欢迎访问 武汉桑拿导航网!

百度自然语言智能处理技术再获国际认可

作者:admin日期:

分类:123/都市资讯/

Recently, the International Conference on International Computational Linguistics (ACL 2019), the world's top academic conference in the field of natural language processing (NLP), announced the results of this year's conference papers. According to ACL 2019 official data, the number of effective submissions for this year's conference reached 2,694, an increase of 75% compared to last year's 1,544. Among them, Baidu, a leader in domestic natural language processing, has 10 papers that have been included in the conference, showing the technological accumulation and international standards in this field. Founded in 1962, the Association for Computational Linguistics (ACL) is one of the most influential and dynamic international academic organizations in the field of natural language processing. Since its inception, the Center for Computational Linguistics has been promoting computational linguistics. And the development of research related to natural language processing and international academic exchanges. Wang Haifeng, senior vice president of Baidu, AI technology platform system (AIG) and basic technology system (TG), was the chairman of ACL in 2013. He is the first Chinese president in ACL's 50-year history and also the ACL Asia Pacific (AACL). Founding chairman, ACL Fellow. Research papers can be accepted by the ACL Academic Annual Meeting, which means that the research results have been recognized by the international academic community.

Baidu's 10 papers have covered many hotspots and frontier research directions in information extraction, machine reading comprehension, dialogue system, video semantic understanding, machine translation, etc., and proposed ARNOR framework (Attention based on attention regularization). Regularization based NOise Reduction), KT-NET model with deep integration of language representation and knowledge representation, multi-granularity cross-modal attention mechanism, and co-finger analysis method based on end-to-end deep reinforcement learning, in human-computer interaction, intelligent customer service, Video comprehension, machine translation and other scenarios have great application value. For Baidu, it is not surprising that it can achieve such results in the international academic community. Dr. Wang Haifeng is one of the international leaders in the field of natural language processing. His research in the field of natural language processing and engineering and scientific achievements have been widely recognized by international counterparts. Baidu's natural language processing team under his leadership has always maintained a leading position in the development and application of natural language processing technology. It has always been regarded as the "first echelon" of natural language processing research, cultivating and attracting many academic circles and industries at home and abroad. Talents in the world.

Baidu's natural language processing technology fully supports Baidu's business. It is the core technology of smart search, information flow, smart home and other products. At the same time, it actively opens up its core capabilities, empowers all walks of life, and helps the industry to intelligently upgrade. As a core component of Baidu's brain open platform, natural language processing related technologies have an average daily call volume of more than 100 billion, and are widely used in the Internet, finance, medical, retail, travel, service and other industries. Under the background of domestic enterprises actively embracing artificial intelligence technology, Baidu natural language processing technology is making an important contribution to China's industrial intelligence process.

Attachment: Baidu is included in the ACL 2019 paper overview 1. ARNOR: Attention Regularization based Noise Reduction for Distant Guided Relation Classification Abstract: The remote supervision automatically obtains the annotation corpus through the knowledge base, which is the key algorithm for relationship extraction. However, far supervision usually introduces a large amount of noise data, that is, the sentence does not express the relationship of automatic labeling. Furthermore, the model based on far-supervised learning is not effective and has poor interpretability, and cannot explain the indications of the relationship.

To this end, we propose the AttentionRegularization based NOise Reduction (NOT Reduction). This method uses the attention mechanism to require the model to pay attention to the indications of the relationship, and then to identify the noise data, and gradually select high-quality annotation data through the bootstrap method to improve the model effect. This method is significantly better than the best enhancement learning algorithm in terms of relationship classification and noise reduction. Application value: It has a wide application value in text information extraction. This method can significantly reduce the dependence on the annotation data, and realize low-cost knowledge base-based automatic relationship learning, which can be used in the information extraction of medical and financial industries in the future. 2.Enhancing Pre-trained Language Representations with Rich Knowledge for Machine Reading Comprehension

Abstract: Machine Reading Comprehension is a problem that lets a machine read text and then answer and read content. This technology enables machines to acquire knowledge and answer questions from text data. It is one of the key technologies for building general artificial intelligence, and has long been widely concerned by academics and industry. In the past two years, the pre-training language representation model has made breakthroughs in machine reading comprehension tasks. By pre-training a sufficiently deep network structure on massive unlabeled text data, the current state-of-the-art language representation model captures complex language phenomena, better understands the language, and answers questions. However, as everyone knows, true reading comprehension requires not only the ability of the machine to understand the language, but also the knowledge of the machine to support complex reasoning. To this end, in the paper "Enhancing Pre-trained Language Representations with Rich Knowledge for Machine Reading Comprehension", Baidu pioneered a deep fusion model of language representation and knowledge representation KT-NET, hoping to further leverage the power of language and knowledge. Improve the effect of machine reading comprehension.

The model architecture of KT-NET is shown below. Firstly, for a given reading content and structured knowledge map, the language representation model and the knowledge representation model are respectively used to encode the two, and the corresponding text representation and knowledge representation are obtained. Next, use the attention mechanism to automatically filter and integrate knowledge that is highly relevant to reading content from the knowledge map. Finally, through the double-layer self-attention matching, the deep integration of text representation and knowledge representation is realized, and the accuracy of the answer boundary prediction is improved. As of the date of publication, KT-NET is still the number one model on the ReCoRD list of common sense inference reading comprehension data sets, and has been the best single model on the SQuAD 1.1 list for a long time.

Application value: This technology can be applied to search questions and answers, smart speakers and other products to directly and accurately locate the answers of user input questions, and present them in the first significant position of the search results or present them to the users through voice announcements. 3.Know More about Each Other: Evolving Dialogue Strategy via Compound Assessment

Abstract: The existing supervised learning-based dialogue system lacks control and planning for multiple rounds of reply directions, which often leads to problems such as repetition and divergence in the dialogue, which makes the user's interactive experience bias. In this paper, we conducted a compound assessment of multiple rounds of dialogue, and based on this evaluation, we used reinforcement learning to optimize two self-playing robots to promote better control of multiple rounds of dialogue during the generation process. direction. Considering that one of the main motivations of the dialogue is to carry out effective information exchange, and for the Persona Chat question (two people talk to each other), we have designed a more complete evaluation system, including two main aspects of the information volume and consistency of the dialogue. . We use composite evaluation as reward, based on the policy gradient, to guide the optimization of the dialogue strategy between two isomorphic dialogue-generating robots. The dialogue strategy dominates the flow of conversation by controlling the choice of knowledge. A comprehensive experiment was conducted on our public dataset, and the results verified the quality of the multi-round dialogue generated by our proposed method, significantly exceeding other best practices.

Application value: It can be applied to dialogue system and intelligent customer service. 4.Proactive Human-Machine Conversation with Explicit Conversation Goal Abstract: The current human-machine dialogue is still at the primary level. Most of the machines are passive dialogues and cannot interact fully like humans. We propose an active dialogue task based on knowledge maps, allowing the machine to actively engage with users as a human. During the dialogue, the machine actively leads the dialogue process according to the knowledge map to complete the set topic (entity) transfer target in advance, and maintains the natural and fluidity of the dialogue. To this end, we manually label 30,000 groups of 270,000 sentences of active dialogue corpus in the field of film and entertainment tasks, and realized two active dialogue baseline models for generation and retrieval.

Application value: It can be applied to the dialogue skills in smart speakers. Based on this, you can develop chat skills and let the machine initiate chat based on knowledge map. 5.Multi-grained Attention with Object-level Grounding for Visual Question Answering

Abstract: Visual Q&A (VQA) is a type of cross-modal information understanding task that requires the system to understand visual picture information and answer textual questions surrounding the content of the picture. This paper proposes a multi-granularity cross-modal attention mechanism, based on the image-sentence granularity attention, proposes a more fine-grained object-level cross-modal information attention mechanism, and gives two effective fine-grained Information understanding enhancement methods. Experiments show that our method helps to identify complex images and small objects, so that the system can more accurately locate the visual information on which the text question is answered, thereby significantly improving the accuracy of VQA. Application value: It can be applied to small video content understanding projects based on multimodal information and knowledge maps. 6.Hubless Nearest Neighbor Search for Bilingual Lexicon Induction

Abstract: This basic study proposes a way to improve nearest neighbor search. This method has a very beautiful theoretical basis, which not only can significantly improve the accuracy of Bilingual Lexicon Induction, but also has a guiding significance for many tasks involving nearest neighbor search. Application value: Machine translation requires a large amount of aligned bilingual text as training data. This requirement cannot be met in some cases, such as small language texts, professional literature. Bilingual dictionary compilation can improve the accuracy of the translation system in this case.

7. STACL: Simultaneous Translation with Implicit Anticipation and Controllable Latency Abstract: Simultaneous translation is one of the most difficult problems recognized in the field of artificial intelligence, which has plagued academics and industry for decades. We present the first simultaneous translation algorithm for predictive and controllable delays in history. Since its release in October last year, it has been widely reported by major technology foreign media, including MIT Technology Review, IEEE Spectrum, Fortune Magazine and so on. The qubit summary report: "This is another new development that has made the technology foreign media so excited since the release of Baidu Deep Speech 2 in 2016."

Application value: The Baidu World Congress in November 2018 adopted this simultaneous interpretation technology, and translated all the speeches of Robin all the time, the delay was only about 3 seconds, and the previous translation technique was delayed by a whole sentence. 10 seconds or more). At the same time, there is no significant decline in the quality of translation. 8.Simultaneous Translation with Flexible Policy via Restricted Imitation Learning Abstract: This article aims to improve the quality of simultaneous translation. The STACL framework we proposed last year (ie, article 7 above) is simple and effective, but sometimes not flexible enough. Now we propose a simultaneous translation algorithm based on imitation learning. By mimicking the dynamic strategy designed in this paper, the model can flexibly decide in real time whether it needs to wait for more information to continue the translation, and then improve the translation while maintaining low latency. quality.

Application value: This technology can be used in simultaneous interpretation systems. 9.Robust Neural Machine Translation with Joint Textual and Phonetic Embedding Abstract: This article aims to improve the robustness of translation, especially the homophone noise. At the input of the translation, we add the pronunciation information corresponding to the input word by means of joint embedding. The experimental results show that the proposed method not only greatly improves the robustness of the translation system under noise, but also greatly improves the performance of the translation system under non-noise conditions. Application value: It can be used for translation, especially speech-to-speech simultaneous interpretation system. One of the main challenges of speech translation is that there are too many errors in speech recognition, and most of these errors are homophones or words with similar pronunciation. This technique can greatly reduce these noises from speech recognition.

10.End-to-end Deep Reinforcement Learning Based Coreference Resolution: Co-referencing is an integral part of the information extraction task. Recent approaches based on end-to-end deep neural networks often lack an understanding of the entire chapter by optimizing the heuristic loss function and making a series of local analytical decisions. This paper first proposes a co-finger analysis method based on end-to-end depth reinforcement learning, completes the referential detection and referential links in the same framework, and directly optimizes the evaluation index of the co-finger analysis, and has achieved good results on OntoNotes.After the wedding ceremony, Wang Hao and Feng Juan of Zhangzhou Jiahe began living together, and a son was born a year later. When Feng Juan was pregnant with a second child, the relationship between the two had a problem and eventually broke up. Feng Juan also chose to abort the second child. The two people who had been in love with each other later went to court, Wang Hao believes that since the breakup, Feng Juan should return the 90,000 yuan gift money that was originally collected. Should Feng Juan return the bride price? Recently, the Jiahe court heard the case. After cohabitation for more than a year, they broke up and passed the second child. In February 2016, Wang Hao and Feng Juan introduced their acquaintances. One year after the exchange, in the first month of 2017, although they did not meet the conditions for the marriage registration formalities, the two held a marriage ceremony, and Wang Hao’s family paid Feng Juan’s gift of more than 90,000 yuan according to custom.

After the wedding ceremony was held, Wang Hao and Feng Juan lived together. On March 16, 2018, Feng Juan gave birth to a boy. However, the differences in personality and concept between the two people became more and more prominent. Feng Juan chose to leave the place where she lived with Wang Hao, when she was pregnant with children. "Wang Hao has a relationship with other women. He is a person who is disloyal and irresponsible to feelings. He is a person who does not trust me and has no family responsibility." In Feng Juan’s view, the two sides broke up and were at fault. It is Wang Hao. Feng Juan submitted a record with Wang Hao's WeChat. "Wang Hao suspects that my child is not a parent, but refuses to pass the paternity test, refuses to sign a miscarriage, refuses to contact me by phone. Forces me to be separated from him." Feng Juan said that in addition to Wang Hao, Wang Hao’s family also gave her a very The big pressure, in addition to forcing her to leave, also kicked her out of the crowd.

The ex-boyfriend sued for the return of the bride price. The contradiction between Wang Hao and Feng Juan was not resolved with the separation of the two. Wang Hao sued the Jiahe court and said that he and Feng Juan had different personalities and did not communicate in too many languages. For this reason, Feng Juan left Wang Hao in May 2018, and the relatives of the two sides repeatedly refused to reconcile. Feng Juan was asked to return the 90,000 yuan gift that was originally paid. "The money is spent on living in a cohabitation. I also gave birth to a child for him, abortion..." Feng Juan could not accept it. She believes that the bride price can be considered at the discretion of the bridegroom before the two parties have no joint expenses, living expenses, and raising children. She and Wang Hao lived together for a long time, raising children, and also understood that the bride price is wrong, and should be recognized as common property.

"The money that has already been spent does not exist and there is no division." Feng Juan believes that the two have lived together for a long time. She gave birth to a child and also had a second child abortion. She has considerable psychological damage. "There are not too many yuan of more than 90,000 yuan, and there is still less damage to make up for me." In view of the above reasons, Feng Juan advocated asking Wang Hao to compensate her 50,000 yuan. The vigorous development of private and small and micro enterprises has provided vitality and creativity for China's economic development. Last year, under the influence of internal and external factors, private and small and micro enterprises once faced difficulties in financing. Since the beginning of this year, the state has taken measures to promote the structural reform of the financial supply side, and directed the guidance of financial and real water “precise drip irrigation” to private and small and micro enterprises, stabilize market expectations, enhance market confidence, and create appropriate supply-side structural reforms and high-quality development. Financial environment. Boosting the confidence of the market More financial living water and moisturizing private enterprises for 8 days, 1 million yuan to account. Qu Ronghua, general manager of Zhejiang Taizhou Unlimited New Energy Technology Co., Ltd. said “I didn’t expect it”: I didn’t expect the patent to be collateral. I didn’t expect Tyrone Bank’s lending to be so efficient. I didn’t expect the troubles that solved the development to be solved. “Before, due to shortage of funds, factory orders were continuously shut down. Now we can concentrate on product development that we are good at.”

This is a microcosm of the increased awareness of financing for many private and small and micro enterprises under a number of favorable policies. Since last year, in the face of corporate financing problems, from the central to the local, from the regulatory authorities to financial institutions, a series of support policies have been introduced: the precise implementation of monetary policy to inject liquidity into the market, the establishment of bailout funds to ease the urgent needs of enterprises, optimize Xiaowei Financial services supervision and assessment methods ... policy effects gradually appear in various financing channels. In the first four months of this year, the cumulative increase in social financing was 9.54 trillion yuan, 1.93 trillion yuan more than the same period of the previous year. In April, the net financing of private enterprise bonds was 15.9 billion yuan, which was the first time since this year. In the first quarter, Pratt & Whitney's small and micro loans increased by 552.9 billion yuan, an increase of 290 billion yuan over the same period last year. Under the guidance of the policy, confidence in all aspects has been boosted, and the risk appetite of banks, markets and enterprises has been reversed. More financial resources have flowed to the real economy, especially private and small and micro enterprises.

Zou Wei, deputy director of the People's Bank of China Marketing Department, said that 19 private enterprises that had difficulties in financing and could not issue bonds had successfully completed bond issuance with the support of creation tools. Later, some enterprises were able to successfully integrate funds without supporting tools. It shows that the confidence of the market and investors has improved significantly. According to a survey released by the central bank, in the first quarter, the entrepreneurial confidence index rose by 1.4 percentage points from the previous quarter. Experts said that the counter-cyclical adjustment of stable monetary policy has achieved positive results, the credit contraction risk has been alleviated, market confidence has been boosted, and social expectations have improved. This has played an important role in promoting a virtuous cycle of financial and real economy and achieving a smooth start of the national economy.

Seize the key links Breaking the shackles of banks supporting private enterprises Since May 15, the People's Bank of China has implemented a low deposit reserve ratio for small and medium-sized banks that focus on local and service counties. About 1,000 rural commercial banks serving the county can enjoy preferential policies, implement an 8% deposit reserve ratio, and release more long-term funds to support the development of private and small and micro enterprises. Reducing the liquidity constraints of small and medium-sized banks through targeted RRR reduction is only one aspect of monetary policy tools' directional regulation and “precise drip irrigation” for private and small and micro enterprises. In response to the social credit contraction that once appeared in the previous period, the People's Bank of China started from the credit supply side, the bank, and adopted market-based means to focus on the constraints of liquidity, capital and interest rates. The head of the Pratt & Whitney Financial Division of a state-owned Daxing Suzhou Branch told the reporter that this year, after adjusting the assessment of the inclusive financial and micro-enterprise loans, the coverage of the inclusive financial fairness of the inclusive finance was expanded, and the liquidity was greatly improved. The funds invested in microfinance have also increased accordingly. This year, the branch is expected to have a loan of over RMB 3 billion for small and micro enterprises, which is several times that of last year.

Since the beginning of this year, the People's Bank of China has flexibly used existing monetary policy tools and actively created new tools to improve the bank's liquidity structure and effectively enhance the credit supply capacity for small and micro enterprises and private enterprises. At the same time, in terms of alleviating capital constraints, following the successful issuance of the first single-period bonds of Bank of China in January this year, many banks such as ICBC and ABC have announced the issuance of perpetual bonds, including “100 billion yuan”. . Behind the bank's active bond issuance of blood, the central bank's bill swap tool (CBS) created by the People's Bank of China provided strong support, improved market acceptance and improved market expectations. With the shackles of supporting private small and micro enterprises to be gradually lifted, the role of “head geese” in large-scale service private enterprises and small and micro enterprises has been further highlighted. The data shows that at the end of the first quarter, the loan balance of the five large banks, Pratt & Whitney Small and Micro Enterprises, was 1.99 trillion yuan, 55.31% of the credit plan was completed at the beginning of the year; the newly issued Pratt & Whitney Small and Micro Enterprise loan interest rate was 4.76%. In the fourth quarter of last year, it fell by 0.13 percentage points.

Improve service level Promote financial supply side reform and move forward to the next step of transformation and upgrading of Derun Electronics Co., Ltd. Chairman Qiu Jianmin has some expectations: the transformation direction has been clear, and only the “dongfeng” of funds is blowing. “Enterprises are in the critical period of innovation and transformation of the industry. The new process of automotive electronics from 'laboratory' to the formation of cash flow requires a long process. In the most difficult stage of R&D and trial production, I hope to get stronger financial support.” Qiu Jianmin A lot of the voices of private entrepreneurs seeking transformation. At present, China is actively deepening the structural reform of the financial supply side and promoting the financial industry to further enhance its market structure, business philosophy, innovation capability and service level to meet the requirements of high-quality economic development.

Dong Xiwei, deputy director of the Chongyang Finance Research Institute of Renmin University of China, said that private enterprises are gradually turning to the development path of relying on technological innovation and refined services. Financial institutions need to reverse the "pawn pawn thinking" and abandon the old road of excessive dependence on land and property mortgages. More "intelligent" and "scientific" risk control measures. In the future, in the face of complex domestic and international situations, China's economy still has downward pressure, and the endogenous growth momentum needs to be further enhanced. Private and small and micro enterprises need a more suitable financing environment to cope with risks and challenges. Experts say that in the face of a private and small and micro enterprise with growth potential, a financing service system covering the entire life cycle of the enterprise should be constructed. From venture capital to bank credit, from bond markets to stock markets, private and small and micro enterprises are given more choices to raise funds. The People's Bank of China said that in the next stage, it is necessary to deepen the structural reform of the financial supply side and improve financial services by optimizing supply and enhancing competition. Appropriate use of structural monetary policy tools, play the role of "precise drip irrigation", and improve support for private enterprises and small and micro enterprises based on market-oriented principles.

According to Article 10 of the Interpretation of the Supreme People's Court on the Application of the Marriage Law of the People's Republic of China (II), if the parties request to return the bride price to be paid according to the custom, the court shall support the following if it is found to be: (1) The two parties have not gone through the formalities for marriage registration; (2) the two parties have gone through the formalities for marriage registration but have not lived together; (3) the pre-marriage payment and the difficulty in life for the payer. The application of the provisions of items (2) and (3) of the preceding paragraph shall be conditional on the divorce of both parties. The court held that the bride price was for the purpose of marriage, and the man paid the woman's large amount of cash and valuables according to local customs. If the purpose of marriage is not reached, the woman will return the bride price. Wang Hao and Feng Juan did not apply for marriage registration, so the bride price was returned. In view of the contractual agreement between the two parties, Wang Hao and Feng Juan are not satisfied with the legal conditions for registration of marriage as stipulated in the Marriage Law of China. The plaintiff knows this and should bear the unfavorable consequences of not registering marriage after partial marriage. Based on the above situation, the court considered that the bride price should be returned at 20% or 18,180 yuan. As for Feng Juan’s request for Wang Hao’s compensation of 50,000 yuan, it is not the same legal relationship with the case, and the court will not handle it.
本文标签:

相关阅读

关键词不能为空
极力推荐

聚合标签

关于我们 联系我们