Intent Recognition Systems

Intent Recognition Systems represent a critical component of AI Discoverability Architecture, serving as the interpretive layer that bridges human communication and machine understanding. These systems analyze user inputs—whether textual, vocal, or behavioral—to identify the underlying purpose or goal behind an interaction, enabling AI systems to respond appropriately and facilitate meaningful discovery of information, services, or capabilities. In the context of AI discoverability, intent recognition transforms ambiguous queries into actionable insights, allowing users to navigate complex AI ecosystems without requiring technical expertise or precise command structures. The significance of these systems has grown exponentially with the proliferation of conversational AI, virtual assistants, and intelligent search platforms, where understanding user intent determines the quality and relevance of discovered resources.

Overview

Intent Recognition Systems constitute a specialized domain within natural language understanding (NLU) that focuses on classifying user utterances into predefined categories representing specific goals or actions ¹². The emergence of these systems traces back to early dialogue systems and command-line interfaces, but their modern form evolved significantly with the advent of machine learning approaches and, more recently, transformer-based architectures like BERT ¹. The fundamental challenge these systems address is the inherent ambiguity and variability of human language—users express identical intents through countless linguistic variations, while superficially similar utterances may represent entirely different objectives.

Traditional keyword-matching approaches proved inadequate for capturing the semantic richness and contextual dependencies of natural language. The evolution toward supervised learning methods using support vector machines and conditional random fields marked an important transition, but the transformative shift occurred with deep learning and pre-trained language models ¹⁷. Modern intent recognition systems leverage contextual embeddings that capture semantic relationships beyond surface-level patterns, enabling generalization to novel phrasings and handling of complex, multi-domain scenarios ³⁶. This evolution has enabled the deployment of sophisticated conversational agents, virtual assistants, and intelligent search systems that understand user objectives with unprecedented accuracy.

Key Concepts

Intent Classes

Intent classes are discrete categories representing specific user objectives or goals that a system is designed to recognize and fulfill ²³. Each intent class defines a particular action or information need, such as "book_flight," "check_weather," or "request_refund." The design of intent taxonomies requires careful consideration of granularity, coverage, and mutual exclusivity to ensure effective classification.

Example: A banking virtual assistant might define intent classes including check_balance, transfer_funds, report_fraud, and locate_atm. When a user says "I need to move money from my savings to checking," the system classifies this utterance as the transfer_funds intent, distinguishing it from superficially similar requests like "I want to see how much is in my savings" (check_balance) or "Can you tell me where the nearest ATM is?" (locate_atm).

Slot Filling

Slot filling is the process of extracting specific parameters or entities from user utterances that are necessary to fulfill the recognized intent ²⁸. While intent classification identifies what the user wants to do, slot filling captures the details of how to execute that action. This component employs named entity recognition techniques to identify and classify relevant information within the input.

Example: For a flight booking system that recognizes the intent "book_flight" from the utterance "I need a flight from Boston to Seattle next Tuesday for two passengers," the slot filling module extracts: origin: Boston, destination: Seattle, date: next Tuesday, and passenger_count: 2. These extracted slots populate a structured representation that the booking system uses to search for appropriate flights.

Confidence Scores

Confidence scores are probabilistic measures that quantify the certainty of intent classification decisions ¹⁰. These scores enable systems to distinguish between high-confidence predictions that can be acted upon immediately and uncertain classifications requiring clarification or human intervention. Proper confidence calibration is essential for maintaining user trust and system reliability.

Example: A customer service chatbot receives the ambiguous query "I have a problem with my account." The intent classifier might assign confidence scores of 0.42 to technical_support, 0.38 to billing_inquiry, and 0.20 to other intents. Recognizing that no single intent exceeds the 0.70 confidence threshold, the system responds with a clarification question: "I'd be happy to help with your account. Are you experiencing a technical issue accessing your account, or do you have a question about billing?"

Context Awareness

Context awareness refers to the system's ability to incorporate conversational history, user preferences, and situational factors when interpreting intent ³⁹. This capability enables resolution of ambiguous references, maintains coherence across multi-turn dialogues, and personalizes intent interpretation based on user-specific patterns. Context management systems track dialogue state and environmental variables that influence meaning.

Example: In a multi-turn conversation with a travel assistant, a user first asks "What's the weather like in Paris?" (informational intent) and then follows up with "Book me a hotel there for next weekend" (transactional intent). The context-aware system resolves "there" to Paris based on the previous utterance, extracting the destination slot without requiring the user to repeat information. Additionally, if the user has a history of preferring boutique hotels, the system might adjust its search parameters accordingly.

Hierarchical Classification

Hierarchical classification structures intent taxonomies as trees or directed acyclic graphs, enabling coarse-to-fine prediction strategies ³⁶. This approach first identifies broad intent categories before refining to specific sub-intents, improving efficiency for large intent spaces and providing graceful degradation when precise classification is uncertain.

Example: An enterprise IT helpdesk system implements a three-level hierarchy. At the top level, it distinguishes between hardware_issue, software_issue, and access_request. Under software_issue, it identifies second-level categories like email_problem, application_error, and network_connectivity. Finally, under email_problem, it recognizes specific intents such as cannot_send_email, missing_emails, and attachment_issue. When a user reports "My Outlook keeps crashing," the system first classifies this as software_issue, then email_problem, and finally application_error, enabling appropriate routing even if the most specific classification is uncertain.

Multi-Intent Recognition

Multi-intent recognition acknowledges that single utterances may express multiple objectives simultaneously ³. Rather than forcing classification into a single category, these systems identify all relevant intents present in an input, enabling more comprehensive responses that address all user needs.

Example: A smart home assistant receives the command "Turn off the living room lights and set the thermostat to 68 degrees." A multi-intent recognition system identifies two distinct intents: control_lighting with slots location: living room and action: off, and adjust_temperature with slot target_temp: 68. The system executes both actions rather than forcing a choice between them or misinterpreting the compound request.

Few-Shot Learning

Few-shot learning methodologies enable intent recognition systems to adapt to new intent categories with minimal training examples ⁴⁵. These approaches leverage semantic similarity between intent descriptions and user utterances, or employ meta-learning frameworks that learn how to learn from limited data, addressing the cold-start problem when launching new capabilities.

Example: A customer service platform needs to add a new subscription_pause intent for a recently introduced feature allowing temporary account suspension. Rather than collecting thousands of labeled examples, the few-shot learning system uses just 10-15 example utterances like "I want to pause my subscription for two months" and "Can I temporarily suspend my account?" Combined with the intent description "User wants to temporarily halt their subscription without canceling," the system generalizes to recognize variations like "I need to freeze my membership while I'm traveling" with reasonable accuracy.

Applications in AI Discoverability

Virtual Assistant Task Automation

Virtual assistants like Amazon Alexa, Google Assistant, and Apple Siri employ intent recognition as the foundation for task automation and information retrieval ⁶. These systems must handle diverse domains—from smart home control to calendar management to general knowledge queries—requiring robust multi-domain intent recognition capabilities. The intent recognition layer determines which backend service or skill to invoke, extracts necessary parameters, and orchestrates complex multi-step workflows.

In practice, when a user tells their smart speaker "Remind me to call the dentist tomorrow at 2 PM," the system recognizes the create_reminder intent, extracts slots for the reminder content ("call the dentist"), date ("tomorrow"), and time ("2 PM"), then interfaces with the calendar service to create the appropriate notification. The discoverability aspect emerges when users explore capabilities through natural language rather than navigating menu structures—asking "What can you do?" triggers intent-based capability discovery.

Customer Service Chatbot Routing

Customer service chatbots leverage intent recognition to route conversations appropriately and extract relevant parameters for issue resolution ²¹⁰. These systems must distinguish between intents like "check order status," "request refund," "technical support," and "product inquiry," each requiring different backend systems and information. Accurate intent recognition directly impacts resolution time, customer satisfaction, and operational efficiency.

A telecommunications company's chatbot receives the message "My internet has been down since yesterday and I need this fixed immediately." The system recognizes the technical_support intent with urgency indicators, extracts the service type (internet) and duration (since yesterday), and routes to the appropriate technical support queue with priority flagging. If the user instead says "I want to cancel my service," the system recognizes the cancel_subscription intent and routes to retention specialists, demonstrating how intent recognition enables intelligent workflow orchestration.

E-Commerce Search and Navigation

E-commerce platforms employ intent recognition to distinguish between different shopping behaviors—product search, price comparison, purchase completion, and post-purchase support ⁶. Understanding whether a user has navigational intent (finding a specific product), informational intent (learning about product categories), or transactional intent (ready to purchase) fundamentally alters the optimal response strategy and interface presentation.

When a user searches for "wireless headphones under $100 with good battery life," the system recognizes a product search intent with specific constraints (price ceiling, feature requirement). The slot filling extracts product_category: wireless headphones, max_price: $100, and desired_feature: good battery life, enabling filtered search results. Conversely, "How do I return my headphones?" triggers a return_request intent, redirecting to customer service workflows rather than product listings, demonstrating how intent recognition enables appropriate resource discovery.

Healthcare Symptom Assessment

Healthcare applications utilize intent recognition in symptom checkers, appointment scheduling systems, and patient communication platforms ⁵. These systems must carefully distinguish between information-seeking intents ("What are the symptoms of flu?"), appointment-related intents ("I need to schedule a checkup"), and urgent medical concerns ("I'm experiencing chest pain"), with the latter requiring immediate escalation protocols.

A patient portal chatbot receives the message "I've had a persistent cough for three weeks and it's getting worse." The system recognizes the symptom_report intent, extracts clinical entities (symptom: cough, duration: three weeks, progression: worsening), and based on the duration and progression, recommends scheduling an appointment rather than providing general information. The intent recognition enables appropriate triage and resource discovery—connecting patients with relevant care pathways based on their underlying needs rather than explicit navigation.

Best Practices

Iterative Intent Taxonomy Refinement

Intent taxonomies should undergo continuous refinement based on production data and user behavior patterns ³⁶. Initial taxonomy design relies on domain expertise and anticipated use cases, but real-world usage inevitably reveals gaps, overlapping categories, and unexpected user needs. Establishing regular review cycles that analyze misclassified utterances, low-confidence predictions, and user reformulations enables data-driven taxonomy evolution.

Implementation Example: A financial services chatbot launches with 25 intent categories based on common customer service scenarios. After three months of production deployment, the team analyzes conversation logs and discovers that 15% of queries relate to cryptocurrency services—a category not represented in the original taxonomy. They also find significant confusion between check_transaction_history and dispute_transaction intents, with many utterances receiving low confidence scores. The team responds by adding a cryptocurrency_inquiry intent, merging the confused categories into a single transaction_inquiry intent with sub-types, and retraining the model with production examples, resulting in a 12% improvement in classification accuracy.

Confidence-Based Clarification Strategies

Systems should implement confidence thresholds that trigger clarification dialogues rather than acting on uncertain predictions ¹⁰. This practice prevents errors that damage user trust while gathering valuable training data. Effective clarification strategies present users with likely interpretations, ask targeted questions to disambiguate, or gracefully escalate to human agents when appropriate.

Implementation Example: An enterprise IT helpdesk chatbot establishes a three-tier confidence strategy: predictions above 0.85 confidence proceed directly to fulfillment, scores between 0.60-0.85 trigger confirmation ("It sounds like you need help resetting your password. Is that correct?"), and scores below 0.60 present multiple options ("I can help with several things: password reset, account unlock, or access request. Which best describes your need?"). This approach reduces error rates by 34% compared to always acting on the top prediction, while the confirmation dialogues provide labeled training data that improves future performance.

Multi-Task Learning for Joint Optimization

Intent classification and slot filling should be jointly optimized through multi-task learning frameworks that exploit their interdependencies ²⁷. The intent context helps disambiguate entity types, while recognized entities provide evidence for intent classification. Joint training enables information sharing between tasks, improving both compared to independent optimization.

Implementation Example: A travel booking system implements a multi-task BERT-based model with shared encoder layers and task-specific output heads for intent classification and slot filling. During training, the model learns that the presence of location entities strongly correlates with navigation or booking intents, while temporal entities appear frequently in scheduling intents. This joint learning improves slot filling F1 score by 8% and intent classification accuracy by 5% compared to separate models, while reducing inference latency by 40% since both tasks share the same encoding pass.

Active Learning for Efficient Data Collection

Implement active learning strategies that strategically select informative examples for annotation, maximizing model improvement per labeled instance ⁴. Uncertainty sampling identifies examples where the current model is most confused, diversity-based selection ensures coverage of the input space, and query-by-committee approaches leverage disagreement among ensemble members to identify valuable training examples.

Implementation Example: A customer service platform needs to expand its intent recognition to a new product line but has limited annotation budget. Rather than randomly sampling utterances for labeling, the team implements an uncertainty-based active learning pipeline that identifies examples where the model's top two predictions have similar confidence scores. They also incorporate diversity sampling to ensure coverage across different linguistic patterns. This approach achieves target performance with 60% fewer labeled examples compared to random sampling, enabling faster deployment while reducing annotation costs from $15,000 to $6,000.

Implementation Considerations

Platform and Framework Selection

Organizations must choose between end-to-end platforms (Rasa, Dialogflow, Amazon Lex, Microsoft LUIS) and custom implementations using ML frameworks ¹⁰. End-to-end platforms offer faster deployment, managed infrastructure, and integrated dialogue management but may limit customization and create vendor lock-in. Custom implementations using Hugging Face Transformers, TensorFlow, or PyTorch provide maximum flexibility but require significant ML expertise and infrastructure investment.

Example: A healthcare startup building a patient communication system evaluates options. They choose Rasa for initial development because it offers healthcare-specific compliance features, on-premise deployment for HIPAA compliance, and sufficient customization for their domain-specific terminology. However, they architect their system with abstraction layers that would enable migration to custom models if they later need specialized medical language understanding beyond Rasa's capabilities. A large technology company with extensive ML infrastructure instead builds a custom system using Hugging Face Transformers, enabling them to leverage proprietary training data and integrate tightly with existing personalization systems.

Domain-Specific Customization

Intent recognition systems require substantial domain adaptation to handle specialized terminology, communication patterns, and user expectations ³⁵. Generic pre-trained models provide strong baseline performance but must be fine-tuned on domain-specific data. This includes not only labeled intent examples but also domain-relevant pre-training on unlabeled text to adapt language representations.

Example: A legal technology company building a contract analysis assistant finds that general-purpose BERT performs poorly on legal language, struggling with archaic terminology, complex sentence structures, and domain-specific entity types. They implement a two-stage adaptation process: first, they continue pre-training BERT on 50,000 legal documents (contracts, case law, regulations) to adapt the language model to legal discourse patterns. Then, they fine-tune for intent classification using 10,000 labeled examples of lawyer queries categorized into intents like clause_search, precedent_lookup, compliance_check, and contract_comparison. This domain adaptation improves intent classification accuracy from 67% to 91% compared to using generic BERT.

Multilingual and Cross-Cultural Deployment

Global deployments require addressing linguistic diversity and cultural communication differences ³. While multilingual pre-trained models (mBERT, XLM-RoBERTa) enable cross-lingual transfer, cultural variations in how intents are expressed necessitate localized training data and evaluation. Direct translation of training data often proves insufficient due to cultural and pragmatic differences.

Example: A global e-commerce platform deploys intent recognition across 15 languages. They use XLM-RoBERTa as the base model, which provides reasonable zero-shot performance in new languages. However, they discover that German users tend to be more direct in customer service interactions ("Ich möchte stornieren" - "I want to cancel"), while Japanese users employ more indirect, polite formulations ("キャンセルについて相談したいのですが" - "I would like to consult about cancellation"). To handle these differences, they collect 2,000 labeled examples per language, focusing on culturally typical phrasings, and fine-tune language-specific models. They also discover that return/refund intents are expressed very differently across cultures, leading them to develop region-specific intent taxonomies rather than forcing a single global structure.

Performance Monitoring and Continuous Improvement

Production deployments require comprehensive monitoring beyond offline accuracy metrics ⁶¹⁰. Key performance indicators include classification confidence distributions, fallback rates, user reformulation patterns, task completion rates, and user satisfaction scores. Implementing A/B testing frameworks enables data-driven evaluation of model updates before full deployment.

Example: A banking virtual assistant implements a monitoring dashboard tracking: (1) daily intent distribution to detect shifts in user needs, (2) low-confidence prediction rates by intent category to identify problematic areas, (3) conversation abandonment rates correlated with intent recognition failures, and (4) user satisfaction ratings segmented by intent type. They discover that the investment_advice intent has a 23% abandonment rate compared to 8% average, investigation reveals the slot filling frequently fails to extract investment amounts and time horizons. They collect additional training data for this intent, implement custom entity extractors for financial amounts, and deploy the update via A/B test to 10% of users. The test shows abandonment drops to 11% and satisfaction increases by 18%, validating full deployment.

Common Challenges and Solutions

Challenge: Data Scarcity for Rare Intents

Many intent categories have limited training examples, particularly for newly introduced capabilities, edge cases, or specialized domains ⁴⁵. This data scarcity leads to poor performance on rare intents, creating an uneven user experience where common requests work well but unusual needs fail. The problem compounds because rare intents often represent high-value scenarios—complex problems that drove users to seek assistance.

Solution:

Implement a multi-pronged approach combining data augmentation, few-shot learning, and strategic data collection ⁴⁵. Use paraphrasing techniques and back-translation to artificially expand training sets for rare intents. Employ few-shot learning methods that leverage semantic similarity between intent descriptions and utterances, enabling reasonable performance with 10-20 examples rather than hundreds. For critical rare intents, implement targeted data collection campaigns—when users reach human agents, have agents label the intents and add successful resolutions to training data.

Example: A technical support chatbot has only 45 examples of the api_authentication_error intent compared to 3,000+ examples for common intents like password_reset. The team implements: (1) paraphrasing augmentation using a T5 model to generate 200 variations of the original 45 examples, (2) a few-shot learning component that matches user utterances against intent descriptions when confidence is low, and (3) a feedback loop where support agents tag escalated conversations with correct intents. Within two months, they accumulate 180 real examples, and the combined approach improves recall for this intent from 34% to 78%.

Challenge: Intent Taxonomy Overlap and Ambiguity

Poorly designed intent taxonomies create overlapping categories where utterances legitimately belong to multiple intents, or boundaries between intents are unclear even to human annotators ³⁶. This manifests as low inter-annotator agreement, inconsistent training labels, and confused models that produce low-confidence predictions for ambiguous regions.

Solution:

Conduct systematic taxonomy audits using inter-annotator agreement metrics and confusion matrix analysis ³. When multiple intents show consistent confusion, consider merging them into a single intent with sub-types, or restructure the taxonomy hierarchically. Develop clear intent definitions with positive and negative examples, and establish annotation guidelines that address edge cases. For genuinely ambiguous cases, implement multi-intent recognition rather than forcing single-label classification.

Example: A customer service system shows persistent confusion between billing_inquiry, payment_problem, and subscription_change intents, with inter-annotator agreement of only 0.62 (below the 0.80 threshold). Analysis reveals that many billing-related conversations involve multiple aspects—users asking about charges often want to modify subscriptions or report payment failures. The team restructures the taxonomy, creating a top-level billing_and_payments category with sub-intents: view_charges, dispute_charge, update_payment_method, modify_subscription, and payment_failure. They also enable multi-intent recognition for this category. This restructuring improves inter-annotator agreement to 0.84 and reduces user reformulation rates by 28%.

Challenge: Context Dependency and Multi-Turn Complexity

Intent recognition in multi-turn conversations requires maintaining context and resolving references to previous utterances ⁹. Users employ pronouns ("Can you change that to Friday?"), ellipsis ("And for two people"), and implicit references that depend on dialogue history. Single-utterance classification fails in these scenarios, producing nonsensical interpretations.

Solution:

Implement dialogue state tracking that maintains conversation history and contextual information ⁷⁹. Use context-aware models that encode previous utterances alongside the current input, enabling resolution of references and maintaining coherent intent sequences. Develop specialized handling for follow-up intents that modify or refine previous requests. Implement entity carryover mechanisms that propagate slot values across turns unless explicitly changed.

Example: A travel booking assistant handles this conversation: User: "I need a flight to Chicago next Monday" (intent: search_flight, slots: destination=Chicago, date=next Monday). User: "Actually, make that Tuesday" (intent: modify_search, slot: date=Tuesday). User: "And I need a hotel too" (intent: add_hotel, inherits: destination=Chicago, date=Tuesday). The system implements a dialogue state tracker that maintains the current search context, recognizes that "that" in the second utterance refers to the date slot, and carries forward the destination and date to the hotel search without requiring repetition. This context-aware approach reduces the average turns-to-booking from 8.3 to 5.1.

Challenge: Handling Out-of-Scope Requests

Users frequently make requests outside the system's capabilities—asking questions the system cannot answer or requesting actions it cannot perform ¹⁰. Poor handling of out-of-scope requests damages user trust and creates frustration. Simply responding "I don't understand" provides no guidance and may cause users to abandon the interaction.

Solution:

Develop a dedicated out-of-scope intent category trained on examples of requests the system cannot handle ¹⁰. Implement confidence thresholds where very low scores across all in-scope intents trigger out-of-scope classification. Create helpful fallback responses that acknowledge the limitation, suggest alternative resources, and offer to escalate to human assistance. Analyze out-of-scope requests to identify common patterns that might justify expanding system capabilities.

Example: A HR chatbot receives requests like "What's the weather going to be like tomorrow?" and "Can you recommend a good restaurant near the office?" The team creates an out_of_scope intent trained on 500 examples of irrelevant requests. When detected, the system responds: "I'm specialized in HR-related questions like benefits, time off, and policies. For other questions, you might try our general company assistant or contact the front desk. Would you like me to connect you with someone who can help?" They also implement monthly analysis of out-of-scope requests, which reveals that 12% involve commute and parking questions—leading them to add a parking_and_commute intent to the system.

Challenge: Performance Degradation Over Time

Intent recognition systems experience performance degradation as user language evolves, new products/services launch, and the distribution of user requests shifts ⁶. Models trained on historical data become increasingly misaligned with current usage patterns, a phenomenon known as concept drift. This degradation often goes unnoticed without proper monitoring until user complaints escalate.

Solution:

Implement continuous monitoring of key performance metrics and establish automated alerts for degradation ⁶. Deploy continuous learning pipelines that regularly retrain models on recent production data with human-in-the-loop validation. Use A/B testing to validate model updates before full deployment. Maintain versioned datasets that enable analysis of performance trends over time and identification of specific intents experiencing degradation.

Example: A retail chatbot's intent recognition accuracy gradually declines from 89% to 81% over six months. Investigation reveals that the company launched a new "buy online, pick up in store" service that users reference with varied terminology not present in training data ("curbside pickup," "BOPIS," "click and collect"). Additionally, pandemic-related language shifts introduced new phrasings ("contactless delivery," "safety protocols"). The team implements: (1) weekly monitoring dashboards tracking accuracy by intent, (2) monthly retraining cycles incorporating the previous month's validated production data, and (3) rapid response protocols for new product launches that include collecting intent examples before deployment. These measures restore accuracy to 88% and establish sustainable maintenance practices.

References

1. Devlin, J., Chang, M., Lee, K., & Toutanova, K. (2018). BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. https://arxiv.org/abs/1810.04805
2. Srinivasan, K., Sengupta, S., & Rudnicky, A. (2020). Intent Classification and Slot Filling for Privacy Policies. https://aclanthology.org/2020.acl-main.442/
3. Xu, P., & Sarikaya, R. (2019). Multi-Domain Intent Recognition in Spoken Language Understanding. https://arxiv.org/abs/1902.10909
4. Sun, C., Qiu, X., Xu, Y., & Huang, X. (2019). Few-Shot Text Classification with Pre-Trained Word Embeddings and a Human in the Loop. https://arxiv.org/abs/1909.00100
5. Zhang, L., Lyu, Q., & Callison-Burch, C. (2021). Intent Detection with WikiHow. https://aclanthology.org/2021.naacl-main.243/
6. Rastogi, P., Gupta, R., & Chen, J. (2019). Towards Scalable Multi-Domain Conversational Agents. https://research.google/pubs/pub48842/
7. Wu, C., Hoi, S., Socher, R., & Xiong, C. (2020). ToD-BERT: Pre-trained Natural Language Understanding for Task-Oriented Dialogue. https://arxiv.org/abs/2005.05909
8. Srinivasan, K., Sengupta, S., & Rudnicky, A. (2021). Intent Classification and Slot Filling for Privacy Policies. https://arxiv.org/abs/2104.08821
9. Rastogi, A., Zang, X., Sunkara, S., Gupta, R., & Khaitan, P. (2020). Schema-Guided Dialogue State Tracking Task at DSTC8. https://aclanthology.org/2020.emnlp-main.660/
10. Braun, D., Mendez, A., Matthes, F., & Langen, M. (2019). Benchmarking Natural Language Understanding Services for Building Conversational Agents. https://arxiv.org/abs/1906.08407