Skip to content

Latest commit

 

History

History
96 lines (84 loc) · 5.56 KB

search-scores-and-weights.md

File metadata and controls

96 lines (84 loc) · 5.56 KB

Scores and weights for search function

At the re-ranking stage, our search system uses scoring to assign values to each document based on different parameters. The interplay between clustering and scoring helps optimize the search process, ensuring that the system considers both the content and context to deliver more accurate and relevant results for the user.

This task becomes challenging when dealing with a large number of documents, creating the need for optimization strategies. A good approach is the utilization of an indexing method called clustering, which categorize documents based on their topics, facilitating a more streamlined and efficient search process. We also use scoring, which involves the assignment of numerical values and weights to documents based on various parameters, influencing their ranking in the search results. Here is our different scores, each serving a unique purpose:

  1. Similarity: Represents the primary signal in our scoring mechanism. It denotes the measurement of how closely a document aligns with the user's query, reflecting the relevance of the document to the search criteria. It is not a static precomputed score but a dynamic metric that is computed on the fly from the search results. This ensures that the relevance is tied to the specific query, going beyond simple keyword matching. Documents with higher similarity scores are considered more relevant. By prioritizing similarity as the first signal in our scoring process, we aim to deliver search results that are more accurate.

    • Scaling: FROM 0.0 = least similar to user query TO 1.0 = most similar to user query

  2. Recency: This score considers the temporal aspect of documents, prioritizing recently added or updated content. A document's recency is crucial in reflecting the latest information available to users.

    • Scaling: FROM 0.0 = oldest document TO 1.0 = most recent document

  3. Traffic: The frequency with which users consult a document influences its score. Popular or frequently accessed documents are given higher scores with the help of web traffic logs, indicating their relevance and importance to users. Warning: The home page is rated really high since it's where every user land at first.

    • Scaling: FROM 0.0 = least consulted document TO 1.0 = most consulted document

  4. Current: This score determines whether a document is currently accessible or if it has been archived. It helps users distinguish between active and inactive content.

    • Scaling: 0.0 = currently accessible document OR 1.0 = archived document

  5. Typicality: This score evaluates how closely the number of site references for a document aligns with the average. Documents with typicality scores reflect a level of correspondence with the average number of references. This ensures that the search results prioritize documents considering how well they conform to the typical reference patterns within the targeted theme.

    • Scaling: FROM 0.0 = least referenced document TO 1.0 = most referenced document

  6. Didactic: This score evaluates the informational value within content chunk. It scores higher based the quality and readability of information provided. Documents with high didactic scores often contain rich textual information, explanations, and details. Implementation of the score looks for signs of the opposite to compute its score - for example, the presence of a large proportion of tabular data which indicate data dumps from spreadsheets or databases.

    • Scaling: FROM 0.0 = mostly tabular data or information that is not expected to be read sequentially by a user TO 1.0 = contains rich textual information, explanations, and details

  7. Guidance: This score pertains to content chunks extracted from guidance-oriented pages, emphasizing their significance and relevance. Guidance pages typically offer comprehensive direction, instruction, or expert advice within a specific domain. As these pages tend to provide crucial information or instructions sought by users, they are given priority to ensure users can readily access the most helpful and directive content.

    • FROM Scaling: 0.0 = doesn't include crucial information or instructions TO 1.0 = includes crucial information or instructions

By incorporating these scoring parameters, we fine-tune the document retrieval process to align with user needs. It allows us to prioritize documents that are not only recent, popular, and representative but also closely related to the user's specific search criteria. This multi-faceted approach enhances the efficiency and effectiveness of our document retrieval system, ensuring a more tailored and user-friendly experience.

Future

In addition to our current considerations, we can explore the integration of thematic context into our scoring system. Thematic context involves a specific focus on the subject or theme related to the user's query, ensuring that the context is taken into account during the initial score calculation. To implement this, we would need to incorporate topic labels for documents, a feature not yet incorporated in our system. Planning for such additional scores allows us to enhance the depth and relevance of our responses by considering the specific themes associated with user queries.