Looking at strategic product work from recall rate

200 queries were randomly selected in the background.

Strategies identified as weather demand: 36

Weather demand: 33

Number of cases with incorrect policy identification: 7

Number of cases with correct strategy identification: 29

Number of cases ignored by policy identification: 4

Recall rate (representing the degree to which the strategy solves the problem)

= Among the cases that want to be covered, the cases actually covered by the policy/the cases that want to be covered by the policy in an ideal state.

= 29 / 33 = 0.88%

Accuracy (representing whether the strategy has caused other injuries)

= Among the cases covered by the policy, all the cases that really want to be covered by the policy.

= 29 / 36 = 0.80

It can accurately analyze the possible user needs of the query and help the next retrieval.

Through manual labeling and analysis of random sampling results, the problems existing in the identification strategy can be summarized as follows:

Case 1: My boyfriend pissed me off today.

There is a continuous word "weather" in this query, but combined with the whole sentence, it obviously has nothing to do with "weather". The existing problem is incorrect word cutting. The correct way to cut words is: boyfriend/today/pisses me off/.

Case 2: Fall in love with your fine weather.

Fall in love with your fine weather, this query is a complete song title. According to the operation of the search results page, it can be judged that the user is more likely to find the song of this query.

Case 1: What clothes should I prepare for Huashan?

Users didn't directly search for the recent weather in Huashan, but what clothes they prepared was closely related to the weather at that time.

This kind of inquiry has clear position information, clear dressing and other demands, and is generally interrogative.

1. When there are clear category words, optimize the parsing rules of search words;

2. Take the feedback historical data of search result click operation as a dimension of query analysis, and grasp the demand more accurately.

3. Optimize the query word segmentation scheme;

According to the above analysis, considering the impact of the problem, the degree of solution and the development cost, Scheme 2 is the project with the highest priority, followed by the scheme of optimizing word segmentation, and finally the project of optimizing the analysis rules of search word segmentation.

Improve the current system's demand identification strategy for search words, and improve the recall and accuracy of queries.

Problem: The query contains the target search term, but from the whole query, there are many weighted search results.

Solution: Take the historical data fed back by the click operation of search results as a dimension of query analysis. By analyzing the historical search result data, it is judged that the user wants to search the whole query.

Problem: Chinese word segmentation can have different word segmentation methods at different granularity, which leads to the deviation of the understanding of search results.

Solution: When the query can produce multiple granularity word segmentation methods, fine-grained word segmentation is used to ensure the recall rate when constructing the index, and coarse-grained word segmentation is used to ensure the accuracy when querying.

Problem: There is no clear search keyword, but it actually reflects the user's need to understand the weather through the brief description.

Solution: In the search word parsing rules, add spoken words as identification. When parsing, if you encounter the recognized spoken words, you will escape the query according to the preset rules.

Difficulties: Because this work involves a large number of rule definitions, it cannot be guaranteed to be completed within a 2-week development cycle. Therefore, this requirement has a low priority and is not included in this development plan.

Core indicators: query recall and accuracy

Observation method: 200 pieces of data are randomly selected from the system and calculated once in the optimized strategy system, and compared with the recall rate and accuracy rate before optimization.