Fine-tuning large language models (LLMs) with domain-specific instructions has emerged as an effective method to enhance their domain-specific understanding. Yet, there is limited work that examines the core characteristics acquired during this process. In this study, we benchmark the fundamental characteristics learned by contact-center (CC) specific instruction fine-tuned LLMs with out-of-the-box (OOB) LLMs via probing tasks encompassing conversational, channel, and automatic speech recognition (ASR) properties. We explore different LLM architectures (Flan-T5 and Llama), sizes (3B, 7B, 11B, 13B), and fine-tuning paradigms (full fine-tuning vs PEFT). Our findings reveal remarkable effectiveness of CC-LLMs on the in-domain downstream tasks, with improvement in response acceptability by over 48% compared to OOB-LLMs. Additionally, we compare the performance of OOB-LLMs and CC-LLMs on the widely used SentEval dataset, and assess their capabilities in terms of surface, syntactic, and semantic information through probing tasks. Intriguingly, we note a relatively consistent performance of probing classifiers on the set of probing tasks. Our observations indicate that CC-LLMs, while outperforming their out-of-the-box counterparts, exhibit a tendency to rely less on encoding surface, syntactic, and semantic properties, highlighting the intricate interplay between domain-specific adaptation and probing task performance opening up opportunities to explore behavior of fine-tuned language models in specialized contexts.
LLM-powered chatbots are becoming widely adopted in applications such as healthcare, personal assistants, industry hiring decisions, etc. In many of these cases, chatbots are fed sensitive, personal information in their prompts, as samples for in-context learning, retrieved records from a database, or as part of the conversation. The information provided in the prompt could directly appear in the output, which might have privacy ramifications if there is sensitive information there. As such, in this paper, we aim to understand the input copying and regurgitation capabilities of these models during inference and how they can be directly instructed to limit this copying by complying with regulations such as HIPAA and GDPR, based on their internal knowledge of them. More specifically, we find that when ChatGPT is prompted to summarize cover letters of a 100 candidates, it would retain personally identifiable information (PII) verbatim in 57.4% of cases, and we find this retention to be non-uniform between different subgroups of people, based on attributes such as gender identity. We then probe ChatGPT's perception of privacy-related policies and privatization mechanisms by directly instructing it to provide compliant outputs and observe a significant omission of PII from output.
The latest WHO report showed that the number of malaria cases climbed to 219 million last year, two million higher than last year. The global efforts to fight malaria have hit a plateau and the most significant underlying reason is international funding has declined. Malaria, which is spread to people through the bites of infected female mosquitoes, occurs in 91 countries but about 90% of the cases and deaths are in sub-Saharan Africa. The disease killed 4,35,000 people last year, the majority of them children under five in Africa. AI-backed technology has revolutionized malaria detection in some regions of Africa and the future impact of such work can be revolutionary. The malaria Cell Image Data-set is taken from the official NIH Website NIH data. The aim of the collection of the dataset was to reduce the burden for microscopists in resource-constrained regions and improve diagnostic accuracy using an AI-based algorithm to detect and segment the red blood cells. The goal of this work is to show that the state of the art accuracy can be obtained even by using 2 layer convolution network and show a new baseline in Malaria detection efforts using AI.
With the future striving toward data-centric decision-making, seamless access to databases is of utmost importance. There is extensive research on creating an efficient text-to-sql (TEXT2SQL) model to access data from the database. Using a Natural language is one of the best interfaces that can bridge the gap between the data and results by accessing the database efficiently, especially for non-technical users. It will open the doors and create tremendous interest among users who are well versed in technical skills or not very skilled in query languages. Even if numerous deep learning-based algorithms are proposed or studied, there still is very challenging to have a generic model to solve the data query issues using natural language in a real-work scenario. The reason is the use of different datasets in different studies, which comes with its limitations and assumptions. At the same time, we do lack a thorough understanding of these proposed models and their limitations with the specific dataset it is trained on. In this paper, we try to present a holistic overview of 24 recent neural network models studied in the last couple of years, including their architectures involving convolutional neural networks, recurrent neural networks, pointer networks, reinforcement learning, generative models, etc. We also give an overview of the 11 datasets that are widely used to train the models for TEXT2SQL technologies. We also discuss the future application possibilities of TEXT2SQL technologies for seamless data queries.
This paper proposes an open source visual analytics tool consisting of several views and perspectives on eye movement data collected during code reading tasks when writing computer programs. Hence the focus of this work is on code and program comprehension. The source code is shown as a visual stimulus. It can be inspected in combination with overlaid scanpaths in which the saccades can be visually encoded in several forms, including straight, curved, and orthogonal lines, modifiable by interaction techniques. The tool supports interaction techniques like filter functions, aggregations, data sampling, and many more. We illustrate the usefulness of our tool by applying it to the eye movements of 216 programmers of multiple expertise levels that were collected during two code comprehension tasks. Our tool helped to analyze the difference between the strategic program comprehension of programmers based on their demographic background, time taken to complete the task, choice of programming task, and expertise.
We explore techniques for eye gaze estimation using machine learning. Eye gaze estimation is a common problem for various behavior analysis and human-computer interfaces. The purpose of this work is to discuss various model types for eye gaze estimation and present the results from predicting gaze direction using eye landmarks in unconstrained settings. In unconstrained real-world settings, feature-based and model-based methods are outperformed by recent appearance-based methods due to factors like illumination changes and other visual artifacts. We discuss a learning-based method for eye region landmark localization trained exclusively on synthetic data. We discuss how to use detected landmarks as input to iterative model-fitting and lightweight learning-based gaze estimation methods and how to use the model for person-independent and personalized gaze estimations.
In Weak Supervised Learning (WSL), a model is trained over noisy labels obtained from semantic rules and task-specific pre-trained models. Rules offer limited generalization over tasks and require significant manual efforts while pre-trained models are available only for limited tasks. In this work, we propose to utilize prompt-based methods as weak sources to obtain the noisy labels on unannotated data. We show that task-agnostic prompts are generalizable and can be used to obtain noisy labels for different Spoken Language Understanding (SLU) tasks such as sentiment classification, disfluency detection and emotion classification. These prompts could additionally be updated to add task-specific contexts, thus providing flexibility to design task-specific prompts. We demonstrate that prompt-based methods generate reliable labels for the above SLU tasks and thus can be used as a universal weak source to train a weak-supervised model (WSM) in absence of labeled data. Our proposed WSL pipeline trained over prompt-based weak source outperforms other competitive low-resource benchmarks on zero and few-shot learning by more than 4% on Macro-F1 on all of the three benchmark SLU datasets. The proposed method also outperforms a conventional rule based WSL pipeline by more than 5% on Macro-F1.
One of the core components of goal-oriented dialog systems is the task of Intent Detection. Few-shot Learning upon Intent Detection is challenging due to the scarcity of available annotated utterances. Although recent works making use of metric-based and optimization-based methods have been proposed, the task is still challenging in large label spaces and much smaller number of shots. Generalized Few-shot learning is more difficult due to the presence of both novel and seen classes during the testing phase. In this work, we propose a simple and effective method based on Natural Language Inference that not only tackles the problem of few shot intent detection, but also proves useful in zero-shot and generalized few shot learning problems. Our extensive experiments on a number of Natural Language Understanding (NLU) and Spoken Language Understanding (SLU) datasets show the effectiveness of our approach. In addition, we highlight the settings in which our NLI based method outperforms the baselines by huge margins.
Language Models (LMs) have been ubiquitously leveraged in various tasks including spoken language understanding (SLU). Spoken language requires careful understanding of speaker interactions, dialog states and speech induced multimodal behaviors to generate a meaningful representation of the conversation. In this work, we propose to dissect SLU into three representative properties:conversational (disfluency, pause, overtalk), channel (speaker-type, turn-tasks) and ASR (insertion, deletion,substitution). We probe BERT based language models (BERT, RoBERTa) trained on spoken transcripts to investigate its ability to understand multifarious properties in absence of any speech cues. Empirical results indicate that LM is surprisingly good at capturing conversational properties such as pause prediction and overtalk detection from lexical tokens. On the downsides, the LM scores low on turn-tasks and ASR errors predictions. Additionally, pre-training the LM on spoken transcripts restrain its linguistic understanding. Finally, we establish the efficacy and transferability of the mentioned properties on two benchmark datasets: Switchboard Dialog Act and Disfluency datasets.
With surge in online platforms, there has been an upsurge in the user engagement on these platforms via comments and reactions. A large portion of such textual comments are abusive, rude and offensive to the audience. With machine learning systems in-place to check such comments coming onto platform, biases present in the training data gets passed onto the classifier leading to discrimination against a set of classes, religion and gender. In this work, we evaluate different classifiers and feature to estimate the bias in these classifiers along with their performance on downstream task of toxicity classification. Results show that improvement in performance of automatic toxic comment detection models is positively correlated to mitigating biases in these models. In our work, LSTM with attention mechanism proved to be a better modelling strategy than a CNN model. Further analysis shows that fasttext embeddings is marginally preferable than glove embeddings on training models for toxicity comment detection. Deeper analysis reveals the findings that such automatic models are particularly biased to specific identity groups even though the model has a high AUC score. Finally, in effort to mitigate bias in toxicity detection models, a multi-task setup trained with auxiliary task of toxicity sub-types proved to be useful leading to upto 0.26% (6% relative) gain in AUC scores.