Data-Efficient Alignment of Large Language Models with Human Feedback Through Natural Language

Learning from human feedback is a prominent technique to align the output of large language models (LLMs) with human expectations. Reinforcement learning from human feedback (RLHF) leverages human preference signals that are in the form of ranking of response pairs to perform this alignment. However, human preference on LLM outputs can come in much richer forms including natural language, which may provide detailed feedback on strengths and weaknesses of a given response. In this work we investigate data efficiency of modeling human feedback that is in natural language. Specifically, we fine-tune an open-source LLM, e.g., Falcon-40B-Instruct, on a relatively small amount (1000 records or even less) of human feedback in natural language in the form of critiques and revisions of responses. We show that this model is able to improve the quality of responses from even some of the strongest LLMs such as ChatGPT, BARD, and Vicuna, through critique and revision of those responses. For instance, through one iteration of revision of ChatGPT responses, the revised responses have 56.6% win rate over the original ones, and this win rate can be further improved to 65.9% after applying the revision for five iterations.

Data Augmentation for Improving Tail-traffic Robustness in Skill-routing for Dialogue Systems

Large-scale conversational systems typically rely on a skill-routing component to route a user request to an appropriate skill and interpretation to serve the request. In such system, the agent is responsible for serving thousands of skills and interpretations which create a long-tail distribution due to the natural frequency of requests. For example, the samples related to play music might be a thousand times more frequent than those asking for theatre show times. Moreover, inputs used for ML-based skill routing are often a heterogeneous mix of strings, embedding vectors, categorical and scalar features which makes employing augmentation-based long-tail learning approaches challenging. To improve the skill-routing robustness, we propose an augmentation of heterogeneous skill-routing data and training targeted for robust operation in long-tail data regimes. We explore a variety of conditional encoder-decoder generative frameworks to perturb original data fields and create synthetic training data. To demonstrate the effectiveness of the proposed method, we conduct extensive experiments using real-world data from a commercial conversational system. Based on the experiment results, the proposed approach improves more than 80% (51 out of 63) of intents with less than 10K of traffic instances in the skill-routing replication task.

Open World Classification with Adaptive Negative Samples

Open world classification is a task in natural language processing with key practical relevance and impact. Since the open or {\em unknown} category data only manifests in the inference phase, finding a model with a suitable decision boundary accommodating for the identification of known classes and discrimination of the open category is challenging. The performance of existing models is limited by the lack of effective open category data during the training stage or the lack of a good mechanism to learn appropriate decision boundaries. We propose an approach based on \underline{a}daptive \underline{n}egative \underline{s}amples (ANS) designed to generate effective synthetic open category samples in the training stage and without requiring any prior knowledge or external datasets. Empirically, we find a significant advantage in using auxiliary one-versus-rest binary classifiers, which effectively utilize the generated negative samples and avoid the complex threshold-seeking stage in previous works. Extensive experiments on three benchmark datasets show that ANS achieves significant improvements over state-of-the-art methods.

Cluster-Guided Label Generation in Extreme Multi-Label Classification

For extreme multi-label classification (XMC), existing classification-based models poorly perform for tail labels and often ignore the semantic relations among labels, like treating "Wikipedia" and "Wiki" as independent and separate labels. In this paper, we cast XMC as a generation task (XLGen), where we benefit from pre-trained text-to-text models. However, generating labels from the extremely large label space is challenging without any constraints or guidance. We, therefore, propose to guide label generation using label cluster information to hierarchically generate lower-level labels. We also find that frequency-based label ordering and using decoding ensemble methods are critical factors for the improvements in XLGen. XLGen with cluster guidance significantly outperforms the classification and generation baselines on tail labels, and also generally improves the overall performance in four popular XMC benchmarks. In human evaluation, we also find XLGen generates unseen but plausible labels. Our code is now available at https://github.com/alexa/xlgen-eacl-2023.

Selective In-Context Data Augmentation for Intent Detection using Pointwise V-Information

This work focuses on in-context data augmentation for intent detection. Having found that augmentation via in-context prompting of large pre-trained language models (PLMs) alone does not improve performance, we introduce a novel approach based on PLMs and pointwise V-information (PVI), a metric that can measure the usefulness of a datapoint for training a model. Our method first fine-tunes a PLM on a small seed of training data and then synthesizes new datapoints - utterances that correspond to given intents. It then employs intent-aware filtering, based on PVI, to remove datapoints that are not helpful to the downstream intent classifier. Our method is thus able to leverage the expressive power of large language models to produce diverse training data. Empirical results demonstrate that our method can produce synthetic training data that achieve state-of-the-art performance on three challenging intent detection datasets under few-shot settings (1.28% absolute improvement in 5-shot and 1.18% absolute in 10-shot, on average) and perform on par with the state-of-the-art in full-shot settings (within 0.01% absolute, on average).

Constrained Policy Optimization for Controlled Self-Learning in Conversational AI Systems

Recently, self-learning methods based on user satisfaction metrics and contextual bandits have shown promising results to enable consistent improvements in conversational AI systems. However, directly targeting such metrics by off-policy bandit learning objectives often increases the risk of making abrupt policy changes that break the current user experience. In this study, we introduce a scalable framework for supporting fine-grained exploration targets for individual domains via user-defined constraints. For example, we may want to ensure fewer policy deviations in business-critical domains such as shopping, while allocating more exploration budget to domains such as music. Furthermore, we present a novel meta-gradient learning approach that is scalable and practical to address this problem. The proposed method adjusts constraint violation penalty terms adaptively through a meta objective that encourages balanced constraint satisfaction across domains. We conduct extensive experiments using data from a real-world conversational AI on a set of realistic constraint benchmarks. Based on the experimental results, we demonstrate that the proposed approach is capable of achieving the best balance between the policy value and constraint satisfaction rate.

Debiasing Neighbor Aggregation for Graph Neural Network in Recommender Systems

Graph neural networks (GNNs) have achieved remarkable success in recommender systems by representing users and items based on their historical interactions. However, little attention was paid to GNN's vulnerability to exposure bias: users are exposed to a limited number of items so that a system only learns a biased view of user preference to result in suboptimal recommendation quality. Although inverse propensity weighting is known to recognize and alleviate exposure bias, it usually works on the final objective with the model outputs, whereas GNN can also be biased during neighbor aggregation. In this paper, we propose a simple but effective approach, neighbor aggregation via inverse propensity (Navip) for GNNs. Specifically, given a user-item bipartite graph, we first derive propensity score of each user-item interaction in the graph. Then, inverse of the propensity score with Laplacian normalization is applied to debias neighbor aggregation from exposure bias. We validate the effectiveness of our approach through our extensive experiments on two public and Amazon Alexa datasets where the performance enhances up to 14.2%.

Learning Personalized Representations using Graph Convolutional Network

Generating representations that precisely reflect customers' behavior is an important task for providing personalized skill routing experience in Alexa. Currently, Dynamic Routing (DR) team, which is responsible for routing Alexa traffic to providers or skills, relies on two features to be served as personal signals: absolute traffic count and normalized traffic count of every skill usage per customer. Neither of them considers the network based structure for interactions between customers and skills, which contain richer information for customer preferences. In this work, we first build a heterogeneous edge attributed graph based customers' past interactions with the invoked skills, in which the user requests (utterances) are modeled as edges. Then we propose a graph convolutional network(GCN) based model, namely Personalized Dynamic Routing Feature Encoder(PDRFE), that generates personalized customer representations learned from the built graph. Compared with existing models, PDRFE is able to further capture contextual information in the graph convolutional function. The performance of our proposed model is evaluated by a downstream task, defect prediction, that predicts the defect label from the learned embeddings of customers and their triggered skills. We observe up to 41% improvements on the cross entropy metric for our proposed models compared to the baselines.

Scalable and Robust Self-Learning for Skill Routing in Large-Scale Conversational AI Systems

Skill routing is an important component in large-scale conversational systems. In contrast to traditional rule-based skill routing, state-of-the-art systems use a model-based approach to enable natural conversations. To provide supervision signal required to train such models, ideas such as human annotation, replication of a rule-based system, relabeling based on user paraphrases, and bandit-based learning were suggested. However, these approaches: (a) do not scale in terms of the number of skills and skill on-boarding, (b) require a very costly expert annotation/rule-design, (c) introduce risks in the user experience with each model update. In this paper, we present a scalable self-learning approach to explore routing alternatives without causing abrupt policy changes that break the user experience, learn from the user interaction, and incrementally improve the routing via frequent model refreshes. To enable such robust frequent model updates, we suggest a simple and effective approach that ensures controlled policy updates for individual domains, followed by an off-policy evaluation for making deployment decisions without any need for lengthy A/B experimentation. We conduct various offline and online A/B experiments on a commercial large-scale conversational system to demonstrate the effectiveness of the proposed method in real-world production settings.

Domain-Aware Contrastive Knowledge Transfer for Multi-domain Imbalanced Data

In many real-world machine learning applications, samples belong to a set of domains e.g., for product reviews each review belongs to a product category. In this paper, we study multi-domain imbalanced learning (MIL), the scenario that there is imbalance not only in classes but also in domains. In the MIL setting, different domains exhibit different patterns and there is a varying degree of similarity and divergence among domains posing opportunities and challenges for transfer learning especially when faced with limited or insufficient training data. We propose a novel domain-aware contrastive knowledge transfer method called DCMI to (1) identify the shared domain knowledge to encourage positive transfer among similar domains (in particular from head domains to tail domains); (2) isolate the domain-specific knowledge to minimize the negative transfer from dissimilar domains. We evaluated the performance of DCMI on three different datasets showing significant improvements in different MIL scenarios.