“This article combines the essence of the answers and discussions of several zhihu leaders, and summarizes the connection and similarities and differences between NLP and recommendation system in terms of question tasks, skill requirements, interview preparation, business landing and future development. It is very worth reading!”

Author: Huang Hanchi

Source: Zhihu column AI hit strange road, has been authorized by the author.

Editor: happyGirl

1. The connection and comparison between NLP and recommendation

1) contact

The output of NLP is often the input of the recommendation system: what NLP does is actually very simple, and the effect is relatively good verification: for example, the emotion recognition of the text, the input is the text sequence, the output is the corresponding emotion, and there are accurate evaluation indicators. The output of NLP can be used as characteristics: whether it is positive, reactionary, or not involved in pornography, etc. – yi da: talk about NLP and zhuanlan.zhihu.com/p/71938647 recommendation system choice

(Audience refuted: NLP has a lot of problems of lack of automatic evaluation indicators, especially in the generation of class NLP. Many problems can only be evaluated by manual, and even by manual, it is difficult to achieve unbiased and convincing evaluation conclusions, such as the open domain chat problem and various controllable text generation problems.)

(2) to compare

On the macro:

Natural language processing is more like a discipline, while recommendation systems are more like an application. So you’re going to use machine learning, you’re going to use data mining, you’re going to use things like natural language processing, and natural language processing may have its own theories, but it’s going to use machine learning, data mining, but it’s not going to involve recommendation systems. Link: Natural language processing and recommendation system, which direction to choose, quasi research? www.zhihu.com/question/64…

(Pay is expensive, engineering is difficult, and from data streaming to feature processing to online services, NLP is not on the same level as a comprehensive system of recommendations. Do not go to NLP after graduation, recommend to do recommendations to improve general engineering skills, if you have the ability to directly engage in RTB advertising that is even better.

Research and Application outline:

NLP divides a sentence or a sequence into: 1. Segmentation of fields; 2. Extraction of keywords; 4. In-depth analysis, such as sentiment analysis, classification and construction of context, etc.; Common methods, simple points, semantic similarity, hidden horse model, difficult points using RNN, LSTM recommendation system birds of a feather flock together: 1. 2. Extracting features of items or users 3. Constructing relevant networks 4. In-depth analysis, user similarity, user classification, user rating. Machine learning minute classification regression model Link: NLP Natural language processing, recommendation systems, or computer vision which has better job prospects? www.zhihu.com/question/26…

Job requirements:

At present, from the number of posts, search advertising recommendation is still the most in demand, not only reflected in the number of posts, but also reflected in the company. As we all know, selling goods advertising as the main income of Internet companies, companies have been increasing the demand for this piece. In Beijing, for example, first-tier and second-tier Internet companies are still hiring, even as Meituan and others downsize. From the point of view of job quality, there are many demands for ordinary engineers, while high-end positions are often not filled by suitable people due to the influence of the team. Meanwhile, there are few high-end demands, which only exist in some high-quality companies, presenting a state of one radish and one pit. In fact, the current demand for NLP is relatively small, probably because the pit is too large, so there are fewer companies doing it. Many NLP candidates may have done low-level algorithm research (scientists), some have done various business systems and user portraits (dialogue, translation, etc.) (a lot of engineers do this), and some have done search recommendations. Link: NLP natural language processing, recommendation systems, or computer vision have better job prospects? www.zhihu.com/question/26…

Follow-up development of the post:

A)NLP positions are actually more suitable for research. Because its evaluation index is relatively single, good is good or bad is not good, and good model effect is not as high as the requirement of recommendation system for real-time algorithm. The model Fancy, the idea work, is basically an article. Compared to support, in the company are basically doing basic function is recommended, meticulous easier to do, because NLP evaluation index is relatively stable, won’t appear offline and online gap is very big (I guess), more energy will be deep digging into the details, if reasonable use advantage, I believe, in the company can also have a lot of harvest. B) The recommended position is a more engineering position. There are all kinds of pits to step on and a lot of dirty work (location feature engineering still plays a very important role at present). AB tests are required before the model is put online, which requires high real-time performance and so on. So engineering capability is still very important in recommendation systems. So it’s more of an engineering position. – yi da: talk about NLP and selection of recommendation system at https://www.zhihu.com/question/268751628/answer/342223389

Audience Refutation 1: There are some factual errors in the article. NLP problems can be divided into three categories: classification, labeling and generation. Bloggers should focus mainly on categorization (matching is also a category), so it may be a bit of a generalization. Many typical classification and matching problems in NLP have been approximately solved, while the really difficult sequence labeling and generation problems in NLP, as well as the classification and matching problems superimposed with some constraints (such as small sample, multi-domain) have been solved. This is where most NLP researchers focus.)

Audience Refute 2: Most NLP systems have a lot of dirty work, such as task completion dialogue, which is much more dirty work and engineering than recommendation systems. The demand for NLP business talent is much greater than for research talent, and it is not accurate to say that NLP is more suitable for research. Few companies are willing to invest in NLP research at this stage, but the effort to incubate NLP products is a key track in the competition among the giants.)

From the perspective of future development, I would like to offer some suggestions. At present, the business and technology of search recommendation are the most mature, which can be said to be the most stable, even in the coldest winter. CV bubble is not small, but positive dividend period, relying on deep learning this metaphysics take off a little serious; Because I do not know the technology, I think THAT NLP may not be as good as CV’s landing in five years. This seems a bit too mysterious and unscientific. In addition, many companies are playing games, so I feel that NLP is a bit miserable at present (not counting those big bull ha). So you have to look at your skills and personality, if you can stay calm I believe the future direction of NLP will not be a small gain.

Link: NLP natural language processing, recommendation systems, or computer vision have better job prospects? www.zhihu.com/question/26…

Interview preparation highlights:

Yi: talk about NLP and selection of recommendation system www.zhihu.com/question/26…

(Audience criticism: When it comes to NLP job hunting, the knowledge mentioned in the article may only be enough to get an offer in the core business department of a big factory, or even the background of the cabbage. The theme model is becoming weaker and weaker in NLP. Word2vec, Elmo and others are only basic knowledge, far from being competitive. The knowledge points you listed are obviously mainly for retrieval and recommendation scenarios, and are far from enough for translation, q&A, dialogue, extraction, and lexical and syntactic analysis, which are more typical NLP problems.)

Comments from other viewers:

CTR as the only target metric recommendation would be a dead end. Toy project. Generally, your agenda metric + stickiness is the first item to ask your PM, and the second item is no more than CTR, viewing duration, page turn rate, retention and other primary KPIs to be weighted. Train a RL agent as an environment and abuse policy network as much as you want.

Do the recommendation first, and then do the representation of NLP, under the condition of the same money, NLP will be much more comfortable, the recommendation may make n months of model, indicators do not rise but fall, of course, everything depends on interest.

NLP will also have inconsistency between online and offline, and the acquisition of NLP annotation data is a huge matter, especially the definition of tasks is directly related to the accuracy of subsequent annotation. Too much dirty work in NLP… If there is no one to help mark the data, you have to get the data yourself. But the NLP standard itself is difficult to draw a clear line. There is often a situation that a is also good, and B is also good. Even people are difficult to divide, and the model is more difficult to make… No recommendation was made, but if you do NLP, unless someone has done all the data for you, it will be a big part of cleaning the data

Goose factory advertising related personnel, unseen temporarily) advertising is one of the best in the business, because advertising is cash flow, most of the company’s first of revenue are advertising and in addition to the goose factory, advertising revenue proportion is very big, such as facebook, 99% are advertising revenues Domestic advertising top players is ali baidu tencent headlines, Alibaba, Phoenix Nest and AMS(formerly known as Guangdiantong) are all famous, but technically speaking, they are not. The so-called success is nothing but failure, and the system is very stable and can be explained if it is linked to cash. Tencent’s advertising revenue is 156 million RMB per day. System unstable hazard ratio model can improve 0.1 points is much more important, so there will be a lot of regularity, and most people do not model, but the characteristics and rules, moreover engineering optimization importance is much higher than the model, it is very fancy in addition, advertising effect and brand, effect of advertising is very similar to recommend, recommendations can be turned, Brand advertising is mostly operations research, it’s far less fancy than machine learning and there’s actually CV NLP, NLP for contextual features that goes without saying, CV in addition to providing image features, Can also provide advertising creative building and recommended replacement, video advertising code, video ads related to recommend (that is, such as a TV appeared you fruit scene will give you push ads) important still groundwork, machine learning are interlinked, and advertising that the practice of a set of recall + fine line + rules with recommended search and has much difference, It’s just a different business and more importantly, write good code

Two: it seems that there are few “unicorn” enterprises in the NLP field at present. Some people also say that natural language processing is a big pit, so what difficulties do you think NLP landing has? (Interview to Zhang Junlin)

A: Various TASKS of NLP are more in the form of background technology platform, which is A kind of capability output, rather than A business form. It is generally the supporting technology to support some business forms. In general, compared with the product form that users can directly perceive, the overall is a little later, and it is difficult for general end users to perceive its existence. In fact, there are few business areas that can directly perceive NLP technology in terms of product form. For example, dialogue robot, intelligent customer service and other business forms are relatively advanced, while most of them are relatively lower. NLP is a technical field, and the product form is not necessarily related. So, I don’t think there should have been a unicorn in NLP in the first place. It would have been more accurate to say it was a unicorn that used NLP heavily. Of course, there are very few unicorns, because there are very few unicorns, and even fewer use NLP technology. Speaking of difficulties in landing, in fact, there is only one: the technology is not mature enough to support good application experience. Of course, this specific sub-field, some fields, such as text classification and clustering, NER named entity recognition and other simple tasks, in fact, the landing effect has been good, has also been widely used. For many of the most difficult applications, the other problems are superficial and the underlying problem is that the technology is not mature enough. The above-mentioned problems are long-term and cannot be solved in a short time. If NLP landing short-term difficulties, such as the new technology of Transformer and Bert this effect is very good, may be because the models are heavier, efficient distributed large-scale training system and fast online services, may be the short-term impediment to large scale at present, but I believe that soon there will be a big company open source out some useful systems, So it’s not really a problem.

Links: perceptions of NLP and the future development trend of recommendation system zhuanlan.zhihu.com/p/79677478

Iii. What technology stack does sina Weibo’s information recommendation system have? What has Sina Weibo done to build a more accurate recommendation system? Are there plans to introduce any new technologies? Or what new technologies are you looking at? (Interview to Zhang Junlin)

A: Because the recommendation scene of Weibo is A typical information flow business, it contains many links. From the background material quality assessment, user interest modeling, weibo content understanding, picture and video understanding and multi-mode fusion, to the recommendation system of the front of the business recall, rough and fine, as well as offline and online large-scale machine learning model training and services. It’s a standard information flow recommendation system. A: In the past two years, the Weibo machine learning team has carried out large-scale technical upgrading in recall and ranking of recommendation system, and achieved obvious business results. At the level of recall, we have realized the large-scale FM unified recall model, which is gradually replacing the traditional multi-way recall model, and has achieved very obvious business effects in various indicators. At the ranking level, we have undergone continuous model upgrades such as LR, large-scale FM, FM+FTRL online model, and each large model upgrade has achieved benefits. At present, deepFM-based deep learning ranking model has also been used in a small amount. In terms of the understanding of materials such as microblog, we are also trying multi-modal technical route and have made some progress. While the business team is constantly upgrading and implementing new technologies, Weibo AI Lab is also constantly trying to recommend some new technologies in important directions in the field and promoting the implementation of these new technologies in the business. For example, the bilinear FFM model we proposed in 2018 reduced the parameter scale by dozens of times under the condition of obtaining similar effects with FFM model. Several new deep learning sorting models, such as FibiNet and FAT-FFM model, were also proposed. These work have been published in international conferences such as Recsys2019 and ICDM2019. I personally pay more attention to the unified recall model technology and the new CTR model. Feel these two in the industry, there is still a lot of room for optimization, is worth further exploration direction.

Links: perceptions of NLP and the future development trend of recommendation system zhuanlan.zhihu.com/p/79677478

Audience comments:

The proposed model is not deep and complicated, and the limitation of latency is an important factor in online inference.

Personally, I think the recommendation of weibo is very poor and weak in terms of user experience. VS depending on the individual, I can’t stop scrolling VS compare Tiktok and Toutiao to see the difference

Iv. The year 2019 is half over, what directions do you think are worth studying in the field of recommendation systems and NLP in the future? (Interview to Zhang Junlin)

A: With the development of recent years, the industrial recommendation system has evolved into the era of Deep learning. There are many representative models, such as Wide& Deep or DeepFM models. However, compared with NLP or image field, it is obvious that deep learning in practical recommendation system application scenarios has not been able to meet people’s expectations. There is no deep learning recommendation model that can greatly improve performance compared with traditional models, and the reasons may be complicated. From a model point of view, I am more optimistic about using Transformer models in the recommendation space. As you know, Transformer is already very popular in NLP, and it is mechanically suitable for recommendation or CTR models. However, it may still need to make some changes according to the characteristics of the recommendation field, and the direct application effect is not reflected, but in general, I am quite optimistic about this model. Of course, from the perspective of practical recommendation engine, there are still many points worth paying attention to, such as multi-modal fusion/multi-objective, multi-task and Transfer Learning, as well as the application of AutoML, which are all promising and worthy of exploration. I just talked about the sorting model and recommendation mechanism. As for the recall stage, another important link in the recommendation field, I think it is an obvious trend to replace the traditional multi-way recall with model unified recall. As for THE NLP field, since the appearance of Bert, it should be said that the various application fields of NLP have entered a new era. Bert has achieved the best results after being used in many NLP application fields, and many fields have significant performance improvement. This heralds the start of a lot of unsupervised data being put to real use by NLP, which is powerful if it can be used. Bert+Transformer’s ability and tendency to unify NLP’s sub-domains is also a very good development. Of course, since Bert and Transformer have only appeared for a short time, people have not yet had a deep understanding of them and the factors that play a role in them have not been sorted out clearly. Therefore, a deep understanding of their mechanism of action and targeted improvement of them are very worthy of further exploration. For example, how to apply Bert in the field of text generation, how to integrate multi-modal information and many other directions have great challenges.

The above is a brief discussion on the connection and comparison between NLP and some problems in the recommendation system, as well as the future development. Click on the bottom of the article to read the original text, you can pay attention to the author’s zhihu, directly communicate with the author.

If you like, click on????