Analytics indexes

LSI is a wholly mathematical approach to indexing documents. Instead of using any outside word lists, such as a dictionary or thesaurus, LSI leverages sophisticated mathematics to discover term correlations and conceptuality within documents. LSI is language-agnostic, meaning that you can index any language and it learns that language. LSI enables Relativity Analytics to learn the language and, ultimately, the conceptuality of each document by first processing a set of data called a training data source. The training data source may be the same as the set of documents that you want to index or categorize. Alternatively, it may be a subset of these documents, or it could be a completely different set of documents. This training data source is used to build a concept space in the Analytics index.

Using LSI, Analytics inspects all the meaningful terms within a document and uses this holistic inspection to give the document a position within a spatial index. The benefits of this approach include the following:

• Analytics learns term correlations, interrelationships, and conceptuality based on the documents being indexed. Therefore, it always is up-to-date in its linguistic understanding.
• Analytics indexes are always in memory when being worked with, so response time is very fast.
• Analytics is inherently language agnostic. It can index most languages and accommodate searches in those same languages without additional training. We recommend creating separate indexes for large populations of different language documents.

Concept space

When you create an Analytics index, Relativity uses the training data source to build a mathematical model called a concept space. The documents you are indexing or categorizing can be mapped into this concept space. While this mathematical concept space is many-dimensional, you can think of it in terms of a three-dimensional space. The training data source enables the system to size the concept space and create the algorithm to map searchable documents into the concept space. In the concept space, documents that are closer together are more conceptually similar than documents that are further from each other.

Concept rank

Throughout Analytics, item similarity is measured using a rank value. Depending on the feature, the rank may be referred to as a coherence score, rank, or threshold. In each scenario, the number represents the same thing.

Because the Analytics engine builds a spatial index, every document has a spatial relationship to every other document. Additionally, every term has a spatial relationship to every other term.

The concept rank is an indication of distance between two items. In the Categorization feature, it indicates the distance between the example document and the resulting document. In Keyword Expansion, it indicates the distance between the two words. The rank does not indicate a percentage of relevance.

For example, when running a concept search, the document that is closest to the query is returned with the highest conceptual score. A higher score means the document is more conceptually related to the query. Remember that this number is not a confidence score or a percentage of shared terms, it is a measurement of distance.

SVM is used by Active Learning and Review Center. With SVM, you do not need to provide the Analytics index with any training text. The system learns from your reviewers and constantly updates the model. SVM predicts the relevance of uncoded documents based on their distance to the hyperplane. This differs from Latent Semantic Indexing (LSI) which is also composed of a multi-dimensional space, but uses a nearest-neighbor approach to predict documents.

Hyperplane

When you create a Classification-type index, Relativity takes reviewer's coding decisions and pulls them into a high-dimensional model. The hyperplane helps differentiate between relevant and not relevant documents.

After the hyperplane is established, all documents without a coding decision are pulled into the model and mapped on either side of the hyperplane based on the model's current understanding of the difference between relevant and not relevant.

Rank

Rank measures the strength or confidence the model has in a document being relevant or not relevant. Rank is measured on a scale from 100 to 0. As documents move farther away from the hyperplane on either side, their score is pushed closer to 100 or 0.

1. Click the Indexing & Analytics tab and select Analytic Indexes.
2. Click New Analytics Index. The Analytics Index Information form appears.
3. Complete the fields on the Analytics Index Information form. See Index creation fields. Fields in orange are required.
4. Click Save.
   
   Note: If no documents appear in the saved search, or if the search contains fields other than the extracted text field, a warning message appears upon clicking Save.

When you save the new index, the Analytics Index console becomes available. See Analytics index console operations.

Index creation fields

The Analytics Index Information form contains the following fields:

Index Information fields

• Name—the name of the index. This value appears in the search drop-down menu after you activate the index.
• Index type—select Conceptual.
• Data source—the document set searched when you use the Analytics index. This set should include all of the documents on which you want to perform any Analytics function. Only documents included in the data source are returned when you run clustering, categorization, or any other Analytics feature. See Data source considerations for more information on creating an optimized data source search.
• Cluster Documents—when checked, this automatically creates a cluster for all indexed documents after the index has been built. This option is set to Yes by default.
  • Yes—creates a conceptual cluster for all documents included in the index. For more information, see Default settings for automatically created clusters.
  • No—no cluster is created when the index builds, but you can still create a cluster manually at any time. For more information, see Clustering.
• Order—the number that represents the position of the index in the search drop-down menu. The lowest-numbered index is at the top. Items that share the same value are sorted in alphanumeric order.
• Email notification recipients (under Optional Settings)—send email notifications when your index successfully completes, fails, or when the index is disabled because it is not being used. Enter the email addresses of the recipients. Separate entries with a semicolon. A message is sent for both automated and manual index building.

Advanced Settings fields

• Repeated content filters to link—the number of repeated content filters to link to the index. When the index runs, it will automatically link the top filters found in the Repeated Content Filters tab, sorted in descending order by number of occurrences times word count. For more information, see Automatically linking repeated content filters.
  • The default value is 200; the maximum value is 1,000. If you also have manually linked filters, those count towards the maximum of 1,000.
  • In order for this setting to take effect, you must have either a structured analytics set with repeated content identification, or manually created repeated content filters. For information on creating these, see Repeated content filters.
• Training data source—the document set from which the Analytics engine learns word relationships to create the concept index. By default, this is the same saved search as the data source. See Training data source considerations for more information including considerations when using Data Grid.
• Optimize training set—when checked, this specifies whether the Analytics engine automatically includes only conceptually valuable documents in the training data source while excluding conceptually irrelevant documents. Enabling this eliminates the need to manually remove bad training documents from the training data source.
  • Yes—includes only quality training documents in the training data source. For more information, see Optimize training set.
  • No—puts all documents from the saved search selected for the Training data source field into the training data source without excluding conceptually-irrelevant documents.
• Dimensions—determines the dimensions of the concept space into which documents are mapped when the index is built; more dimensions increase the conceptual values applied to documents and refine the relationships between documents. The default setting is 100 dimensions.
  
  Note: A larger number of dimensions can lead to more nuances due to more subtle correlations that the system detects between documents. However, higher dimensionality has a diminishing return on results once you exceed 300 or more dimensions.
• Remove English email signatures and footers—removes signatures and footers from emails containing English characters only. By default, this is set to Yes for new indexes and No for existing ones. Enabling this tells the Analytics engine to auto-detect and ignore email confidentiality footers and signatures. This helps the index focus on authored content.
  • This process can slow down index population, but we strongly recommend leaving it set to Yes.
  • Setting this to No makes it possible to turn off the Enable email header filter setting.
    
• Enable email header filter—removes common header fields, such as To, From, and Date, and reply-indicator lines. It does not remove content from the Subject line or the attachment header. This helps the index focus on authored content instead of making unhelpful correlations among standard words such as To or From.
  • By default, this is set to Yes. We recommend keeping this filter set to Yes on all indexes.
  • To set this option to No, you must set Remove English email signatures and footers to No.
    
    Note: Relativity recognizes words as being part of the email header only when they are at the beginning of a new line and followed by a colon. The email filter also filters out subsequent lines, but only if the new line begins with white space. For example, if the email contains a BCC section that is multiple lines, each subsequent line would have to begin with a space, otherwise it is not filtered out.
• Stop words—determines the words you want the conceptual index to suppress. You can add or remove stop/noise words from the list. Separate each word with a hard return. If you modify this list after the index builds, you will need to re-build the index.
  Stop/noise words are common terms that are filtered from the Analytics index in order to improve quality. By default, this field contains only English words. If you are indexing documents in another language, you can customize the list to include words in the language you want. For multiple languages, we recommend creating multiple indexes in the workspace template for each commonly indexed language. These indexes are copied over to any new workspace that uses that template.
  To create a stop/noise word list in another language, there are several options:
  • Translating the English stop/noise words list to the desired language manually.
  • Ranks NL (http://www.ranks.nl/stopwords) - 40 languages are available.
  • Microsoft SQL Server Stop Lists - 33 languages are available. Use the following query:
    
    Copy

    SELECT LCID, Name FROM sys.syslanguages
    SELECT * FROM sys.fulltext_system_stopwords WHERE language_id = ####

Note: Classification indexes are used only by Active Learning projects.

1. Click the Indexing & Analytics tab and select Analytic Indexes.
2. Click New Analytics Index. The Analytics Index Information form appears.
3. Complete the fields on the Analytics Index Information form. See Index creation fields. Fields in orange are required.
4. Click Save.
   
   Note: If no documents appear in the saved search, or if the search contains fields that would cause the index to error, a warning message appears upon clicking Save.
   

When you save the new index, the Analytics Index console becomes available. See Analytics index console operations.

Index creation fields

The Analytics Index Information form contains the following fields:

Index Information fields

• Name—the name of the index. This value appears in the search drop-down menu after you activate the index.
• Index type—select Classification.
• Data source—the document set searched when you use the Analytics index. This set should include all of the documents on which you want to perform any Analytics function. Only documents included in the data source are returned when you run clustering, categorization, or any other Analytics feature. See Data source considerations for more information on creating an optimized data source search.
  Notes:

  • For best results, we recommend no more than 9 million documents in the data source.
  • If you want to use family-based review in Active Learning, parent documents and their family must all be added to the data source.
• Trained Model (optional)—select a trained model to jump-start the relevance predictions for the new project. For more information, see Reusing trained models.
• Order—the number that represents the position of the index in the search drop-down menu. The lowest-numbered index is at the top. Items that share the same value are sorted in alphanumeric order.
• Email notification recipients (under Optional Settings)—send email notifications when your index successfully completes, fails, or when the index is disabled because it is not being used. Enter the email addresses of the recipients. Separate entries with a semicolon. A message is sent for both automated and manual index building.

Training data source considerations

Note: Training data source considerations only apply to conceptual indexes.

A training data source is a set of documents that the system uses to learn the language, the correlation between terms, and the conceptual value of documents. This data source formulates the mapping scheme of all documents into the concept space. Because the system uses this data source to learn, include only the highest quality documents when creating the training data source. The system needs authored content with conceptually relevant text to learn the language and concepts.

Use the following settings when creating a saved search to use as a training data source:

• Only the authored content of the document should be indexed. Typically, the Extracted Text field is the only field returned as a column in a data source. If there is authored content in another long text field, such as OCR Text, Translation, or others, then this field should be included as well. You should never train on any sort of metadata fields like Custodian or Email Author or system fields like Control Number. Including these fields skews your results when trying to find conceptually related documents.
  
  Note: Do not index single choice, multiple choice, multiple object, or any number fields, such as Extracted Text Size.
  
• Documents that do not contain authored content or conceptual data are useless for training the Analytics index. The following types of documents should be excluded from the training data source:
  • Compressed files, like ZIP or RAR files
  • System files
  • Excel files with mostly numbers
  • Image files
  • CAD drawings
  • Maps
  • Calendar items
  • Documents with OCR errors, such as OCR not processed
  • Documents with poor quality OCR, such as OCR on handwritten documents.
  • Documents with little or no text, less than 0.2 KB
• During population, words beginning with a number are not included in the index. For example, 4ward. Words ending in a number, mp3, or with numbers embedded, passw0rd, are indexed as written.

Optimize training set

When you select the Optimize training set feature on an Analytics index, you improve the quality of that index by excluding documents that could result in inaccurate term correlations due to their low conceptual value, such as:

• Very short documents
• Very long documents
• Lists containing a significant amount of numbers
• Spreadsheet-like documents
• System log files
• Text resulting from processing errors

To perform this automatic removal of bad documents from the training data source, the Analytics engine evaluates documents based on:

• Word count
• Uniqueness
• Number count
• Punctuation marks
• Words with many characters, 50 or more

If the optimization excludes a document, the following results are possible:

• If the document is designated to be populated as both a training and searchable document, Relativity populates it as a searchable document only. The document could be returned in a concept search, assuming it meets the minimum rank.
• If the document is designated to be populated only as a training document, Relativity does not populate it into the index at all.
• If the document is designated to be populated as a searchable document only, the Analytics engine doesn't examine it.

Data source considerations

For conceptual indexes, the data source is the collection of documents to be clustered, categorized, or returned in a concept query. The data source is typically larger than the training data source. There are fewer documents culled from the data source.

For classification indexes, the data source is the collection of documents to be ranked by the Active Learning model. The data source must contain example documents to train the model.

Use the following settings when creating a saved search to use as a data source:

• Index only the authored content of the document. We recommend returning as few fields as possible. Typically, the Extracted Text field is the only field returned as a column in a data source. If there is authored content in another long text field, such as OCR Text, Translation, and others, include this field as well. Returning email header fields in your search may cause recipient names and email address components to appear in clusters.
  
  Note: Do not index single choice, multiple choice, multiple object, or any number fields, such as Extracted Text Size.
• Documents that do not contain any conceptual data cannot be returned by any conceptual analytics operations. Consider excluding the following types of documents from the data source:
  • Compressed files, like ZIP or RAR files
  • System files
  • Excel files with mostly numbers
  • Image files
  • CAD drawings
  • Maps
• During population, words beginning with a number are not included in the index. For example, 4ward. Words ending in a number, mp3, or with numbers embedded, passw0rd, are indexed as written.

Identifying data source and training data source documents

When you populate the index, the Conceptual Index multi-choice field on the Document object lists whether a document is included in the data source, training data source, or both. This field is populated every time the index is populated with a full or incremental population. You can use this field as a condition in a saved search to return only training or data source documents.

You can also find data source and training data source documents in the field tree, as well as those which were excluded from training when you enabled the Optimize training set field on the index.