Larry Heck

Anirudh S Sundar, Larry Heck · gatech

Optimizing accuracy and performance while eliminating hallucinations of open-domain conversational large language models (LLMs) is an open research challenge. A particularly promising direction is to augment and ground LLMs with information from structured sources. This paper introduces Conversational Tables (cTBLS), a three-step architecture to retrieve and generate dialogue responses grounded on retrieved tabular information. cTBLS uses Transformer encoder embeddings for Dense Table Retrieval and obtains up to 125% relative improvement over the retriever in the previous state-of-the-art system on the HyrbiDialogue dataset. cTBLS then uses a shared process between encoder and decoder models to perform a coarse+fine tabular knowledge (e.g., cell) ranking combined with a GPT-3.5 LLM response generator to yield a 2x relative improvement in ROUGE scores. Finally, human evaluators prefer cTBLs +80% of the time (coherency, fluency) and judge informativeness to be 4x better than the previous state-of-the-art.

Christopher Richardson, Larry Heck · gatech

Large, transformer-based pretrained language models like BERT, GPT, and T5 have demonstrated a deep understanding of contextual semantics and language syntax. Their success has enabled significant advances in conversational AI, including the development of open-dialogue systems capable of coherent, salient conversations which can answer questions, chat casually, and complete tasks. However, state-of-the-art models still struggle with tasks that involve higher levels of reasoning - including commonsense reasoning that humans find trivial. This paper presents a survey of recent conversational AI research focused on commonsense reasoning. The paper lists relevant training datasets and describes the primary approaches to include commonsense in conversational AI. The paper also discusses benchmarks used for evaluating commonsense in conversational AI problems. Finally, the paper presents preliminary observations of the limited commonsense capabilities of two state-of-the-art open dialogue models, BlenderBot3 and LaMDA, and its negative effect on natural interactions. These observations further motivate research on commonsense reasoning in conversational AI.

Christopher Richardson, Anirudh Sundar, Larry Heck · gatech

Commonsense reasoning is a critical aspect of human communication. Despite recent advances in conversational AI driven by large language models, commonsense reasoning remains a challenging task. In this work, we introduce SYNDICOM - a method for improving commonsense in dialogue response generation. SYNDICOM consists of two components. The first component is a dataset composed of commonsense dialogues created from a knowledge graph and synthesized into natural language. This dataset includes both valid and invalid responses to dialogue contexts, along with natural language feedback (NLF) for the invalid responses. The second contribution is a two-step procedure: training a model to predict natural language feedback (NLF) for invalid responses, and then training a response generation model conditioned on the predicted NLF, the invalid response, and the dialogue. SYNDICOM is scalable and does not require reinforcement learning. Empirical results on three tasks are evaluated using a broad range of metrics. SYNDICOM achieves a relative improvement of 53% over ChatGPT on ROUGE1, and human evaluators prefer SYNDICOM over ChatGPT 57% of the time. We will publicly release the code and the full dataset.

Benjamin Reichman, Adar Avsian, Kartik Talamadupula et al. · ibm-research

Queries to large language models (LLMs) can be divided into two parts: the instruction/question and the accompanying context. The context for retrieval-augmented generation (RAG) systems in most benchmarks comes from Wikipedia-like texts written in a neutral and factual tone. However, real-world RAG applications often retrieve internet-based text with diverse tones and linguistic styles, posing challenges for downstream tasks. This paper introduces (a) a dataset that transforms RAG-retrieved passages into emotionally inflected and sarcastic text, (b) an emotion translation model for adapting text to different tones, and (c) a prompt-based method to improve LLMs' pragmatic interpretation of retrieved text.

Anirudh Sundar, Larry Heck

As humans, we experience the world with all our senses or modalities (sound, sight, touch, smell, and taste). We use these modalities, particularly sight and touch, to convey and interpret specific meanings. Multimodal expressions are central to conversations; a rich set of modalities amplify and often compensate for each other. A multimodal conversational AI system answers questions, fulfills tasks, and emulates human conversations by understanding and expressing itself via multiple modalities. This paper motivates, defines, and mathematically formulates the multimodal conversational research objective. We provide a taxonomy of research required to solve the objective: multimodal representation, fusion, alignment, translation, and co-learning. We survey state-of-the-art datasets and approaches for each research area and highlight their limiting assumptions. Finally, we identify multimodal co-learning as a promising direction for multimodal conversational AI research.

Benjamin Reichman, Xiaofan Yu, Lanxiang Hu et al.

With the rapid growth in sensor data, effectively interpreting and interfacing with these data in a human-understandable way has become crucial. While existing research primarily focuses on learning classification models, fewer studies have explored how end users can actively extract useful insights from sensor data, often hindered by the lack of a proper dataset. To address this gap, we introduce SensorQA, the first human-created question-answering (QA) dataset for long-term time-series sensor data for daily life monitoring. SensorQA is created by human workers and includes 5.6K diverse and practical queries that reflect genuine human interests, paired with accurate answers derived from sensor data. We further establish benchmarks for state-of-the-art AI models on this dataset and evaluate their performance on typical edge devices. Our results reveal a gap between current models and optimal QA performance and efficiency, highlighting the need for new contributions. The dataset and code are available at: https://github.com/benjamin-reichman/SensorQA.

Benjamin Reichman, Adar Avasian, Samuel Webster et al.

Large language models are routinely deployed on text that varies widely in emotional tone, yet their reasoning behavior is typically evaluated without accounting for emotion as a source of representational variation. Prior work has largely treated emotion as a prediction target, for example in sentiment analysis or emotion classification. In contrast, we study emotion as a latent factor that shapes how models attend to and reason over text. We analyze how emotional tone systematically alters attention geometry in transformer models, showing that metrics such as locality, center-of-mass distance, and entropy vary across emotions and correlate with downstream question-answering performance. To facilitate controlled study of these effects, we introduce Affect-Uniform ReAding QA (AURA-QA), a question-answering dataset with emotionally balanced, human-authored context passages. Finally, an emotional regularization framework is proposed that constrains emotion-conditioned representational drift during training. Experiments across multiple QA benchmarks demonstrate that this approach improves reading comprehension in both emotionally-varying and non-emotionally varying datasets, yielding consistent gains under distribution shift and in-domain improvements on several benchmarks.

Adar Avsian, Christopher Richardson, Anirudh Sundar et al.

Language models have become effective at a wide range of tasks, from math problem solving to open-domain question answering. However, they still make mistakes, and these mistakes are often repeated across related queries. Natural language explanations can help correct these errors, but collecting them at scale may be infeasible, particularly in domains where expert annotators are required. To address this issue, we introduce FLEx ($\textbf{F}$ew-shot $\textbf{L}$anguage $\textbf{Ex}$planations), a method for improving model behavior using a small number of explanatory examples. FLEx selects representative model errors using embedding-based clustering, verifies that the associated explanations correct those errors, and summarizes them into a prompt prefix that is prepended at inference-time. This summary guides the model to avoid similar errors on new inputs, without modifying model weights. We evaluate FLEx on CounterBench, GSM8K, and ReasonIF. We find that FLEx consistently outperforms chain-of-thought (CoT) prompting across all three datasets and reduces up to 83\% of CoT's remaining errors.

Benjamin Reichman, Constantin Patsch, Jack Truxal et al.

In outside knowledge visual question answering (OK-VQA), the model must identify relevant visual information within an image and incorporate external knowledge to accurately respond to a question. Extending this task to a visually grounded dialogue setting based on videos, a conversational model must both recognize pertinent visual details over time and answer questions where the required information is not necessarily present in the visual information. Moreover, the context of the overall conversation must be considered for the subsequent dialogue. To explore this task, we introduce a dataset comprised of $2,017$ videos with $5,986$ human-annotated dialogues consisting of $40,954$ interleaved dialogue turns. While the dialogue context is visually grounded in specific video segments, the questions further require external knowledge that is not visually present. Thus, the model not only has to identify relevant video parts but also leverage external knowledge to converse within the dialogue. We further provide several baselines evaluated on our dataset and show future challenges associated with this task. The dataset is made publicly available here: https://github.com/c-patsch/OKCV.

Saif Punjwani, Larry Heck

As virtual agents become increasingly prevalent in human-computer interaction, generating realistic and contextually appropriate gestures in real-time remains a significant challenge. While neural rendering techniques have made substantial progress with static scripts, their applicability to human-computer interactions remains limited. To address this, we introduce Large Body Language Models (LBLMs) and present LBLM-AVA, a novel LBLM architecture that combines a Transformer-XL large language model with a parallelized diffusion model to generate human-like gestures from multimodal inputs (text, audio, and video). LBLM-AVA incorporates several key components enhancing its gesture generation capabilities, such as multimodal-to-pose embeddings, enhanced sequence-to-sequence mapping with redefined attention mechanisms, a temporal smoothing module for gesture sequence coherence, and an attention-based refinement module for enhanced realism. The model is trained on our large-scale proprietary open-source dataset Allo-AVA. LBLM-AVA achieves state-of-the-art performance in generating lifelike and contextually appropriate gestures with a 30% reduction in Fréchet Gesture Distance (FGD), and a 25% improvement in Fréchet Inception Distance compared to existing approaches.

Tyler Lizzo, Larry Heck

Given the prevalence of large language models (LLMs) and the prohibitive cost of training these models from scratch, dynamically forgetting specific knowledge e.g., private or proprietary, without retraining the model has become an important capability. This paper proposes a novel method to achieve this objective called UNLEARN. The approach builds upon subspace methods to identify and specifically target the removal of knowledge without adversely affecting other knowledge in the LLM. Results demonstrate 96% of targeted knowledge can be forgotten while maintaining performance on other knowledge within 2.5% of the original model, significantly outperforming the discriminatory abilities of the previous state-of-the-art. A dual method called LEARN is also proposed for targeted knowledge addition. Results show LEARN can match the fine-tuning accuracy of Low-Rank Adaptation (LoRA) without adversely affecting similar tasks.

Chao-Han Huck Yang, Andreas Stolcke, Larry Heck

Spoken conversational agents are converging toward voice-native LLMs. This tutorial distills the path from cascaded ASR/NLU to end-to-end, retrieval-and vision-grounded systems. We frame adaptation of text LLMs to audio, cross-modal alignment, and joint speech-text training; review datasets, metrics, and robustness across accents and compare design choices (cascaded vs. E2E, post-ASR correction, streaming). We link industrial assistants to current open-domain and task-oriented agents, highlight reproducible baselines, and outline open problems in privacy, safety, and evaluation. Attendees leave with practical recipes and a clear systems-level roadmap.

Anirudh Sundar, Christopher Richardson, Larry Heck

Distilling large, unstructured text into a structured, condensed form such as tables is an open research problem. One of the primary challenges in automatically generating tables is ensuring their syntactic validity. Prior approaches address this challenge by including additional parameters in the Transformer's attention mechanism to attend to specific rows and column headers. In contrast to this single-stage method, this paper presents a two-stage approach called Generative Tables (gTBLS). The first stage infers table structure (row and column headers) from the text. The second stage formulates questions using these headers and fine-tunes a causal language model to answer them. Furthermore, the gTBLS approach is amenable to the utilization of pre-trained Large Language Models in a zero-shot configuration, presenting a solution for table generation in situations where fine-tuning is not feasible. gTBLS improves prior approaches by up to 10% in BERTScore on the table construction task and up to 20% on the table content generation task of the E2E, WikiTableText, WikiBio, and RotoWire datasets.

Benjamin Reichman, Larry Heck

Dense passage retrieval (DPR) is the first step in the retrieval augmented generation (RAG) paradigm for improving the performance of large language models (LLM). DPR fine-tunes pre-trained networks to enhance the alignment of the embeddings between queries and relevant textual data. A deeper understanding of DPR fine-tuning will be required to fundamentally unlock the full potential of this approach. In this work, we explore DPR-trained models mechanistically by using a combination of probing, layer activation analysis, and model editing. Our experiments show that DPR training decentralizes how knowledge is stored in the network, creating multiple access pathways to the same information. We also uncover a limitation in this training style: the internal knowledge of the pre-trained model bounds what the retrieval model can retrieve. These findings suggest a few possible directions for dense retrieval: (1) expose the DPR training process to more knowledge so more can be decentralized, (2) inject facts as decentralized representations, (3) model and incorporate knowledge uncertainty in the retrieval process, and (4) directly map internal model knowledge to a knowledge base.

Xiaofan Yu, Lanxiang Hu, Benjamin Reichman et al.

Natural language interaction with sensing systems is crucial for addressing users' personal concerns and providing health-related insights into their daily lives. When a user asks a question, the system automatically analyzes the full history of sensor data, extracts relevant information, and generates an appropriate response. However, existing systems are limited to short-duration (e.g., one minute) or low-frequency (e.g., daily step count) sensor data. In addition, they struggle with quantitative questions that require precise numerical answers. In this work, we introduce SensorChat, the first end-to-end QA system designed for daily life monitoring using long-duration, high-frequency time series data. Given raw sensor signals spanning multiple days and a user-defined natural language question, SensorChat generates semantically meaningful responses that directly address user concerns. SensorChat effectively handles both quantitative questions that require numerical precision and qualitative questions that require high-level reasoning to infer subjective insights. To achieve this, SensorChat uses an innovative three-stage pipeline including question decomposition, sensor data query, and answer assembly. The first and third stages leverage Large Language Models (LLMs) to interpret human queries and generate responses. The intermediate querying stage extracts relevant information from the complete sensor data history. Real-world implementations demonstrate SensorChat's capability for real-time interactions on a cloud server while also being able to run entirely on edge platforms after quantization. Comprehensive QA evaluations show that SensorChat achieves 93% higher answer accuracy than the best performing state-of-the-art systems on quantitative questions. Furthermore, a user study with eight volunteers highlights SensorChat's effectiveness in answering qualitative questions.

Adar Avsian, Larry Heck

Large language models (LLMs) are increasingly deployed in multi-agent settings where communication must balance informativeness and secrecy. In such settings, an agent may need to signal information to collaborators while preventing an adversary from inferring sensitive details. However, existing LLM benchmarks primarily evaluate capabilities such as reasoning, factual knowledge, or instruction following, and do not directly measure strategic communication under asymmetric information. We introduce SNEAK (Secret-aware Natural language Evaluation for Adversarial Knowledge), a benchmark for evaluating selective information sharing in language models. In SNEAK, a model is given a semantic category, a candidate set of words, and a secret word, and must generate a message that indicates knowledge of the secret without revealing it too clearly. We evaluate generated messages using two simulated agents with different information states: an ally, who knows the secret and must identify the intended message, and a chameleon, who does not know the secret and attempts to infer it from the message. This yields two complementary metrics: utility, measuring how well the message communicates to collaborators, and leakage, measuring how much information it reveals to an adversary. Using this framework, we analyze the trade-off between informativeness and secrecy in modern language models and show that strategic communication under asymmetric information remains a challenging capability for current systems. Notably, human participants outperform all evaluated models by a large margin, achieving up to four times higher scores.

Saif Punjwani, Larry Heck

The scarcity of high-quality, multimodal training data severely hinders the creation of lifelike avatar animations for conversational AI in virtual environments. Existing datasets often lack the intricate synchronization between speech, facial expressions, and body movements that characterize natural human communication. To address this critical gap, we introduce Allo-AVA, a large-scale dataset specifically designed for text and audio-driven avatar gesture animation in an allocentric (third person point-of-view) context. Allo-AVA consists of $\sim$1,250 hours of diverse video content, complete with audio, transcripts, and extracted keypoints. Allo-AVA uniquely maps these keypoints to precise timestamps, enabling accurate replication of human movements (body and facial gestures) in synchronization with speech. This comprehensive resource enables the development and evaluation of more natural, context-aware avatar animation models, potentially transforming applications ranging from virtual reality to digital assistants.

Toshish Jawale, Chaitanya Animesh, Sekhar Vallath et al. · ibm-research

This study analyzes changes in the attention mechanisms of large language models (LLMs) when used to understand natural conversations between humans (human-human). We analyze three use cases of LLMs: interactions over web content, code, and mathematical texts. By analyzing attention distance, dispersion, and interdependency across these domains, we highlight the unique challenges posed by conversational data. Notably, conversations require nuanced handling of long-term contextual relationships and exhibit higher complexity through their attention patterns. Our findings reveal that while language models exhibit domain-specific attention behaviors, there is a significant gap in their ability to specialize in human conversations. Through detailed attention entropy analysis and t-SNE visualizations, we demonstrate the need for models trained with a diverse array of high-quality conversational data to enhance understanding and generation of human-like dialogue. This research highlights the importance of domain specialization in language models and suggests pathways for future advancement in modeling human conversational nuances.

Anirudh Sundar, Christopher Richardson, Adar Avsian et al.

This paper introduces Interactive Tables (iTBLS), a dataset of interactive conversations that focuses on natural-language manipulation of tabular information sourced from academic pre-prints on ArXiv. The iTBLS dataset consists of three types of tabular tasks -- interpretation, modification, and generation. Interpretation focuses on tabular understanding, modification focuses on manipulating tabular information, and generation focuses on the addition of new natural-language evidence. In addition, the paper presents a novel framework that reformulates tabular operations as question-answering, where an appropriate question is formulated based on the nature of interaction and the question is answered using the user request as evidence. The developed approach results in an improvement on all tasks on a sequence-to-sequence modeling baseline on iTBLS. In addition, the question-answering-based reformulation is applied to datasets from prior work for the text-to-table task where textual paragraphs are summarized into tables. The novel approach results in up to 13% improvement in Exact-Match accuracy and up to 16% improvement in BERTScores compared to the prior state-of-the-art.

Atahan Dokme, Benjamin Reichman, Larry Heck

Large language models are trained and evaluated on quantitative reasoning tasks written in clean, emotionally neutral language. However, real-world queries are often wrapped in frustration, urgency or enthusiasm. Does emotional framing alone degrade reasoning when all numerical content is preserved? To investigate this, a controlled emotion translation framework is developed that rewrites problems into emotional variants while preserving all quantities and relationships. Using this framework, Temper-5400 (5,400 semantically verified emotion--neutral pairs) is constructed across GSM8K, MultiArith, and ARC-Challenge, and evaluated on eighteen models (1B to frontier scale). Two core results emerge: First, emotional framing reduces accuracy by 2-10 percentage points even though all numerical content is preserved. Second, neutralizing emotional variants recovers most of the lost performance, showing both that the degradation is tied to emotional style rather than content corruption and that neutralization can serve as a lightweight inference-time mitigation. Non-emotional paraphrases cause no such degradation, implicating emotional content rather than surface-level changes. Beyond emotion specifically, the benchmark construction procedure provides a general framework for controlled stylistic translation and robustness evaluation.

Benjamin Reichman, Adar Avsian, Larry Heck · gatech

Queries to large language models (LLMs) can be divided into two parts: the instruction/question and the accompanying context. The context for retrieval-augmented generation (RAG) systems in most benchmarks comes from Wikipedia or Wikipedia-like texts which are written in a neutral and factual tone. However, when RAG systems retrieve internet-based content, they encounter text with diverse tones and linguistic styles, introducing challenges for downstream tasks. The Reading with Intent task addresses this issue by evaluating how varying tones in context passages affect model performance. Building on prior work that focused on sarcasm, we extend this paradigm by constructing a dataset where context passages are transformed to $11$ distinct emotions using a better synthetic data generation approach. Using this dataset, we train an emotion translation model to systematically adapt passages to specified emotional tones. The human evaluation shows that the LLM fine-tuned to become the emotion-translator benefited from the synthetically generated data. Finally, the emotion-translator is used in the Reading with Intent task to transform the passages to a neutral tone. By neutralizing the passages, it mitigates the challenges posed by sarcastic passages and improves overall results on this task by about $3\%$.

Tyler Lizzo, Larry Heck

Large language models (LLMs) have achieved remarkable success across natural language processing tasks, yet their widespread deployment raises pressing concerns around privacy, copyright, security, and bias. Machine unlearning has emerged as a promising paradigm for selectively removing knowledge or data from trained models without full retraining. In this survey, we provide a structured overview of unlearning methods for LLMs, categorizing existing approaches into data-centric, parameter-centric, architecture-centric, hybrid, and other strategies. We also review the evaluation ecosystem, including benchmarks, metrics, and datasets designed to measure forgetting effectiveness, knowledge retention, and robustness. Finally, we outline key challenges and open problems, such as scalable efficiency, formal guarantees, cross-language and multimodal unlearning, and robustness against adversarial relearning. By synthesizing current progress and highlighting open directions, this paper aims to serve as a roadmap for developing reliable and responsible unlearning techniques in large language models.

Benjamin Reichman, Adar Avsian, Larry Heck · gatech

This work investigates how large language models (LLMs) internally represent emotion by analyzing the geometry of their hidden-state space. The paper identifies a low-dimensional emotional manifold and shows that emotional representations are directionally encoded, distributed across layers, and aligned with interpretable dimensions. These structures are stable across depth and generalize to eight real-world emotion datasets spanning five languages. Cross-domain alignment yields low error and strong linear probe performance, indicating a universal emotional subspace. Within this space, internal emotion perception can be steered while preserving semantics using a learned intervention module, with especially strong control for basic emotions across languages. These findings reveal a consistent and manipulable affective geometry in LLMs and offer insight into how they internalize and process emotion.

Shrenik Bhansali, Larry Heck

Autoregressive (AR) decoding is a major latency bottleneck for large language models. Speculative decoding (SD) accelerates AR by letting a drafter propose multi-token blocks that a verifier accepts or rejects. However, many SD systems require heavy offline training or extra components. These choices raise data/compute cost and can yield brittle drafters under distribution drift. We introduce \emph{Draft, Verify, \& Improve (DVI)}, a training-aware self-speculative framework that combines inference with continual online learning. We partition an LLM into a drafter and a verifier, and during generation, verifier accept/reject decisions are converted into supervision signals and used to update the drafter head. A simple \emph{KL$\rightarrow$RL} schedule bootstraps calibration via online distillation and then adds reward-masked cross-entropy with a on-policy policy-gradient term, preserving lossless, single model deployment. On Spec-Bench, DVI achieves a $2.16\times$ wall-time speedup, on par with SoTA approaches like EAGLE-2, while orders of magnitude less data for training, and ablations show that DVI outperforms KL-only online distillation. DVI demonstrates that \emph{training-aware} self-speculation can deliver state-of-the-art, lossless speedups with minimal training overhead.

Saif Punjwani, Larry Heck

Large language models (LLMs) have demonstrated remarkable reasoning capabilities when prompted with strategies such as Chain-of-Thought (CoT). However, these approaches focus on token-level output without considering internal weight dynamics. We introduce Weight-of-Thought (WoT) reasoning, a novel approach that examines neural network weights before inference to identify reasoning pathways. Unlike existing methods, WoT explores the weight space through graph-based message passing, multi-step reasoning processes, and attention mechanisms. Our implementation creates an interconnected graph of reasoning nodes. Experiments on diverse reasoning tasks (syllogistic, mathematical, algebraic, combinatorial, and geometric) demonstrate that WoT achieves superior performance compared to traditional methods, particularly for complex problems. This approach leads to both improved performance and greater interpretability of the reasoning process, offering a promising direction for enhancing LLM reasoning capabilities.

Shrenik Bhansali, Alwin Jin, Tyler Lizzo et al.

Large language models (LLMs) are essential in natural language processing (NLP) but are costly in data collection, pre-training, fine-tuning, and inference. Task-specific small language models (SLMs) offer a cheaper alternative but lack robustness and generalization. This paper proposes LEGO, a novel technique to extract SLMs from an LLM and recombine them. Using state-of-the-art LLM pruning strategies, we can create task- and user-specific SLM building blocks that are efficient for fine-tuning and inference while also preserving user data privacy. LEGO utilizes Federated Learning and a novel aggregation scheme for the LLM reconstruction, maintaining robustness without high costs and preserving user data privacy. We experimentally demonstrate the versatility of LEGO, showing its ability to enable model heterogeneity and mitigate the effects of data heterogeneity while maintaining LLM robustness.

Abhinav Narayan Harish, Larry Heck, Josiah P. Hanna et al.

We present Reinforcement Learning via Auxiliary Task Distillation (AuxDistill), a new method that enables reinforcement learning (RL) to perform long-horizon robot control problems by distilling behaviors from auxiliary RL tasks. AuxDistill achieves this by concurrently carrying out multi-task RL with auxiliary tasks, which are easier to learn and relevant to the main task. A weighted distillation loss transfers behaviors from these auxiliary tasks to solve the main task. We demonstrate that AuxDistill can learn a pixels-to-actions policy for a challenging multi-stage embodied object rearrangement task from the environment reward without demonstrations, a learning curriculum, or pre-trained skills. AuxDistill achieves $2.3 \times$ higher success than the previous state-of-the-art baseline in the Habitat Object Rearrangement benchmark and outperforms methods that use pre-trained skills and expert demonstrations.

Anirudh Sundar, Jin Xu, William Gay et al.

An emerging area of research in situated and multimodal interactive conversations (SIMMC) includes interactions in scientific papers. Since scientific papers are primarily composed of text, equations, figures, and tables, SIMMC methods must be developed specifically for each component to support the depth of inquiry and interactions required by research scientists. This work introduces Conversational Papers (cPAPERS), a dataset of conversational question-answer pairs from reviews of academic papers grounded in these paper components and their associated references from scientific documents available on arXiv. We present a data collection strategy to collect these question-answer pairs from OpenReview and associate them with contextual information from LaTeX source files. Additionally, we present a series of baseline approaches utilizing Large Language Models (LLMs) in both zero-shot and fine-tuned configurations to address the cPAPERS dataset.

Nancy Xu, Sam Masling, Michael Du et al.

Grounding natural language instructions on the web to perform previously unseen tasks enables accessibility and automation. We introduce a task and dataset to train AI agents from open-domain, step-by-step instructions originally written for people. We build RUSS (Rapid Universal Support Service) to tackle this problem. RUSS consists of two models: First, a BERT-LSTM with pointers parses instructions to ThingTalk, a domain-specific language we design for grounding natural language on the web. Then, a grounding model retrieves the unique IDs of any webpage elements requested in ThingTalk. RUSS may interact with the user through a dialogue (e.g. ask for an address) or execute a web operation (e.g. click a button) inside the web runtime. To augment training, we synthesize natural language instructions mapped to ThingTalk. Our dataset consists of 80 different customer service problems from help websites, with a total of 741 step-by-step instructions and their corresponding actions. RUSS achieves 76.7% end-to-end accuracy predicting agent actions from single instructions. It outperforms state-of-the-art models that directly map instructions to actions without ThingTalk. Our user study shows that RUSS is preferred by actual users over web navigation.

Larry Heck, Simon Heck, Anirudh Sundar

This paper presents a new approach to form-filling by reformulating the task as multimodal natural language Question Answering (QA). The reformulation is achieved by first translating the elements on the GUI form (text fields, buttons, icons, etc.) to natural language questions, where these questions capture the element's multimodal semantics. After a match is determined between the form element (Question) and the user utterance (Answer), the form element is filled through a pre-trained extractive QA system. By leveraging pre-trained QA models and not requiring form-specific training, this approach to form-filling is zero-shot. The paper also presents an approach to further refine the form-filling by using multi-task training to incorporate a potentially large number of successive tasks. Finally, the paper introduces a multimodal natural language form-filling dataset Multimodal Forms (mForms), as well as a multimodal extension of the popular ATIS dataset to support future research and experimentation. Results show the new approach not only maintains robust accuracy for sparse training conditions but achieves state-of-the-art F1 of 0.97 on ATIS with approximately 1/10th of the training data.

Chanwoo Kim, Sungsoo Kim, Kwangyoun Kim et al.

In this paper, we present an end-to-end training framework for building state-of-the-art end-to-end speech recognition systems. Our training system utilizes a cluster of Central Processing Units(CPUs) and Graphics Processing Units (GPUs). The entire data reading, large scale data augmentation, neural network parameter updates are all performed "on-the-fly". We use vocal tract length perturbation [1] and an acoustic simulator [2] for data augmentation. The processed features and labels are sent to the GPU cluster. The Horovod allreduce approach is employed to train neural network parameters. We evaluated the effectiveness of our system on the standard Librispeech corpus [3] and the 10,000-hr anonymized Bixby English dataset. Our end-to-end speech recognition system built using this training infrastructure showed a 2.44 % WER on test-clean of the LibriSpeech test set after applying shallow fusion with a Transformer language model (LM). For the proprietary English Bixby open domain test set, we obtained a WER of 7.92 % using a Bidirectional Full Attention (BFA) end-to-end model after applying shallow fusion with an RNN-LM. When the monotonic chunckwise attention (MoCha) based approach is employed for streaming speech recognition, we obtained a WER of 9.95 % on the same Bixby open domain test set.

Dawei Li, Serafettin Tasci, Shalini Ghosh et al.

Object detection models shipped with camera-equipped edge devices cannot cover the objects of interest for every user. Therefore, the incremental learning capability is a critical feature for a robust and personalized object detection system that many applications would rely on. In this paper, we present an efficient yet practical system, RILOD, to incrementally train an existing object detection model such that it can detect new object classes without losing its capability to detect old classes. The key component of RILOD is a novel incremental learning algorithm that trains end-to-end for one-stage deep object detection models only using training data of new object classes. Specifically to avoid catastrophic forgetting, the algorithm distills three types of knowledge from the old model to mimic the old model's behavior on object classification, bounding box regression and feature extraction. In addition, since the training data for the new classes may not be available, a real-time dataset construction pipeline is designed to collect training images on-the-fly and automatically label the images with both category and bounding box annotations. We have implemented RILOD under both edge-cloud and edge-only setups. Experiment results show that the proposed system can learn to detect a new object class in just a few minutes, including both dataset construction and model training. In comparison, traditional fine-tuning based method may take a few hours for training, and in most cases would also need a tedious and costly manual dataset labeling step.

Junting Zhang, Jie Zhang, Shalini Ghosh et al.

Deep neural networks (DNNs) often suffer from "catastrophic forgetting" during incremental learning (IL) --- an abrupt degradation of performance on the original set of classes when the training objective is adapted to a newly added set of classes. Existing IL approaches tend to produce a model that is biased towards either the old classes or new classes, unless with the help of exemplars of the old data. To address this issue, we propose a class-incremental learning paradigm called Deep Model Consolidation (DMC), which works well even when the original training data is not available. The idea is to first train a separate model only for the new classes, and then combine the two individual models trained on data of two distinct set of classes (old classes and new classes) via a novel double distillation training objective. The two existing models are consolidated by exploiting publicly available unlabeled auxiliary data. This overcomes the potential difficulties due to the unavailability of original training data. Compared to the state-of-the-art techniques, DMC demonstrates significantly better performance in image classification (CIFAR-100 and CUB-200) and object detection (PASCAL VOC 2007) in the single-headed IL setting.

Heming Zhang, Shalini Ghosh, Larry Heck et al.

The key challenge of generative Visual Dialogue (VD) systems is to respond to human queries with informative answers in natural and contiguous conversation flow. Traditional Maximum Likelihood Estimation (MLE)-based methods only learn from positive responses but ignore the negative responses, and consequently tend to yield safe or generic responses. To address this issue, we propose a novel training scheme in conjunction with weighted likelihood estimation (WLE) method. Furthermore, an adaptive multi-modal reasoning module is designed, to accommodate various dialogue scenarios automatically and select relevant information accordingly. The experimental results on the VisDial benchmark demonstrate the superiority of our proposed algorithm over other state-of-the-art approaches, with an improvement of 5.81% on recall@10.

Ramprasaath R. Selvaraju, Stefan Lee, Yilin Shen et al.

Many vision and language models suffer from poor visual grounding - often falling back on easy-to-learn language priors rather than basing their decisions on visual concepts in the image. In this work, we propose a generic approach called Human Importance-aware Network Tuning (HINT) that effectively leverages human demonstrations to improve visual grounding. HINT encourages deep networks to be sensitive to the same input regions as humans. Our approach optimizes the alignment between human attention maps and gradient-based network importances - ensuring that models learn not just to look at but rather rely on visual concepts that humans found relevant for a task when making predictions. We apply HINT to Visual Question Answering and Image Captioning tasks, outperforming top approaches on splits that penalize over-reliance on language priors (VQA-CP and robust captioning) using human attention demonstrations for just 6% of the training data.

Bing Liu, Gokhan Tur, Dilek Hakkani-Tur et al.

In this work, we present a hybrid learning method for training task-oriented dialogue systems through online user interactions. Popular methods for learning task-oriented dialogues include applying reinforcement learning with user feedback on supervised pre-training models. Efficiency of such learning method may suffer from the mismatch of dialogue state distribution between offline training and online interactive learning stages. To address this challenge, we propose a hybrid imitation and reinforcement learning method, with which a dialogue agent can effectively learn from its interaction with users by learning from human teaching and feedback. We design a neural network based task-oriented dialogue agent that can be optimized end-to-end with the proposed learning method. Experimental results show that our end-to-end dialogue agent can learn effectively from the mistake it makes via imitation learning from user teaching. Applying reinforcement learning with user feedback after the imitation learning stage further improves the agent's capability in successfully completing a task.

Pararth Shah, Dilek Hakkani-Tür, Gokhan Tür et al.

We propose Machines Talking To Machines (M2M), a framework combining automation and crowdsourcing to rapidly bootstrap end-to-end dialogue agents for goal-oriented dialogues in arbitrary domains. M2M scales to new tasks with just a task schema and an API client from the dialogue system developer, but it is also customizable to cater to task-specific interactions. Compared to the Wizard-of-Oz approach for data collection, M2M achieves greater diversity and coverage of salient dialogue flows while maintaining the naturalness of individual utterances. In the first phase, a simulated user bot and a domain-agnostic system bot converse to exhaustively generate dialogue "outlines", i.e. sequences of template utterances and their semantic parses. In the second phase, crowd workers provide contextual rewrites of the dialogues to make the utterances more natural while preserving their meaning. The entire process can finish within a few hours. We propose a new corpus of 3,000 dialogues spanning 2 domains collected with M2M, and present comparisons with popular dialogue datasets on the quality and diversity of the surface forms and dialogue flows.

Abhinav Rastogi, Dilek Hakkani-Tur, Larry Heck

Dialogue state tracking (DST) is a key component of task-oriented dialogue systems. DST estimates the user's goal at each user turn given the interaction until then. State of the art approaches for state tracking rely on deep learning methods, and represent dialogue state as a distribution over all possible slot values for each slot present in the ontology. Such a representation is not scalable when the set of possible values are unbounded (e.g., date, time or location) or dynamic (e.g., movies or usernames). Furthermore, training of such models requires labeled data, where each user turn is annotated with the dialogue state, which makes building models for new domains challenging. In this paper, we present a scalable multi-domain deep learning based approach for DST. We introduce a novel framework for state tracking which is independent of the slot value set, and represent the dialogue state as a distribution over a set of values of interest (candidate set) derived from the dialogue history or knowledge. Restricting these candidate sets to be bounded in size addresses the problem of slot-scalability. Furthermore, by leveraging the slot-independent architecture and transfer learning, we show that our proposed approach facilitates quick adaptation to new domains.

Saurabh Kumar, Pararth Shah, Dilek Hakkani-Tur et al.

We present a framework combining hierarchical and multi-agent deep reinforcement learning approaches to solve coordination problems among a multitude of agents using a semi-decentralized model. The framework extends the multi-agent learning setup by introducing a meta-controller that guides the communication between agent pairs, enabling agents to focus on communicating with only one other agent at any step. This hierarchical decomposition of the task allows for efficient exploration to learn policies that identify globally optimal solutions even as the number of collaborating agents increases. We show promising initial experimental results on a simulated distributed scheduling problem.

Bing Liu, Gokhan Tur, Dilek Hakkani-Tur et al.

In this paper, we present a neural network based task-oriented dialogue system that can be optimized end-to-end with deep reinforcement learning (RL). The system is able to track dialogue state, interface with knowledge bases, and incorporate query results into agent's responses to successfully complete task-oriented dialogues. Dialogue policy learning is conducted with a hybrid supervised and deep RL methods. We first train the dialogue agent in a supervised manner by learning directly from task-oriented dialogue corpora, and further optimize it with deep RL during its interaction with users. In the experiments on two different dialogue task domains, our model demonstrates robust performance in tracking dialogue state and producing reasonable system responses. We show that deep RL based optimization leads to significant improvement on task success rate and reduction in dialogue length comparing to supervised training model. We further show benefits of training task-oriented dialogue model end-to-end comparing to component-wise optimization with experiment results on dialogue simulations and human evaluations.

Ankur Bapna, Gokhan Tur, Dilek Hakkani-Tur et al.

State-of-the-art slot filling models for goal-oriented human/machine conversational language understanding systems rely on deep learning methods. While multi-task training of such models alleviates the need for large in-domain annotated datasets, bootstrapping a semantic parsing model for a new domain using only the semantic frame, such as the back-end API or knowledge graph schema, is still one of the holy grail tasks of language understanding for dialogue systems. This paper proposes a deep learning based approach that can utilize only the slot description in context without the need for any labeled or unlabeled in-domain examples, to quickly bootstrap a new domain. The main idea of this paper is to leverage the encoding of the slot names and descriptions within a multi-task deep learned slot filling model, to implicitly align slots across domains. The proposed approach is promising for solving the domain scaling problem and eliminating the need for any manually annotated data or explicit schema alignment. Furthermore, our experiments on multiple domains show that this approach results in significantly better slot-filling performance when compared to using only in-domain data, especially in the low data regime.

Mason Bretan, Sageev Oore, Doug Eck et al.

In this work we describe and evaluate methods to learn musical embeddings. Each embedding is a vector that represents four contiguous beats of music and is derived from a symbolic representation. We consider autoencoding-based methods including denoising autoencoders, and context reconstruction, and evaluate the resulting embeddings on a forward prediction and a classification task.

Ankur Bapna, Gokhan Tur, Dilek Hakkani-Tur et al.

Spoken Language Understanding (SLU) is a key component of goal oriented dialogue systems that would parse user utterances into semantic frame representations. Traditionally SLU does not utilize the dialogue history beyond the previous system turn and contextual ambiguities are resolved by the downstream components. In this paper, we explore novel approaches for modeling dialogue context in a recurrent neural network (RNN) based language understanding system. We propose the Sequential Dialogue Encoder Network, that allows encoding context from the dialogue history in chronological order. We compare the performance of our proposed architecture with two context models, one that uses just the previous turn context and another that encodes dialogue context in a memory network, but loses the order of utterances in the dialogue history. Experiments with a multi-domain dialogue dataset demonstrate that the proposed architecture results in reduced semantic frame error rates.

Mason Bretan, Gil Weinberg, Larry Heck

Several methods exist for a computer to generate music based on data including Markov chains, recurrent neural networks, recombinancy, and grammars. We explore the use of unit selection and concatenation as a means of generating music using a procedure based on ranking, where, we consider a unit to be a variable length number of measures of music. We first examine whether a unit selection method, that is restricted to a finite size unit library, can be sufficient for encompassing a wide spectrum of music. We do this by developing a deep autoencoder that encodes a musical input and reconstructs the input by selecting from the library. We then describe a generative model that combines a deep structured semantic model (DSSM) with an LSTM to predict the next unit, where units consist of four, two, and one measures of music. We evaluate the generative model using objective metrics including mean rank and accuracy and with a subjective listening test in which expert musicians are asked to complete a forced-choiced ranking task. We compare our model to a note-level generative baseline that consists of a stacked LSTM trained to predict forward by one note.

Lu Wang, Larry Heck, Dilek Hakkani-Tur

Training statistical dialog models in spoken dialog systems (SDS) requires large amounts of annotated data. The lack of scalable methods for data mining and annotation poses a significant hurdle for state-of-the-art statistical dialog managers. This paper presents an approach that directly leverage billions of web search and browse sessions to overcome this hurdle. The key insight is that task completion through web search and browse sessions is (a) predictable and (b) generalizes to spoken dialog task completion. The new method automatically mines behavioral search and browse patterns from web logs and translates them into spoken dialog models. We experiment with naturally occurring spoken dialogs and large scale web logs. Our session-based models outperform the state-of-the-art method for entity extraction task in SDS. We also achieve better performance for both entity and relation extraction on web search queries when compared with nontrivial baselines.

Aaron Jaech, Larry Heck, Mari Ostendorf

The goal of this paper is to use multi-task learning to efficiently scale slot filling models for natural language understanding to handle multiple target tasks or domains. The key to scalability is reducing the amount of training data needed to learn a model for a new task. The proposed multi-task model delivers better performance with less data by leveraging patterns that it learns from the other tasks. The approach supports an open vocabulary, which allows the models to generalize to unseen words, which is particularly important when very little training data is used. A newly collected crowd-sourced data set, covering four different domains, is used to demonstrate the effectiveness of the domain adaptation and open vocabulary techniques.

Shalini Ghosh, Oriol Vinyals, Brian Strope et al.

Documents exhibit sequential structure at multiple levels of abstraction (e.g., sentences, paragraphs, sections). These abstractions constitute a natural hierarchy for representing the context in which to infer the meaning of words and larger fragments of text. In this paper, we present CLSTM (Contextual LSTM), an extension of the recurrent neural network LSTM (Long-Short Term Memory) model, where we incorporate contextual features (e.g., topics) into the model. We evaluate CLSTM on three specific NLP tasks: word prediction, next sentence selection, and sentence topic prediction. Results from experiments run on two corpora, English documents in Wikipedia and a subset of articles from a recent snapshot of English Google News, indicate that using both words and topics as features improves performance of the CLSTM models over baseline LSTM models for these tasks. For example on the next sentence selection task, we get relative accuracy improvements of 21% for the Wikipedia dataset and 18% for the Google News dataset. This clearly demonstrates the significant benefit of using context appropriately in natural language (NL) tasks. This has implications for a wide variety of NL applications like question answering, sentence completion, paraphrase generation, and next utterance prediction in dialog systems.

Hongzhao Huang, Larry Heck, Heng Ji

Entity Disambiguation aims to link mentions of ambiguous entities to a knowledge base (e.g., Wikipedia). Modeling topical coherence is crucial for this task based on the assumption that information from the same semantic context tends to belong to the same topic. This paper presents a novel deep semantic relatedness model (DSRM) based on deep neural networks (DNN) and semantic knowledge graphs (KGs) to measure entity semantic relatedness for topical coherence modeling. The DSRM is directly trained on large-scale KGs and it maps heterogeneous types of knowledge of an entity from KGs to numerical feature vectors in a latent space such that the distance between two semantically-related entities is minimized. Compared with the state-of-the-art relatedness approach proposed by (Milne and Witten, 2008a), the DSRM obtains 19.4% and 24.5% reductions in entity disambiguation errors on two publicly available datasets respectively.

Yann N. Dauphin, Gokhan Tur, Dilek Hakkani-Tur et al.

We propose a novel zero-shot learning method for semantic utterance classification (SUC). It learns a classifier $f: X \to Y$ for problems where none of the semantic categories $Y$ are present in the training set. The framework uncovers the link between categories and utterances using a semantic space. We show that this semantic space can be learned by deep neural networks trained on large amounts of search engine query log data. More precisely, we propose a novel method that can learn discriminative semantic features without supervision. It uses the zero-shot learning framework to guide the learning of the semantic features. We demonstrate the effectiveness of the zero-shot semantic learning algorithm on the SUC dataset collected by (Tur, 2012). Furthermore, we achieve state-of-the-art results by combining the semantic features with a supervised method.