Project Presentations

1

GRADED IN CLASS ACTIVITY:

Go to https://forms.gle/qE76x6CUEAXVWSwu9 to enter feedback for your classmates.

(If you don’t have a computer, write your feedback on a piece of paper and hand it in after class.)

LMFAOH

Language Modeling For the Analysis of Harry Potter

1. Inspiration
1. Beyond Canonical Texts: A Computational Analysis of Fanfictionhttps://www.aclweb.org/anthology/D16-1218.pdf
2. Personality Profiling of Fictional Characters using Sense-Level Links between Lexical Resources

http://www.emnlp2015.org/proceedings/EMNLP/pdf/EMNLP208.pdf

• Motivations / Applications
• FBI profiling
• Recommendation Systems
• Identifying nodes and vertices in real time that model human discourse and bias
• Focus
• Personality Extraction
• Network Analysis

Object Tracking From D&D Games

• Tracking objects from sensor data is old problem in robotics
• Data normally come from sensors (TOF, Lidars) or in more recent years Computer Vision models (YOLO)
• Tools use to parse/filter/track/associate this data such as Particle Filters, EKF, DPMM etc are not dependant on the data to come from vision
• Can you use a NLP pipeline instead to extract this “data” from text and use the same old statistical tools on that data ?
• Would that help for example tracking weapons in a D&D text?

Poetry Generator

• A tool that can generate different forms of poetry
• Haikus, Limericks, Sonnets
• Meter and rhyming scheme are two key aspects of poetry
• CMU has a pronouncing dictionary
• Meter can be tracked through syllables
• Myriad of other variables to play with
• Length, repetition, alliteration, sentiment

Wikihow Dungeon

• An AI Dungeon inspired game trained using Wikihow to generate puzzles
• Given an article, such as “Put On A Dog Harness” or “How To Resolve an Identity Crisis”, generate a short text adventure game around the task, using the steps listed in the article as the actions needed to be taken to complete the puzzle
• Components:
• Article → Game component
• GPT-2 fine tuned to generate new articles
• Example:
• “You are Daphne. You are in your kitchen. You’re surrounded by a dog harness, and a cute doggo. You want to put the harness on the dog. What do you do?”
• a) Place the unbuckled harness on the ground
• b) Pick up the right paw and move it forward into the appropriate loop
• c) Reward you pup with a treat and praise

​

​

​

​

NLP Visual Novel

• A visual-text adventure game with NLP elements expanding on other features of the game
• Sentiment analysis to raise an affection level, and display emotions on characters
• BookNLP to try to associate characters with other characters or ideas positively or negatively for generating random maps, items, NPCs
• Language model trained on fanfiction archive to cater to audience of fandom
• Potential Ideas
• Storyline and quotes are NLP generated with minor editing/picking to make them flow together
• Completely random-generated adventure

​

​

​

​

Genesis

• An interface for game architects to author text-based games
• Example of commands:
• “Create great dining hall”
• “Connect great dining hall to stairs to the dungeon”
• Structure:
• A parser interface that interprets the game architect’s command
• A game building backend that stores and updates relevant parameters
• A map that shows the current game logic being built
• A playing system that allows the architect to instantly play the current game
• Model / algorithm:
• Pre-trained models such as AllenNLP
• Fine tuning based on text adventure game commands dataset

​

​

​

IDLI-PPLM

Increasing Diversity of Lexicon In Plug and Play Language Models

1. PPLM:
1. Conditional language generation without fine-tuning
2. Use attribute models to control attributes of the text
3. E.g., GPT-2 + Bag of Words model for Legal documents
2. Problem and Motivation:
• Generated text has skewed word distribution (for words in the topic model)
• Uniform word distribution is desirable for tasks like data augmentation

Music Lyrics Generation

• Example
• Given genre such as “country”, “pop”, “Hip hop”...
• Given title such as “Back to December”, “Everything I Wanted”...
• Compare the machine generation with human writing to get pairwise similarity
• Potential Methods
• LSTM used for sequence data.
• GAN used for generate similar style data.
• Transfer Learning, using this lyrics model for other rhythm text, such as poems.
• Like using this model to write a “country” style poems..

​

​

Text adventure game with control code

1. Project Description: More controllable generation of text in text adventure game

Ex: Player >>> %Horror% take the knife.

Generated >>> John pulls the knife out. Blood drips off its blade. His eyes widen in horror.

• Use the CTRL model with fine-tuning.

Similar to how they use GPT-2 to make AI-dungeon, but use CTRL instead of GPT-2.

​

Analysis of State Changes in Generated Text

We want to analyze what are the typical errors made by transformer based generation models

Automatic Metrics

• Coreference resolution models
• Natural language Inference models
• D&D Dataset which has state associated

Human Evaluation

• Turkers label text (alternatively, lure students using food)
• We perturb text to make it internally inconsistent (train a model to detect inconsistent text)