Actions for Dialogue State Tracking with Universal Adversarial Triggers

Email RIS file
Bookmark
Report an Issue

Dialogue State Tracking with Universal Adversarial Triggers

Author
Zhao, Tianyang
Published
[University Park, Pennsylvania] : Pennsylvania State University, 2022.
Physical Description
1 electronic document
Additional Creators
Das, Chita R.
Access Online

Availability

I Want It
I Want It

Finding items...

Graduate Program
Restrictions on Access
Restricted (PSU Only).
Summary
Dialogue State Tracking (DST) is one of the more complex subtopics in Natural Language Processing (NLP) as the development of a conversational agent often requires maintaining a history of relevant information across multiple input domains and topics. Universal Adversarial Triggers (UATs) are input-agnostic text sequences that have already been applied to a variety of NLP tasks. UATs aim to confuse NLP models into making an intended incorrect prediction. In this work, we apply UATs to the DST and observe their characteristics. We focus on the TripPy model and obtain triggers to attack each slot in the MultiWOZ 2.1 dataset. We find that UATs generally perform well at triggering the target value and perturbing correct predictions toward TripPy. We further discover that UATs' effectiveness is influenced by their length, the placement of the target value in them, and where they are inserted in DST input data. Then, we make use of UATs to generate additional training examples and retrain the model. We find that augmenting with UAT examples improves the model's performance but not as much as examples generated by certain other adversarial attack method. In the end, we explore the transferability of UATs to other DST models, specifically SimpleTOD and TRADE. We notice that UATs become much less effective on these models thus fail to transfer to other models when applied to DST. We conclude by discussing what potential future works can be done with UATs on DST.
Other Subject(s)
Genre(s)
Dissertation Note
M.S. Pennsylvania State University 2022.
Technical Details
The full text of the dissertation is available as an Adobe Acrobat .pdf file ; Adobe Acrobat Reader required to view the file.
View MARC record | catkey: 38970334