Exclusive interview with ByteDance Wang Mingxuan: Machine translation and manual translation are essentially two tracks | T Frontline

The advancement of technology often means that the evolution of the industry has found a new direction. The translation industry is no exception. As the process of globalization continues to accelerate, people cannot do without cross-language communication when conducting foreign-related activities. The emergence of machine translation has greatly expanded the application scenarios of translation. Although it is far from perfect, it has taken a solid step towards mankind's challenge to the Tower of Babel. 51CTO specially invited Wang Mingxuan, head of machine translation at ByteDance AI Lab, to talk about the development of machine translation over the years.

From rule-based, to statistical model-based, to neural network-based

The development of machine translation is closely related to the development of computer technology, information theory, linguistics and other disciplines. After entering the 21st century, with the improvement of hardware capabilities and optimization of algorithms, machine translation technology has ushered in an unprecedented leap forward, and has successfully stepped out of the ivory tower and embarked on the road to inclusiveness.

51CTO: Throughout history, what important development nodes has machine translation experienced?

Wang Mingxuan: Machine translation is essentially a very old problem. The history of machine translation can be traced back to the "universal" proposed by philosophers such as Descartes and Leibniz in the 17th century. language". After the official birth of computers in 1946, people expected computers to be able to translate one language into another. American scientist Warren Weaver formally defined the concept and ideas of machine translation in "Translation Memo". During this period, in the context of the Cold War, the United States and the Soviet Union also invested a lot of money in machine translation-related research based on the need to collect information.

At the beginning, everyone was relatively optimistic, thinking that this matter would be settled soon. The first version of the translation system was very simple, mainly based on dictionaries, such as translating "sun" into "sun". However, this word-to-word translation quickly encountered a bottleneck because there were many polysemy phenomena, such as "Bank" can be either "bank" or "river bank". In specific contexts, you will face many word selection dilemmas. Combining the semantic rules formulated by linguists can resolve some ambiguities, but in the later stages of development, the more rules there are, the more conflicts there will be, and the system will become more and more complex, but the problem still cannot be solved.

In 1966, the United States published the report "Language and Machine", which comprehensively denied the feasibility of machine translation and recommended that financial support for machine translation projects be stopped. Affected by this, machine translation has fallen into a low ebb.

Until the 1990s, IBM proposed a translation model based on word alignment, marking the birth of modern statistical machine translation methods. The principle of machine translation based on statistics is very simple. For example, if you want to determine whether bank should be translated into "bank" or "river bank" in the context, then conduct a large amount of relevant corpus statistics, and you will find that there is something related to "money" in the context, then It is more likely to be translated as "bank". If the context mentions "river", it is more likely to correspond to "river bank". In this way, does not use dictionaries and grammatical rules, but judges the semantics in specific scenarios based on probability. This is an epoch-making change, and the quality of machine translation has been greatly improved. Soon, machine translation began to be implemented in many practical scenarios.

The period from 1993 to 2014 basically belonged to the era of statistics. However, although it was based on statistics, it still required manual work to define many features and templates, and then further design details. Therefore, it was not very flexible, and the power of the model was not very powerful.

Then in the neural network era, neural network translation mainly includes encoders and decoders from a model perspective. The encoder represents the source language into a high-dimensional vector after a series of neural network transformations, and the decoder is responsible for re-decoding this high-dimensional vector into the target language. With the introduction of Seq2Seq in 2014, neural network translation gradually began to do better than statistical machine translation.

By 2017, Google proposed Transformer, which has a larger model, more flexible structure, and higher degree of parallelization, which further improved the quality of translation. In the same year, AlphaGo’s victory also made everyone more confident in artificial intelligence. It was also after 2017 that the industrialization of machine translation ushered in an explosive period. Until now, the overall framework has not changed much, but there have been many innovations in small details.

Challenging the "Tower of Babel"

From dictionary matching, to rule translation combined with the knowledge of linguistic experts, to corpus-based statistical machine translation, and the current mainstream neural network machine translation, Compared with before, the quality of machine translation has improved dramatically, but it still faces many challenges.

51CTO: What are the main challenges currently facing machine translation?

Wang Mingxuan: There are actually many challenges.

First, How to do machine translation of scarce languages. This is a problem that machine translation has faced since its inception. The smaller the language, the smaller the amount of data, and the scarcity of corpus will be a long-term challenge.

Second, How to do multi-modal machine translation. In recent years, we often need to do voice translation and video translation. In fact, this kind of translation requires AI to do some pre-processing before translation. If AI processing is wrong, translation errors may occur. For another example, in the scenario of simultaneous interpretation, it is usually done while speaking, and complete contextual information cannot be obtained. This is a common problem in multimodal translation.

Third, The most essential problem is that the current machine translation is still based on data-driven and has not gone deeper in understanding. Model learning still relies on the contribution of language rather than true understanding of semantics. This greatly limits the upper limit of machine translation.

51CTO: As a machine translation brand under Bytedance, how does Huoshan Translation deal with the problem of sparse corpus?

Wang Mingxuan: There are two more direct methods.

The first is to expand the corpus and strive to make the scarce corpus "no longer scarce". This idea is to use some models to obtain corpus from the Internet as much as possible. For example, in Icelandic, we can collect a large amount of Icelandic monolingual corpus. On the Internet, we can collect English texts that are similar to the monolingual corpus. We look for such corpus that may be aligned to form bilingual pairs. Of course, we sometimes use manual annotation, but more often we rely on intelligent methods to add it ourselves.

The second is to use the commonality of language. We all live on the same planet. Although we use different languages, we are actually describing the same world. Therefore, languages ​​have many commonalities at a high level. We will use some transfer learning or pre-training methods to solve this type of problem, such as letting the English model help the French model, or letting the German model help the French model. Mainly these two ideas.

51CTO: In order to reduce noise interference in multi-modal machine translation, what strategies has been adopted by Huoshan Translation?

Wang Mingxuan: To deal with noise interference, first of all, we conducted joint modeling of multiple modes. We will use voice signals and text signals together to do downstream tasks, so that error transmission will be reduced a lot. Currently, building multi-modal unified semantics is also a very hot topic in academia, so we will also absorb a lot of things from other fields.

Secondly, we will also do a lot of robustness training in the text area, trying to ensure that the model can still ensure correct output even if there are incorrect inputs, or not amplify this error, which is equivalent to Automatic error correction and machine translation are integrated into one model. Because people actually have this kind of automatic error correction ability. For example, human translators will automatically correct themselves when they hear wrong information, so we will also consider this information in the model.

51CTO: Simultaneous interpretation has very high requirements on delay. However, if there is no contextual context or complete semantics, it is difficult to guarantee accuracy. How does machine translation balance the contradiction between the two?

Wang Mingxuan: This is very challenging in the industrial world, because it is not only about the trade off of delay and accuracy, but actually requires more optimization.

For example, in some conference scenarios, translated subtitles need to be displayed on the big screen. The speed at which the audience accepts the subtitles is also one of the key issues, including the length of each subtitle display and the frequency of subtitle pop-up. How to read more comfortably. There are many details that require us to repeatedly communicate with the product manager and conduct in-depth user surveys to see overall satisfaction. Therefore, This is not just a matter of accuracy. The actual user experience must be taken into consideration before adjusting the model.

In addition, Latency may be one of the indicators of user satisfaction, but the shorter the delay, the better. Usually it is better to have a suitable gap. Because if the delay is very short, the subtitles will pop up very quickly, and the user's acceptance effect will not be very good. In this regard, we will also learn from many mature practices in the industry, such as dynamically controlling the interval of subtitle translation. Overall, this is a very engineering and product-oriented problem.

Future Trends

Machine translation is still not perfect, but practitioners are working hard to make it higher quality, more usable, and more applicable. Let’s take a look at its development trends, especially when machine translation “collides” with professional translators, what chemical reactions will occur in the translation service scene.

51CTO: With the development of technology, will machine translation derive more interesting application scenarios?

Wang Mingxuan: The volcano translation AR glasses we launched before were a similar attempt. The AR translation glasses released at the finale of this year's Google I/O conference are also a very interesting application. After wearing them, users can see the interlocutor's translation in real time, similar to the subtitle effect.

This actually reflects a relatively simple ideal: We hope that everyone can live in a world with barrier-free communication. For example: When traveling abroad, you can understand text prompts in any language by wearing glasses. The street signs you see are in German, but the ones displayed on the glasses are in Chinese. During daily communication, when someone talks to you, the conversation information is automatically turned into text you understand and displayed under the glasses. These are all scenarios where information can be obtained more effectively.

51CTO: In the long run, how will machine translation develop?

Wang Mingxuan: In terms of application, I think machine translation may be more closely integrated with multi-modal applications, such as video content and audio content. The demand for translation will increase. In addition, machine translation may be more associated with overseas business and cultural overseas expansion. Because many domestic companies are actively expanding their overseas business, I think this field will be of great help to the development of machine translation.

In terms of technology

, the trends I can see that are already happening are: First, the training of big data and large models. There are more and more people engaged in this field, the models are getting larger and larger, and the amount of data is also increasing. Many people believe that this change may bring about a qualitative change in the capabilities of machine translation. The second is the combination of translation and modality. Not only in terms of translation, many people in the industry are trying to build a unified semantic representation of different modalities. In the past few years, the boundaries between different modalities were relatively clear, and there was relatively little communication. Today, models are increasingly consistent. In the future, there may be a model that can do both text translation, voice translation, and even video translation.

51CTO: In the future, is it possible for machine translation to completely replace human translation in specific scenarios?

Wang Mingxuan:

According to the current practice, it certainly cannot replace labor. However, I think machine translation and human translation may not belong to the same track. The characteristics of machine translation are that it is very fast and can be scaled up, so it is suitable for processing massive amounts of information that needs to be processed in a timely manner. For example, if there are 10 million videos that need to be translated from English to French, it is impossible to do it purely manually, but machines can do it. This allows the machine to play a very important role in its track, which is beneficial in the long run, because it broadens the entire market and makes the cross-language market larger.

But for very precise translation scenarios, machine translation may not be able to handle it. As someone mentioned, can machine translation translate "A Dream of Red Mansions"? In my opinion, this does not fall within the scope of machine translation tasks. Translation of novels or poems, this type of translation must rely on experts. There are also high-standard conference simultaneous interpretations, which definitely require professional translators, not machines. But in some meetings that are not very important, the cost advantage of machine translation will be revealed.

Machine translation and professional translators belong to different tracks, and the distinction is still very clear. However, to some extent, the two also have a relationship of mutual help.

This is reflected in: On the one hand, the corpus required for machine translation is produced by professional translators. Professional translators continue to produce a large amount of corpus during their work, and these corpus can continue to help machine translation improve its capabilities. On the other hand, machine translation can also help reduce the burden on people and handle less demanding tasks. Nowadays, many translators are doing post-translation editing. Many translation companies let machines do the translation first, and the translators do the editing later. This can greatly improve efficiency. Guest introduction

Wang Mingxuan is the head of ByteDance AI-Lab machine translation team. His research directions are mainly machine translation and natural language processing. In the field of machine translation, he has published more than 40 papers at top conferences such as ACL and EMNLP, and has won first place in international translation evaluation competitions such as WMT many times. At the same time, he also serves as the sponsoring chair of EMNLP2022 and the chair of conference areas such as NeurIPS 2022, NLPCC 2022, and AACL2022.

Column Introduction

"T Frontline" is one of the in-depth interview columns specially opened by the 51CTO Content Center for technical figures. By inviting business people in the technology industry Leaders, senior architects, senior technical experts, etc. provide in-depth interpretation and insight into current technology hot spots, technology practices and technology trends, and promote the dissemination and development of cutting-edge technology.

The above is the detailed content of Exclusive interview with ByteDance Wang Mingxuan: Machine translation and manual translation are essentially two tracks | T Frontline. For more information, please follow other related articles on the PHP Chinese website!