In virtual assistant (VA) systems it is important to reject or redirect user queries that fall outside the scope of the system. One of the most accurate approaches for out-of-scope (OOS) rejection is to combine it with the task of intent classification on in-scope queries, and to use methods based on the similarity of embeddings produced by transformer-based sentence encoders. Typically, such encoders are fine-tuned for the intent-classification task, using cross-entropy loss. Recent work has shown that while this produces suitable embeddings for the intent-classification task, it also tends to disperse in-scope embeddings over the full sentence embedding space. This causes the in-scope embeddings to potentially overlap with OOS embeddings, thereby making OOS rejection difficult. This is compounded when OOS data is unknown. To mitigate this issue our work proposes to regularize the cross-entropy loss with an in-scope embedding reconstruction loss learned using an auto-encoder. Our method achieves a 1-4% improvement in the area under the precision-recall curve for rejecting out-of-sample (OOS) instances, without compromising intent classification performance.
This study evaluates the application of large language models (LLMs) for intent classification within a chatbot with predetermined responses designed for banking industry websites. Specifically, the research examines the effectiveness of fine-tuning SlovakBERT compared to employing multilingual generative models, such as Llama 8b instruct and Gemma 7b instruct, in both their pre-trained and fine-tuned versions. The findings indicate that SlovakBERT outperforms the other models in terms of in-scope accuracy and out-of-scope false positive rate, establishing it as the benchmark for this application.
Classification is a core NLP task architecture with many potential applications. While large language models (LLMs) have brought substantial advancements in text generation, their potential for enhancing classification tasks remains underexplored. To address this gap, we propose a framework for thoroughly investigating fine-tuning LLMs for classification, including both generation- and encoding-based approaches. We instantiate this framework in edit intent classification (EIC), a challenging and underexplored classification task. Our extensive experiments and systematic comparisons with various training approaches and a representative selection of LLMs yield new insights into their application for EIC. We investigate the generalizability of these findings on five further classification tasks. To demonstrate the proposed methods and address the data shortage for empirical edit analysis, we use our best-performing EIC model to create Re3-Sci2.0, a new large-scale dataset of 1,780 scientific document revisions with over 94k labeled edits. The quality of the dataset is assessed through human evaluation. The new dataset enables an in-depth empirical study of human editing behavior in academic writing. We make our experimental framework, models and data publicly available.
Large language models (LLMs) have enabled a range of applications in zero-shot and few-shot learning settings, including the generation of synthetic datasets for training and testing. However, to reliably use these synthetic datasets, it is essential to understand how representative they are of real-world data. We investigate this by assessing the effectiveness of generating synthetic data through LLM and using it as a benchmark for various NLP tasks. Our experiments across six datasets, and three different tasks, show that while synthetic data can effectively capture performance of various methods for simpler tasks, such as intent classification, it falls short for more complex tasks like named entity recognition. Additionally, we propose a new metric called the bias factor, which evaluates the biases introduced when the same LLM is used to both generate benchmarking data and to perform the tasks. We find that smaller LLMs exhibit biases towards their own generated data, whereas larger models do not. Overall, our findings suggest that the effectiveness of synthetic data as a benchmark varies depending on the task, and that practitioners should rely on data generated from multiple larger models whenever possible.
In the realm of task-oriented dialogue systems, a robust intent detection mechanism must effectively handle malformed utterances encountered in real-world scenarios. This study presents a novel fine-tuning framework for large language models (LLMs) aimed at enhancing in-distribution (ID) intent classification and out-of-distribution (OOD) intent detection, which utilizes semantic matching with prototypes derived from ID class names. By harnessing the highly distinguishable representations of LLMs, we construct semantic prototypes for each ID class using a diversity-grounded prompt tuning approach. We rigorously test our framework in a challenging OOD context, where ID and OOD classes are semantically close yet distinct, referred to as \emph{near} OOD detection. For a thorough assessment, we benchmark our method against the prevalent fine-tuning approaches. The experimental findings reveal that our method demonstrates superior performance in both few-shot ID intent classification and near-OOD intent detection tasks.
With the increasing popularity of daily information sharing and acquisition on the Internet, this paper introduces an innovative approach for intent classification in Bangla language, focusing on social media posts where individuals share their thoughts and opinions. The proposed method leverages multimodal data with particular emphasis on authorship identification, aiming to understand the underlying purpose behind textual content, especially in the context of varied user-generated posts on social media. Current methods often face challenges in low-resource languages like Bangla, particularly when author traits intricately link with intent, as observed in social media posts. To address this, we present the Multimodal-based Author Bangla Intent Classification (MABIC) framework, utilizing text and images to gain deeper insights into the conveyed intentions. We have created a dataset named "Uddessho," comprising 3,048 instances sourced from social media. Our methodology comprises two approaches for classifying textual intent and multimodal author intent, incorporating early fusion and late fusion techniques. In our experiments, the unimodal approach achieved an accuracy of 64.53% in interpreting Bangla textual intent. In contrast, our multimodal approach significantly outperformed traditional unimodal methods, achieving an accuracy of 76.19%. This represents an improvement of 11.66%. To our best knowledge, this is the first research work on multimodal-based author intent classification for low-resource Bangla language social media posts.
Streaming services have reshaped how we discover and engage with digital entertainment. Despite these advancements, effectively understanding the wide spectrum of user search queries continues to pose a significant challenge. An accurate query understanding system that can handle a variety of entities that represent different user intents is essential for delivering an enhanced user experience. We can build such a system by training a natural language understanding (NLU) model; however, obtaining high-quality labeled training data in this specialized domain is a substantial obstacle. Manual annotation is costly and impractical for capturing users' vast vocabulary variations. To address this, we introduce a novel approach that leverages large language models (LLMs) through weak supervision to automatically annotate a vast collection of user search queries. Using prompt engineering and a diverse set of LLM personas, we generate training data that matches human annotator expectations. By incorporating domain knowledge via Chain of Thought and In-Context Learning, our approach leverages the labeled data to train low-latency models optimized for real-time inference. Extensive evaluations demonstrated that our approach outperformed the baseline with an average relative gain of 113% in recall. Furthermore, our novel prompt engineering framework yields higher quality LLM-generated data to be used for weak supervision; we observed 47.60% improvement over baseline in agreement rate between LLM predictions and human annotations with respect to F1 score, weighted according to the distribution of occurrences of the search queries. Our persona selection routing mechanism further adds an additional 3.67% increase in weighted F1 score on top of our novel prompt engineering framework.
As Large Language Models become more ubiquitous across domains, it becomes important to examine their inherent limitations critically. This work argues that hallucinations in language models are not just occasional errors but an inevitable feature of these systems. We demonstrate that hallucinations stem from the fundamental mathematical and logical structure of LLMs. It is, therefore, impossible to eliminate them through architectural improvements, dataset enhancements, or fact-checking mechanisms. Our analysis draws on computational theory and Godel's First Incompleteness Theorem, which references the undecidability of problems like the Halting, Emptiness, and Acceptance Problems. We demonstrate that every stage of the LLM process-from training data compilation to fact retrieval, intent classification, and text generation-will have a non-zero probability of producing hallucinations. This work introduces the concept of Structural Hallucination as an intrinsic nature of these systems. By establishing the mathematical certainty of hallucinations, we challenge the prevailing notion that they can be fully mitigated.
In an enterprise Virtual Assistant (VA) system, intent classification is the crucial component that determines how a user input is handled based on what the user wants. The VA system is expected to be a cost-efficient SaaS service with low training and inference time while achieving high accuracy even with a small number of training samples. We pretrain a transformer-based sentence embedding model with a contrastive learning objective and leverage the embedding of the model as features when training intent classification models. Our approach achieves the state-of-the-art results for few-shot scenarios and performs better than other commercial solutions on popular intent classification benchmarks. However, generating features via a transformer-based model increases the inference time, especially for longer user inputs, due to the quadratic runtime of the transformer's attention mechanism. On top of model distillation, we introduce a practical multi-task adaptation approach that configures dynamic token pruning without the need for task-specific training for intent classification. We demonstrate that this approach improves the inference speed of popular sentence transformer models without affecting model performance.
Query intent classification is an essential module for customers to find desired products on the e-commerce application quickly. Most existing query intent classification methods rely on the users' click behavior as a supervised signal to construct training samples. However, these methods based entirely on posterior labels may lead to serious category imbalance problems because of the Matthew effect in click samples. Compared with popular categories, it is difficult for products under long-tail categories to obtain traffic and user clicks, which makes the models unable to detect users' intent for products under long-tail categories. This in turn aggravates the problem that long-tail categories cannot obtain traffic, forming a vicious circle. In addition, due to the randomness of the user's click, the posterior label is unstable for the query with similar semantics, which makes the model very sensitive to the input, leading to an unstable and incomplete recall of categories. In this paper, we propose a novel Semi-supervised Multi-channel Graph Convolutional Network (SMGCN) to address the above problems from the perspective of label association and semi-supervised learning. SMGCN extends category information and enhances the posterior label by utilizing the similarity score between the query and categories. Furthermore, it leverages the co-occurrence and semantic similarity graph of categories to strengthen the relations among labels and weaken the influence of posterior label instability. We conduct extensive offline and online A/B experiments, and the experimental results show that SMGCN significantly outperforms the strong baselines, which shows its effectiveness and practicality.
In this work, we introduce several schemes to leverage description-augmented embedding similarity for dataless intent classification using current state-of-the-art (SOTA) text embedding models. We report results of our methods on four commonly used intent classification datasets and compare against previous works of a similar nature. Our work shows promising results for dataless classification scaling to a large number of unseen intents. We show competitive results and significant improvements (+6.12\% Avg.) over strong zero-shot baselines, all without training on labelled or task-specific data. Furthermore, we provide qualitative error analysis of the shortfalls of this methodology to help guide future research in this area.
Identifying the reason for which an author cites another work is essential to understand the nature of scientific contributions and to assess their impact. Citations are one of the pillars of scholarly communication and most metrics employed to analyze these conceptual links are based on quantitative observations. Behind the act of referencing another scholarly work there is a whole world of meanings that needs to be proficiently and effectively revealed. This study emphasizes the importance of trustfully classifying citation intents to provide more comprehensive and insightful analyses in research assessment. We address this task by presenting a study utilizing advanced Ensemble Strategies for Citation Intent Classification (CIC) incorporating Language Models (LMs) and employing Explainable AI (XAI) techniques to enhance the interpretability and trustworthiness of models' predictions. Our approach involves two ensemble classifiers that utilize fine-tuned SciBERT and XLNet LMs as baselines. We further demonstrate the critical role of section titles as a feature in improving models' performances. The study also introduces a web application developed with Flask and currently available at http://137.204.64.4:81/cic/classifier, aimed at classifying citation intents. One of our models sets as a new state-of-the-art (SOTA) with an 89.46% Macro-F1 score on the SciCite benchmark. The integration of XAI techniques provides insights into the decision-making processes, highlighting the contributions of individual words for level-0 classifications, and of individual models for the metaclassification. The findings suggest that the inclusion of section titles significantly enhances classification performances in the CIC task. Our contributions provide useful insights for developing more robust datasets and methodologies, thus fostering a deeper understanding of scholarly communication.
The growing enrollments in computer science courses and increase in class sizes necessitate scalable, automated tutoring solutions to adequately support student learning. While Large Language Models (LLMs) like GPT-4 have demonstrated potential in assisting students through question-answering, educators express concerns over student overreliance, miscomprehension of generated code, and the risk of inaccurate answers. Rather than banning these tools outright, we advocate for a constructive approach that harnesses the capabilities of AI while mitigating potential risks. This poster introduces CourseAssist, a novel LLM-based tutoring system tailored for computer science education. Unlike generic LLM systems, CourseAssist uses retrieval-augmented generation, user intent classification, and question decomposition to align AI responses with specific course materials and learning objectives, thereby ensuring pedagogical appropriateness of LLMs in educational settings. We evaluated CourseAssist against a baseline of GPT-4 using a dataset of 50 question-answer pairs from a programming languages course, focusing on the criteria of usefulness, accuracy, and pedagogical appropriateness. Evaluation results show that CourseAssist significantly outperforms the baseline, demonstrating its potential to serve as an effective learning assistant. We have also deployed CourseAssist in 6 computer science courses at a large public R1 research university reaching over 500 students. Interviews with 20 student users show that CourseAssist improves computer science instruction by increasing the accessibility of course-specific tutoring help and shortening the feedback loop on their programming assignments. Future work will include extensive pilot testing at more universities and exploring better collaborative relationships between students, educators, and AI that improve computer science learning experiences.
This paper presents a novel and comprehensive solution to enhance both the robustness and efficiency of question answering (QA) systems through supervised contrastive learning (SCL). Training a high-performance QA system has become straightforward with pre-trained language models, requiring only a small amount of data and simple fine-tuning. However, despite recent advances, existing QA systems still exhibit significant deficiencies in functionality and training efficiency. We address the functionality issue by defining four key tasks: user input intent classification, out-of-domain input detection, new intent discovery, and continual learning. We then leverage a unified SCL-based representation learning method to efficiently build an intra-class compact and inter-class scattered feature space, facilitating both known intent classification and unknown intent detection and discovery. Consequently, with minimal additional tuning on downstream tasks, our approach significantly improves model efficiency and achieves new state-of-the-art performance across all tasks.
Large language models (LLM) have achieved remarkable success in natural language generation but lesser focus has been given to their applicability in decision making tasks such as classification. We show that LLMs like LLaMa can achieve high performance on large multi-class classification tasks but still make classification errors and worse, generate out-of-vocabulary class labels. To address these critical issues, we introduce Paraphrase and AGgregate (PAG)-LLM approach wherein an LLM generates multiple paraphrases of the input query (parallel queries), performs multi-class classification for the original query and each paraphrase, and at the end aggregate all the classification labels based on their confidence scores. We evaluate PAG-LLM on two large multi-class classication datasets: CLINC, and Banking and show 22.7% and 15.1% error reduction. We show that PAG-LLM is especially effective for hard examples where LLM is uncertain, and reduces the critical misclassification and hallucinated label generation errors
Self-supervised learning models have revolutionized the field of speech processing. However, the process of fine-tuning these models on downstream tasks requires substantial computational resources, particularly when dealing with multiple speech-processing tasks. In this paper, we explore the potential of adapter-based fine-tuning in developing a unified model capable of effectively handling multiple spoken language processing tasks. The tasks we investigate are Automatic Speech Recognition, Phoneme Recognition, Intent Classification, Slot Filling, and Spoken Emotion Recognition. We validate our approach through a series of experiments on the SUPERB benchmark, and our results indicate that adapter-based fine-tuning enables a single encoder-decoder model to perform multiple speech processing tasks with an average improvement of 18.4% across the five target tasks while staying efficient in terms of parameter updates.
Discrete audio tokens have recently gained considerable attention for their potential to connect audio and language processing, enabling the creation of modern multimodal large language models. Ideal audio tokens must effectively preserve phonetic and semantic content along with paralinguistic information, speaker identity, and other details. While several types of audio tokens have been recently proposed, identifying the optimal tokenizer for various tasks is challenging due to the inconsistent evaluation settings in existing studies. To address this gap, we release the Discrete Audio and Speech Benchmark (DASB), a comprehensive leaderboard for benchmarking discrete audio tokens across a wide range of discriminative tasks, including speech recognition, speaker identification and verification, emotion recognition, keyword spotting, and intent classification, as well as generative tasks such as speech enhancement, separation, and text-to-speech. Our results show that, on average, semantic tokens outperform compression tokens across most discriminative and generative tasks. However, the performance gap between semantic tokens and standard continuous representations remains substantial, highlighting the need for further research in this field.
Recently, multi-task spoken language understanding (SLU) models have emerged, designed to address various speech processing tasks. However, these models often rely on a large number of parameters. Also, they often encounter difficulties in adapting to new data for a specific task without experiencing catastrophic forgetting of previously trained tasks. In this study, we propose finding task-specific subnetworks within a multi-task SLU model via neural network pruning. In addition to model compression, we expect that the forgetting of previously trained tasks can be mitigated by updating only a task-specific subnetwork. We conduct experiments on top of the state-of-the-art multi-task SLU model ``UniverSLU'', trained for several tasks such as emotion recognition (ER), intent classification (IC), and automatic speech recognition (ASR). We show that pruned models were successful in adapting to additional ASR or IC data with minimal performance degradation on previously trained tasks.
Self-supervised speech models (S3Ms) have become an effective backbone for speech applications. Various analyses suggest that S3Ms encode linguistic properties. In this work, we seek a more fine-grained analysis of the word-level linguistic properties encoded in S3Ms. Specifically, we curate a novel dataset of near homophone (phonetically similar) and synonym (semantically similar) word pairs and measure the similarities between S3M word representation pairs. Our study reveals that S3M representations consistently and significantly exhibit more phonetic than semantic similarity. Further, we question whether widely used intent classification datasets such as Fluent Speech Commands and Snips Smartlights are adequate for measuring semantic abilities. Our simple baseline, using only the word identity, surpasses S3M-based models. This corroborates our findings and suggests that high scores on these datasets do not necessarily guarantee the presence of semantic content.
The advent of Internet of Things (IoT) technology has generated massive interest in voice-controlled smart homes. While many voice-controlled smart home systems are designed to understand and support widely spoken languages like English, speakers of low-resource languages like Luganda may need more support. This research project aimed to develop a Luganda speech intent classification system for IoT applications to integrate local languages into smart home environments. The project uses hardware components such as Raspberry Pi, Wio Terminal, and ESP32 nodes as microcontrollers. The Raspberry Pi processes Luganda voice commands, the Wio Terminal is a display device, and the ESP32 nodes control the IoT devices. The ultimate objective of this work was to enable voice control using Luganda, which was accomplished through a natural language processing (NLP) model deployed on the Raspberry Pi. The NLP model utilized Mel Frequency Cepstral Coefficients (MFCCs) as acoustic features and a Convolutional Neural Network (Conv2D) architecture for speech intent classification. A dataset of Luganda voice commands was curated for this purpose and this has been made open-source. This work addresses the localization challenges and linguistic diversity in IoT applications by incorporating Luganda voice commands, enabling users to interact with smart home devices without English proficiency, especially in regions where local languages are predominant.