Software/Data
Software
-
NGT（Neighborhood Graph and Tree for Indexing）
Overview
The software being provided performs high-speed searches for data near to the vectors indicated as queries from among large volumes of high dimensional vector data.
How to get
-
big3store
Overview
This is a prototype system for storing and searching large-scale knowledge graph data efficiently. The scalability of storage system and query processing system towards Peta triples is currently possible by using large-scale distribution of data into shared-nothing clusters.
How to get
-
AnnexML: Approximate Nearest Neighbor Search for Extreme Multi-Label Classification
Overview
AnnexML is a multi-label classifier designed for extremely large label space (10^4 to 10^6). The prediction of AnnexML is efficiently performed by using an approximate nearest neighbor search method.
How to get
-
yskip: Incremental Skip-gram Model with Negative Sampling
Overview
C++ implementation of incremental algorithm for learning skip-gram model with negative sampling.
Paper : Incremental Skip-gram Model with Negative Sampling(external link)
How to get
-
japanese-large-lm
Overview
A Japanese language model with 3.6 billion and 1.7 billion parameters, trained and published by LINE.
For more information, please see our blog post (only in Japanese): https://engineering.linecorp.com/ja/blog/3.6-billion-parameter-japanese-language-modelHow to get
-
japanese-large-lm-instruction-sft
Overview
A model of japanese-large-lm with improved interaction performance using the Instruction Tuning technique.
For more information, please see our blog post (only in Japanese): https://engineering.linecorp.com/ja/blog/3.6b-japanese-language-model-with-improved-dialog-performance-by-instruction-tuningHow to get
-
LINE DistilBERT Japanese
Overview
A high-performance, fast, and lightweight Japanese language model. It outperforms existing Japanese DistilBERTs such as Laboro DistilBERT and BandaiNamco DistilBERT on all JGLUE tasks, the benchmark for Japanese natural language understanding.
For more information, please see our blog post (only in Japanese): https://engineering.linecorp.com/ja/blog/line-distilbert-high-performance-fast-lightweight-japanese-language-modelHow to get
-
Meta AI Video Similarity Challenge - 3rd Place Solution
Overview
The solution code for the third place winner in the Descriptor Track / Matching Track of the Meta AI Video Similarity Challenge, a competition for video copy detection accuracy.
How to get
Data
-
VFD Dataset (Japanese)
Overview
We propose a visually-grounded first-person dialogue (VFD) dataset with verbal and non-verbal responses. The VFD dataset provides manually annotated (1) first-person images of agents, (2) utterances of human speakers, (3) eye-gaze locations of the speakers, and (4) the agents' verbal and non-verbal responses. All utterances and responses are represented in Japanese. The images with eye-gaze locations are available at GazeFollow (MIT).
How to get
-
Yahoo! Chiebukuro Data (Ver. 3)
Overview
Yahoo! Chiebukuro is the largest community-driven question answering service in Japan. It connects users with questions to those users who may have the answer, enabling people to share information and knowledge with each other. The data being provided consists of resolved questions and answers extracted from the Chiebukuro database for the period as below.
Period :April 2016 –March 2019
Number of Questions :about 2.6 million
Number of Answers :about 6.7 million
How to get
This data is available for download through the National Institute of Informatics (NII) (external link) homepage. Please refer to the NII’s Yahoo! Chiebukuro Data (Ver. 3) Usage Procedures page (external link) for details regarding applying for and using the data.
-
Yahoo! Search Query Data
Overview
The data is composed of a set of related queries to the topic queries of the 12th NTCIR (NTCIR-12) tasks. By using three different techniques, related queries were extracted from search logs of Yahoo! Search for the period as below. The data does not contain any personal information such as operation history, personal identifiers and context.
Period :July 2009 – June 2013
How to get
This data is provided to NTCIR (NII Testbeds and Community for Information access Research) (external link) Evaluation of Information Access Technologies Workshop participants, and can be used for free by research groups taking part in the workshop.
For details, please check the NTCIR (external link) web page.
※ Applications to participate in the task that will use the data provided by Yahoo! JAPAN are no longer being accepted. -
YJ Captions Dataset
Overview
We have developed a Japanese version of the MS COCO caption dataset (external link), which we call YJ Captions 26k Dataset. It is created to facilitate the development of image captioning in Japanese language. Each Japanese caption describes the specified image provided in MS COCO dataset and each image has 5 captions.
How to get
-
YJ Chat Detection Dataset
Overview
This is the chat detection dataset introduced in (Akasaki and Kaji ACL 2017) (external link).
How to get
The dataset is available for research purposes only. Please fill in Application for Use of Yahoo’s Speech Transcription Data on Chat Detection Study and send it to ml-lyresearch-data "at" lycorp.co.jp as a pdf file. Qualified applicants include academic or industrial researchers. Students can use the data, but are not qualified as applicants.
-
Japanese Visual Genome VQA Dataset
Overview
We have created a Japanese visual question answering (VQA) dataset by using Yahoo! Crowdsourcing, based on the images from the Visual Genome dataset(external link). Our dataset is meant to be comparable to the freeform QA part of Visual Genome dataset. The dataset consists of 99,208 images, together with 793,664 QA pairs in Japanese with every image having eight QA pairs.
How to get
-
Visual Scenes with Utterances Dataset
Overview
With the widespread use of intelligent systems, more and more people expect such systems to understand complicated social scenes. To facilitate development of intelligent systems, we created a mock dataset called Visual Scenes with Utterances (VSU) that contains a vast body of image variations in visual scenes with an annotated utterance and a corresponding addressee. Our dataset is based on images and annotations from the GazeFollow dataset (Recasens et al., 2015). The GazeFollow dataset consists of (1) the original image, (2) cropped speaker image with head location annotated, and (3) gaze. To create our dataset, we further annotated (4) utterances in texts, and (5) to whom an utterance is addressed. The images are available at http://gazefollow.csail.mit.edu/ .
How to get
-
Experimental Dataset for Post-Ensemble Methods
Overview
This is the dataset including 128 summarization models and their outputs used for comparing post-ensemble methods in the following paper.
Paper :Frustratingly Easy Model Ensemble for Abstractive Summarization (EMNLP 2018)
How to get
-
Yahoo! Bousai Crowd Data
Overview
This data is the "Yahoo! Bousai Crowd Data" (data representing urban dynamic analyzed from Yahoo! JAPAN disaster application) used in the following paper.
Paper :DeepCrowd: A Deep Model for Large-Scale Citywide Crowd Density and Flow Prediction (IEEE TKDE)
Period :From April 1st, 2017 to July 9th, 2017 （100days）
Area :Tokyo, Osaka
Mesh Size :about 450m grid
* score is normalized, and follows k-anonymity.How to get
The dataset is available for research purposes only.
Please fill in Application for Use of Yahoo! Bousai Crowd Data and send it to ml-lyresearch-data "at" lycorp.co.jp as a pdf file. The fields below the underline on the application form do not need to be filled in.
Qualified applicants include academic or industrial researchers.
Please ask your supervisor or other responsible teacher to apply, since students are not qualified as applicants. -
JGLUE: Japanese General Language Understanding Evaluation
Overview
This data is a Japanese language comprehension benchmark that can be used for training and evaluating models. It includes a text classification task, a sentence pair classification task, and a question answering task. JGLUE has been constructed in a joint research project between Yahoo Japan Corporation and Kawahara Lab at Waseda University.
How to get
-
YJ Covid-19 Prediction Data
Overview
This data is the “YJ Covid-19 Prediction Data” used in the following paper.
Paper：Multiwave COVID-19 Prediction from Social Awareness using Web Search and Mobility Data (KDD2022)
mobility data
Period: From February 2020 to June, 2021
Area: 23 wards of Tokyo only
search data
Period: From February 2020 to June, 2021
Query: Covid-19 Symptoms query (44 queries written in the paper)
How to get
The dataset is available for research purposes only.
Please fill in Application for Use of “YJ Covid-19 Prediction Data” and send it to ml-lyresearch-data "at" lycorp.co.jp as a pdf file.
Qualified applicants include academic or industrial researchers. Please ask your supervisor or other responsible teacher to apply, since students are not qualified as applicants.
-
LibriTTS-P
Overview
This dataset comes from the paper "LibriTTS-P: A Corpus with Speaking Style and Speaker Identity Prompts for Text-to-Speech and Style Captioning," accepted at INTERSPEECH 2024. LibriTTS-P adds prompts describing speaking styles and speaker Identity to the publicly available LibriTTS-R dataset, which has 585 hours of speech. The LibriTTS-P includes 373,868 prompts for 2,443 speakers.
How to get
-
YJ AmbigDialogue
Overview
This is the ambiguous utterance detection dataset introduced in (Akasaki and Sassano EMNLP 2024) (external link).
How to get