Personalization is well studied in search and recommendation, but personalized question answering remains underexplored due to challenges in inferring preferences from long, noisy, implicit contexts and generating responses that are both accurate and aligned with user expectations. To address this, we propose Pathways of Thoughts (PoT), an inference-stage method that applies to any large language model (LLM) without task-specific fine-tuning. PoT models the thinking as an iterative decision process, where the model dynamically selects among cognitive operations such as reasoning, revision, personalization, and clarification. This enables exploration of multiple reasoning trajectories, producing diverse candidate responses that capture different perspectives. PoT then aggregates and reweights these candidates according to inferred user preferences, yielding a final personalized response that benefits from the complementary strengths of diverse reasoning paths. Experiments on the LaMP-QA benchmark show that PoT consistently outperforms competitive baselines, achieving up to a 10.8% relative improvement. Human evaluation further validates these improvements, with annotators preferring PoT in 66% of cases compared to the best-performing baseline and reporting ties in 15% of cases.

This is the second workshop in the series which discusses the cutting-edge developments in research and applications of personalizing large language models (LLMs) and adapting them to the demands of diverse user populations and societal needs. The full-day workshop plan includes several keynotes and invited talks, a poster session and a panel discussion.

This workshop discusses the cutting-edge developments in research and applications of personalizing large language models (LLMs) and adapting them to the demands of diverse user populations and societal needs. The full-day workshop includes several keynotes and invited talks, a poster session and a panel discussion.

Facilitated by large language models (LLMs), personalized text generation has become a rapidly growing research direction. Most existing studies focus on designing specialized models for a particular domain, or they require fine-tuning the LLMs to generate personalized text. We consider a typical scenario in which the large language model, which generates personalized output, is frozen and can only be accessed through APIs. Under this constraint, all one can do is to improve the input text (i.e., text prompts) sent to the LLM, a procedure that is usually done manually. In this paper, we propose a novel method to automatically revise prompts for personalized text generation. The proposed method takes the initial prompts generated by a state-of-the-art, multistage framework for personalized generation and rewrites a few critical components that summarize and synthesize the personal context. The prompt rewriter employs a training paradigm that chains together supervised learning (SL) and reinforcement learning (RL), where SL reduces the search space of RL and RL facilitates end-to-end training of the rewriter. Using datasets from three representative domains, we demonstrate that the rewritten prompts outperform both the original prompts and the prompts optimized via supervised learning or reinforcement learning alone. In-depth analysis of the rewritten prompts shows that they are not only human readable, but also able to guide manual revision of prompts when there is limited resource to employ reinforcement learning to train the prompt rewriter, or when it is costly to deploy an automatic prompt rewriter for inference.

A recent line of work in first-stage Neural Information Retrieval has focused on learning sparse lexical representations instead of dense embeddings. One such work is SPLADE, which has been shown to lead to state-of-the-art results in both the in-domain and zero-shot settings, can leverage inverted indices for efficient retrieval, and offers enhanced interpretability. However, existing SPLADE models are fundamentally limited to learning a sparse representation based on the native BERT WordPiece vocabulary.

Bag-of-words based lexical retrieval systems are still the most commonly used methods for real-world search applications. Recently deep learning methods have shown promising results to improve this retrieval performance but are expensive to run in an online fashion, non-trivial to integrate into existing production systems, and might not generalize well in out-of-domain retrieval scenarios. Instead, we build on top of lexical retrievers by proposing a Term Weighting BERT (TW-BERT) model. TW-BERT learns to predict the weight for individual n-gram (e.g., uni-grams and bi-grams) query input terms. These inferred weights and terms can be used directly by a retrieval system to perform a query search. To optimize these term weights, TW-BERT incorporates the scoring function used by the search engine, such as BM25, to score query-document pairs. Given sample query-document pairs we can compute a ranking loss over these matching scores, optimizing the learned query term weights in an end-to-end fashion. Aligning TW-BERT with search engine scorers minimizes the changes needed to integrate it into existing production applications, whereas existing deep learning based search methods would require further infrastructure optimization and hardware requirements. The learned weights can be easily utilized by standard lexical retrievers and by other retrieval techniques such as query expansion. We show that TW-BERT improves retrieval performance over strong term weighting baselines within MSMARCO and in out-of-domain retrieval on TREC datasets.

In dense retrieval, prior work has largely improved retrieval effectiveness using multi-vector dense representations, exemplified by ColBERT. In sparse retrieval, more recent work, such as SPLADE, demonstrated that one can also learn sparse lexical representations to achieve comparable effectiveness while enjoying better interpretability. In this work, we combine the strengths of both the sparse and dense representations for first-stage retrieval. Specifically, we propose SparseEmbed - a novel retrieval model that learns sparse lexical representations with contextual embeddings. Compared with SPLADE, our model leverages the contextual embeddings to improve model expressiveness. Compared with ColBERT, our sparse representations are trained end-to-end to optimize both efficiency and effectiveness.

Prior work in Dense Retrieval usually encodes queries and documents using single-vector representations (also called embeddings) and performs retrieval in the embedding space using approximate nearest neighbor search. This paradigm enables efficient semantic retrieval. However, the single-vector representations can be ineffective at capturing different aspects of the queries and documents in relevance matching, especially for some vertical domains. For example, in e-commerce search, these aspects could be category, brand and color. Given a query ''white nike socks", a Dense Retrieval model may mistakenly retrieve some ''white adidas socks" while missing out the intended brand. We propose to explicitly represent multiple aspects using one embedding per aspect. We introduce an aspect prediction task to teach the model to capture aspect information with particular aspect embeddings. We design a lightweight network to fuse the aspect embeddings for representing queries and documents. Our evaluation using an e-commerce dataset shows impressive improvements over strong Dense Retrieval baselines. We also discover that the proposed aspect embeddings can enhance the interpretability of Dense Retrieval models as a byproduct.

Many natural language processing and information retrieval problems can be formalized as the task of semantic matching. Existing work in this area has been largely focused on matching between short texts (e.g., question answering), or between a short and a long text (e.g., ad-hoc retrieval). Semantic matching between long-form documents, which has many important applications like news recommendation, related article recommendation and document clustering, is relatively less explored and needs more research effort. In recent years, self-attention based models like Transformers and BERT have achieved state-of-the-art performance in the task of text matching. These models, however, are still limited to short text like a few sentences or one paragraph due to the quadratic computational complexity of self-attention with respect to input text length. In this paper, we address the issue by proposing the Siamese Multi-depth Transformer-based Hierarchical (SMITH) Encoder for long-form document matching. Our model contains several innovations to adapt self-attention models for longer text input. We propose a transformer based hierarchical encoder to capture the document structure information. In order to better capture sentence level semantic relations within a document, we pre-train the model with a novel masked sentence block language modeling task in addition to the masked word language modeling task used by BERT. Our experimental results on several benchmark data sets for long-form document matching show that our proposed SMITH model outperforms the previous state-of-the-art models including hierarchical attention, multi-depth attention-based hierarchical recurrent neural network, and BERT. Comparing to BERT based baselines, our model is able to increase maximum input text length from 512 to 2048. We will open source a Wikipedia based benchmark data set, code and a pre-trained model to accelerate future research on long-form document matching.

Learning-to-Rank deals with maximizing the utility of a list of examples presented to the user, with items of higher relevance being prioritized. It has several practical applications such as large-scale search, recommender systems, document summarization and question answering. While there is widespread support for classification and regression based learning, support for learning-to-rank in deep learning has been limited. We introduce TensorFlow Ranking, the first open source library for solving large-scale ranking problems in a deep learning framework. It is highly configurable and provides easy-to-use APIs to support different scoring mechanisms, loss functions and evaluation metrics in the learning-to-rank setting. Our library is developed on top of TensorFlow and can thus fully leverage the advantages of this platform. TensorFlow Ranking has been deployed in production systems within Google; it is highly scalable, both in training and in inference, and can be used to learn ranking models over massive amounts of user activity data, which can include heterogeneous dense and sparse features. We empirically demonstrate the effectiveness of our library in learning ranking functions for large-scale search and recommendation applications in Gmail and Google Drive. We also show that ranking models built using our model scale well for distributed training, without significant impact on metrics. The proposed library is available to the open source community, with the hope that it facilitates further academic research and industrial applications in the field of learning-to-rank.

Synonym expansion is a technique that adds related words to search queries, which may lead to more relevant documents being retrieved, thus improving recall. There is extensive prior work on synonym expansion for web search, however very few studies have tackled its application for email search. Synonym expansion for private corpora like emails poses several unique research challenges. First, the emails are not shared across users, which precludes us from directly employing query-document bipartite graphs, which are standard in web search synonym expansion. Second, user search queries are of personal nature, and may not be generalizable across users. Third, the size of the underlying corpora from which the synonyms may be mined is relatively small (i.e., user's private email inbox) compared to the size of the web corpus. Therefore, in this paper, we propose a solution tailored to the challenges of synonym expansion for email search. We formulate it as a multi-view learning problem, and propose a novel embedding-based model that joins information from multiple sources to obtain the optimal synonym candidates. To demonstrate the effectiveness of the proposed technique, we evaluate our model using both explicit human ratings as well as a live experiment using the Gmail Search service, one of the world's largest email search engines.

Presentation bias is one of the key challenges when learning from implicit feedback in search engines, as it confounds the relevance signal. While it was recently shown how counterfactual learning-to-rank (LTR) approaches \citeJoachims/etal/17a can provably overcome presentation bias when observation propensities are known, it remains to show how to effectively estimate these propensities. In this paper, we propose the first method for producing consistent propensity estimates without manual relevance judgments, disruptive interventions, or restrictive relevance modeling assumptions. First, we show how to harvest a specific type of intervention data from historic feedback logs of multiple different ranking functions, and show that this data is sufficient for consistent propensity estimation in the position-based model. Second, we propose a new extremum estimator that makes effective use of this data. In an empirical evaluation, we find that the new estimator provides superior propensity estimates in two real-world systems -- Arxiv Full-text Search and Google Drive Search. Beyond these two points, we find that the method is robust to a wide range of settings in simulation studies.

Existing unbiased learning-to-rank models use counterfactual inference, notably Inverse Propensity Scoring (IPS), to learn a ranking function from biased click data. They handle the click incompleteness bias, but usually assume that the clicks are noise-free, i.e., a clicked document is always assumed to be relevant. In this paper, we relax this unrealistic assumption and study click noise explicitly in the unbiased learning-to-rank setting. Specifically, we model the noise as the position-dependent trust bias and propose a noise-aware Position-Based Model, named TrustPBM, to better capture user click behavior. We propose an Expectation-Maximization algorithm to estimate both examination and trust bias from click data in TrustPBM. Furthermore, we show that it is difficult to use a pure IPS method to incorporate click noise and thus propose a novel method that combines a Bayes rule application with IPS for unbiased learning-to-rank. We evaluate our proposed methods on three personal search data sets and demonstrate that our proposed model can significantly outperform the existing unbiased learning-to-rank methods.

Semantic text matching is one of the most important research problems in many domains, including, but not limited to, information retrieval, question answering, and recommendation. Among the different types of semantic text matching, long-document-to-long-document text matching has many applications, but has rarely been studied. Most existing approaches for semantic text matching have limited success in this setting, due to their inability to capture and distill the main ideas and topics from long-form text.

How to optimize ranking metrics such as Normalized Discounted Cumulative Gain (NDCG) is an important but challenging problem, because ranking metrics are either flat or discontinuous everywhere, which makes them hard to be optimized directly. Among existing approaches, LambdaRank is a novel algorithm that incorporates ranking metrics into its learning procedure. Though empirically effective, it still lacks theoretical justification. For example, the underlying loss that LambdaRank optimizes for remains unknown until now. Due to this, there is no principled way to advance the LambdaRank algorithm further. In this paper, we present LambdaLoss, a probabilistic framework for ranking metric optimization. We show that LambdaRank is a special configuration with a well-defined loss in the LambdaLoss framework, and thus provide theoretical justification for it. More importantly, the LambdaLoss framework allows us to define metric-driven loss functions that have clear connection to different ranking metrics. We show a few cases in this paper and evaluate them on three publicly available data sets. Experimental results show that our metric-driven loss functions can significantly improve the state-of-the-art learning-to-rank algorithms.

Much work has been done recently on learning word embeddings from large corpora, which attempts to find the coordinates of words in a static and high dimensional semantic space. In reality, such corpora often span a sufficiently long time period, during which the meanings of many words may have changed. The co-evolution of word meanings may also result in a distortion of the semantic space, making these static embeddings unable to accurately represent the dynamics of semantics. In this paper, we present a novel computational method to capture such changes and to model the evolution of word semantics. Distinct from existing approaches that learn word embeddings independently from time periods and then align them, our method explicitly establishes the stable topological structure of word semantics and identifies the surprising changes in the semantic space over time through a principled statistical method. Empirical experiments on large-scale real-world corpora demonstrate the effectiveness of the proposed approach, which outperforms the state-of-the-art by a large margin.

How to improve the quality of conversations in online communities has attracted considerable attention recently. Having engaged, civil, and reactive online conversations has a critical effect on the social life of Internet users. In this study, we are particularly interested in identifying a post in a multi-party conversation that is unlikely to be further replied to, which therefore kills that thread of the conversation. For this purpose, we propose a deep learning model called the ConverNet. ConverNet is attractive due to its capability of modeling the internal structure of a long conversation and its appropriate encoding of the contextual information of the conversation, through effective integration of attention mechanisms. Empirical experiments on real-world datasets demonstrate the effectiveness of the proposed model. For the widely concerned topic, our analysis also offers implications for how to improve the quality and user experience of online conversations, or how to engage users in a conversation with a chatbot.

We consider the problem of discovering local events on the web, where events are entities extracted from webpages. Examples of such local events include small venue concerts, farmers markets, sports activities, etc. Given an event entity, we propose a graph-based framework for retrieving a ranked list of related events that a user is likely to be interested in attending. Due to the difficulty of obtaining ground-truth labels for event entities, which are temporal and are constrained by location, our retrieval framework is unsupervised, and its graph-based formulation addresses (a) the challenge of feature sparseness and noisiness, and (b) the semantic mismatch problem in a self-contained and principled manner.

We consider the task of identifying attitudes towards a given set of entities from text. Conventionally, this task is decomposed into two separate subtasks: target detection that identifies whether each entity is mentioned in the text, either explicitly or implicitly, and polarity classification that classifies the exact sentiment towards an identified entity (the target) into positive, negative, or neutral.

Information cascades, effectively facilitated by most social network platforms, are recognized as a major factor in almost every social success and disaster in these networks. Can cascades be predicted? While many believe that they are inherently unpredictable, recent work has shown that some key properties of information cascades, such as size, growth, and shape, can be predicted by a machine learning algorithm that combines many features. These predictors all depend on a bag of hand-crafting features to represent the cascade network and the global network structures. Such features, always carefully and sometimes mysteriously designed, are not easy to extend or to generalize to a different platform or domain.

Existing retrieval systems rely on a single active query to pull documents from the index. Relevance feedback may be used to iteratively refine the query, but only one query is active at a time. If the user's information need has multiple aspects, the query must represent the union of these aspects. We consider a new paradigm of retrieval where multiple queries are kept ``active'' simultaneously. In the presence of rate limits, the active queries take turns accessing the index to retrieve another ``page'' of results. Turns are assigned by a multi-armed bandit based on user feedback. This allows the system to explore which queries return more relevant results and to exploit the best ones. In empirical tests, query pools outperform solo, combined queries. Significant improvement is observed both when the subtopic queries are known in advance and when the queries are generated in a user-interactive process.

We present the problem of click-through prediction for advertising in Twitter timeline, which displays a stream of Tweets from accounts a user choose to follow. Traditional computational advertising usually appears in two forms: sponsored search that places ads onto the search result page when a query is issued to a search engine, and contextual advertising that places ads onto a regular, usually static Web page. Compared with these two paradigms, placing ads into a Tweet stream is particularly challenging given the nature of the data stream: the context into which an ad can be placed updates dynamically and never replicates. Every ad is therefore placed into a unique context. This makes the information available for training a machine learning model extremely sparse. In this study, we propose a learning-to-rank method which not only addresses the sparsity of training signals but also can be trained and updated online. The proposed method is evaluated using both offline experiments and online A/B tests, which involve very large collections of Twitter data and real Twitter users. Results of the experiments prove the effectiveness and efficiency of our solution, and its superiority over the current production model adopted by Twitter.

We consider a scenario where a searcher requires both high precision and high recall from an interactive retrieval process. Such scenarios are very common in real life, exemplified by medical search, legal search, market research, and literature review. When access to the entire data set is available, an active learning loop could be used to ask for additional relevance feedback labels in order to refine a classifier. When data is accessed via search services, however, only limited subsets of the corpus can be considered, subsets defined by queries. In that setting, relevance feedback has been used in a query enhancement loop that updates a query.