Eliot Brenner

Eliot Brenner, Dominic Seyler, Manjunath Hegde et al.

Despite advances in generative large language models (LLMs), practical application of specialized conversational AI agents remains constrained by computation costs, latency requirements, and the need for precise domain-specific relevance measures. While existing embedding models address the first two constraints, they underperform on information retrieval in specialized domains like finance. This paper introduces a scalable pipeline that trains specialized models from an unlabeled corpus using a general purpose retrieval embedding model as foundation. Our method yields an average of 27.7% improvement in MRR$\texttt{@}$5, 44.6% improvement in mean DCG$\texttt{@}$5 across 14 financial filing types measured over 21,800 query-document pairs, and improved NDCG on 3 of 4 document classes in FinanceBench. We adapt retrieval embeddings (bi-encoder) for RAG, not LLM generators, using LLM-judged relevance to distill domain knowledge into a compact retriever. There are prior works which pair synthetically generated queries with real passages to directly fine-tune the retrieval model. Our pipeline differs from these by introducing interaction between student and teacher models that interleaves retrieval-based mining of hard positive/negative examples from the unlabeled corpus with iterative retraining of the student model's weights using these examples. Each retrieval iteration uses the refined student model to mine the corpus for progressively harder training examples for the subsequent training iteration. The methodology provides a cost-effective solution to bridging the gap between general-purpose models and specialized domains without requiring labor-intensive human annotation.

Ahsaas Bajaj, Pavitra Dangati, Kalpesh Krishna et al.

Abstractive summarization is the task of compressing a long document into a coherent short document while retaining salient information. Modern abstractive summarization methods are based on deep neural networks which often require large training datasets. Since collecting summarization datasets is an expensive and time-consuming task, practical industrial settings are usually low-resource. In this paper, we study a challenging low-resource setting of summarizing long legal briefs with an average source document length of 4268 words and only 120 available (document, summary) pairs. To account for data scarcity, we used a modern pretrained abstractive summarizer BART (Lewis et al., 2020), which only achieves 17.9 ROUGE-L as it struggles with long documents. We thus attempt to compress these long documents by identifying salient sentences in the source which best ground the summary, using a novel algorithm based on GPT-2 (Radford et al., 2019) language model perplexity scores, that operates within the low resource regime. On feeding the compressed documents to BART, we observe a 6.0 ROUGE-L improvement. Our method also beats several competitive salience detection baselines. Furthermore, the identified salient sentences tend to agree with an independent human labeling by domain experts.

Eliot Brenner, Jun Zhao, Aliasgar Kutiyanawala et al.

We consider the problem of retrieving and ranking items in an eCommerce catalog, often called SKUs, in order of relevance to a user-issued query. The input data for the ranking are the texts of the queries and textual fields of the SKUs indexed in the catalog. We review the ways in which this problem both resembles and differs from the problems of IR in the context of web search. The differences between the product-search problem and the IR problem of web search necessitate a different approach in terms of both models and datasets. We first review the recent state-of-the-art models for web search IR, distinguishing between two distinct types of model which we call the distributed type and the local-interaction type. The different types of relevance models developed for IR have complementary advantages and disadvantages when applied to eCommerce product search. Further, we explain why the conventional methods for dataset construction employed in the IR literature fail to produce data which suffices for training or evaluation of models for eCommerce product search. We explain how our own approach, applying task modeling techniques to the click-through logs of an eCommerce site, enables the construction of a large-scale dataset for training and robust benchmarking of relevance models. Our experiments consist of applying several of the models from the IR literature to our own dataset. Empirically, we have established that, when applied to our dataset, certain models of local-interaction type reduce ranking errors by one-third compared to the baseline tf-idf. Applied to our dataset, the distributed models fail to outperform the baseline. As a basis for a deployed system, the distributed models have several advantages, computationally, over the local-interaction models. This motivates an ongoing program of work, which we outline at the conclusion of the paper.

Eliot Brenner, David Sontag

We give a new consistent scoring function for structure learning of Bayesian networks. In contrast to traditional approaches to score-based structure learning, such as BDeu or MDL, the complexity penalty that we propose is data-dependent and is given by the probability that a conditional independence test correctly shows that an edge cannot exist. What really distinguishes this new scoring function from earlier work is that it has the property of becoming computationally easier to maximize as the amount of data increases. We prove a polynomial sample complexity result, showing that maximizing this score is guaranteed to correctly learn a structure with no false edges and a distribution close to the generating distribution, whenever there exists a Bayesian network which is a perfect map for the data generating distribution. Although the new score can be used with any search algorithm, in our related UAI 2013 paper [BS13], we have given empirical results showing that it is particularly effective when used together with a linear programming relaxation approach to Bayesian network structure learning. The present paper contains all details of the proofs of the finite-sample complexity results in [BS13] as well as detailed explanation of the computation of the certain error probabilities called beta-values, whose precomputation and tabulation is necessary for the implementation of the algorithm in [BS13].

Eliot Brenner, David Sontag

We give a new consistent scoring function for structure learning of Bayesian networks. In contrast to traditional approaches to scorebased structure learning, such as BDeu or MDL, the complexity penalty that we propose is data-dependent and is given by the probability that a conditional independence test correctly shows that an edge cannot exist. What really distinguishes this new scoring function from earlier work is that it has the property of becoming computationally easier to maximize as the amount of data increases. We prove a polynomial sample complexity result, showing that maximizing this score is guaranteed to correctly learn a structure with no false edges and a distribution close to the generating distribution, whenever there exists a Bayesian network which is a perfect map for the data generating distribution. Although the new score can be used with any search algorithm, we give empirical results showing that it is particularly effective when used together with a linear programming relaxation approach to Bayesian network structure learning.