AUDRIS: The Army Unclassified Doctrine and Regulations Interaction System

This concept was developed for the User Interface Software and Technology course for the Master of Human-Computer Interaction and Design program (MHCI+D). It was inspired by CPT James Tollefson’s piece for the Military Review, “Fixing Army Doctrine: A Network Approach”.

Long before the Amazon Echo permeated our homes, psychologist and computer scientist J. C. R. Licklider envisioned a future where humans and computers could interact through voice user interfaces (VUIs). In his piece, “Man-Computer Symbiosis,” he explains that since one could “hardly take a military commander or a corporation president away from his work to teach him to type,” it was important for communication to be available to “top-level decision makers … via the most natural means, even at considerable cost” (Licklider, 1960).

Today, typing on a keyboard is one of the most common ways to interact with a computer; however, with the popularity of “thumb” typing, swipe input, predictive text and speech recognition, our time with physical keyboards has become fragmented. This pattern of change is also not withheld from Licklider’s original target stakeholder — military members — whose computer habits reflect general trends among Americans (Edwards-Stewart et al., 2016). While Licklider originally envisioned a military commander who sought assistance from a VUI to direct a battle, there is a more common scenario where all military members — from lower enlisted to field grade officers — would benefit on a regular basis: interpreting Army doctrine and regulations.

The Army’s doctrinal database is currently maintained by the Army Publishing Directorate (APD) in Fort Belvoir, VA, and while accessing these regulations may seem like a non-issue, the task of finding answers to inquiries is anything but straightforward. With over 540 publications (Tollefson, 2018), one can search through Army Doctrine Publications (ADPs), Army Doctrine Reference Publications (ADPRs), Field Manuals (FMs), Army Regulations (ARs), Joint Publications (JRs), and Army Tactics/Techniques Publications (ATPs/ATTPs) to find answers to their questions. Nearly all of these documents have been updated within the last couple of years. This also doesn’t include Training and Doctrine Command Regulations (TRs), Soldier Handbooks (SHs), All Army Activities (ALARACTs), Army Directives, Student Texts (STs), Supplemental Manuals (SMs), and unit-level Standard Operating Procedures (SOPs), which offer guidance.

Recently, these documents were reorganized from the Doctrine 2015 effort, and in “Fixing Army Doctrine: A Network Approach” for the Military Review, Captain James Tollefson of the Alaska Army National Guard explains that though the intent of this reorganization was to help maintain up-to-date regulations by clustering information into smaller chunks, it has continued to cause confusion. CPT Tollefson tells of a field grade officer who referred to a regulation that, unbeknownst to him, was superseded by a field manual with considerable changes (Tollefson, 2018). I personally experienced this confusion while attending the Human Resources Senior Leaders Course where confusion affected exams, and the Army’s new physical fitness test has been swarmed with misinformation as new regulations are released at different intervals and commands receive new information from those who attended the most current training.

While Doctrine 2015 has created a better system that allows supporting regulations to stay up to date without affecting lasting guidelines, CPT Tollefson advocates in his article for a networked approach to maintaining publications, as well as a user-friendly tool that would help Soldiers find appropriate and consistent guidance (Tollefson, 2018). His suggestions are the main inspirations for the solution detailed in this paper.

Army Training Doctrine: Past & Future
Retrieved from https://www.slideserve.com/ojal/training-units-and-developing-leaders-adp-and-adrp-7-0-doctrine-update

Technological Overview and Limitations

The Army Unclassified Doctrine and Regulations Interaction System (AUDRIS) is a collection of all current unclassified volumes published by the Army Publishing Directorate (APD). It is accessed primarily through a voice user interface via an Android application. For the scope of this paper, several limitations are considered:

  • AUDRIS will only maintain unclassified data. Environments that use the Secret Internet Protocol Router Network (SIPRNet) and Joint Worldwide Intelligence Communications System (JWICS) will not be supported as mobile phones are restricted from these areas. Access to protected databases such as Command Strength Management Modules and Digital Training Management Systems will also not be supported to protect Personally Identifiable Information that would be exposed through voice interaction to others who are not on a need-to-know basis.
  • AUDRIS will be accessed through an Android mobile app with voice as the primary interaction. However, preferably it would be accessed through a website interface and on iPhones as well. Android is currently the most popular platform for Soldiers (Edwards-Stewart et al., 2016). Next steps include these interfaces as well as support for communication devices used in forward operations.
  • The database will be fed only current Army-wide doctrine and updates from the Army Publishing Directorate. These include ADPs, ADRPs, FMs, ATTPs, ATPs, ARs, JPs, SHs, and Army General Orders (AGOs). Support for ALARACTs, TRADOC regulations, STs, SMs, SOPs, and other types of orders would require multi-organization support and unit-level personalization that is outside the scope of this paper.
  • AUDRIS will not feature “always on” capability. This is intended to reduce the frequency of accidental inquiries. It also means AUDRIS will not use a wake word.
  • AUDRIS will initially support English, since this is the official language of US military documents; however, future language support will include Spanish for our units in Puerto Rico, as well as Korean, German, Arabic, Kurdish, Japanese, Italian, and others.
  • AUDRIS will offer answers based on textual information. Audio description for graphs and images would require a computer vision machine learning (ML) model, which is intended for later versions.
  • If this was a DoD effort, a bidding process would be conducted to determine a service provider. Since Microsoft is a current military contractor, Azure APIs will be used as examples for speech recognition, processing and generation. Other ML models such as individual speaker recognition, sentiment analysis, translation, and web search are outside the scope of this solution. A method of automatic summarization will be provided by the SQuAD question answering service explained below.

Technical Elements

AUDRIS consists of many parts:

  • A database managed by the Army Publishing Directorate on Azure servers.
  • A mobile app written in Java for Android devices which includes a basic visual user interface that can confirm spoken inquiries and display information found within the regulations.
  • Azure Cognitive Services including:
  • Speech-to-text, speech recognition and text-to-speech through the Java SDK.
  • Language Understanding Intelligent Service (LUIS), which creates intents, entities, utterances and phrase lists from spoken inquiries, as well as syntactic processing of part-of-speech tagging and noun phrase chunking.
  • Text Analytics API for semantic processes like named entity recognition (NER) and keyphrase extraction.
  • A question answering service. Though QnA Maker is a Microsoft product, it is currently a supervised chatbot that retrieves information from FAQs and pre-made Excel sheets. AUDRIS will instead use the Stanford Question Answering Dataset (SQuAD) (Rajpurkar, P. et al., 2016), which offers span-based information retrieval, list and cloze-style (fill-in-the-blank) reading comprehension, and confidence scores based on supervised neural networks trained on a closed-domain dataset through crowdsourced questions and answers provided by the Army’s thousands of Human Resources experts.

Example Scenario

SPC Nguyen has been on a temporary profile (a documented limitation of physical duties) and is seeking guidance from his team leader SSG Vasquez about whether he should receive a permanent profile and go to a medical review board. SSG Vasquez begins by searching AR 40–501 “Standards of Medical Fitness” and Army Directives 2016–07 “Redesign of Personnel Readiness and Medical Deployability,” though her access to armypubs.army.mil may be limited if she is using a personal computer. As she reads through AR 40–501, she sees it was updated last year which resulted in half of its information moving to AR 40–502 “Medical Readiness” and DA Pam 40–502 “Medical Readiness Procedures.” After locating the correct regulations, she finds her answer: once SPC Nguyen is on profile for one year due to the same injury, he will need a permanent profile from his doctor, and he will reach an administrative Medical Retention Determination Point (DoA, 2019). His prospects of going to a military medical review board will be determined by many other factors which will require SSG Vasquez to spend at least another hour searching through several other regulations that are referenced by the three she has already found.

If SSG Vasquez used AUDRIS for assistance, her process could instead be as follows:

  • SSG Vasquez opens the AUDRIS app on her phone.
  • She presses the Listen Button which allows the app to record her question. She asks, “How long can a Soldier have a temporary profile before it becomes a permanent profile?”
  • The recording of her voice is sent to the Azure server where it is parsed into phonemes and converted into words. The LUIS and Text Analytics APIs parse her question into parts of speech and keywords and utilizes semantic rules to find named entities and relationships to extract the question’s intent.
  • SQuAD then uses terms like “temporary profile” and “permanent profile” to search through the regulations and discovers the ones SSG Vasquez found above. It narrows down the search by translating “how long” into compatible search terms such as time periods and durations. Eventually, it lands on possible answers from DA Pam 40–502. Using ML algorithms for answer extraction and ranking, it determines a confidence score for each answer candidate and chooses the highest one.
  • The answer is generated into speech through Azure’s text-to-speech Java SDK, responding with, “The maximum duration of temporary profiles is 12 months for the same medical condition. This information was found in DA Pam 40–502 paragraph 4–4, subparagraph c. (1) (DoA, 2019). Does this answer your question?” The page of this regulation will also be displayed on SSG Vasquez’s mobile screen. To support multi-turn conversations, the app will continue to record the conversation.
  • SSG Vasquez then asks, “Is this the only regulation that mentions profiles?”
  • AUDRIS will respond with information found from the previous search: “Regulations that discuss medical profiles are AR 40–501, AR 40–502 and DA Pam 40–502. Do you need any other information?”
  • SSG Vasquez can either pause AUDRIS to talk to SPC Nguyen or continue the conversation. She asks, “When will someone with a permanent profile need to go to a medical review board?”
  • AUDRIS will once again parse the regulations for the term “medical review board” while remembering the previous information. It will find no results for “medical review board” and search for synonyms instead. It will find terms like “medical readiness classification,” “medical retention determination point” and “medical deployability” within the glossaries of these regulations. By searching for the previous keywords and these synonyms, it will find information that contains a lower confidence score and announce this to SSG Vasquez:
  • “There is no specified timeline, but if the Medical Retention Determination Point is met, and the Soldier meets the medical retention standards of AR 40–501, the permanent profile must address the requirements of the Military Occupational Specialty or Area of Concentration and may indicate referral to an administrative board in accordance with AR 635–40. This information was found in DA Pam 40–502 paragraph 4–4, subparagraph c. (2) a. (DoA, 2019). Would you like to know more about medical retention standards?”
  • From there, SSG Vasquez can continue to delve deeper into the regulations with guided questions offered by AUDRIS. Her conversation also allows AUDRIS to actively learn what users find important to prioritize this information, as well as log terms that are not in the regulations but are still used in conversations. Once SSG Vasquez has finished, she can close the application, and it will stop listening.

Use Cases

There are several use cases for this type of interaction with the Army’s doctrine. Personnel Readiness and Training Officers and Sergeants work with these regulations on a daily basis, and while many of these professionals have memorized the names of the publications they use most, rarer questions can lead to time-consuming research. Especially for Unit Administrators who oversee the readiness and training of several subordinate units, these leaders are often overloaded and unable to answer individual questions until they have completed more pressing matters around Soldiers’ pay, schools, and families.

At the unit level, there are several people who could benefit from this system, especially when they are familiar only with Army regulations that affect their occupational specialty. Company Commanders, First Sergeants and team leaders may need assistance with guiding Soldiers in their careers and answering questions about medical, training and administrative issues like the above scenario. Lower enlisted Soldiers who are studying for a promotion board or Best Warrior Competition may want to use AUDRIS as a study guide to help them test their knowledge. Mechanics may want eyes-free assistance with checklists when performing preventative maintenance checks and services. Soldiers who are re-enlisting could receive up-to-date information about bonuses, deadlines and open positions, which would free up phone calls to Retention staff (this information would be implemented in the next steps). Drill Sergeants, who are expected to have in-depth knowledge of the regulations, could receive a monthly summary of doctrinal changes to stay up to date with training standards.

Benefits

In addition to the use cases above, AUDRIS has several other benefits:

  • It allows Soldiers to easily access unclassified regulations at any time while allowing them to concentrate on the question at hand and expend less time and clerical labor in seeking the answer.
  • It offers guided instruction on how to navigate the regulations.
  • It maintains a networked database that stays up to date with changes provided by the Army Publishing Directorate and flags discrepancies between new and old policies.
  • It gives Soldiers at-home access to regulations since the APD website is blocked from personal computers.
  • It relies on a closed-domain question answering system that allows AUDRIS to become well-versed in military jargon and actively learn from Soldiers’ questions about what information is most sought out.
  • It lets Soldiers choose to hear the regulations with either their many acronyms or with these terms spoken in their complete forms.

Harms Assessment and Value Tension

Beyond the limitations mentioned at the beginning of this paper, there are several harms and value tensions that could arise from this interaction model. These harms fall into two groups of concerns that are based on military social norms and Army-wide access:

Social Norms

  • Since questions are asked out loud rather than searched privately, this may prevent some military members from using the system since it may expose their lack of knowledge.
  • Interpreting regulations requires understanding of processes, interpersonal relationships, and decision-making norms. Human intervention and expertise will always be required for important matters.
  • Access to an AI expert may cause Soldiers to not ask others for help when they should. It may also cause some to constantly second-guess others or settle debates in the “Google style” since AUDRIS is readily available on their phones.
  • It may cause people who are unsure of their knowledge to treat AUDRIS like a safety net and second guess themselves in times when they are required to make decisions.
  • If the training model is incorrect, it could add gasoline to an already inflamed situation.

Access

  • Since it accesses Azure servers, it would require a reliable internet connection, which would be limited on Forward Operating Bases.
  • The Army Publishing Directorate does not control all published documentation; therefore, to support all publications, AUDRIS would require coordination with multiple organizations.
  • Several members of the military are hard of hearing, so information would need to be displayed visually as well. However, mobile device screens may not be large enough for users to sufficiently see the references.
  • Since AUDRIS is a text-based Natural Language Processing system, it would not audibly describe graphs, maps, etc. A computer vision machine learning model would need to be implemented in future versions to supply this information.

Conclusion

When focusing on technology that affects the military, we often — just like Licklider — jump to visions of war. However, Soldiers’ abilities to access and understand the regulations affect them on a daily basis, and low morale can be tied to one’s lack of understanding, which causes people to miss important updates for promotions, training, readiness and other administrative areas. AUDRIS can help bring this data to one easy-to-access place, hold its discrepancies accountable, save time in correspondence, and offer access and guidance for those who want to learn but don’t know where to start.

References

Edwards-Stewart, A., Smolenski, D.J., Reger, G.M., Bush, N., & Workman, D.E. (2017). An Analysis of Personal Technology Use by Service Members and Military Behavioral Health Providers. Military Medical, vol. 181, iss. 7, pp. 701–709.

Licklider, J. C. R. (1960, March). Man-Computer Symbiosis. IRE Transactions on Human Factors in Electronics, vol. HFE-1, no. 1, pp. 4–11.

Rajpurkar, P., Zhang, J., Lopyrev, K., & Liang, P. (2016, October 11). SQuAD: 100,000+ Questions for Machine Comprehension of Text. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing (EMNLP), version 3. Stanford University.

Tollefson, James. (2018). Fixing Army Doctrine: A Network Approach. Military Review, ed. January-February 2018, pp. 72. Army University Press.

U.S. Department of the Army. (2019, June 27). Department of the Army Pamphlet 40–502: Medical Readiness Procedures, pp. 15. Department of the Army Headquarters, Washington, DC.

Additional Resources

Manning, C. (2019, March 21). Stanford CS224N: NLP with Deep Learning | Winter 2019 | Lecture 10 — Question Answering. Stanford. Retrieved on March 16, 2020 from https://www.youtube.com/watch?v=yIdF-17HwSk.

­­Microsoft. (2020). Azure Cognitive Services. Retrieved on March 16, 2020 from https://docs.microsoft.com/en-us/azure/cognitive-services/.

Sharma, Y. & Gupta, S. (2018). Deep Learning Approaches for Question Answering System. Procedia Computer Science 132, pp. 785–794. Elsevier.

Wikipedia. (2020). Question Answering. Retrieved on March 16, 2020 from https://en.wikipedia.org/wiki/Question_answering.

Accessibility researcher