Introduction: Read your mind
Author: Idle fish technology – have eureurs
I. Background:
In his book how to Win People, Carnegie said, “The only way in the world to influence someone is to talk about what he wants, and to tell him how to get it.” Thus, good communication can greatly shorten the relationship between the two sides, and then influence each other. Xianyu news is an indispensable part of buyers to understand second-hand information, we have reason to infer that a good chat can have a positive impact on the transaction. The inference needs data to support. Through correlation analysis of previous chat transaction data, we get the following conclusions:
- The more the seller talks than the buyer, the more likely the transaction is
- Compared with operators, individual sellers’ interactive conversion rate is lower
We speculate that one of the reasons for the low conversion rate of individual sellers is that they cannot chat well. Through sampling cases, it is found that some sellers reply with cold, abrupt and emotional content, which leads to the embarrassment of both parties and the abrupt stop of the chat. Therefore, our exploration direction is how to guide sellers and buyers to have a good chat, through algorithmic intention recognition, to provide sellers with friendly suggestions in the form of chat assistant, so that the communication is no longer stiff.
Two, start from bargaining
1. Why bargaining
The main dialogue scenes of trading chat include opening inquiry, price question, postage question, commodity information, etc. In the conversation, the price takes up 30% to 40%, and the conversation about price is usually at the later stage of the transaction. If the buyer and seller communicate smoothly on the price issue, it is expected to effectively promote the transaction. Therefore, we consider starting from the bargaining scene, minimum closed loop online, observation data to support conjecture, and then gradually online other dialogue scene strategies.
2. Scheme design
In order not to affect the main link of the message, the identification process must go through the asynchronous MQ consumption process. At the same time, in order to reduce the pressure of algorithm identification, the message type, sender, keyword filtering, session fatigue control and other rules are preliminaries screened. In addition, we also linked the price force data of commodities. The price force data can provide the selling price and recommendation price of commodities of the same kind and the same type. The proposed price of commodities can be given by comprehensive selection of the algorithm. In this way, regardless of whether the buyer’s offer is fair or not, and whether the seller accepts it, offer the seller a friendly bargaining option, so that the buyer and the seller still have a chance to continue the conversation, and there is still hope for a deal. If the buyer does not make a reasonable bid, they will also issue a card with an unreasonable bargaining price to the buyer, and give a reference price of similar goods to equalize the buyer’s price expectation.
3. Product effect
The seller side:
Buyer side:
Third, continue to explore
1. Is the baby still there?
I believe this is almost every idle fish sellers are common chat opening lines, but many sellers see more but not willing to reply to this kind of nutrition problems, inner OS: “not still hanging what???” . In order to make the conversation between buyers and sellers have a pleasant start, we consider optimizing this scenario from two aspects:
- Provide more variety and more valuable opening questions for buyers
- Provide quick reply to greeting messages for sellers
2. Weighted random prologue based on class
Not similar to commodity buyers concern point in chat, we based on offline data analysis, back to get a number of different orders, the buyer is the most concern of the list, these problems according to the first category classification, secondly according to the focus (second-hand property) classification, plus trigger rule conditions, weight, formed a set of opening questions library, can add the modified by product self-help. Flow chart:
As shown in the above, when the session is created, according to the first category list read the library get the corresponding problems, according to some of the properties of the goods itself, such as whether to pack mail, whether new, according to the list of questions to rule out the condition, according to the weight weighted random points to satisfy the conditions of the problem, finally get some second-hand property problem. In order to ensure the diversity of the problem content, each second hand problem item will have different copywriting but the same semantic expression, and finally express a specific problem randomly again, then the suggestion card can be issued.
3. Intention identification process framework
The algorithm recognition of bargaining intention has been realized before. This time, we need to add a new recognition of greeting. Do we need to re-develop the whole process? Obviously don’t need, the early stage of the negotiation in order to see the effect fast iterative online, there is no abstract design into a common intention recognition process, the new second intention recognition, it is necessary to pay his debts redesigned, any good system design is a continuously iterative reconstruction, will never be achieved overnight, the beginning of the project under the condition of the demand is not clear if you consider too much, This often leads to overdesign and then rework if requirements change. Therefore, a general intention recognition process framework is abstracted and designed.
Process design:
Class diagram:
As shown in the figure above, each intent needs to implement the IntentProcessor interface to implement its own filtering and processing logic. Different intents have different initial screening logic and extension parameters. The filterAndCompleteContext method is used to filter and supplement the intent recognition extension parameters into the process context. If the initial screening condition is met, the intent type is added to the list of possible intents, and the final result is identified by the algorithm. If multiple intents exist in a message, the algorithm selects the most relevant intent according to the priority rules. After the intent identification result is obtained, the process method of the corresponding IntentProcessor is called to complete the specific business logic processing. For example, in the bargaining scenario, different documents are assembled and suggestion cards are issued according to the price rules.
4. Product effect
Prologue:
Say “hello” :
Fourth, take the chestnut from the question and answer
1. Why do you keep asking me the same question?
Observation data, we found that the chat coverage rate 35% ~ 40% commodity information in the scene, because idle fish unique commodity trading second-hand property, the seller may face the same products the same problem will be more buyers asked for many times, so as to repeat the answer, sellers have bitterness could not say, only the heart OS “why always asking me the same question?” . In order to optimize the seller experience, improve the efficiency of the seller to reply, we have decided to identify answers to in chat and then behind the question and answer to insert boot tip, the seller side can choose to question and answer to added the details of goods, if the question and answer of commodity structured second-hand attribute information is contained in (such as colour, presence of repairing, brand, etc.), It also identifies and guides the seller to supplement the structured attributes of the product.
2. The extended properties of the general message are changed
The intent recognition process framework described above is immediately available, and the entire functionality can be easily implemented by adding an IntentProcessor that implements the filtering and processing logic for the scenario. But new problems come, bargaining, say “hello” scenario we advise sellers side guide is a kind of card, the card is a new message, mixed with other messages, and trigger the source message correlation is not strong, even if there is delay, the position of the card is inserted into the partial after effect also is not very big. But in the question and answer scenario, a guide tip is delivered, which is strongly related to the message to which the answer belongs. The guide tip must follow the answer message, and if it is not right, it will greatly affect the experience. The message list of Idfish is sorted according to the sending time. If the message is inserted in the form of previous new messages, it cannot be strictly guaranteed that the sending time follows a certain message. If the message sending time is manually modified, the semantics of the message sending time field will be damaged. The other way to think about it is that this tip must be following a piece of news, so why not merge the two? Consider the tip as an extended property of the message, so we decided to introduce the ability to generically change the extended property of the message, which is sent to the client through the event, and then parsedand displayed by the client according to the agreed protocol, as shown in the figure below
If the extended properties of a message are changed initiated by a service, you can set the storage server message base and update the session view, for example, tip in the Q&A scenario. The change is time-sensitive and only needs to be transmitted to the client through, and the server does not need to store the change. The solution also opens up the possibility of personalized changes and presentation of messages in the future.
3. Product effect
V. Summary and prospect
So far, our chat assistant has developed a set of general intention identification process framework, which can realize the three intention identification of bargaining, greeting and question and answer, guide the majority of individual sellers to have a good chat with the buyers, and help the sellers to quickly add more detailed commodity information. Function of online chat little helper, utilization rate is higher, compared to the experimental barrels contrast barrels, response rates relative increase of 4%, in the negotiation scenario, the card issued by experimental bucket compared to the control of barrel relative increase of 4%, clinch a deal the conversion from data, it also proves that we really before, lead sales chat, has positive effect to the transaction, It provides the possibility for the continued evolution of the project.
I believe that friendly communication can warm people’s hearts and bring them closer. This is true in the real world as well as online communication. In the future, we will continue to iteratively optimize the chat assistant, and explore more intention identification scenes in the horizontal direction, such as free mail, place of shipment, etc. At the same time, we will further explore the commodity information identification scenes that are most involved in the chat, so as to help sellers better improve and supplement commodity information. In short, the chatterbox assistant of the future will have a richer and more practical skill tree, making her smarter and better at understanding you.
The original link
This article is ali Cloud original content, shall not be reproduced without permission.