Elasticsearch and Apache Lucene: Fundamentals Behind the Relevance Score

I still remember those days of using search engines in different portals and not getting even one relevant result. That could be the reason Google became so popular, and Google certainly resolved that problem.

As the amount of content is growing daily and with an increased pace, giving such powerful search capabilities to users is getting more important as well. Which is why I want to share some fundamental information on the topic.

Currently, Elasticsearch is ranked as the most popular search engine according to DB-Engines. This article explores fundamental Elasticsearch concepts such as indexes, documents, and inverted indexes, and how these concepts work together to provide storage and relevance scoring. Additionally, I will present a case where these Elasticsearch features were evaluated and Elasticsearch was proposed as the main data repository for an internal project.

FUNDAMENTAL CONCEPTS OF THE APACHE LUCENE LIBRARY

The core of Elasticsearch is the Apache Lucene library, which includes features for indexing, searching, retrieving and updating documents, and text analysis. The figure below depicts the integration between Elasticsearch and Lucene, and how they interact with external systems:

The fundamental concepts required to understand the theory behind Apache Lucene are indexes, documents, inverted indexes, scoring, and tokenisation. I will briefly describe these concepts below.

An index is a collection of documents sharing conceptual and logical similarities. You can think of an index as a folder with multiple related documents. Given that Elasticsearch is a distributed system and clusters can be added on demand, there is virtually no limit to the number of documents an Elasticsearch server can store.

A document is a record containing information related to the index. Documents do not have a specific scheme and every document pushed into the index is tagged with a unique identifier. A document contains a set of fields and is generally stored in JSON format. Fields are the minimal unit of storage in the Lucene ecosystem. A simple document is illustrated in the following figure:

An inverted index is a data structure storing information in a complex HashMap, aiming to facilitate the search of terms contained within the fields of the documents. The concept of the inverted index is close to the concept of a book index. When you need to search for a specific term in the book, you use the book index and check which pages contain information about the queried term. The inverted index is composed of two substructures:

• the term dictionary which groups all the terms included in the documents in a sorted list.
• the postings list which creates a list of each term, indicating the documents where the term appears.

The following table shows an example of an inverted index substructures build based on a set of documents:

Scoring is the process that compares the user input against the stored documents and assigns relevance values for each result. Similar to how the Google search engine works, the outcome of the scoring process is a sorted array containing all the search matches ordered by score.

Three aspects should be considered for the scoring process:

• relevance,
• user satisfaction, and
• algorithms.

The first aspect is related to the relevance definition itself. Even though different users perform the same search, the relevance results could be different for each of them; therefore, the relevance should be aligned to the user expectations.

The second aspect concerns knowledge about a user's satisfaction with the relevance score provided in the search. In fact, giving feedback to the algorithm on the user's perception is a very fascinating field related to reinforcement learning techniques, which by itself would warrant a separate article.

The third aspect encompasses the algorithms which calculate the scores. There are two main algorithms used for scoring: Term Frequency-Inverse Document Frequency (TD-IDF) and Best Match 25 (BM25). Both algorithms are rooted in the concept of tokenisation.

Tokenisation is a fundamental concept of the Natural Language Processing (NLP) field, which is also being applied to search engines. Tokenisation consists of splitting the text string or documents into tokens, or smaller chunks. Although tokenisation libraries provide a set of methods for splitting text, users can implement their own rules using regular expressions. Common tokenisation methods include breaking up the text by words or by punctuation marks.

Although tokenisation is commonly applied to text, in certain situations, special techniques allow for the tokenisation of other kinds of data, such as numbers or geographical coordinates. Consequently, each specific use case should be taken into consideration during the configuration of Elasticsearch tokenisation capabilities.

HOW THE FUNDAMENTAL CONCEPTS WORK TOGETHER

The previously reviewed concepts work together to establish relevance scores from semi-structured data. When a document is saved in Elasticsearch, two processes happen in parallel: in the first, the raw document is stored in the index; in the second, each document passes through an analysis phase before being stored in the inverted index. This is described in the following figure:

The analysis layer is a configurable step which, by default, includes three actions: character filter, tokeniser, and token filters. These actions facilitate the search algorithms. Additionally, each document pushed into the search engine goes through a metadata enrichment process, which can be as simple as the counting of terms or as complex as the calculation of a sentiment trend.

Naturally, when the query API is used for retrieving documents, the search terms pass through the same actions used when documents were stored. That way, search terms can be matched against the inverted index, and the search server can calculate the scores accordingly and return the results sorted by relevance. This process is depicted in the next figure:

The following example will clarify the process described above. Let us say we have the next document from a review:

{
  'description': 'Barcelona has beautifullllll beaches.'
}

After applying the analysis step, the information in the inverted index looks something like this:

['barcelona', 'beautiful', 'beach']

When a user inputs the search string 'Barcelona's beaches' and the analysis layer is applied, the search string is transformed into something like this:

['barcelona', 'beach']

The resulting terms can now be matched against the inverted index, and the result can be obtained by direct comparison.

GETTING THE BEST OUT OF THIS STACK

For Endava's Innovation Lab, an internal competition where Endavans develop innovative tech solutions, we proposed a product using Elasticsearch as data storage. The goal was to provide the recruiting team with a list of the most relevant candidates for a specific opening.

To achieve this, the system extracts professional profiles from various sources and then creates a cherry-picked selection of candidates for the recruiting team, based on NLP techniques leveraged by Elasticsearch's built-in algorithms.

All the features explained before fed the idea of using Elasticsearch as principal storage. However, there were two design decisions remaining for the final architectural approach: the principal storage mechanism and the deployment method (cloud or on-premises).

For the principal storage decision,

• the first option was to use Elasticsearch just as a search engine returning indexes to retrieve content from a dedicated database. This dedicated database would be a relational database, like PostgreSQL or MySQL, or a more robust document database, like Mongo or DynamoDB.
• the other option was to use Elasticsearch as main storage as well.

Ultimately, Elasticsearch was selected as the principal storage mechanism, due to the medium-scale storage requirement for the entire solution (estimated at just 36 Gigabytes) and the non-transactional nature of the business requirements.

We deployed this solution in the AWS cloud because it provides a fully maintained Elasticsearch service, and it can be easily integrated with other services within the AWS ecosystem.

The following diagram reflects the final design, including the forementioned decisions:

CONCLUSIONS

The purpose of this article was to review how the main components of Elasticsearch and Apache Lucene interact to provide an effortless way to resolve relevance problems with search results.

Apache Lucene is the heart of Elasticsearch and provides an interface which helps with abstracting the complexity and algorithms behind the scenes.

For most business requirements, a default configuration of Elasticsearch will be sufficient. However, some cases may require improvements in how documents are scored. This kind of score tuning is possible, but such implementations require a deep understanding of the relevance algorithms used in Apache Lucene.

Public link

Please use the above public link if you want to share this noodl on another website.