Igor Shalyminov

Hwanjun Song, Hang Su, Igor Shalyminov et al. · amazon-science

Automated evaluation is crucial for streamlining text summarization benchmarking and model development, given the costly and time-consuming nature of human evaluation. Traditional methods like ROUGE do not correlate well with human judgment, while recently proposed LLM-based metrics provide only summary-level assessment using Likert-scale scores. This limits deeper model analysis, e.g., we can only assign one hallucination score at the summary level, while at the sentence level, we can count sentences containing hallucinations. To remedy those limitations, we propose FineSurE, a fine-grained evaluator specifically tailored for the summarization task using large language models (LLMs). It also employs completeness and conciseness criteria, in addition to faithfulness, enabling multi-dimensional assessment. We compare various open-source and proprietary LLMs as backbones for FineSurE. In addition, we conduct extensive benchmarking of FineSurE against SOTA methods including NLI-, QA-, and LLM-based methods, showing improved performance especially on the completeness and conciseness dimensions. The code is available at https://github.com/DISL-Lab/FineSurE-ACL24.

Shunqi Mao, Chaoyi Zhang, Hang Su et al.

Contextualized Image Captioning (CIC) evolves traditional image captioning into a more complex domain, necessitating the ability for multimodal reasoning. It aims to generate image captions given specific contextual information. This paper further introduces a novel domain of Controllable Contextualized Image Captioning (Ctrl-CIC). Unlike CIC, which solely relies on broad context, Ctrl-CIC accentuates a user-defined highlight, compelling the model to tailor captions that resonate with the highlighted aspects of the context. We present two approaches, Prompting-based Controller (P-Ctrl) and Recalibration-based Controller (R-Ctrl), to generate focused captions. P-Ctrl conditions the model generation on highlight by prepending captions with highlight-driven prefixes, whereas R-Ctrl tunes the model to selectively recalibrate the encoder embeddings for highlighted tokens. Additionally, we design a GPT-4V empowered evaluator to assess the quality of the controlled captions alongside standard assessment methods. Extensive experimental results demonstrate the efficient and effective controllability of our method, charting a new direction in achieving user-adaptive image captioning. Code is available at https://github.com/ShunqiM/Ctrl-CIC .

Hwanjun Song, Igor Shalyminov, Hang Su et al.

Sequence-level knowledge distillation reduces the size of Seq2Seq models for more efficient abstractive summarization. However, it often leads to a loss of abstractiveness in summarization. In this paper, we propose a novel approach named DisCal to enhance the level of abstractiveness (measured by n-gram overlap) without sacrificing the informativeness (measured by ROUGE) of generated summaries. DisCal exposes diverse pseudo summaries with two supervision to the student model. Firstly, the best pseudo summary is identified in terms of abstractiveness and informativeness and used for sequence-level distillation. Secondly, their ranks are used to ensure the student model to assign higher prediction scores to summaries with higher ranks. Our experiments show that DisCal outperforms prior methods in abstractive summarization distillation, producing highly abstractive and informative summaries.

Jianfeng He, Hang Su, Jason Cai et al. · amazon-science

Semi-supervised dialogue summarization (SSDS) leverages model-generated summaries to reduce reliance on human-labeled data and improve the performance of summarization models. While addressing label noise, previous works on semi-supervised learning primarily focus on natural language understanding tasks, assuming each sample has a unique label. However, these methods are not directly applicable to SSDS, as it is a generative task, and each dialogue can be summarized in different ways. In this work, we propose a novel scoring approach, SiCF, which encapsulates three primary dimensions of summarization model quality: Semantic invariance (indicative of model confidence), Coverage (factual recall), and Faithfulness (factual precision). Using the SiCF score, we select unlabeled dialogues with high-quality generated summaries to train summarization models. Comprehensive experiments on three public datasets demonstrate the effectiveness of SiCF scores in uncertainty estimation and semi-supervised learning for dialogue summarization tasks. Our code is available at \url{https://github.com/amazon-science/summarization-sicf-score}.

Liyan Tang, Igor Shalyminov, Amy Wing-mei Wong et al.

Single document news summarization has seen substantial progress on faithfulness in recent years, driven by research on the evaluation of factual consistency, or hallucinations. We ask whether these advances carry over to other text summarization domains. We propose a new evaluation benchmark on topic-focused dialogue summarization, generated by LLMs of varying sizes. We provide binary sentence-level human annotations of the factual consistency of these summaries along with detailed explanations of factually inconsistent sentences. Our analysis shows that existing LLMs hallucinate significant amounts of factual errors in the dialogue domain, regardless of the model's size. On the other hand, when LLMs, including GPT-4, serve as binary factual evaluators, they perform poorly and can be outperformed by prevailing state-of-the-art specialized factuality evaluation metrics. Finally, we conducted an analysis of hallucination types with a curated error taxonomy. We find that there are diverse errors and error distributions in model-generated summaries and that non-LLM based metrics can capture all error types better than LLM-based evaluators.

Hossein Aboutalebi, Hwanjun Song, Yusheng Xie et al.

Development of multimodal interactive systems is hindered by the lack of rich, multimodal (text, images) conversational data, which is needed in large quantities for LLMs. Previous approaches augment textual dialogues with retrieved images, posing privacy, diversity, and quality constraints. In this work, we introduce Multimodal Augmented Generative Images Dialogues (MAGID), a framework to augment text-only dialogues with diverse and high-quality images. Subsequently, a diffusion model is applied to craft corresponding images, ensuring alignment with the identified text. Finally, MAGID incorporates an innovative feedback loop between an image description generation module (textual LLM) and image quality modules (addressing aesthetics, image-text matching, and safety), that work in tandem to generate high-quality and multi-modal dialogues. We compare MAGID to other SOTA baselines on three dialogue datasets, using automated and human evaluation. Our results show that MAGID is comparable to or better than baselines, with significant improvements in human evaluation, especially against retrieval baselines where the image database is small.

Mahnaz Koupaee, Jake W. Vincent, Saab Mansour et al.

Faithfulness evaluators based on large language models (LLMs) are often fooled by the fluency of the text and struggle with identifying errors in the summaries. We propose an approach to summary faithfulness evaluation in which multiple LLM-based agents are assigned initial stances (regardless of what their belief might be) and forced to come up with a reason to justify the imposed belief, thus engaging in a multi-round debate to reach an agreement. The uniformly distributed initial assignments result in a greater diversity of stances leading to more meaningful debates and ultimately more errors identified. Furthermore, by analyzing the recent faithfulness evaluation datasets, we observe that naturally, it is not always the case for a summary to be either faithful to the source document or not. We therefore introduce a new dimension, ambiguity, and a detailed taxonomy to identify such special cases. Experiments demonstrate our approach can help identify ambiguities, and have even a stronger performance on non-ambiguous summaries.

Yuwei Zhang, Siffi Singh, Sailik Sengupta et al.

Conversational systems often rely on embedding models for intent classification and intent clustering tasks. The advent of Large Language Models (LLMs), which enable instructional embeddings allowing one to adjust semantics over the embedding space using prompts, are being viewed as a panacea for these downstream conversational tasks. However, traditional evaluation benchmarks rely solely on task metrics that don't particularly measure gaps related to semantic understanding. Thus, we propose an intent semantic toolkit that gives a more holistic view of intent embedding models by considering three tasks -- (1) intent classification, (2) intent clustering, and (3) a novel triplet task. The triplet task gauges the model's understanding of two semantic concepts paramount in real-world conversational systems -- negation and implicature. We observe that current embedding models fare poorly in semantic understanding of these concepts. To address this, we propose a pre-training approach to improve the embedding model by leveraging augmentation with data generated by an auto-regressive model and a contrastive loss term. Our approach improves the semantic understanding of the intent embedding model on the aforementioned linguistic dimensions while slightly effecting their performance on downstream task metrics.

Jason Cai, Hang Su, Monica Sunkara et al.

Large Language Models (LLMs) are powerful models for generation tasks, but they may not generate good quality outputs in their first attempt. Apart from model fine-tuning, existing approaches to improve prediction accuracy and quality typically involve LLM self-improvement / self-reflection that incorporate feedback from models themselves. Despite their effectiveness, these methods are hindered by their high computational cost and lack of scalability. In this work, we propose CERET, a method for refining text generations by considering semantic stability, entailment and inter-sample uncertainty measures. Experimental results show that CERET outperforms Self-consistency and Self-rerank baselines consistently under various task setups, by ~1.6% in Rouge-1 for abstractive summarization and ~3.5% in hit rate for question answering. Compared to LLM Self-rerank method, our approach only requires 9.4% of its latency and is more cost-effective.

Igor Shalyminov

Conversational User Interface (CUI) has become ubiquitous in everyday life, in consumer-focused products like Siri and Alexa or business-oriented solutions. Deep learning underlies many recent breakthroughs in dialogue systems but requires very large amounts of training data, often annotated by experts. Trained with smaller data, these methods end up severely lacking robustness (e.g. to disfluencies and out-of-domain input), and often just have too little generalisation power. In this thesis, we address the above issues by introducing a series of methods for training robust dialogue systems from minimal data. Firstly, we study two orthogonal approaches to dialogue: linguistically informed and machine learning-based - from the data efficiency perspective. We outline the steps to obtain data-efficient solutions with either approach. We then introduce two data-efficient models for dialogue response generation: the Dialogue Knowledge Transfer Network based on latent variable dialogue representations, and the hybrid Generative-Retrieval Transformer model (ranked first at the DSTC 8 Fast Domain Adaptation task). Next, we address the problem of robustness given minimal data. As such, propose a multitask LSTM-based model for domain-general disfluency detection. For the problem of out-of-domain input, we present Turn Dropout, a data augmentation technique for anomaly detection only using in-domain data, and introduce autoencoder-augmented models for efficient training with Turn Dropout. Finally, we focus on social dialogue and introduce a neural model for response ranking in social conversation used in Alana, the 3rd place winner in the Amazon Alexa Prize 2017 and 2018. We employ a novel technique of predicting the dialogue length as the main ranking objective and show that this approach improves upon the ratings-based counterpart in terms of data efficiency while matching it in performance.

Igor Shalyminov, Alessandro Sordoni, Adam Atkinson et al.

Domain adaptation has recently become a key problem in dialogue systems research. Deep learning, while being the preferred technique for modeling such systems, works best given massive training data. However, in the real-world scenario, such resources aren't available for every new domain, so the ability to train with a few dialogue examples can be considered essential. Pre-training on large data sources and adapting to the target data has become the standard method for few-shot problems within the deep learning framework. In this paper, we present the winning entry at the fast domain adaptation task of DSTC8, a hybrid generative-retrieval model based on GPT-2 fine-tuned to the multi-domain MetaLWOz dataset. Robust and diverse in response generation, our model uses retrieval logic as a fallback, being SoTA on MetaLWOz in human evaluation (>4% improvement over the 2nd place system) and attaining competitive generalization performance in adaptation to the unseen MultiWOZ dataset.

Igor Shalyminov, Sungjin Lee, Arash Eshghi et al.

Goal-oriented dialogue systems are now being widely adopted in industry where it is of key importance to maintain a rapid prototyping cycle for new products and domains. Data-driven dialogue system development has to be adapted to meet this requirement --- therefore, reducing the amount of data and annotations necessary for training such systems is a central research problem. In this paper, we present the Dialogue Knowledge Transfer Network (DiKTNet), a state-of-the-art approach to goal-oriented dialogue generation which only uses a few example dialogues (i.e. few-shot learning), none of which has to be annotated. We achieve this by performing a 2-stage training. Firstly, we perform unsupervised dialogue representation pre-training on a large source of goal-oriented dialogues in multiple domains, the MetaLWOz corpus. Secondly, at the transfer stage, we train DiKTNet using this representation together with 2 other textual knowledge sources with different levels of generality: ELMo encoder and the main dataset's source domains. Our main dataset is the Stanford Multi-Domain dialogue corpus. We evaluate our model on it in terms of BLEU and Entity F1 scores, and show that our approach significantly and consistently improves upon a series of baseline models as well as over the previous state-of-the-art dialogue generation model, ZSDG. The improvement upon the latter --- up to 10% in Entity F1 and the average of 3% in BLEU score --- is achieved using only the equivalent of 10% of ZSDG's in-domain training data.

Igor Shalyminov, Sungjin Lee, Arash Eshghi et al.

Learning with minimal data is one of the key challenges in the development of practical, production-ready goal-oriented dialogue systems. In a real-world enterprise setting where dialogue systems are developed rapidly and are expected to work robustly for an ever-growing variety of domains, products, and scenarios, efficient learning from a limited number of examples becomes indispensable. In this paper, we introduce a technique to achieve state-of-the-art dialogue generation performance in a few-shot setup, without using any annotated data. We do this by leveraging background knowledge from a larger, more highly represented dialogue source --- namely, the MetaLWOz dataset. We evaluate our model on the Stanford Multi-Domain Dialogue Dataset, consisting of human-human goal-oriented dialogues in in-car navigation, appointment scheduling, and weather information domains. We show that our few-shot approach achieves state-of-the art results on that dataset by consistently outperforming the previous best model in terms of BLEU and Entity F1 scores, while being more data-efficient by not requiring any data annotation.

Sungjin Lee, Igor Shalyminov

Neural dialog models often lack robustness to anomalous user input and produce inappropriate responses which leads to frustrating user experience. Although there are a set of prior approaches to out-of-domain (OOD) utterance detection, they share a few restrictions: they rely on OOD data or multiple sub-domains, and their OOD detection is context-independent which leads to suboptimal performance in a dialog. The goal of this paper is to propose a novel OOD detection method that does not require OOD data by utilizing counterfeit OOD turns in the context of a dialog. For the sake of fostering further research, we also release new dialog datasets which are 3 publicly available dialog corpora augmented with OOD turns in a controllable way. Our method outperforms state-of-the-art dialog models equipped with a conventional OOD detection mechanism by a large margin in the presence of OOD utterances.

Igor Shalyminov, Sungjin Lee

Neural network-based dialog models often lack robustness to anomalous, out-of-domain (OOD) user input which leads to unexpected dialog behavior and thus considerably limits such models' usage in mission-critical production environments. The problem is especially relevant in the setting of dialog system bootstrapping with limited training data and no access to OOD examples. In this paper, we explore the problem of robustness of such systems to anomalous input and the associated to it trade-off in accuracies on seen and unseen data. We present a new dataset for studying the robustness of dialog systems to OOD input, which is bAbI Dialog Task 6 augmented with OOD content in a controlled way. We then present turn dropout, a simple yet efficient negative sampling-based technique for improving robustness of neural dialog models. We demonstrate its effectiveness applied to Hybrid Code Network-family models (HCNs) which reach state-of-the-art results on our OOD-augmented dataset as well as the original one. Specifically, an HCN trained with turn dropout achieves state-of-the-art performance of more than 75% per-utterance accuracy on the augmented dataset's OOD turns and 74% F1-score as an OOD detector. Furthermore, we introduce a Variational HCN enhanced with turn dropout which achieves more than 56.5% accuracy on the original bAbI Task 6 dataset, thus outperforming the initially reported HCN's result.

Igor Shalyminov, Arash Eshghi, Oliver Lemon

Spontaneous spoken dialogue is often disfluent, containing pauses, hesitations, self-corrections and false starts. Processing such phenomena is essential in understanding a speaker's intended meaning and controlling the flow of the conversation. Furthermore, this processing needs to be word-by-word incremental to allow further downstream processing to begin as early as possible in order to handle real spontaneous human conversational behaviour. In addition, from a developer's point of view, it is highly desirable to be able to develop systems which can be trained from `clean' examples while also able to generalise to the very diverse disfluent variations on the same data -- thereby enhancing both data-efficiency and robustness. In this paper, we present a multi-task LSTM-based model for incremental detection of disfluency structure, which can be hooked up to any component for incremental interpretation (e.g. an incremental semantic parser), or else simply used to `clean up' the current utterance as it is being produced. We train the system on the Switchboard Dialogue Acts (SWDA) corpus and present its accuracy on this dataset. Our model outperforms prior neural network-based incremental approaches by about 10 percentage points on SWDA while employing a simpler architecture. To test the model's generalisation potential, we evaluate the same model on the bAbI+ dataset, without any additional training. bAbI+ is a dataset of synthesised goal-oriented dialogues where we control the distribution of disfluencies and their types. This shows that our approach has good generalisation potential, and sheds more light on which types of disfluency might be amenable to domain-general processing.

Ioannis Papaioannou, Amanda Cercas Curry, Jose L. Part et al.

Open-domain social dialogue is one of the long-standing goals of Artificial Intelligence. This year, the Amazon Alexa Prize challenge was announced for the first time, where real customers get to rate systems developed by leading universities worldwide. The aim of the challenge is to converse "coherently and engagingly with humans on popular topics for 20 minutes". We describe our Alexa Prize system (called 'Alana') consisting of an ensemble of bots, combining rule-based and machine learning systems, and using a contextual ranking mechanism to choose a system response. The ranker was trained on real user feedback received during the competition, where we address the problem of how to train on the noisy and sparse feedback obtained during the competition.

Arash Eshghi, Igor Shalyminov, Oliver Lemon

We investigate an end-to-end method for automatically inducing task-based dialogue systems from small amounts of unannotated dialogue data. It combines an incremental semantic grammar - Dynamic Syntax and Type Theory with Records (DS-TTR) - with Reinforcement Learning (RL), where language generation and dialogue management are a joint decision problem. The systems thus produced are incremental: dialogues are processed word-by-word, shown previously to be essential in supporting natural, spontaneous dialogue. We hypothesised that the rich linguistic knowledge within the grammar should enable a combinatorially large number of dialogue variations to be processed, even when trained on very few dialogues. Our experiments show that our model can process 74% of the Facebook AI bAbI dataset even when trained on only 0.13% of the data (5 dialogues). It can in addition process 65% of bAbI+, a corpus we created by systematically adding incremental dialogue phenomena such as restarts and self-corrections to bAbI. We compare our model with a state-of-the-art retrieval model, MemN2N. We find that, in terms of semantic accuracy, MemN2N shows very poor robustness to the bAbI+ transformations even when trained on the full bAbI dataset.

Igor Shalyminov, Arash Eshghi, Oliver Lemon

Natural, spontaneous dialogue proceeds incrementally on a word-by-word basis; and it contains many sorts of disfluency such as mid-utterance/sentence hesitations, interruptions, and self-corrections. But training data for machine learning approaches to dialogue processing is often either cleaned-up or wholly synthetic in order to avoid such phenomena. The question then arises of how well systems trained on such clean data generalise to real spontaneous dialogue, or indeed whether they are trainable at all on naturally occurring dialogue data. To answer this question, we created a new corpus called bAbI+ by systematically adding natural spontaneous incremental dialogue phenomena such as restarts and self-corrections to the Facebook AI Research's bAbI dialogues dataset. We then explore the performance of a state-of-the-art retrieval model, MemN2N, on this more natural dataset. Results show that the semantic accuracy of the MemN2N model drops drastically; and that although it is in principle able to learn to process the constructions in bAbI+, it needs an impractical amount of training data to do so. Finally, we go on to show that an incremental, semantic parser -- DyLan -- shows 100% semantic accuracy on both bAbI and bAbI+, highlighting the generalisation properties of linguistically informed dialogue models.