Industry Track

Part of my series of notes from NAACL-HLT 2019 in Minneapolis.

Locale-agnostic Universal Domain Classification

• ambiguities in large-scale NLU in Alexa
  • multiple domains can properly handle an utterance
  • e.g. “I want a pepperoni pizza” – order takeout, recipe, restaurant reservation?
• shortlister Kim et al., NAACL 2018 model to solve this problem
• expanding to other locales
• challenges
  • maintaining separate model for each locale
  • lack of training data (no usage history)
• locale / domain maturity – how long since model deployed, how much data collected
• locale specificity – overlapping domains between locales, but can have locale-specific slots, etc.
  • sharing data is tricky
• both shared and locale-specific encoders for utterances (BiLSTM)
• shared encoder tuned to be locale-invariant – adversarial locale prediction loss
• route utterance to locale-specific encoder without knowing domain using attention
• then concat representations and predict domain
• I don’t really understand how this addresses the challenges?
  • maintaining different models – still have to do this I guess?
  • lack of training data – presumably can rely on the shared encoder
• experiments with different locales, different amounts of data, different kinds of domains, different encoder architectures

Practical Semantic Parsing for Spoken Language Understanding

• paper here
• executable semantic parsing – sentences -> logical forms
  • unify Alexa’s primary problems of SLU and Q&A
• SLU Alexa data – annotated for intent/slot tagging
  • think this is proprietary, sadly
• Q&A data – Overnight, NLmaps
  • much more low-resource than SLU datasets
• convert this data to trees
  • allows more complex requests / multiple intents

• transition-based parser [Cheng et al. 2017] to parse into trees
  • add character embeddings and mechanism to copy directly from input
• evaluate on Q&A and SLU tasks, including ablation studies
• this is training on each domain independently
• transfer learning from high- to low-resource domains
  • tried fine-tuning and multi-task learning, both seem to help
• transfer from SLU to Q&A also seems to work (preliminary)
• avoided more data-hungry architectures, e.g. seq2seq
• different annotation schemes among datasets
• full match metric is quite harsh

Fast Prototyping Dialogue Comprehension System for Nurse-Patient Conversations

• paper here
  • this one was really interesting and really relevant to my work, so I may or may not have taken photos of like every single slide

• nurses do follow-up calls after discharge
• want to extract clinical info from unstructured spoken interactions

• attributes of these dypes of dialogues
  • co-references across turns & speakers (zero anaphora)
  • thinking aloud (private self-talk, backchanneling)
  • self-contradiction – within a single utterance or across multiple turns
  • topic drift
  • dialect / discourse particles

• dialogue comprehension task
• existing datasets – reading comprehension, very limited data on open-domain dialogues
• need for data augmentation
  • seed with real-world data
  • annotate with inquiry types & response types

• from analysis of seed data, create templates to generate new data
  • pair inquiry and response types
  • entity value pool
  • enrich verbal expressions and get them linguistically / clinically validated

• model – standard neural architecture thing
• error analysis
  • nurses and patients do chit-chat, even within task-oriented dialogue
  • patients speculate on potential causal relations

• proof-of-concept that even without a lot of data you can get a reasonable prototype
  • lots of work to be done!

Graph Convolution for Multimodal Information Extraction

• IE in the real world isn’t just text – can be visually rich
• for these documents visual attributes (font size, color, etc.) are very relevant
• need multimodal IE

• diverse tasks with low-development efficiency – techniques for one might not transfer well
• contributions: visual text embeddings and unified model
• documents are graphs of text segments
• combine graph convolution with BiLSTM+CRF
  • text embeddings for node features, edge features from relative positions
  • self-attention for neighbourhoods