News classification
Contact us
- Add: No. 9, North Fourth Ring Road, Haidian District, Beijing. It mainly includes face recognition, living detection, ID card recognition, bank card recognition, business card recognition, license plate recognition, OCR recognition, and intelligent recognition technology.
- Tel: 13146317170 廖经理
- Fax:
- Email: 398017534@qq.com
Machine translation based on AI artificial intelligence method
Machine translation based on AI artificial intelligence method
The earliest machine translation was rule based machine translation (RBMT). Since it can't handle the multiple meanings of words and the diversity of sentence structure, this method is no longer used.
A few years ago, the mainstream approach to the machine translation community was the Phrased-Based Machine Translation (PBMT), which was also based on the algorithm of Google Framework. The so-called Phrased-based, that is, the smallest unit of translation is a phrase that is composed of arbitrary words (Word). It is essentially a statistical machine translation (SMT), a mapping function that learns source speech to target speech based on probability statistics rather than rules. The IBM model proposed in the 1990s is a classic translation model in statistical machine translation, and is also the root of the word-based statistical machine translation system. The IBM translation model has five statistical translation models with increasing complexity. IBM model1 is the simplest model and the basis for other models to stop calculation. IBM Model 1 only considers the probability of word-to-word translation, Model 2 introduces the probability of position change of words, and Model 3 participates in the probability of translating a word into multiple words. In the whole translation process, SMT calls various other lower-level NLP algorithms in turn, such as Chinese word segmentation, part-of-speech tagging, syntactic construction, etc., to finally generate the correct translation. In this way, the translation method like the pipeline is a loop, and there is a mistake in the middle of a link. This kind of error solution is propagated and the final result is wrong.
Deep neural networks advocate end-to-end learning, which skips the various sub-NLP steps in the middle, and uses deep network construction to directly learn the probability of fitting the source language to the target language. The latest results come from the paper: "Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation." The evaluation of this achievement is: "The results of the comprehensive accumulation of various algorithms are the integration of the natural language disposition in the past few years. The integration of Seq2Seq + Attention + Stack LSTM + Minimize Sentence Loss, the dedication of the method is not much, the experiment The dedication of experience is much more than that." It is essentially an improvement on the basis of NMT (Neural Machine Translation).
Among the many automatic evaluation methods for machine translation, the 2002-based BLEU method based on N-gram accuracy rate is the most common. The principle of the BLEU measure is to calculate the interval between the translation to be evaluated and one or more reference translations. The interval is the uniformity of n-element similarity between texts, n=1, 2, 3 (higher values do not seem to matter), that is, the number of n-ary words that are simultaneously presented in the system translation and the reference translation, and finally the matching The number of n-words obtained is divided by the number of words in the system translation, and the evaluation results are obtained. That is to say, if the 2-yuan (continuous word pair) or 3-yuan similarity of the candidate translation and the reference translation is higher, the score of the translation is higher. Generally speaking, people have a BLEU value between 50 and 70.
(Revelation: In the experience metric, the initiative considers a method and mechanism to stop the effect of the automatic evaluation model, which can be independent of the network device side, just as an automatic response)
Although SMT is now replaced by NMT, the gap is not very large from the current accuracy rate, so it still has reference value.
Analysis of thoughts
SMT
Business modeling
Based on Bayesian theory to think about the translation problem, it is to find the most probable result under a given condition. This assumption fits the practice of the speech category. For example, human beings always tell their own words according to the grammar and rules of their native language. The words and associations in each sentence constitute a probability spread. In detail, in the figure below, f denotes the source language grammar, and e denotes the target language English, the purpose is to find e with p(e) as max. P(e) is the possibility that the sentence is a legal English sentence, also called the speech model; P(f|e) is the similarity between the meaning of the translated English sentence expression and the meaning expressed by the original law sentence. Or, to express the meaning of several methods, also called the translation model. The following formula graphically depicts three questions in the entire translation process—speech model, translation model, and solution search.
A few years ago, the mainstream approach to the machine translation community was the Phrased-Based Machine Translation (PBMT), which was also based on the algorithm of Google Framework. The so-called Phrased-based, that is, the smallest unit of translation is a phrase that is composed of arbitrary words (Word). It is essentially a statistical machine translation (SMT), a mapping function that learns source speech to target speech based on probability statistics rather than rules. The IBM model proposed in the 1990s is a classic translation model in statistical machine translation, and is also the root of the word-based statistical machine translation system. The IBM translation model has five statistical translation models with increasing complexity. IBM model1 is the simplest model and the basis for other models to stop calculation. IBM Model 1 only considers the probability of word-to-word translation, Model 2 introduces the probability of position change of words, and Model 3 participates in the probability of translating a word into multiple words. In the whole translation process, SMT calls various other lower-level NLP algorithms in turn, such as Chinese word segmentation, part-of-speech tagging, syntactic construction, etc., to finally generate the correct translation. In this way, the translation method like the pipeline is a loop, and there is a mistake in the middle of a link. This kind of error solution is propagated and the final result is wrong.
Deep neural networks advocate end-to-end learning, which skips the various sub-NLP steps in the middle, and uses deep network construction to directly learn the probability of fitting the source language to the target language. The latest results come from the paper: "Google’s Neural Machine Translation System: Bridging the Gap between Human and Machine Translation." The evaluation of this achievement is: "The results of the comprehensive accumulation of various algorithms are the integration of the natural language disposition in the past few years. The integration of Seq2Seq + Attention + Stack LSTM + Minimize Sentence Loss, the dedication of the method is not much, the experiment The dedication of experience is much more than that." It is essentially an improvement on the basis of NMT (Neural Machine Translation).
Among the many automatic evaluation methods for machine translation, the 2002-based BLEU method based on N-gram accuracy rate is the most common. The principle of the BLEU measure is to calculate the interval between the translation to be evaluated and one or more reference translations. The interval is the uniformity of n-element similarity between texts, n=1, 2, 3 (higher values do not seem to matter), that is, the number of n-ary words that are simultaneously presented in the system translation and the reference translation, and finally the matching The number of n-words obtained is divided by the number of words in the system translation, and the evaluation results are obtained. That is to say, if the 2-yuan (continuous word pair) or 3-yuan similarity of the candidate translation and the reference translation is higher, the score of the translation is higher. Generally speaking, people have a BLEU value between 50 and 70.
(Revelation: In the experience metric, the initiative considers a method and mechanism to stop the effect of the automatic evaluation model, which can be independent of the network device side, just as an automatic response)
Although SMT is now replaced by NMT, the gap is not very large from the current accuracy rate, so it still has reference value.
Analysis of thoughts
SMT
Business modeling
Based on Bayesian theory to think about the translation problem, it is to find the most probable result under a given condition. This assumption fits the practice of the speech category. For example, human beings always tell their own words according to the grammar and rules of their native language. The words and associations in each sentence constitute a probability spread. In detail, in the figure below, f denotes the source language grammar, and e denotes the target language English, the purpose is to find e with p(e) as max. P(e) is the possibility that the sentence is a legal English sentence, also called the speech model; P(f|e) is the similarity between the meaning of the translated English sentence expression and the meaning expressed by the original law sentence. Or, to express the meaning of several methods, also called the translation model. The following formula graphically depicts three questions in the entire translation process—speech model, translation model, and solution search.