skip to main content
US FlagAn official website of the United States government
dot gov icon
Official websites use .gov
A .gov website belongs to an official government organization in the United States.
https lock icon
Secure .gov websites use HTTPS
A lock ( lock ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites.

Explore Research Products in the PAR
It may take a few hours for recently added research products to appear in PAR search results.


Title: Augmenting Training Data for a Virtual Character Using GPT-3.5
This paper compares different methods of using a large language model (GPT-3.5) for creating synthetic training data for a retrieval-based conversational character. The training data are in the form of linked questions and answers, which allow a classifier to retrieve a pre-recorded answer to an unseen question; the intuition is that a large language model could predict what human users might ask, thus saving the effort of collecting real user questions as training data. Results show small improvements in test performance for all synthetic datasets. However, a classifier trained on only small amounts of collected user data resulted in a higher F-score than the classifiers trained on much larger amounts of synthetic data generated using GPT-3.5. Based on these results, we see a potential in using large language models for generating training data, but at this point it is not as valuable as collecting actual user data for training.  more » « less
Award ID(s):
2150187
PAR ID:
10566489
Author(s) / Creator(s):
;
Publisher / Repository:
The Florida Artificial Intelligence Research Society
Date Published:
Journal Name:
The International FLAIRS Conference Proceedings
Volume:
37
ISSN:
2334-0762
Format(s):
Medium: X
Sponsoring Org:
National Science Foundation
More Like this
  1. ; ; ; (, Springer, Cham.)
    Wang, N. (Ed.)
    In education, intelligent learning environments allow students to choose how to tackle open-ended tasks while monitoring performance and behavior, allowing for the creation of adaptive support to help students overcome challenges. Timely feedback is critical to aid students’ progression toward learning and improved problem-solving. Feedback on text-based student responses can be delayed when teachers are overloaded with work. Automated evaluation can provide quick student feedback while easing the manual evaluation burden for teachers in areas with a high teacher-to-student ratio. Current methods of evaluating student essay responses to questions have included transformer-based natural language processing models with varying degrees of success. One main challenge in training these models is the scarcity of data for student-generated data. Larger volumes of training data are needed to create models that perform at a sufficient level of accuracy. Some studies have vast data, but large quantities are difficult to obtain when educational studies involve student-generated text. To overcome this data scarcity issue, text augmentation techniques have been employed to balance and expand the data set so that models can be trained with higher accuracy, leading to more reliable evaluation and categorization of student answers to aid teachers in the student’s learning progression. This paper examines the text-generating AI model, GPT-3.5, to determine if prompt-based text-generation methods are viable for generating additional text to supplement small sets of student responses for machine learning model training. We augmented student responses across two domains using GPT-3.5 completions and used that data to train a multilingual BERT model. Our results show that text generation can improve model performance on small data sets over simple self-augmentation. 
    more » « less
  2. ; ; (, Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024))
    Calzolari, Nicoletta; Kan, Min-Yen; Hoste, Veronique; Lenci, Alessandro; Sakti, Sakriani; Xue, Nianwen (Ed.)
    In this work, we revisit the problem of semi-supervised named entity recognition (NER) focusing on extremely light supervision, consisting of a lexicon containing only 10 examples per class. We introduce ELLEN, a simple, fully modular, neuro-symbolic method that blends fine-tuned language models with linguistic rules. These rules include insights such as ''One Sense Per Discourse'', using a Masked Language Model as an unsupervised NER, leveraging part-of-speech tags to identify and eliminate unlabeled entities as false negatives, and other intuitions about classifier confidence scores in local and global context. ELLEN achieves very strong performance on the CoNLL-2003 dataset when using the minimal supervision from the lexicon above. It also outperforms most existing (and considerably more complex) semi-supervised NER methods under the same supervision settings commonly used in the literature (i.e., 5% of the training data). Further, we evaluate our CoNLL-2003 model in a zero-shot scenario on WNUT-17 where we find that it outperforms GPT-3.5 and achieves comparable performance to GPT-4. In a zero-shot setting, ELLEN also achieves over 75% of the performance of a strong, fully supervised model trained on gold data. Our code is available at: https://github.com/hriaz17/ELLEN 
    more » « less
  3. ; ; ; ; (, https://doi.org/10.48550/arXiv.2407.16607)
    The pretraining data of today's strongest language models is opaque; in particular, little is known about the proportions of various domains or languages represented. In this work, we tackle a task which we call data mixture inference, which aims to uncover the distributional make-up of training data. We introduce a novel attack based on a previously overlooked source of information: byte-pair encoding (BPE) tokenizers, used by the vast majority of modern language models. Our key insight is that the ordered list of merge rules learned by a BPE tokenizer naturally reveals information about the token frequencies in its training data. Given a tokenizer's merge list along with example data for each category of interest, we formulate a linear program that solves for the proportion of each category in the tokenizer's training set. In controlled experiments, we show that our attack recovers mixture ratios with high precision for tokenizers trained on known mixtures of natural languages, programming languages, and data sources. We then apply our approach to off-the-shelf tokenizers released with recent LMs. We confirm much publicly disclosed information about these models, and also make several new inferences: GPT-4o and Mistral NeMo's tokenizers are much more multilingual than their predecessors, training on 39% and 47% non-English language data, respectively; Llama 3 extends GPT-3.5's tokenizer primarily for multilingual (48%) use; GPT-3.5's and Claude's tokenizers are trained on predominantly code (~60%). We hope our work sheds light on current design practices for pretraining data, and inspires continued research into data mixture inference for LMs. 
    more » « less
  4. ; (, Proceedings of the Thirteenth Language Resources and Evaluation Conference)
    In this paper, we compare two different approaches to language understanding for a human-robot interaction domain in which a human commander gives navigation instructions to a robot. We contrast a relevance-based classifier with a GPT-2 model, using about 2000 input-output examples as training data. With this level of training data, the relevance-based model outperforms the GPT-2 based model 79% to 8%. We also present a taxonomy of types of errors made by each model, indicating that they have somewhat different strengths and weaknesses, so we also examine the potential for a combined model. 
    more » « less
  5. ; ; ; ; ; (, OpenReview.net)
    Code Large Language Models (Code LLMs) have excelled at tasks like code completion but often miss deeper semantics such as execution effects and dynamic states. This paper aims to bridge the gap between Code LLMs' reliance on static text data and the need for semantic understanding for complex tasks like debugging and program repair. We introduce a novel strategy, monologue reasoning, to train Code LLMs to reason comprehensive semantics, encompassing high-level functional descriptions, local execution effects of individual statements, and overall input/output behavior, thereby linking static code text with dynamic execution states. We begin by collecting PyX, a clean Python corpus of fully executable code samples with functional descriptions and test cases. We propose training Code LLMs not only to write code but also to understand code semantics by reasoning about key properties, constraints, and execution behaviors using natural language, mimicking human verbal debugging, i.e., rubber-duck debugging. This approach led to the development of SemCoder, a Code LLM with only 6.7B parameters, which shows competitive performance with GPT-3.5-turbo on code generation and execution reasoning tasks. SemCoder achieves 79.3% on HumanEval (GPT-3.5-turbo: 76.8%), 63.6% on CRUXEval-I (GPT-3.5-turbo: 50.3%), and 63.9% on CRUXEval-O (GPT-3.5-turbo: 59.0%). We also study the effectiveness of SemCoder's monologue-style execution reasoning compared to concrete scratchpad reasoning, showing that our approach integrates semantics from multiple dimensions more smoothly. Finally, we demonstrate the potential of applying learned semantics to improve Code LLMs' debugging and self-refining capabilities. Our data, code, and models are available at: https://github.com/ARiSE-Lab/SemCoder. 
    more » « less
  • Citation Formats
  • MLA
  • APA
  • Chicago
  • BibTeX
  • Save / Share this Record
  • Send to Email