Entropy Guided Extrapolative Decoding to Improve Factuality in Large Language Models

Large language models (LLMs) exhibit impressive natural language capabilities but suffer from hallucination -- generating content ungrounded in the realities of training data. Recent work has focused on decoding techniques to improve factuality during inference by leveraging LLMs' hierarchical representation of factual knowledge, manipulating the predicted distributions at inference time. Current state-of-the-art approaches refine decoding by contrasting early-exit distributions from a lower layer with the final layer to exploit information related to factuality within the model forward procedure. However, such methods often assume the final layer is the most reliable and the lower layer selection process depends on it. In this work, we first propose extrapolation of critical token probabilities beyond the last layer for more accurate contrasting. We additionally employ layer-wise entropy-guided lower layer selection, decoupling the selection process from the final layer. Experiments demonstrate strong performance - surpassing state-of-the-art on multiple different datasets by large margins. Analyses show different kinds of prompts respond to different selection strategies.

Improving Dialog Safety using Socially Aware Contrastive Learning

State-of-the-art conversational AI systems raise concerns due to their potential risks of generating unsafe, toxic, unethical, or dangerous content. Previous works have developed datasets to teach conversational agents the appropriate social paradigms to respond effectively to specifically designed hazardous content. However, models trained on these adversarial datasets still struggle to recognize subtle unsafe situations that appear naturally in conversations or introduce an inappropriate response in a casual context. To understand the extent of this problem, we study prosociality in both adversarial and casual dialog contexts and audit the response quality of general-purpose language models in terms of propensity to produce unsafe content. We propose a dual-step fine-tuning process to address these issues using a socially aware n-pair contrastive loss. Subsequently, we train a base model that integrates prosocial behavior by leveraging datasets like Moral Integrity Corpus (MIC) and ProsocialDialog. Experimental results on several dialog datasets demonstrate the effectiveness of our approach in generating socially appropriate responses.

Diving Deep into Modes of Fact Hallucinations in Dialogue Systems

Knowledge Graph(KG) grounded conversations often use large pre-trained models and usually suffer from fact hallucination. Frequently entities with no references in knowledge sources and conversation history are introduced into responses, thus hindering the flow of the conversation -- existing work attempt to overcome this issue by tweaking the training procedure or using a multi-step refining method. However, minimal effort is put into constructing an entity-level hallucination detection system, which would provide fine-grained signals that control fallacious content while generating responses. As a first step to address this issue, we dive deep to identify various modes of hallucination in KG-grounded chatbots through human feedback analysis. Secondly, we propose a series of perturbation strategies to create a synthetic dataset named FADE (FActual Dialogue Hallucination DEtection Dataset). Finally, we conduct comprehensive data analyses and create multiple baseline models for hallucination detection to compare against human-verified data and already established benchmarks.

Using Multi-Encoder Fusion Strategies to Improve Personalized Response Selection

Personalized response selection systems are generally grounded on persona. However, there exists a co-relation between persona and empathy, which is not explored well in these systems. Also, faithfulness to the conversation context plunges when a contradictory or an off-topic response is selected. This paper attempts to address these issues by proposing a suite of fusion strategies that capture the interaction between persona, emotion, and entailment information of the utterances. Ablation studies on the Persona-Chat dataset show that incorporating emotion and entailment improves the accuracy of response selection. We combine our fusion strategies and concept-flow encoding to train a BERT-based model which outperforms the previous methods by margins larger than 2.3 % on original personas and 1.9 % on revised personas in terms of hits@1 (top-1 accuracy), achieving a new state-of-the-art performance on the Persona-Chat dataset.

Proto: A Neural Cocktail for Generating Appealing Conversations

In this paper, we present our Alexa Prize Grand Challenge 4 socialbot: Proto. Leveraging diverse sources of world knowledge, and powered by a suite of neural and rule-based natural language understanding modules, state-of-the-art neural generators, novel state-based deterministic generators, an ensemble of neural re-rankers, a robust post-processing algorithm, and an efficient overall conversation strategy, Proto strives to be able to converse coherently about a diverse range of topics of interest to humans, and provide a memorable experience to the user. In this paper we dissect and analyze the different components and conversation strategies implemented by our socialbot, which enables us to generate colloquial, empathetic, engaging, self-rectifying, factually correct, and on-topic response, which has helped us achieve consistent scores throughout the competition.

Similarity Based Label Smoothing For Dialogue Generation

Generative neural conversational systems are generally trained with the objective of minimizing the entropy loss between the training "hard" targets and the predicted logits. Often, performance gains and improved generalization can be achieved by using regularization techniques like label smoothing, which converts the training "hard" targets to "soft" targets. However, label smoothing enforces a data independent uniform distribution on the incorrect training targets, which leads to an incorrect assumption of equi-probable incorrect targets for each correct target. In this paper we propose and experiment with incorporating data dependent word similarity based weighing methods to transforms the uniform distribution of the incorrect target probabilities in label smoothing, to a more natural distribution based on semantics. We introduce hyperparameters to control the incorrect target distribution, and report significant performance gains over networks trained using standard label smoothing based loss, on two standard open domain dialogue corpora.

Medical Literature Mining and Retrieval in a Conversational Setting

The Covid-19 pandemic has caused a spur in the medical research literature. With new research advances in understanding the virus, there is a need for robust text mining tools which can process, extract and present answers from the literature in a concise and consumable way. With a DialoGPT based multi-turn conversation generation module, and BM-25 \& neural embeddings based ensemble information retrieval module, in this paper we present a conversational system, which can retrieve and answer coronavirus-related queries from the rich medical literature, and present it in a conversational setting with the user. We further perform experiments to compare neural embedding-based document retrieval and the traditional BM25 retrieval algorithm and report the results.

A multi-agent evolutionary robotics framework to train spiking neural networks

A novel multi-agent evolutionary robotics (ER) based framework, inspired by competitive evolutionary environments in nature, is demonstrated for training Spiking Neural Networks (SNN). The weights of a population of SNNs along with morphological parameters of bots they control in the ER environment are treated as phenotypes. Rules of the framework select certain bots and their SNNs for reproduction and others for elimination based on their efficacy in capturing food in a competitive environment. While the bots and their SNNs are given no explicit reward to survive or reproduce via any loss function, these drives emerge implicitly as they evolve to hunt food and survive within these rules. Their efficiency in capturing food as a function of generations exhibit the evolutionary signature of punctuated equilibria. Two evolutionary inheritance algorithms on the phenotypes, Mutation and Crossover with Mutation, are demonstrated. Performances of these algorithms are compared using ensembles of 100 experiments for each algorithm. We find that Crossover with Mutation promotes 40% faster learning in the SNN than mere Mutation with a statistically significant margin.