Analyzing machine learning model performance

Descriptions of fixes

• Add dictionaries

  A dictionary contains example surface forms of a given entity type. You might need to add a new dictionary or add more entries to an existing dictionary if training statistics show that annotations of the type occur infrequently in the training data. If you know that the entity type is key to the domain and does occur often, then the omission might indicate that the surface forms associated with the type which are present in the training data are not being recognized by the machine learning model. Providing more surface form examples can help to solve this problem.

• Edit document sets

  Ensure that you have enough overall training data. To learn, a machine learning model requires sufficient content. Aim to provide around 300,000 words. (You can run a word count tool on your source documents to check the amount.) If your training set is too small, add documents to the corpus. It is not enough to have a handful of documents that exercise each important entity type or relationship; you want many documents that illustrate how those types are used in typical domain literature.

  • Ensure that the assortment of documents that is used for testing and training is similar. For example, do not use documents from one data source as the test data set and documents from another data source as the training data set. Different data sources might represent entity and relation type information in completely different ways. It is best to use a combination of documents from all of the data sources in both data sets to get the most comprehensive set of usage examples available from the industry literature. If the overall performances scores for a test run differ greatly from the training run, it might be due to inconsistencies in the data sets used for each.
  • Don't waste time with incorrectly formatted documents. Text documents that you add to the corpus must be in UTF-8 format. If you converted document from other formats to use UTF-8 encoding, you might see diacritic marks and other issues with character normalization. Such incorrect formatting can lead to inaccurate token representation. If the word tokens used to represent a mention differ between documents, it weakens the example and negatively impacts machine learning.

• Add type-specific documents

  If you have a low percentage of documents that contain a certain type, it might mean that your corpus is not fully representative. The machine learning model needs lots of examples to learn from. Adding more documents to the corpus might help. If your type system and documents are truly representative of a domain, you would expect to see that any randomly chosen documents contain a reasonable subset of the types. Although not true in all cases, this situation is a signal for you to investigate your type system and the quality of the documents in the corpus. You might need to find more domain documents that better exercise the types that are under-performing. If recall is low, it is often an indication that you need to add more documents.

• Annotate corpus more

  If the machine learning model is having trouble finding instances of certain types then it might be because the ground truth does not contain enough examples of the type's usage. Human annotators can sometimes be thorough about labeling entity type mentions, but be less diligent about annotating relation types and coreferences. For any key types that have a low percentage of corpus density, you might want to focus on finding more annotations of those type in the source documents. But, don't worry too much about coreference and relation accuracy if mention accuracy is deficient. Relation mentions between entities and coreferences of entities cannot be accurate unless entity mentions are accurate to begin with.

• Fix human annotations

  Check to see whether your training data is consistently and fully annotated. A machine learning model learns from your ground truth annotations. For example, if a sentence contains the phrase Obama family, and you label "Obama" as PERSON in one sentence and "Obama family" as PEOPLE in another sentence, the inconsistency means that the machine learning model cannot learn the correct annotation. Likewise, if you label "Obama" as PERSON in one sentence but do not label his name at all in another sentence, the annotation effort is incomplete and the machine learning model will be improperly trained. This type of inconsistent and partial labeling is sometimes referred to as type confusion. In most cases, just the act of having multiple human annotators review an overlapping set of documents will surface occurrences of type confusion-related mistakes. Pay attention to the issues that surface during document conflict resolution because they can provide insight into deeper issues with the type system itself. If there is no room to further improve or refine the type system, then it might be necessary to update the annotation guidelines and include examples. You can provide illustrations of common mistakes and how to annotate mentions properly under a given set of circumstances.

  Another indicator of annotation inconsistency is if you have enough annotations, but the corpus density is low. Density can be impacted when a mention that is significant in the domain literature occurs often, but is annotated as different types across the document set.

  Low precision is often an indication that you need to improve annotation consistency. To do so, review the annotation guidelines, better train the human annotators, and ensure that human annotators are working in concert and not in isolation from one another.

  Check the inter-annotator agreement score. This score, which measures the degree of agreement between different annotators' output on the same document, is a valuable number. Not only does this score tell you the quality of the ground truth documents that will be used to train the machine learning model, but it also indicates the upper bound of machine learning model performance. A model that is trained on these documents is unlikely to outperform the best agreement that humans can reach. For example, if performance persists at 75 and does not go higher, take a look at the inter-annotator agreement results. If the inter-annotator agreement score is 80, take actions to better train the human annotators and ensure that conflicts are correctly resolved (according to the annotation guidelines) during adjudication. If humans can't agree on how something should be labeled, then it's unlikely that a machine learning model will apply the correct labels.

• Enhance human annotator guidelines

  Clear and comprehensive annotator guidelines are a crucial part of a harmonious and successful annotation development effort. Human annotators have a tough job to do. There can be nuances in assigning entity and relation types that are hard to anticipate until you start to work with the domain documents. The guidelines can provide a sanity check to human annotators as they evaluate documents. The guidelines should be a living and changing document, especially at the beginning of the annotation process. They provide a key feedback loop because a human annotator can capture things she learned while annotating a few documents, then as she or someone else annotates a few more documents, new tips and gotchas can be added to the guideline, and so on. Be sure to include examples of difficult decisions and their preferred resolutions. The best way to determine what you need to add to the annotation guidelines is to carefully review document conflicts. Real examples of annotations that real people disagreed upon and how they were resolved can be a great help to human annotators as they tackle the annotation of new documents.

• Update type system

  You might need to update the type system for these reasons:

  • The documents that comprise the training data have references to concepts that are important types in the domain but are not represented anywhere in the type system. This suggests that you might need to add types that capture the missing concepts or relationships. Be careful not to try to define a type for every concept in a field, or every entity that occurs in domain documents; the type system should be limited to only the most fundamental types.
  • An existing type is being consistently misused by human annotators. If a type consistently causes confusion, then you might need to rename it or eliminate it if it is redundant.
  • An existing type is never used by human annotators because it is never referenced in the documents. If the type is unlikely to ever be used in literature from this domain, then remove it from the type system.
  • Two types are often interchanged when human annotators annotate documents. Consider whether the two types could be consolidated into one type that accurately represents the concept or relationship. For example, if the type system contains both PERSON and PEOPLE, which are often used interchangeably, it might be best to use one type named PERSONPEOPLE that covers both cases instead of two separate types.

  Use caution when you update the type system. If you update it after human annotators have evaluated documentation sets that were associated with the old type system, the human annotators will have to re-evaluate the documentation sets. Be sure that the changes you need to make are significant enough to warrant such rework.

• Investigate further

  If your type system, dictionary, and source documents are complete, and human annotation has been done well, but the machine learning model still performs poorly, then something might have gone wrong in the model training process. For example, you should always see high overall scores (over 95%) when testing on the training data.

Symptoms

An F1 score reaches its best value at 1 and worst value at 0. A low F1 score is an indication of both poor precision and poor recall. The machine learning model generates erroneous annotations and fails to find annotations that it should have found.

Causes

Low F1 scores can occur for many different reasons that depend on the domain, type system complexity, appropriateness of training documents, human annotator skills, and other factors.

Resolving the problem

Tune the performance of your machine learning model by performing one or more of following steps, and then retraining your model:

1. Identify commonly occurring types with low accuracy.

   When analyzing relations, look at both the F1 score of the relation type itself and at the F1 score of each of the two entities that participates in the relationship.

2. Identify commonly confused types. This information can be found by looking at the numbers that are off the diagonal in the confusion matrix.

3. Review errors where the machine learning model has high confidence.

4. Find patterns in the false negatives and false positives in the confusion matrix.

5. If certain types occur infrequently in the training data, add training data that contains those types.

   You can determine the frequency of occurrence by checking the percentage statistics (there are three: % of annotations, % of corpus density, and % of documents) for the type.

6. If certain types have low F1 scores, review the clarity of the annotation guidelines that apply to those types.

7. Add a dictionary for types that occur infrequently in the training data.

Figure 1. How to resolve low F1 scores

Symptoms

A precision score reaches its best value at 1 and worst value at 0. A low precision score indicates that the machine learning model generated incorrect annotations.

Causes

Low precision scores can occur for many different reasons that depend on the domain, type system complexity, appropriateness of training documents, human annotator skills, and other factors.

Resolving the problem

Tune the performance of your machine learning model by performing one or more of following steps then retrain your model:

1. Identify commonly occurring types with low precision.
2. Identify commonly confused types. This information can be found by looking at the numbers that are off the diagonal in the confusion matrix.
3. Review errors where the machine learning model has high confidence.
4. Find patterns in the false negatives in the confusion matrix.
5. If certain types have low precision scores, review the clarity of the annotation guidelines that apply to those types.

Figure 2. How to resolve low precision scores

Symptoms

A recall score reaches its best value at 1 and worst value at 0. A low recall score indicates that the machine learning model failed to create annotations that it should have created.

Causes

Low recall scores can occur for many different reasons that depend on the domain, type system complexity, appropriateness of training documents, human annotator skills, and other factors.

Resolving the problem

Tune the performance of your machine learning model by performing one or more of following steps then retrain your model:

1. Identify commonly occurring types with low recall.
2. Identify commonly confused types. This information can be found by looking at the numbers that are off the diagonal in the confusion matrix.
3. Review errors where the machine learning model has high confidence.
4. Find patterns in the false positives in the confusion matrix.
5. If certain types have low recall scores, review the clarity of the annotation guidelines that apply to those types.

Figure 3. How to resolve low recall scores