Lucian Popa

Sara Rosenthal, Yannis Katsis, Vraj Shah et al. · ibm-research

We present MTRAG-UN, a benchmark for exploring open challenges in multi-turn retrieval augmented generation, a popular use of large language models. We release a benchmark of 666 tasks containing over 2,800 conversation turns across 6 domains with accompanying corpora. Our experiments show that retrieval and generation models continue to struggle on conversations with UNanswerable, UNderspecified, and NONstandalone questions and UNclear responses. Our benchmark is available at https://github.com/IBM/mt-rag-benchmark

Yu Zhang, Yunyi Zhang, Yucheng Jiang et al. · amazon-science

Given a few seed entities of a certain type (e.g., Software or Programming Language), entity set expansion aims to discover an extensive set of entities that share the same type as the seeds. Entity set expansion in software-related domains such as StackOverflow can benefit many downstream tasks (e.g., software knowledge graph construction) and facilitate better IT operations and service management. Meanwhile, existing approaches are less concerned with two problems: (1) How to deal with multiple types of seed entities simultaneously? (2) How to leverage the power of pre-trained language models (PLMs)? Being aware of these two problems, in this paper, we study the entity set co-expansion task in StackOverflow, which extracts Library, OS, Application, and Language entities from StackOverflow question-answer threads. During the co-expansion process, we use PLMs to derive embeddings of candidate entities for calculating similarities between entities. Experimental results show that our proposed SECoExpan framework outperforms previous approaches significantly.

Ronald Fagin, Phokion G. Kolaitis, Domenico Lembo et al.

We propose a new framework for combining entity resolution and query answering in knowledge bases (KBs) with tuple-generating dependencies (tgds) and equality-generating dependencies (egds) as rules. We define the semantics of the KB in terms of special instances that involve equivalence classes of entities and sets of values. Intuitively, the former collect all entities denoting the same real-world object, while the latter collect all alternative values for an attribute. This approach allows us to both resolve entities and bypass possible inconsistencies in the data. We then design a chase procedure that is tailored to this new framework and has the feature that it never fails; moreover, when the chase procedure terminates, it produces a universal solution, which in turn can be used to obtain the certain answers to conjunctive queries. We finally discuss challenges arising when the chase does not terminate.

Krishnateja Killamsetty, Guttu Sai Abhishek, Aakriti et al.

Deep neural networks have seen great success in recent years; however, training a deep model is often challenging as its performance heavily depends on the hyper-parameters used. In addition, finding the optimal hyper-parameter configuration, even with state-of-the-art (SOTA) hyper-parameter optimization (HPO) algorithms, can be time-consuming, requiring multiple training runs over the entire dataset for different possible sets of hyper-parameters. Our central insight is that using an informative subset of the dataset for model training runs involved in hyper-parameter optimization, allows us to find the optimal hyper-parameter configuration significantly faster. In this work, we propose AUTOMATA, a gradient-based subset selection framework for hyper-parameter tuning. We empirically evaluate the effectiveness of AUTOMATA in hyper-parameter tuning through several experiments on real-world datasets in the text, vision, and tabular domains. Our experiments show that using gradient-based data subsets for hyper-parameter tuning achieves significantly faster turnaround times and speedups of 3$\times$-30$\times$ while achieving comparable performance to the hyper-parameters found using the entire dataset.

Krishnateja Killamsetty, Alexandre V. Evfimievski, Tejaswini Pedapati et al.

Training deep networks and tuning hyperparameters on large datasets is computationally intensive. One of the primary research directions for efficient training is to reduce training costs by selecting well-generalizable subsets of training data. Compared to simple adaptive random subset selection baselines, existing intelligent subset selection approaches are not competitive due to the time-consuming subset selection step, which involves computing model-dependent gradients and feature embeddings and applies greedy maximization of submodular objectives. Our key insight is that removing the reliance on downstream model parameters enables subset selection as a pre-processing step and enables one to train multiple models at no additional cost. In this work, we propose MILO, a model-agnostic subset selection framework that decouples the subset selection from model training while enabling superior model convergence and performance by using an easy-to-hard curriculum. Our empirical results indicate that MILO can train models $3\times - 10 \times$ faster and tune hyperparameters $20\times - 75 \times$ faster than full-dataset training or tuning without compromising performance.

Pritom Saha Akash, Jie Huang, Kevin Chen-Chuan Chang et al.

We propose a probabilistic approach to select a subset of a \textit{target domain representative keywords} from a candidate set, contrasting with a context domain. Such a task is crucial for many downstream tasks in natural language processing. To contrast the target domain and the context domain, we adapt the \textit{two-component mixture model} concept to generate a distribution of candidate keywords. It provides more importance to the \textit{distinctive} keywords of the target domain than common keywords contrasting with the context domain. To support the \textit{representativeness} of the selected keywords towards the target domain, we introduce an \textit{optimization algorithm} for selecting the subset from the generated candidate distribution. We have shown that the optimization algorithm can be efficiently implemented with a near-optimal approximation guarantee. Finally, extensive experiments on multiple domains demonstrate the superiority of our approach over other baselines for the tasks of keyword summary generation and trending keywords selection.

Yannis Katsis, Sara Rosenthal, Kshitij Fadnis et al. · ibm-research

Retrieval-augmented generation (RAG) has recently become a very popular task for Large Language Models (LLMs). Evaluating them on multi-turn RAG conversations, where the system is asked to generate a response to a question in the context of a preceding conversation is an important and often overlooked task with several additional challenges. We present MTRAG: an end-to-end human-generated multi-turn RAG benchmark that reflects several real-world properties across diverse dimensions for evaluating the full RAG pipeline. MTRAG contains 110 conversations averaging 7.7 turns each across four domains for a total of 842 tasks. We also explore automation paths via synthetic data and LLM-as-a-Judge evaluation. Our human and automatic evaluations show that even state-of-the-art LLM RAG systems struggle on MTRAG. We demonstrate the need for strong retrieval and generation systems that can handle later turns, unanswerable questions, non-standalone questions, and multiple domains. MTRAG is available at https://github.com/ibm/mt-rag-benchmark.

Pritom Saha Akash, Kashob Kumar Roy, Lucian Popa et al.

Long-form question answering (LFQA) poses a challenge as it involves generating detailed answers in the form of paragraphs, which go beyond simple yes/no responses or short factual answers. While existing QA models excel in questions with concise answers, LFQA requires handling multiple topics and their intricate relationships, demanding comprehensive explanations. Previous attempts at LFQA focused on generating long-form answers by utilizing relevant contexts from a corpus, relying solely on the question itself. However, they overlooked the possibility that the question alone might not provide sufficient information to identify the relevant contexts. Additionally, generating detailed long-form answers often entails aggregating knowledge from diverse sources. To address these limitations, we propose an LFQA model with iterative Planning, Retrieval, and Generation. This iterative process continues until a complete answer is generated for the given question. From an extensive experiment on both an open domain and a technical domain QA dataset, we find that our model outperforms the state-of-the-art models on various textual and factual metrics for the LFQA task.

Kristjan Greenewald, Luis Lastras, Thomas Parnell et al.

Low-Rank Adaptation (LoRA) has emerged as a highly efficient framework for finetuning the weights of large foundation models, and has become the go-to method for data-driven customization of LLMs. Despite the promise of highly customized behaviors and capabilities, switching between relevant LoRAs in a multiturn setting is inefficient, as the key-value (KV) cache of the entire turn history must be recomputed with the LoRA weights before generation can begin. To address this problem, we propose Activated LoRA (aLoRA), an adapter architecture which modifies the LoRA framework to only adapt weights for the tokens in the sequence after the aLoRA is invoked. This change crucially allows aLoRA to accept the base model's KV cache of the input string, meaning that aLoRA can be instantly activated whenever needed in a chain without recomputing the prior keys and values. This enables building what we call intrinsics, i.e. specialized models invoked to perform well-defined operations on portions of an input chain or conversation that otherwise uses the base model by default. We train a set of aLoRA-based intrinsics models, demonstrating competitive accuracy with standard LoRA while significantly improving inference efficiency. We contributed our Activated LoRA implementation to the Huggingface PEFT library https://github.com/huggingface/peft.

Ishika Agarwal, Krishnateja Killamsetty, Lucian Popa et al.

Fine-tuning large language models (LLMs) is essential for enhancing their performance on specific tasks but is often resource-intensive due to redundant or uninformative data. To address this inefficiency, we introduce DELIFT (Data Efficient Language model Instruction Fine-Tuning), a novel algorithm that systematically optimizes data selection across the three key stages of fine-tuning: (1) instruction tuning, (2) task-specific fine-tuning (e.g., reasoning, question-answering), and (3) continual fine-tuning (e.g., incorporating new data versions). Unlike existing methods that focus on single-stage optimization or rely on computationally intensive gradient calculations, DELIFT operates efficiently across all stages. Central to our approach is a pairwise utility metric that quantifies how beneficial a data sample is for improving the model's responses to other samples, effectively measuring the informational value relative to the model's current capabilities. By leveraging different submodular functions applied to this metric, DELIFT selects diverse and optimal subsets that are useful across all stages of fine-tuning. Experiments across various tasks and model scales demonstrate that DELIFT can reduce the fine-tuning data size by up to 70% without compromising performance, offering significant computational savings and outperforming existing methods in both efficiency and efficacy.

Yu Zhang, Yunyi Zhang, Yanzhen Shen et al. · amazon-science

Accurately typing entity mentions from text segments is a fundamental task for various natural language processing applications. Many previous approaches rely on massive human-annotated data to perform entity typing. Nevertheless, collecting such data in highly specialized science and engineering domains (e.g., software engineering and security) can be time-consuming and costly, without mentioning the domain gaps between training and inference data if the model needs to be applied to confidential datasets. In this paper, we study the task of seed-guided fine-grained entity typing in science and engineering domains, which takes the name and a few seed entities for each entity type as the only supervision and aims to classify new entity mentions into both seen and unseen types (i.e., those without seed entities). To solve this problem, we propose SEType which first enriches the weak supervision by finding more entities for each seen type from an unlabeled corpus using the contextualized representations of pre-trained language models. It then matches the enriched entities to unlabeled text to get pseudo-labeled samples and trains a textual entailment model that can make inferences for both seen and unseen types. Extensive experiments on two datasets covering four domains demonstrate the effectiveness of SEType in comparison with various baselines.

Marina Danilevsky, Kristjan Greenewald, Chulaka Gunasekara et al. · ibm-research

In the developer community for large language models (LLMs), there is not yet a clean pattern analogous to a software library, to support very large scale collaboration. Even for the commonplace use case of Retrieval-Augmented Generation (RAG), it is not currently possible to write a RAG application against a well-defined set of APIs that are agreed upon by different LLM providers. Inspired by the idea of compiler intrinsics, we propose some elements of such a concept through introducing a library of LLM Intrinsics for RAG. An LLM intrinsic is defined as a capability that can be invoked through a well-defined API that is reasonably stable and independent of how the LLM intrinsic itself is implemented. The intrinsics in our library are released as LoRA adapters on HuggingFace, and through a software interface with clear structured input/output characteristics on top of vLLM as an inference platform, accompanied in both places with documentation and code. This article describes the intended usage, training details, and evaluations for each intrinsic, as well as compositions of multiple intrinsics.

Bingsheng Yao, Ishan Jindal, Lucian Popa et al.

Real-world domain experts (e.g., doctors) rarely annotate only a decision label in their day-to-day workflow without providing explanations. Yet, existing low-resource learning techniques, such as Active Learning (AL), that aim to support human annotators mostly focus on the label while neglecting the natural language explanation of a data point. This work proposes a novel AL architecture to support experts' real-world need for label and explanation annotations in low-resource scenarios. Our AL architecture leverages an explanation-generation model to produce explanations guided by human explanations, a prediction model that utilizes generated explanations toward prediction faithfully, and a novel data diversity-based AL sampling strategy that benefits from the explanation annotations. Automated and human evaluations demonstrate the effectiveness of incorporating explanations into AL sampling and the improved human annotation efficiency and trustworthiness with our AL architecture. Additional ablation studies illustrate the potential of our AL architecture for transfer learning, generalizability, and integration with large language models (LLMs). While LLMs exhibit exceptional explanation-generation capabilities for relatively simple tasks, their effectiveness in complex real-world tasks warrants further in-depth study.

Bingsheng Yao, Prithviraj Sen, Lucian Popa et al.

Human-annotated labels and explanations are critical for training explainable NLP models. However, unlike human-annotated labels whose quality is easier to calibrate (e.g., with a majority vote), human-crafted free-form explanations can be quite subjective. Before blindly using them as ground truth to train ML models, a vital question needs to be asked: How do we evaluate a human-annotated explanation's quality? In this paper, we build on the view that the quality of a human-annotated explanation can be measured based on its helpfulness (or impairment) to the ML models' performance for the desired NLP tasks for which the annotations were collected. In comparison to the commonly used Simulatability score, we define a new metric that can take into consideration the helpfulness of an explanation for model performance at both fine-tuning and inference. With the help of a unified dataset format, we evaluated the proposed metric on five datasets (e.g., e-SNLI) against two model architectures (T5 and BART), and the results show that our proposed metric can objectively evaluate the quality of human-annotated explanations, while Simulatability falls short.

Ayush Maheshwari, Krishnateja Killamsetty, Ganesh Ramakrishnan et al.

A critical bottleneck in supervised machine learning is the need for large amounts of labeled data which is expensive and time consuming to obtain. However, it has been shown that a small amount of labeled data, while insufficient to re-train a model, can be effectively used to generate human-interpretable labeling functions (LFs). These LFs, in turn, have been used to generate a large amount of additional noisy labeled data, in a paradigm that is now commonly referred to as data programming. However, previous approaches to automatically generate LFs make no attempt to further use the given labeled data for model training, thus giving up opportunities for improved performance. Moreover, since the LFs are generated from a relatively small labeled dataset, they are prone to being noisy, and naively aggregating these LFs can lead to very poor performance in practice. In this work, we propose an LF based reweighting framework \ouralgo{} to solve these two critical limitations. Our algorithm learns a joint model on the (same) labeled dataset used for LF induction along with any unlabeled data in a semi-supervised manner, and more critically, reweighs each LF according to its goodness, influencing its contribution to the semi-supervised loss using a robust bi-level optimization algorithm. We show that our algorithm significantly outperforms prior approaches on several text classification datasets.

Hang Jiang, Sairam Gurajada, Qiuhao Lu et al.

Entity linking (EL), the task of disambiguating mentions in text by linking them to entities in a knowledge graph, is crucial for text understanding, question answering or conversational systems. Entity linking on short text (e.g., single sentence or question) poses particular challenges due to limited context. While prior approaches use either heuristics or black-box neural methods, here we propose LNN-EL, a neuro-symbolic approach that combines the advantages of using interpretable rules based on first-order logic with the performance of neural learning. Even though constrained to using rules, LNN-EL performs competitively against SotA black-box neural approaches, with the added benefits of extensibility and transferability. In particular, we show that we can easily blend existing rule templates given by a human expert, with multiple types of features (priors, BERT encodings, box embeddings, etc), and even scores resulting from previous EL methods, thus improving on such methods. For instance, on the LC-QuAD-1.0 dataset, we show more than $4$\% increase in F1 score over previous SotA. Finally, we show that the inductive bias offered by using logic results in learned rules that transfer well across datasets, even without fine tuning, while maintaining high accuracy.

Pavan Kapanipathi, Ibrahim Abdelaziz, Srinivas Ravishankar et al.

Knowledge base question answering (KBQA)is an important task in Natural Language Processing. Existing approaches face significant challenges including complex question understanding, necessity for reasoning, and lack of large end-to-end training datasets. In this work, we propose Neuro-Symbolic Question Answering (NSQA), a modular KBQA system, that leverages (1) Abstract Meaning Representation (AMR) parses for task-independent question understanding; (2) a simple yet effective graph transformation approach to convert AMR parses into candidate logical queries that are aligned to the KB; (3) a pipeline-based approach which integrates multiple, reusable modules that are trained specifically for their individual tasks (semantic parser, entity andrelationship linkers, and neuro-symbolic reasoner) and do not require end-to-end training data. NSQA achieves state-of-the-art performance on two prominent KBQA datasets based on DBpedia (QALD-9 and LC-QuAD1.0). Furthermore, our analysis emphasizes that AMR is a powerful tool for KBQA systems.

Kun Qian, Poornima Chozhiyath Raman, Yunyao Li et al.

Structured representations of entity names are useful for many entity-related tasks such as entity normalization and variant generation. Learning the implicit structured representations of entity names without context and external knowledge is particularly challenging. In this paper, we present a novel learning framework that combines active learning and weak supervision to solve this problem. Our experimental evaluation show that this framework enables the learning of high-quality models from merely a dozen or so labeled examples.

Xinyi Zheng, Doug Burdick, Lucian Popa et al.

Documents are often used for knowledge sharing and preservation in business and science, within which are tables that capture most of the critical data. Unfortunately, most documents are stored and distributed as PDF or scanned images, which fail to preserve logical table structure. Recent vision-based deep learning approaches have been proposed to address this gap, but most still cannot achieve state-of-the-art results. We present Global Table Extractor (GTE), a vision-guided systematic framework for joint table detection and cell structured recognition, which could be built on top of any object detection model. With GTE-Table, we invent a new penalty based on the natural cell containment constraint of tables to train our table network aided by cell location predictions. GTE-Cell is a new hierarchical cell detection network that leverages table styles. Further, we design a method to automatically label table and cell structure in existing documents to cheaply create a large corpus of training and test data. We use this to enhance PubTabNet with cell labels and create FinTabNet, real-world and complex scientific and financial datasets with detailed table structure annotations to help train and test structure recognition. Our framework surpasses previous state-of-the-art results on the ICDAR 2013 and ICDAR 2019 table competition in both table detection and cell structure recognition with a significant 5.8% improvement in the full table extraction system. Further experiments demonstrate a greater than 45% improvement in cell structure recognition when compared to a vanilla RetinaNet object detection model in our new out-of-domain FinTabNet.

Venkata Vamsikrishna Meduri, Lucian Popa, Prithviraj Sen et al.

Entity Matching (EM) is a core data cleaning task, aiming to identify different mentions of the same real-world entity. Active learning is one way to address the challenge of scarce labeled data in practice, by dynamically collecting the necessary examples to be labeled by an Oracle and refining the learned model (classifier) upon them. In this paper, we build a unified active learning benchmark framework for EM that allows users to easily combine different learning algorithms with applicable example selection algorithms. The goal of the framework is to enable concrete guidelines for practitioners as to what active learning combinations will work well for EM. Towards this, we perform comprehensive experiments on publicly available EM datasets from product and publication domains to evaluate active learning methods, using a variety of metrics including EM quality, #labels and example selection latencies. Our most surprising result finds that active learning with fewer labels can learn a classifier of comparable quality as supervised learning. In fact, for several of the datasets, we show that there is an active learning combination that beats the state-of-the-art supervised learning result. Our framework also includes novel optimizations that improve the quality of the learned model by roughly 9% in terms of F1-score and reduce example selection latencies by up to 10x without affecting the quality of the model.

Bogdan Groza, Lucian Popa, Pal-Stefan Murvay

Previous research works have endorsed the use of delays and clock skews for detecting intrusions or fingerprinting ECUs on the CAN bus. Similar techniques have been also proposed for establishing a time-covert cryptographic authentication channel, in this way cleverly removing the need for cryptographic material inside the limited payload of CAN frames. The main shortcoming of such works is the limited security level that can be achieved under normal CAN-bus traffic. In this work we endeavour to test the limits of the achievable security level by relying on optimization algorithms for scheduling CAN frames. Under practical bus allocations that are based on real-world scenarios, we are able to extract around 4--5 bits of authentication data from each frame which leads to an efficient intrusion detection and authentication mechanism. By accumulating covert channel data over several consecutive frames, we can achieve higher security levels that are in line with current security demands. To prove the correctness of our approach, we present experiments on state-of-the-art automotive-grade controllers (Infineon Aurix) and bus measurements with the use of industry standard tools, i.e., CANoe.

Jungo Kasai, Kun Qian, Sairam Gurajada et al.

Entity resolution (ER) is the task of identifying different representations of the same real-world entities across databases. It is a key step for knowledge base creation and text mining. Recent adaptation of deep learning methods for ER mitigates the need for dataset-specific feature engineering by constructing distributed representations of entity records. While these methods achieve state-of-the-art performance over benchmark data, they require large amounts of labeled data, which are typically unavailable in realistic ER applications. In this paper, we develop a deep learning-based method that targets low-resource settings for ER through a novel combination of transfer learning and active learning. We design an architecture that allows us to learn a transferable model from a high-resource setting to a low-resource one. To further adapt to the target dataset, we incorporate active learning that carefully selects a few informative examples to fine-tune the transferred model. Empirical evaluation demonstrates that our method achieves comparable, if not better, performance compared to state-of-the-art learning-based methods while using an order of magnitude fewer labels.

Phokion G. Kolaitis, Lucian Popa, Kun Qian

In several different applications, including data transformation and entity resolution, rules are used to capture aspects of knowledge about the application at hand. Often, a large set of such rules is generated automatically or semi-automatically, and the challenge is to refine the encapsulated knowledge by selecting a subset of rules based on the expected operational behavior of the rules on available data. In this paper, we carry out a systematic complexity-theoretic investigation of the following rule selection problem: given a set of rules specified by Horn formulas, and a pair of an input database and an output database, find a subset of the rules that minimizes the total error, that is, the number of false positive and false negative errors arising from the selected rules. We first establish computational hardness results for the decision problems underlying this minimization problem, as well as upper and lower bounds for its approximability. We then investigate a bi-objective optimization version of the rule selection problem in which both the total error and the size of the selected rules are taken into account. We show that testing for membership in the Pareto front of this bi-objective optimization problem is DP-complete. Finally, we show that a similar DP-completeness result holds for a bi-level optimization version of the rule selection problem, where one minimizes first the total error and then the size.