Technical Difficulties and Challenges of Natural Language Processing_Analysis of Development Status

With the development of mobile Internet technology, deep learning technology in machine learning, and the accumulation of big data corpus, natural language processing technology has undergone rapid changes. More and more technology giants have begun to see the value of this potential “big cake”, and further expand their natural language processing by recruiting, cooperating, and M&A to expand their business scope in the field of natural language processing research. The dominant position in the entire company. At the same time, there are also emerging technology companies emerging, proposing their own solutions in intelligent interaction, speech recognition, machine translation, etc., trying to divide their own territory and establish their own in the vast blue sea of ​​natural language processing. Benchmarking.

Artificial intelligence is already a familiar term for most ordinary people, and most of the understanding of natural language processing technology is still at the surface stage. By reviewing the development history of natural language processing, this paper interprets the major changes in the entire natural language processing industry in 2015, and then proposes the development bottleneck of natural language processing technology in the new era, as well as the challenges raised by natural language processing, and the natural language processing future. The direction of development.

Since artificial intelligence was first proposed at the Dartmouth meeting in 1956, letting the machine do more intellectual work has become the direction of scientists' efforts. One of the important goals is to hope that the machine can communicate with humans more naturally and efficiently. It is hoped that the machine will understand the profound language of human beings and interact in a way we are used to. The key technology to solve this problem is natural language processing.

Especially in the past 20 years, with the development of the Internet, there has been a strong demand for this technology. While this technology has been greatly developed, it is also strongly promoting the enhancement of the core capabilities of the Internet. For example, one of the basic capabilities provided by the Internet today is information retrieval. When people enter keywords in a search engine, they can get relevant information. Twenty years ago, at the beginning of the Internet, the search engine entered “kimono”. The returned results probably contain information about companies that produce and sell “shoes and clothing”. Nowadays, this kind of error has been relatively small, and one of the cores to promote the continuous improvement of its quality is to adopt the improved natural language understanding technology. The "Internet" natural language understanding has become a consensus in the development of the Internet and is constantly deepening.

In recent years, many technology giants are laying out in this regard. In 2013, Google acquired Wavii, a newsreading application developer, for more than $30 million. Wavii is good at natural language processing technology, can scan the Internet to find news, and give a summary of the sentence; Microsoft applied natural language processing technology to the smart assistant Xiao Bing, Cortana, and achieved good results, through the machine translation to make Skype The real-time translation function; natural language processing technology is one of the core technologies behind Facebook's intelligent assistant M. The person in charge of the product said, “What we do with M can let us better understand natural language processing.” At the end of last year, the natural language processing cloud platform was released, and voice synthesis products were launched very early, and there was a deep accumulation in natural language processing and speech synthesis in the Chinese field. It can be seen that as early as in the past few years, many technology giants and domestic IT vendors have been ignoring the potential broad market of natural language processing for a long time, and they are beginning to flex their muscles and prepare to transfer natural language processing technology to the company's core business. For business line transformation, new products are proposed to brew bigger moves.

2015 is a year of further development of natural language processing technology. Since the mainstream technology of natural language processing is mainly based on statistical machine learning, the performance of these technologies depends on two factors: one is the statistical model and optimization algorithm for different tasks, and the other is the corresponding large-scale corpus. In 2015, thanks to the rapid progress of deep learning algorithms and the continuous accumulation of large-scale social text data and corpus data, natural language processing technology has developed by leaps and bounds. In this year, major manufacturers are committed to solving more complex and difficult problems in the fields of speech recognition, semantic understanding, intelligent interaction, search optimization, etc., and continuously optimize and innovate the algorithms and models of the original products.

In terms of new products, 2015 brought us too many surprises. During the college entrance examination, Baidu launched a small robot, which is unique. At the end of October, the Rokid robot launched by the Rokid team also met with the public. These physical robots can not only use anthropomorphic thinking to identify semantics, try to establish deeper and continuous communication with users, but also imitate people's "thinking", first extract information from massive Internet content, and then follow the human thinking logic to information. Conduct inferential analysis and screening to get the answer. Not only that, excellent physical robots can also be deeply integrated with smart home devices, and access to music, news and other content, but also give the camera a gesture wake-up and far-field recognition.

In 2015, far-field speech recognition technology broke through the bottleneck of 5 meters, greatly improved the freedom of voice interaction, and refreshed the industry expectations again. Using the microphone array, echo cancellation and other techniques to enhance the voice of the target speaker, and suppress/eliminate noise and echo, thereby performing voice front-end processing; in the voice recognition engine, collecting and training the data processed by the microphone array to Optimize the far field effect. At present, a new 4-microphone array solution is integrated in China, which utilizes the spatial filtering characteristics of the microphone array to form a beeping beam in the direction of the target speaker to suppress noise outside the beam. Combined with a unique de-reverberation algorithm, the maximum degree is Absorbs the reflected sound to achieve the purpose of removing reverberation. Among them, Chinese speech recognition technology has also made a major breakthrough: the recognition relative error rate is reduced by more than 15% compared with the prior art, and the recognition rate of Mandarin speech recognition in Chinese quiet environment is close to 97%. Through the multi-layer unidirectional LSTM based Chinese vowel master modeling technology, the connection timing classification (CTC) training technology is successfully embedded into the traditional speech recognition modeling framework, combined with decision tree clustering and cross- Techniques such as word decoding and discrimination training have greatly improved the performance of online speech recognition products and are a framework innovation. In 2015, the "root embedding" proposal puts "the root as the smallest unit of Chinese language punishment", and the effect of the machine in dealing with Chinese word segmentation, short text classification and web page sorting is greatly improved, which can effectively promote the machine. Deep learning of the user's Chinese ideology makes the search engine more intelligent and more "understand" users. With the continuous development of big data technology and deep learning algorithms, in 2015, the data-driven natural language dialogue system also opened up new ideas for us: by submitting two matching models, Deep Match CNN and Deep Match Tree, and Neural Responding Machine. (NRM) dialogue generation model, and deep exploration of large-scale dialogue data, it is easy to build an automatic generation dialogue system, the accuracy rate is improved from 26% to 76% compared with the traditional machine translation model, and the dialogue is very natural and smooth.

In 2015, many vendors also opened up their own patents and technical patents for natural language processing and machine learning. The Facebook Artificial Intelligence Institute (FAIR) announced the open source of a set of deep learning tools, mainly for the Torch machine learning framework plug-ins, including iTorch, fbcunn, fbnn, fbcuda and fblualib. These plugins can greatly enhance the speed of deep learning and can be used in scenarios such as computer vision and natural language processing. Torch has been adopted by companies such as Google, Twitter, Intel, AMD, and NVIDIA. Google, Microsoft, and IBM have released and open their own machine learning tools, TensorFlow, DMTK, and SystemML, respectively. Google has used TensorFlow for products such as GMail (SmartReply), Search (RankBrain), Images (IncepTIon Image ClassificaTIon Model), and Translator (Character Recognition). DMTK's features and positioning are more inclined to natural language processing, such as text classification and clustering, topic recognition and sentiment analysis. SystemML is a machine learning technology that IBM has developed for more than a decade. Watson integrated many of the machine learning features of SystemML in a large number of events a few years ago. SoundHound.inc, a well-known voice recognition company, also opened its own "Houndify" platform at the end of the year, integrating industry data with various traditional industry vendors to integrate various aspects of industry data: such as cooperation with Expedia.com, integrating corpus data on hotels and flights Working with Xignite to integrate financial market corpus data, intended to build a broader recognition platform through the voice "recognize everything".  

Pet Products

Sensen Group Co., Ltd.   , https://www.sunsunglobal.com