Xiaoyu Shen, Hui Su, Yanran Li, Wenjie Shuzi Niu, Yang Zhao, Akiko Aizawa, Guoping Long. Proceedings of the 55th Annual Meeting Association for Computational Linguistics (Volume 2: Short Papers). 2017.

We propose a novel adversarial multi-task learning scheme, aiming at actively curtailing the inter-talker feature variability while maximizing its senone discriminability so as to enhance performance of deep neural network (DNN) based ASR system. call scheme speaker-invariant training (SIT). In SIT, DNN acoustic model and speaker classifier are jointly optimized minimize (tied triphone state) classification loss, simultaneously mini-maximize loss. A senone-discriminative is learned through...

Machine translation is an important and challenging task that aims at automatically translating natural language sentences from one into another. Recently, Transformer-based neural machine (NMT) has achieved great break-throughs become a new mainstream method in both methodology applications. In this article, we conduct overview of NMT its extension to other tasks. Specifically, first introduce the framework Transformer, discuss main challenges list representative methods for each challenge....

Document-level neural machine translation has yielded attractive improvements. However, majority of existing methods roughly use all context sentences in a fixed scope. They neglect the fact that different source need sizes context. To address this problem, we propose an effective approach to select dynamic so document-level model can utilize more useful selected produce better translations. Specifically, introduce selection module is independent score each candidate sentence. Then, two...

Modulation recognition is a major task in many wireless communication systems including cognitive radio and signal reconnaissance. The diversification of modulation schemes the increased complexity channel environment put higher requirements on correct identification modulated signals. Deep learning (DL) considered as potential solution to solve these problems due superior big data processing classification capabilities. This paper proposes an efficient digital method based deep neural...

Knowledge graphs (KGs) store much structured information on various entities, many of which are not covered by the parallel sentence pairs neural machine translation (NMT). To improve quality these in this paper we propose a novel KGs enhanced NMT method. Specifically, first induce new results entities transforming source and target into unified semantic space. We then generate adequate pseudo that contain induced entity pairs. Finally, model is jointly trained original The extensive...

The attention mechanism has become the <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">de facto</i> standard component in neural sequence to tasks, such as machine translation and abstractive summarization. It dynamically determines which parts input sentence should be focused on when generating each word output sequence. Ideally, only few relevant words attended at decoding time step weight distribution sparse sharp. However, previous...

We herein present a language-model-based evaluator for deletion-based sentence compression and view this task as series of deletion-and-evaluation operations using the evaluator. More specifically, is syntactic neural language model that first built by learning structural collocation among words. Subsequently, trial-and-error deletion are conducted on source sentences via reinforcement framework to obtain best target compression. An empirical study shows proposed can effectively generate...

Many automatic evaluation metrics have been proposed to score the overall quality of a response in open-domain dialogue. Generally, is comprised various aspects, such as relevancy, specificity, and empathy, importance each aspect differs according task. For instance, specificity mandatory food-ordering dialogue task, whereas fluency preferred language-teaching system. However, existing are not designed cope with flexibility. example, BLEU fundamentally relies only on word overlapping,...

Filter bank multicarrier with offset quadrature amplitude modulation (FBMC/OQAM) is considered as a powerful supplementary waveform for future wireless communications. However, FBMC systems have the same high peak-to-average power ratio (PAPR) problem other multi-carrier systems, and PAPR reduction methods designed orthogonal frequency division multiplexing (OFDM) cannot be directly applied to due unique overlapping structure of signals. Therefore, some suppression schemes tailored been...

Shock wave focusing is considered a new type of initiation method that does not require external ignition devices. For this method, understanding the shock wave/flame interaction accurately one critical issues for revealing and triggering mechanisms during processes. In study, numerical simulations were carried out to investigate detailed flow field evolution process with kerosene as fuel. Two detonation modes found processes, which direct mode reflected collision mode, respectively. At...

One of the weaknesses Neural Machine Translation (NMT) is in handling lowfrequency and ambiguous words, which we refer as troublesome words. To address this problem, propose a novel memoryenhanced NMT method. First, investigate different strategies to define detect Then, contextual memory constructed memorize target words should be produced what situations. Finally, design hybrid model dynamically access so correctly translate The extensive experiments on Chinese-to-English English-to-German...

Xiaoyu Shen, Yang Zhao, Hui Su, Dietrich Klakow. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and 9th International Joint (EMNLP-IJCNLP). 2019.

Neural Machine Translation (NMT) has drawn much attention due to its promising translation performance recently. However, several studies indicate that NMT often generates fluent but unfaithful translations. In this paper, we propose a method alleviate problem by using phrase table as recommendation memory. The main idea is add bonus words worthy of recommendation, so can make correct predictions. Specifically, first derive prefix tree accommodate all the candidate target phrases searching...

Document Image Translation (DIT) aims to translate texts on document images from one language another. It is a multi-modal task involving cooperation of text and layout. Current approaches either handle layout translation as separate processes, risking accumulative errors, or use vanilla end-to-end encoder-decoder models capture implicitly, often suffering inadequate incorporation. We argue that favorable framework should explicitly engage layout-specific modules properly organize them...

Simultaneous Machine Translation (SiMT) generates translations while receiving streaming source inputs. This requires the SiMT model to learn a read/write policy, deciding when translate and wait for more input. Numerous linguistic studies indicate that audiences in scenarios have distinct preferences, such as accurate translations, simpler syntax, no unnecessary latency. Aligning models with these human preferences is crucial improve their performances. However, this issue still remains...

The Tanimoto distance is a widely used metric in cheminformatics that satisfies the triangle inequality. Several methods have been proposed to prove this inequality; however, we note there has no discussion regarding conditions under which equality holds, can be crucial specific contexts. In paper, provide concise and intuitive proof of inequality for and, based on proof, derive corresponding conditions.

Natural language in the maintenance data of high speed railway system is big challenge for fault diagnosis due to its unstructual feature and uncertainty semantics. In this paper, a text mining based method vehicle on-board equipment (VOBE) has been proposed, which, topic model used extract from records with arbitrary nature. addition, Bayesian network (BN) also adapt complexity VOBE. Furthermore, that fully utilizes domain expert knowledge presented derive an appropriate BN structure At...

Previous studies combining knowledge graph (KG) with neural machine translation (NMT) have two problems: i) Knowledge under-utilization: they only focus on the entities that appear in both KG and training sentence pairs, making much unable to be fully utilized. ii) Granularity mismatch: current methods utilize entity as basic granularity, while NMT utilizes sub-word different utilized NMT. To alleviate above problems, we propose a multi-task learning method sub-entity granularity....

Remaining useful life estimation of the prognostics and health management technique is a complicated difficult research question for maintenance. In this article, we consider problem modeling turbofan engine under circumstances propose kernel principal component analysis–based degradation model remaining method such aircraft engine. We first analyze output data created by thermodynamic simulation that based on analysis then distinguish qualitative quantitative relationships between key...