Kevin Small

Yilun Hua, Giuseppe Castellucci, Peter Schulam et al.

Retrieval Augmented Generation (RAG)'s success depends on the utility the LLM derives from the content used for grounding. Quantifying content utility does not have a definitive specification and existing metrics ignore model-specific capabilities and/or rely on costly annotations. In this paper, we propose Grounding Generation Utility (GroGU), a model-specific and reference-free metric that defines utility as a function of the downstream LLM's generation confidence based on entropy. Despite having no annotation requirements, GroGU is largely faithful in distinguishing ground-truth documents while capturing nuances ignored by LLM-agnostic metrics. We apply GroGU to train a query-rewriter for RAG by identifying high-utility preference data for Direct Preference Optimization. Experiments show improvements by up to 18.2 points in Mean Reciprocal Rank and up to 9.4 points in answer accuracy.

Ashutosh Baheti, Alan Ritter, Kevin Small

Question answering (QA) is an important aspect of open-domain conversational agents, garnering specific research focus in the conversational QA (ConvQA) subtask. One notable limitation of recent ConvQA efforts is the response being answer span extraction from the target corpus, thus ignoring the natural language generation (NLG) aspect of high-quality conversational agents. In this work, we propose a method for situating QA responses within a SEQ2SEQ NLG approach to generate fluent grammatical answer responses while maintaining correctness. From a technical perspective, we use data augmentation to generate training data for an end-to-end system. Specifically, we develop Syntactic Transformations (STs) to produce question-specific candidate answer responses and rank them using a BERT-based classifier (Devlin et al., 2019). Human evaluation on SQuAD 2.0 data (Rajpurkar et al., 2018) demonstrate that the proposed model outperforms baseline CoQA and QuAC models in generating conversational responses. We further show our model's scalability by conducting tests on the CoQA dataset. The code and data are available at https://github.com/abaheti95/QADialogSystem.

Li Zhou, Kevin Small

Multi-domain dialogue state tracking (DST) is a critical component for conversational AI systems. The domain ontology (i.e., specification of domains, slots, and values) of a conversational AI system is generally incomplete, making the capability for DST models to generalize to new slots, values, and domains during inference imperative. In this paper, we propose to model multi-domain DST as a question answering problem, referred to as Dialogue State Tracking via Question Answering (DSTQA). Within DSTQA, each turn generates a question asking for the value of a (domain, slot) pair, thus making it naturally extensible to unseen domains, slots, and values. Additionally, we use a dynamically-evolving knowledge graph to explicitly learn relationships between (domain, slot) pairs. Our model has a 5.80% and 12.21% relative improvement over the current state-of-the-art model on MultiWOZ 2.0 and MultiWOZ 2.1 datasets, respectively. Additionally, our model consistently outperforms the state-of-the-art model in domain adaptation settings. (Code is released at https://github.com/alexa/dstqa )

Li Zhou, Kevin Small, Yong Zhang et al.

Motivated by suggested question generation in conversational news recommendation systems, we propose a model for generating question-answer pairs (QA pairs) with self-contained, summary-centric questions and length-constrained, article-summarizing answers. We begin by collecting a new dataset of news articles with questions as titles and pairing them with summaries of varying length. This dataset is used to learn a QA pair generation model producing summaries as answers that balance brevity with sufficiency jointly with their corresponding questions. We then reinforce the QA pair generation process with a differentiable reward function to mitigate exposure bias, a common problem in natural language generation. Both automatic metrics and human evaluation demonstrate these QA pairs successfully capture the central gists of the articles and achieve high answer accuracy.

Xusen Yin, Li Zhou, Kevin Small et al.

Users frequently ask simple factoid questions for question answering (QA) systems, attenuating the impact of myriad recent works that support more complex questions. Prompting users with automatically generated suggested questions (SQs) can improve user understanding of QA system capabilities and thus facilitate more effective use. We aim to produce self-explanatory questions that focus on main document topics and are answerable with variable length passages as appropriate. We satisfy these requirements by using a BERT-based Pointer-Generator Network trained on the Natural Questions (NQ) dataset. Our model shows SOTA performance of SQ generation on the NQ dataset (20.1 BLEU-4). We further apply our model on out-of-domain news articles, evaluating with a QA system due to the lack of gold questions and demonstrate that our model produces better SQs for news articles -- with further confirmation via a human evaluation.

Li Zhou, Kevin Small

Humans generally use natural language to communicate task requirements to each other. Ideally, natural language should also be usable for communicating goals to autonomous machines (e.g., robots) to minimize friction in task specification. However, understanding and mapping natural language goals to sequences of states and actions is challenging. Specifically, existing work along these lines has encountered difficulty in generalizing learned policies to new natural language goals and environments. In this paper, we propose a novel adversarial inverse reinforcement learning algorithm to learn a language-conditioned policy and reward function. To improve generalization of the learned policy and reward function, we use a variational goal generator to relabel trajectories and sample diverse goals during training. Our algorithm outperforms multiple baselines by a large margin on a vision-based natural language instruction following dataset (Room-2-Room), demonstrating a promising advance in enabling the use of natural language instructions in specifying agent goals.

Yuheng Bu, Kevin Small

While recommendation systems generally observe user behavior passively, there has been an increased interest in directly querying users to learn their specific preferences. In such settings, considering queries at different levels of granularity to optimize user information acquisition is crucial to efficiently providing a good user experience. In this work, we study the active learning problem with multi-level user preferences within the collective matrix factorization (CMF) framework. CMF jointly captures multi-level user preferences with respect to items and relations between items (e.g., book genre, cuisine type), generally resulting in improved predictions. Motivated by finite-sample analysis of the CMF model, we propose a theoretically optimal active learning strategy based on the Fisher information matrix and use this to derive a realizable approximation algorithm for practical recommendations. Experiments are conducted using both the Yelp dataset directly and an illustrative synthetic dataset in the three settings of personalized active learning, cold-start recommendations, and noisy data -- demonstrating strong improvements over several widely used active learning methods.

Li Zhou, Kevin Small, Oleg Rokhlenko et al.

Learning a goal-oriented dialog policy is generally performed offline with supervised learning algorithms or online with reinforcement learning (RL). Additionally, as companies accumulate massive quantities of dialog transcripts between customers and trained human agents, encoder-decoder methods have gained popularity as agent utterances can be directly treated as supervision without the need for utterance-level annotations. However, one potential drawback of such approaches is that they myopically generate the next agent utterance without regard for dialog-level considerations. To resolve this concern, this paper describes an offline RL method for learning from unannotated corpora that can optimize a goal-oriented policy at both the utterance and dialog level. We introduce a novel reward function and use both on-policy and off-policy policy gradient to learn a policy offline without requiring online user interaction or an explicit state space definition.