Recent approaches to Open-domain Question Answering refer to an external knowledge base using a retriever model, optionally rerank passages with a separate reranker model and generate an answer using another reader model.Despite performing related tasks, the models have separate parameters and are weaklycoupled during training.We propose casting the retriever and the reranker as internal passage-wise attention mechanisms applied sequentially within the transformer architecture and feeding computed representations to the reader, with the hidden representations progressively refined at each stage.This allows us to use a single question answering model trained end-to-end, which is a more efficient use of model capacity and also leads to better gradient flow.We present a pre-training method to effectively train this architecture and evaluate our model on the Natural Questions and TriviaQA open datasets.For a fixed parameter budget, our model outperforms the previous state-of-the-art model by 1.0 and 0.7 exact match scores.Reading Layer Reranking Layer Retrieval Layer Encoder Query Passage 1 Passage 2 Passage 3 Passage 4 Passage n Encoder Encoder Encoder Encoder Encoder Encoder … … Passage-wise Disjoint Attention Encoder Encoder Encoder Encoder Passage-wise Joint Attention Concat Query and Passage Answer Concat All Decoder Encoder Encoder Encoder … Decoder

We propose an adapter based multi-domain Transformer based language model (LM) for Transformer ASR. The model consists of a big size common LM and small size adapters. The model can perform multi-domain adaptation with only the small size adapters and its related layers. The proposed model can reuse the full fine-tuned LM which is fine-tuned using all layers of an original model. The proposed LM can be expanded to new domains by adding about 2% of parameters for a first domain and 13% parameters for after second domain. The proposed model is also effective in reducing the model maintenance cost because it is possible to omit the costly and time-consuming common LM pre-training process. Using proposed adapter based approach, we observed that a general LM with adapter can outperform a dedicated music domain LM in terms of word error rate (WER).

Home-based monitoring of the changes on patients' state using non-invasive sensing device such as actigraph is important in the aspect of treatment, management and prevention of many chronic diseases. Although the characteristics of the internal structures (i.e., the fluctuation or the trend) inherent in the actigraphic signal can significantly represent the disease state, there has been no method to extract and analyze such information. In this paper, we have proposed a novel approach to determine the disease state based on features extracted from the structural information of patient's actigraphic data by utilizing a mathematical time-series modeling method, autoregressive (AR)-generalized conditional heteroskedascity (GARCH) models. Our approach consists of three steps: finding structural breaks as the disease state transitions, building a time-series model for each local segment, and analyzing the features from each model to classify the severity of the disease. We have applied the proposed method to the actigraphic data of the patients with depressive disorder, and the experimental results showed that features extracted by our modeling method have played an important role to discriminate disease severities.

Wireless sensor networks have been widely used for environment monitoring and event surveillance. In recent years, the skyline query, as an important operator for multiple-criteria decision making, has received much attention because of its many real-world applications. In this paper, we propose a new filter-based method, called SKYFILTER, for skyline query processing. Our method provides an enhanced efficiency by reduction of the total wireless communication between sensor nodes. In addition, the proposed method also provides a better performance than other existing methods with respect to the computational efficiency of skyline computation in sensor nodes. In this paper, we first present a new multiple filter-based technique to reduce wireless communication. Then, we provide a new merge sort-based algorithm for efficient skyline computation in sensor nodes. We show through performance experiments and algorithm analysis that the proposed method generally provides a better performance than other existing methods in terms of the wireless communication cost and computational efficiency.

With the widespread of the internet, text-based data sources have become ubiquitous and the demand of effective support for string matching queries becomes ever increasing. The relational query language SQL also supports LIKE clause over string data to handle substring matching queries. Due to popularity of such substring matching queries, there have been a lot of study on designing efficient indexes to support the LIKE clause in SQL. Among them, q-gram based indexes have been studied extensively. However, how to process substring matching queries efficiently with such indexes has received very little attention until recently. In this paper, we show that the optimal execution of intersecting posting lists of q-grams for substring matching queries should be decided judiciously. Then we present the optimal and approximate algorithms based on cost estimation for substring matching queries. Performance study confirms that our techniques improve query execution time with q-gram indexes significantly compared to the traditional algorithms.

Nowadays, due to the increased user requirements of the fast and reliable data management operation for mobile applications, major device vendors use embedded DBMS for their mobile devices such as MP3 players, mobile phones, digital cameras and PDAs. However, database logging is the major bottleneck against the fast response time. There has been a lot of work minimizing logging overhead but no single recovery method provides the best performance to a variety of database workloads. In this paper, we present a novel recovery method called adaptive logging which can switch the logging method from ARIES to shadow paging adaptively at a page level according to the update state of each page on run time. Also, we propose a log compaction method called deferred logging which removes redundant logs by deferring to create log records until the updated data page is flushed or until the transaction commits. Deferred logging is coupled with adaptive logging seamlessly so that it boosts the performance of adaptive logging by reducing the typical overhead of hybrid methods. We have implemented the proposed approaches to our embedded DBMS which was deployed to more than 10 million mobile devices and evaluated them through a real world application on a mobile device. The result shows that our approaches can reduce logging overhead significantly and consequently can improve the response time of both small update transaction and large update transaction effectively.