We present RML, the first known general reinforcement learning framework for relevance feedback that directly optimizes any desired retrieval metric, including precision-oriented, recall-oriented, and even diversity metrics: RML can be easily extended to optimize arbitrary user satisfaction signal. Using framework, we select effective terms weight them appropriately, improving on past methods fit parameters algorithms using heuristic approaches or do not performance. Learning an model is...

Pseudo-relevance feedback (PRF) has been proven to be an effective query expansion strategy improve retrieval performance. Several PRF methods have so far proposed for many models. Recent theoretical studies of show that most the do not satisfy all necessary constraints. Among all, log-logistic model shown method satisfies In this paper, we first introduce two new We further analyze and it does these constraints as well previously "relevance effect" constraint. then modify formulation...

Pseudo-relevance feedback (PRF) refers to a query expansion strategy based on top-retrieved documents, which has been shown be highly effective in many retrieval models. Previous work introduced set of constraints (axioms) that should satisfied by any PRF model. In this paper, we propose three additional the proximity terms documents. As case study, consider log-logistic model, state-of-the-art model proven successful method satisfying existing constraints, and show it does not satisfy...

Number of terms in a query is query-specific constant that typically ignored retrieval functions. However, previous studies have shown the performance models varies for different lengths, and it usually degrades when length increases. A possible reason this issue can be extraneous longer queries makes challenge to distinguish between key complementary concepts query. As signal understand importance term, inverse document frequency (IDF) used discriminate terms. In paper, we propose...

Several neural networks have been developed for end-to-end training of information retrieval models. These differ in many aspects including architecture, data, data representations, and loss functions. However, only pointwise pairwise functions are employed ranking models without human-engineered features. do not consider the ranks documents estimation over data. Because this limitation, conventional learning-to-rank using or generally shown lower performance compared to those listwise...

Axiomatic analysis is a well-defined theoretical framework for analytical evaluation of information retrieval models. The current studies in axiomatic implicitly assume that the constraints (axioms) are independent. In this paper, we revisit assumption and hypothesize there might be interdependence relationships between existing constraints. As preliminary study, focus on pseudo-relevance feedback (PRF) models have been theoretically studied using approach. introduce two novel interdependent...

We propose AC-CRS, a novel conversational recommendation system based on reinforcement learning that better models user interaction compared to prior work. Interactive recommender systems expect an initial request from and then iterate by asking questions or recommending potential matching items, continuing until some stopping criterion is achieved. Unlike most existing works stop as soon item recommended, we model the more realistic expectation will continue if not appropriate. Using this...

Conversational search and recommendation systems can ask clarifying questions through the conversation collect valuable information from users. However, an important question remains: how we extract relevant user's utterances use it in retrieval or next turn of conversation? Utilizing users' leads system to better results at end conversation. In this paper, propose a model based on reinforcement learning, namely RelInCo, which takes context classifies each word as belonging non-relevant...

Search result diversification based on topic proportionality considers a document as bag of weighted topics and aims to reorder or down-sample ranked list in way that maintains proportionality. The goal is show the distribution from an ambiguous query at all points revised list, hoping satisfy users expectation. One effective approach, PM-2, greedily selects best each ranking position then represents topic. From theoretical perspective, this approach does not provide any guarantee holds...

We propose RelQuest, a conversational product search model based on reinforcement learning to generate questions from descriptions in each round of the conversation, directly maximizing any desired metrics (i.e., ultimate goal conversation), objectives, or even an arbitrary user satisfaction signal. By enabling systems ask about needs, has gained increasing attention recent years. Asking right through conversations helps system collect valuable feedback create better experiences and...

Term discrimination value is among the three basic heuristics exploited, directly or indirectly, in almost all ranking models for ad-hoc Information Retrieval (IR). Query term monolingual IR usually estimated based on document collection frequency of terms. In query translation approach CLIR, a needs to be frequencies its translations, which more challenging. We show that existing estimation do not correctly estimate and adequately reflect difference between power terms, hurts retrieval...

The rise in popularity of mobile and voice search has led to a shift focus from document retrieval short answer passage for non-factoid questions. Some the questions have multiple answers, aim is retrieve set relevant passages, which covers all these alternatives. Compared documents, answers are more specific typically form defined types or groups. Grouping passages based on strong similarity measures may provide means identifying types. Typically, kNN clustering combination with term-based...

Conversational recommender systems (CRSs) are improving rapidly, according to the standard recommendation accuracy metrics. However, it is essential make sure that these robust in interacting with users including regular and malicious who want attack system by feeding modified input data. In this paper, we propose an adversarial evaluation scheme four scenarios two categories automatically generate examples evaluate robustness of face different By executing can compare ability conversational...

When information retrieval systems return a ranked list of results in response to query, they may be choosing from large set candidate that are equally useful and relevant. This means we might able identify difference between rankers A B, where ranker systematically prefers certain type relevant results. Ranker have this systematic (different "vibe") without having better or worse according standard metrics. We first show vibe can exist, comparing two publicly available rankers, the one is...