Search engine user image retrieval needs to meet how to achieve?

First, what is the demand satisfaction

1.1 What is the demand satisfaction

Users search for "Octopus Paul," and in terms of textual relevancy, search engines simply return results that relate to the "Octopus" content, so are users satisfied?

User A: I heard that the octopus Emperor hung up to see the latest results, how are all in August, the next page ...

User B: Today, colleagues discuss the octopus brother hung, octopus brother is what? I am out, to search the octopus life story is what, how are all the latest results, there is no introduction of octopus brother ah, transform a query Look

User C: I am a hardcore fan, after reading the octopus brother, look at the football-related it, Rooney, Steven Gerrard scored again, how even a related recommendation did not, I have to personally enter.

User Ding: find an octopus brother's picture use it, must pull the wind, how the whole result is no square chart, how to use such a flat chart ah

User E: Another octopus brother's wallpaper, maybe the next time you buy a lottery can make a fortune, 咦, how are all small size map ...

Generally speaking, the user to express his needs to the search engine, search engine to understand user needs, provide different needs under the resources, the entire process can be collectively referred to as demand satisfaction. Simply put, rank work in addition to the basic text relevancy belongs to the category of demand satisfaction, that is, the search results provided to the user, not only requires literally and user-entered text but also Meet the different needs of users.

1.2 why need needs to meet

Users express their own needs through the query, and for most of the query, especially with implicit demand for the query, only the literal matching query results may not be able to meet their needs. At present, our ranking system is mainly based on the dimension of text relevance. The weights reflect the correlation between terms and obj in query. Under this system, the related results may not be able to meet the needs of users.

For example, the aforementioned example of "Octopus Paul", it is clear that these needs are difficult to solve in the context of textual relevance, especially with regard to sudden and time-critical needs and pan-demand. 1.3 Requirements to meet the inclusion of what the work From the above example, we can see that needs to meet the needs of addressing timeliness, multi-demand issues, related recommendations, size requirements, material needs, browse guide and other issues. Rank strategies other than basic textual relevance, as well as query analytics for these purposes, can be considered work-for-need, as well as front-end results presentation and user-guided browsing.

Image needs to meet, according to different dimensions, can be divided into the following aspects:

a. Needs identification

b. Resource building

c. demand transfer of power

d. Results organization and recommendation

e. User-directed interaction

Second, the demand to meet how to do

Needs to meet the core issues to be addressed:

Demand recognition

Resource building

Demand transfer of power

2.1 Requirements Identification

2.1.1 The type of requirement

Identify query what needs, as well as the strength of demand, is the most basic work. First of all, there is a need of the system, a complete description of the various needs, followed by how to identify these needs, the needs of each query corresponds to this system.

Statistics-based demand recognition

Through a large number of statistical analysis of data, you can identify what kind of common query. There are many data available for analysis, such as user behavior data, click feedback, search results and more.

For example: search for "Octopus Paul wallpaper", by clicking the image user's length and width data and found that the user clicks on the picture, most of the larger aspect ratio picture, and "Octopus Paul Avatar" is exactly the opposite.

Another example is the analysis of a large number of users of the search data and found that a considerable part of the user in the retrieval of "Octopus Paul", and retrieve the football-related keywords, then you can identify the "Octopus Paul" and football have a high degree of relevance, Users recommend related search, you can insert football-related recommendations.

Name & Demand Word

Judgment query contains keywords or requirement words and other keywords, is the most direct way. For example, "Octopus Paul Avatar", the user displays the query in the expression of the avatar needs, which includes the size of the needs of the avatar is the need for small size map, as shown on the right, this time out of a large picture It does not meet the needs of users.

Timeliness of demand

Timely identification of needs, mainly through the user search of the sudden burst and the number of resources to determine the burst.

Bursts of retrieval volume can be calculated by accumulating the daily retrieval frequency of each query, searching the retrieval frequency by the user for many days, calculating the retrieval quantity and the history comparison of the current day, whether there is a burst, and then judging whether the requirement of timeliness is strong weak. Judgment of the number of resources can be tapped in a similar way.

For example, "Octopus Paul", during the World Cup, the query, as well as the related query query volume, before the World Cup, the user search volume has been an explosive growth, and continued to maintain high search volume of the state, can be considered as having Timely needs of the query.

2.2 demand satisfaction

Identify query what needs, the next step is to provide the appropriate resources.

2.2.1 resource mining

How to get the resources to meet the demand is another core issue of demand satisfaction. Resources, through one or a combination of features, can meet the requirements of the resources and resources do not meet the requirements to distinguish between resources to find the needs of users demand, remove the resources do not meet the requirements, is the main task.

Content attribute characteristics

For the content attribute dimension, it can be divided into the underlying physical features, the object recognition in the middle layer, and the semantic features in the upper layers.

For the bottom of the physical characteristics, is relatively simple, including the size, color, format, sharpness and clarity, etc., the middle of the characteristics of human and non-human, erotic images, vehicle identification, cell phone picture identification, etc .; for high-level Semantic features, including scene recognition, picture style recognition, emotion recognition, such as indoor or outdoor, whether the non-mainstream style, can be used as a resource screening features.

Topic attribute dimension

The topic attribute dimension refers to a variety of topics such as animals, plants, handsome, beautiful, military, sports and so on. We hope that the picture can be divided according to such a classification.

For example, through this classification, we can know which pictures are avatars, which are wallpaper, and which are football sports related. Users search for "Octopus Paul", you can recommend football-related resources.

Timeliness of resources included

Timeliness of resources, can be easily judged by included time, and non-time-sensitive resources to distinguish. The sources of timeliness resources generally include news sites, major forums, bbs and other community sites.

2.2.2 demand transfer of power

Clear the needs of the query, mining resources to meet the demand, then how to meet the needs of the resource rank to the front?

For a variety of different demand dimensions, have their own strategy of transfer of power. For example, "Octopus Paul wallpaper", we identify the size of the demand, then the larger size of the picture can be weighted; Another example is the demand for timeliness, you can directly insert the first three pages of the results of the time-sensitive library, which This is because the requirement of timeliness is a strong demand dimension, simple weighting, and can not guarantee that the result is adjusted to the first three pages.

At present, this strategy directly superposition of the right to adjust the way, the advantage is simple, direct, disadvantages are more, the biggest is uncontrollable, a dimension of the right to adjust, will have much effect on the final result, he said the weight of how much ,do not know.

Third, the conclusion

For the needs to meet the future, we must continue to develop in the direction of intelligence, automation and diversification. Our ultimate goal is to meet the needs of this direction did not do, demand mining, resources to meet all the automation, so that "no sword in the hands of a sword."