Issue link: https://content.lighthouseglobal.com/i/1412483
WHITE PAPER: TAR + ADVANCED AI – THE FUTURE IS NOW | 8 the train station," and "I'm going to train at the gym," are classified the same way – because statistically, there is not much difference between how the word "train" is placed within both sentences. However, newer tools that combine supervised machine learning with deep learning and NLP can learn the context of when the word "train" is used to mean the noun "train," and when the word is being used as a verb. Additionally, ediscovery tools that leverage deep learning and NLP can analyze more than just the text of the document – they can analyze the document from a variety of angles, including the metadata and data source when processing and learning from data. Thus, rather than throwing all data into a blender and extracting word meanings from statistics, tools that utilize deep learning and NLP can recognize that a word used in a chat platform may produce different results than the same word used over email. This ability can be especially useful in identifying privilege, as attorneys may speak differently over an informal chat system while discussing fantasy baseball with a large group of co-workers than within an email conveying legal advice to one person within the company – even if statistically, the word usage is similar. The context of the data source and how words are used matters, and an advanced AI tool that leverages a combination of technologies can better understand that context. Given the above, the real conversation can then shift from TAR vs. AI to "what kind of AI is best to use to execute a TAR workflow?" And we'll do just that in the next section where we outline the "when" and "why" of how to apply various subsets of AI in TAR workflows. Using advanced AI for TAR While advanced AI can be more effective, and often necessary – particularly for large and/or complex matters – many attorneys remain hesitant to move away from traditional TAR technology. As previously noted, there are a variety of factors behind this hesitancy: familiarity and comfort with the older technology, distrust from lack of technical understanding, fear of the inability to get stakeholders on board, concerns about a learning curve, etc. However, we believe that most of these factors can be overcome by education, transparency, and support. Part of this education is understanding that the process of using TAR in combination with more advanced AI is not much different from an end-user perspective. TAR 1.0 and TAR 2.0 workflows, wherein subject matter experts or reviewers code documents and the tool learns from those decisions to classify the unreviewed documents, can still be applied with a tool that uses advanced AI. The difference is in the results – as a combination of TAR workflows plus advanced AI can provide results more accurately and quickly with increased flexibility. Again, this is because a tool that uses NLP, deep learning, and machine learning can analyze each document from a variety of different angles, including understanding the context of how language is used rather than simply the definition of one-off words. Thus, for buy-in from other stakeholders, the key is finding the right partner who can not only support the implementation of the AI tool, but can also explain how the technology works, as well as the statistical defensibility of the results it yields. DATA REUSE Often companies have hundreds of thousands of previously coded documents sitting unused and dormant, effectively locked in archived or inactive databases. When this data can be leveraged to train a more advanced AI tool, it allows legal teams in the current case to immediately reduce the need for eyes-on review of larger portions of documents. In this scenario, a tool that uses advanced AI can amass a wealth of knowledge by ingesting and analyzing the previous attorney review decisions on each document, as well the metadata, language use, data source type, etc., prior to making the traditional TAR statistical predictions in the current case. The outcome is much more accurate predictions on a much wider variety of classifications than had previously been seen with technology within the ediscovery space. This means that not only are advanced AI tools that can reuse data from past matters better at detecting responsiveness, they can also more accurately detect attorney- client privilege, personal information, key documents, and trade secret information. For example, when a process is in place that allows for the ingestion and analysis of previous attorney-client privilege decisions, advanced AI tools can not only immediately pinpoint similar or identical documents to those that were coded