In this paper, we study the generation quality of interpolation-based retrieval-augmented language models (LMs). These methods, best exemplified by the KNN-LM, interpolate the LM's predicted distribution of the next word with a distribution formed from the most relevant retrievals for a given prefix. While the KNN-LM and related methods yield impressive decreases in perplexity, we discover that they do not exhibit corresponding improvements in open-ended generation quality, as measured by both automatic evaluation metrics (e.g., MAUVE) and human evaluations. Digging deeper, we find that interpolating with a retrieval distribution actually increases perplexity compared to a baseline Transformer LM for the majority of tokens in the WikiText-103 test set, even though the overall perplexity is lower due to a smaller number of tokens for which perplexity dramatically decreases after interpolation. However, when decoding a long sequence at inference time, significant improvements on this smaller subset of tokens are washed out by slightly worse predictions on most tokens. Furthermore, we discover that the entropy of the retrieval distribution increases faster than that of the base LM as the generated sequence becomes longer, which indicates that retrieval is less reliable when using model-generated text as queries (i.e., is subject to exposure bias). We hope that our analysis spurs future work on improved decoding algorithms and interpolation strategies for retrieval-augmented language models.

Suppose we want to train text prediction models in email clients or word processors. The models must preserve the privacy of user data and adhere to a specific fixed size to meet memory and inference time requirements. We introduce a generic framework to solve this problem. Specifically, we are given a public dataset $D_\text{pub}$ and a private dataset $D_\text{priv}$ corresponding to a downstream task $T$. How should we pre-train a fixed-size model $M$ on $D_\text{pub}$ and fine-tune it on $D_\text{priv}$ such that performance of $M$ with respect to $T$ is maximized and $M$ satisfies differential privacy with respect to $D_\text{priv}$? We show that pre-training on a {\em subset} of dataset $D_\text{pub}$ that brings the public distribution closer to the private distribution is a crucial ingredient to maximize the transfer learning abilities of $M$ after pre-training, especially in the regimes where model sizes are relatively small. Besides performance improvements, our framework also shows that with careful pre-training and private fine-tuning, {\em smaller models} can match the performance of much larger models, highlighting the promise of differentially private training as a tool for model compression and efficiency.

Multi-vector retrieval models such as ColBERT [Khattab and Zaharia, 2020] allow token-level interactions between queries and documents, and hence achieve state of the art on many information retrieval benchmarks. However, their non-linear scoring function cannot be scaled to millions of documents, necessitating a three-stage process for inference: retrieving initial candidates via token retrieval, accessing all token vectors, and scoring the initial candidate documents. The non-linear scoring function is applied over all token vectors of each candidate document, making the inference process complicated and slow. In this paper, we aim to simplify the multi-vector retrieval by rethinking the role of token retrieval. We present XTR, ConteXtualized Token Retriever, which introduces a simple, yet novel, objective function that encourages the model to retrieve the most important document tokens first. The improvement to token retrieval allows XTR to rank candidates only using the retrieved tokens rather than all tokens in the document, and enables a newly designed scoring stage that is two-to-three orders of magnitude cheaper than that of ColBERT. On the popular BEIR benchmark, XTR advances the state-of-the-art by 2.8 nDCG@10 without any distillation. Detailed analysis confirms our decision to revisit the token retrieval stage, as XTR demonstrates much better recall of the token retrieval stage compared to ColBERT.

In this paper, we discuss the mechanics and planning algorithms to slide an object on a horizontal planar surface via frictional patch contact made with its top surface. Here, we propose an asymmetric dual limit surface model to determine slip boundary conditions for both the top and bottom contact. With this model, we obtain a range of twists that can keep the object in sticking contact with the robot end-effector while slipping on the supporting plane. Based on these constraints, we derive a planning algorithm to slide objects with only top contact to arbitrary goal poses without slippage between end effector and the object. We validate the proposed model empirically and demonstrate its predictive accuracy on a variety of object geometries and motions. We also evaluate the planning algorithm over a variety of objects and goals demonstrate an orientation error improvement of 90\% when compared to methods naive to linear path planners.

Counterspeech has been demonstrated to be an efficacious approach for combating hate speech. While various conventional and controlled approaches have been studied in recent years to generate counterspeech, a counterspeech with a certain intent may not be sufficient in every scenario. Due to the complex and multifaceted nature of hate speech, utilizing multiple forms of counter-narratives with varying intents may be advantageous in different circumstances. In this paper, we explore intent-conditioned counterspeech generation. At first, we develop IntentCONAN, a diversified intent-specific counterspeech dataset with 6831 counterspeeches conditioned on five intents, i.e., informative, denouncing, question, positive, and humour. Subsequently, we propose QUARC, a two-stage framework for intent-conditioned counterspeech generation. QUARC leverages vector-quantized representations learned for each intent category along with PerFuMe, a novel fusion module to incorporate intent-specific information into the model. Our evaluation demonstrates that QUARC outperforms several baselines by an average of 10% across evaluation metrics. An extensive human evaluation supplements our hypothesis of better and more appropriate responses than comparative systems.

We propose ChatGPT-EDSS, an empathetic dialogue speech synthesis (EDSS) method using ChatGPT for extracting dialogue context. ChatGPT is a chatbot that can deeply understand the content and purpose of an input prompt and appropriately respond to the user's request. We focus on ChatGPT's reading comprehension and introduce it to EDSS, a task of synthesizing speech that can empathize with the interlocutor's emotion. Our method first gives chat history to ChatGPT and asks it to generate three words representing the intention, emotion, and speaking style for each line in the chat. Then, it trains an EDSS model using the embeddings of ChatGPT-derived context words as the conditioning features. The experimental results demonstrate that our method performs comparably to ones using emotion labels or neural network-derived context embeddings learned from chat histories. The collected ChatGPT-derived context information is available at https://sarulab-speech.github.io/demo_ChatGPT_EDSS/.

Large language models have shown impressive performance across a wide variety of tasks, including text summarization. In this paper, we show that this strong performance extends to opinion summarization. We explore several pipeline methods for applying GPT-3.5 to summarize a large collection of user reviews in a prompted fashion. To handle arbitrarily large numbers of user reviews, we explore recursive summarization as well as methods for selecting salient content to summarize through supervised clustering or extraction. On two datasets, an aspect-oriented summarization dataset of hotel reviews (SPACE) and a generic summarization dataset of Amazon and Yelp reviews (FewSum), we show that GPT-3.5 models achieve very strong performance in human evaluation. We argue that standard evaluation metrics do not reflect this, and introduce three new metrics targeting faithfulness, factuality, and genericity to contrast these different methods.

Consider two $D$-dimensional data vectors (e.g., embeddings): $u, v$. In many embedding-based retrieval (EBR) applications where the vectors are generated from trained models, $D=256\sim 1024$ are common. In this paper, OPORP (one permutation + one random projection) uses a variant of the ``count-sketch'' type of data structures for achieving data reduction/compression. With OPORP, we first apply a permutation on the data vectors. A random vector $r$ is generated i.i.d. with moments: $E(r_i) = 0, E(r_i^2)=1, E(r_i^3) =0, E(r_i^4)=s$. We multiply (as dot product) $r$ with all permuted data vectors. Then we break the $D$ columns into $k$ equal-length bins and aggregate (i.e., sum) the values in each bin to obtain $k$ samples from each data vector. One crucial step is to normalize the $k$ samples to the unit $l_2$ norm. We show that the estimation variance is essentially: $(s-1)A + \frac{D-k}{D-1}\frac{1}{k}\left[ (1-\rho^2)^2 -2A\right]$, where $A\geq 0$ is a function of the data ($u,v$). This formula reveals several key properties: (1) We need $s=1$. (2) The factor $\frac{D-k}{D-1}$ can be highly beneficial in reducing variances. (3) The term $\frac{1}{k}(1-\rho^2)^2$ is a substantial improvement compared with $\frac{1}{k}(1+\rho^2)$, which corresponds to the un-normalized estimator. We illustrate that by letting the $k$ in OPORP to be $k=1$ and repeat the procedure $m$ times, we exactly recover the work of ``very spars random projections'' (VSRP). This immediately leads to a normalized estimator for VSRP which substantially improves the original estimator of VSRP. In summary, with OPORP, the two key steps: (i) the normalization and (ii) the fixed-length binning scheme, have considerably improved the accuracy in estimating the cosine similarity, which is a routine (and crucial) task in modern embedding-based retrieval (EBR) applications.

We present QLoRA, an efficient finetuning approach that reduces memory usage enough to finetune a 65B parameter model on a single 48GB GPU while preserving full 16-bit finetuning task performance. QLoRA backpropagates gradients through a frozen, 4-bit quantized pretrained language model into Low Rank Adapters~(LoRA). Our best model family, which we name Guanaco, outperforms all previous openly released models on the Vicuna benchmark, reaching 99.3% of the performance level of ChatGPT while only requiring 24 hours of finetuning on a single GPU. QLoRA introduces a number of innovations to save memory without sacrificing performance: (a) 4-bit NormalFloat (NF4), a new data type that is information theoretically optimal for normally distributed weights (b) double quantization to reduce the average memory footprint by quantizing the quantization constants, and (c) paged optimziers to manage memory spikes. We use QLoRA to finetune more than 1,000 models, providing a detailed analysis of instruction following and chatbot performance across 8 instruction datasets, multiple model types (LLaMA, T5), and model scales that would be infeasible to run with regular finetuning (e.g. 33B and 65B parameter models). Our results show that QLoRA finetuning on a small high-quality dataset leads to state-of-the-art results, even when using smaller models than the previous SoTA. We provide a detailed analysis of chatbot performance based on both human and GPT-4 evaluations showing that GPT-4 evaluations are a cheap and reasonable alternative to human evaluation. Furthermore, we find that current chatbot benchmarks are not trustworthy to accurately evaluate the performance levels of chatbots. A lemon-picked analysis demonstrates where Guanaco fails compared to ChatGPT. We release all of our models and code, including CUDA kernels for 4-bit training.

Learning high quality sentence embeddings from dialogues has drawn increasing attentions as it is essential to solve a variety of dialogue-oriented tasks with low annotation cost. However, directly annotating and gathering utterance relationships in conversations are difficult, while token-level annotations, \eg, entities, slots and templates, are much easier to obtain. General sentence embedding methods are usually sentence-level self-supervised frameworks and cannot utilize token-level extra knowledge. In this paper, we introduce Template-aware Dialogue Sentence Embedding (TaDSE), a novel augmentation method that utilizes template information to effectively learn utterance representation via self-supervised contrastive learning framework. TaDSE augments each sentence with its corresponding template and then conducts pairwise contrastive learning over both sentence and template. We further enhance the effect with a synthetically augmented dataset that enhances utterance-template relation, in which entity detection (slot-filling) is a preliminary step. We evaluate TaDSE performance on five downstream benchmark datasets. The experiment results show that TaDSE achieves significant improvements over previous SOTA methods, along with a consistent Intent Classification task performance improvement margin. We further introduce a novel analytic instrument of Semantic Compression method, for which we discover a correlation with uniformity and alignment. Our code will be released soon.