Papers by Paulo Cavalin

5 papers
Improving Out-of-Scope Detection in Intent Classification by Using Embeddings of the Word Graph Space of the Classes (2020.emnlp-main)

Copied to clipboard

Challenge: Existing methods for intent classification use one-class classification or inverse dictionary.
Approach: They propose to represent class labels as a vector space where word graphs are mapped . they use inverse dictionary to take in account inter-class similarities provided by repeated occurrences .
Outcome: The proposed method beats the state-of-the-art method in the Larson dataset by about 31 percentage points.
Using Meta-Knowledge Mined from Identifiers to Improve Intent Recognition in Conversational Systems (2021.acl-long)

Copied to clipboard

Challenge: a recent study shows that meta-knowledge embedded in intent identifiers improves intent recognition in conversational systems . a meta-learning approach is used to classify sentences into discrete sets of classes . classification is a key part of professional conversational system implementations .
Approach: They use meta-knowledge embedded in intent identifiers to improve intent recognition . authors found that meta-knowledge improved accuracy in conversational systems .
Outcome: The meta-knowledge enabled improved intent recognition in conversational systems . the meta-learning improved the false acceptance rate in two thirds of the chatbots .
Theoretical and Empirical Advantages of Dense-Vector to One-Hot Encoding of Intent Classes in Open-World Scenarios (2024.lrec-main)

Copied to clipboard

Challenge: Existing methods for encoding intents as one-hot are not used for zero-shot learning.
Approach: They show that one-hot encodings can produce much richer topologies for OOS detection . they argue that such gains are likely due to advantages of knowledge-free encoded intents .
Outcome: The proposed method can produce better systems for open-world classification tasks, the authors show . they show that knowledge-free, randomly generated dense-vector encodings can produce 20% gains over one-hot encodes.
Fixing Rogue Memorization in Many-to-One Multilingual Translators of Extremely-Low-Resource Languages by Rephrasing Training Samples (2024.naacl-long)

Copied to clipboard

Challenge: Existing fine-tuning of large high-resource language models into multilingual machine translators is difficult for extremely lowresource languages.
Approach: They propose to fine-tune large high-resource language models into multilingual machine translators for extremely-lowresource languages such as endangered Indigenous languages.
Outcome: The proposed model halls are reformulated to improve translation accuracy and improve translation quality.
From Disjoint Sets to Parallel Data to Train Seq2Seq Models for Sentiment Transfer (2020.findings-emnlp)

Copied to clipboard

Challenge: Existing methods for sentiment transfer have relied on unsupervised methods due to lack of parallel corpora.
Approach: They propose a method for creating parallel data to train Seq2Seq neural networks for sentiment transfer.
Outcome: The proposed method outperforms existing unsupervised methods in sentiment transfer tasks.

What is GenGO?

GenGO is an NLP powered publication search system. It currenctly indexes 30k+ papers from ACL Anthology, and implements multi-aspect summarization, semantic search, and more!

Information

About
Limitations