 22, No. 8, pp. 26–30, Aug. 2024. https://doi.org/10.53829/ntr202408fa2 NTT’s LLM “tsuzumi”: Capable of Comprehending Graphical DocumentsAbstractLarge language models (LLMs) are being applied to fields such as healthcare, customer support, and office digital transformation. Information handled in such fields includes not only text but also a variety of visual content such as figures and diagrams. To develop LLMs as the core of artificial intelligence, their capabilities must be expanded so they can comprehend visual information. NTT Human Informatics Laboratories has been researching and developing NTT’s LLM called “tsuzumi.” In this article, we discuss our efforts related to tsuzumi’s visual machine reading comprehension technology for comprehending the content of a document from visual information. Keywords: tsuzumi, LLM, visual machine reading comprehension  1. Visual machine reading comprehension technology for visual document understandingBesides text, the documents we handle every day include visual elements (e.g., icons, figures, and diagrams). Such information exists in a variety of forms and layouts. Developing technology to read and comprehend real-world documents is one of the most urgent challenges in the field of artificial intelligence (AI). Many AI models, including large language models (LLMs), which have the ability to achieve general-purpose language understanding and generation, have been developed. Although AI capabilities have expanded vastly, surpassing, for example, humans’ reading ability, AI models still face the limitation of being able to understand only textual information in documents. To address this issue, NTT has proposed visual machine reading comprehension technology (Fig. 1). This technology enables comprehension of documents from visual information in a manner similar to how humans understand information.
We constructed datasets, such as VisualMRC [1] and SlideVQA [2], to make this technology possible. These datasets contain question-answering pairs on single and multiple document images, such as screenshots of webpages and presentation materials. Comprehending document images requires comprehension of not only linguistic information but also visual information such as the size and color of characters, figures and diagrams, graphs, and layout. We proposed LayoutT5 [1], a visual machine reading comprehension model that integrates two sets of inputs. It first applies object-recognition technology to extract regions in a document (titles, paragraphs, images, captions, lists, etc.) and applies text-recognition technology to extract the position and appearance information of text as additional input. We also proposed M3D [2], which understands the relationships between multiple document images. These models, which take into account visual information, perform better than models that only handle text, confirming the effectiveness of this technology inspired by human information processing. Using the knowledge we obtained from constructing the datasets and models, we participated in the Document Visual Question Answering (DocVQA) competition at the International Conference on Document Analysis and Recognition in 2021 (ICDAR 2021). This challenge tests a model’s ability to answer questions presented in infographics containing visual representations of information, data, and knowledge. Examples of questions given at this competition and their answers are shown in Fig. 2. To answer Q1 in the figure, the model must understand that the icon shown in the center right of the document represents women. To answer Q2, the model must be able to extract the numerical values from the document and calculate “40% + 39% = 79%”. Addressing these challenging questions requires a wide range of capabilities: understanding both textual content and visual information (e.g., icons, figures, and diagrams), comprehending the spatial relationships between text and visual elements, and executing arithmetic operations. For this competition, we thus proposed a model for answering infographic questions called IG-BERT [3]. We introduced a new method for learning layout relationships between text and visual objects in document images and a data-augmentation method for generating reasoning processes. IG-BERT achieved the highest performance among models of similar size while curbing the amount of pre-training data needed to 1/22 that of conventional models. It won second place out of 337 models submitted by 18 teams.
2. Issues with conventional visual machine reading comprehension technologyConventional techniques in visual machine reading comprehension struggled with handling diverse tasks, such as extracting information from invoices. Typically, achieving high performance on specific tasks required extensive training on task-specific datasets, resulting in high data creation and training costs. This approach created barriers to developing models that could effectively meet user needs across different applications. We thus sought to develop a visual machine reading comprehension model that is effective in following instructions by using an LLM, which is endowed with the general-purpose ability to understand and generate language. Our model would be able to respond to questions without training on each comprehension task. Specifically, our developmental challenge was to determine how to integrate visual information such as figures and diagrams contained in document images with text so the data can be understood by the LLM, which can only understand text information, without degrading the LLM’s reasoning capability. 3. LLM-based visual machine reading comprehension technologyTo enable an LLM to comprehend visual information such as figures and diagrams included in document images, visual information represented as a collection of pixels must be converted into a form that the LLM can easily process. Regarding visual document understanding of NTT’s LLM “tsuzumi,” we achieved the addition of visual understanding while maintaining text comprehension skills by combining an image encoder and lightweight adapter, as shown in Fig. 3 [4]. The image encoder maps a set of pixels to textual meaning. The adapter transforms the meaning so it can be processed by tsuzumi.
3.1 Image encoder capable of understanding Japanese images characteristicsAn image encoder processes the visual information of what appears in an image. We prepared an image encoder that converts images to vectors as well as a text encoder that converts text sequences to vectors. We train the image encoder so that the distance between an image vector and vector of the text representing the content of the image is close, and the distance is far when the image and text have no relationship. Therefore, visual information obtained by the image encoder can be connected to textual information. To train the image encoder, we constructed a dataset of several hundred million pairs of text and images. We collected not only general images and English captions but also images particular to Japan and their Japanese captions. The Japan-specific images contain, for example, Japanese writing and landscapes found only in Japan. By purposefully including Japanese captions during training, we enable the model to learn expressions particular to the Japanese language, such as “aoi shingo (blue traffic light)” and “makkana taiyo (bright red sun).” We are also engaged in constructing models that robustly support both English and Japanese by developing techniques that allow encoders trained on English text and images to adapt to the Japanese language [5]. 3.2 Training adapter for understanding visual documentsAn adapter has the role of translating a document representation into the “words” understood by the LLM. An LLM splits a text sequence into several strings called tokens. To input tokens into the neural network, they are converted to vectors. The vector corresponding to the token, called an embedding, is the “word” received by the LLM. The adapter communicates the image to the LLM by converting the output of the image encoder to an embedding. Because the adapter is a neural network with a small number of parameters, training is necessary. In tsuzumi’s visual machine reading comprehension, we maintain tsuzumi’s reasoning capability by fixing its parameters and those of the image encoder and training only the parameters of the adapter. An adapter suited for images is also achieved by carrying out multi-stage training. Initially, large-scale image-caption pairs are used to train tsuzumi to recognize general visual concepts such as objects, scenery, and locations. Subsequently, tsuzumi is further trained using optical character recognition (OCR) from image and textual instructions as inputs, with answers as outputs. To this end, we created the dataset InstructDoc, which is a large-scale collection of 30 publicly available visual document understanding datasets, each with diverse instructions in a unified format [4]. For example, it includes tasks such as information extraction, wherein given a question like “Tell the total amount” and a invoice image, the system provides answers such as “5000 yen.” We experimentally verified that an LLM trained on InstructDoc can achieve a high rate of success for unseen tasks (tasks not included in the training data). 3.3 Output examplesAn example of tsuzumi’s visual machine reading comprehension is shown in Fig. 4. The left side shows a question-answering task on the figure. The figure depicts the Innovative Optical and Wireless Network (IOWN) roadmap and the question “When are the chip-to-chip devices scheduled to be introduced?” is posed. Our model responds with “FY2029,” the correct answer. This demonstrates that it is capable of reading the visual structure of the diagram, in which the roadmap is divided into columns showing fiscal years. Because the training dataset contains images of various figures and diagrams, our model understands standard visual layouts, thus could answer the question. The right side of the figure shows a photo-recognition task. A photo of a cat is shown, and the instruction “Summarize this picture into a JSON format with the keys ‘salient_objects’ and ‘caption’” is given. JSON is a standard text-based format for representing structured data. Our model responded with “{"salient_objects": ["white cat", "laptop", "sofa"], "caption": "A white cat lying on a sofa, while using a laptop."}.” The model could not only output text in JSON format with the given keys, it could also extract content from the image that matches the meaning of each key. Controlling output format based on text-image understanding can be used for diverse applications such as image tagging. Therefore, the flexibility of tsuzumi’s visual machine reading comprehension connects text and image understanding to accomplish tasks that meet the needs of users.
4. Future goalsWe seek to further expand the capabilities of the current document comprehension model. We will also expand the application range of tsuzumi by connecting it with other modalities besides the visual modality, with the goal of advancing research and development and commercialization to ultimately achieve a society where humans and AI coexist. References
|
|||||||||||
