Fact-checked by Grok 2 weeks ago

Rocchio algorithm

The Rocchio algorithm is a for in systems that refines an initial user query by adjusting its vector representation in the , incorporating examples of relevant and non-relevant documents to enhance retrieval . Developed as part of the (Salton’s Magical Automatic Retriever of Text) system, it enables iterative query optimization through user interaction, where on retrieved documents guides the or contraction of query terms. Introduced by J.J. Rocchio in 1965, the algorithm emerged from early experiments in automatic document processing and has since become a foundational technique in , influencing modern search engines and text classification methods. It assumes documents and queries are represented as term-weight vectors, typically using term frequency-inverse document frequency (TF-IDF) weighting, to compute similarities via cosine measures. Early evaluations demonstrated significant performance gains, with average precision improvements of up to 20-30% after feedback iterations. The core of the algorithm lies in its query modification formula, which blends the original query with averaged vectors from relevant and non-relevant document sets:
\vec{q}_m = \alpha \vec{q}_0 + \beta \sum_{d_j \in D_r} \frac{\vec{d}_j}{|D_r|} - \gamma \sum_{d_j \in D_{nr}} \frac{\vec{d}_j}{|D_{nr}|}
where \vec{q}_0 is the initial query, D_r and D_{nr} are the sets of relevant and non-relevant documents, and \alpha, \beta, \gamma are tunable parameters (often set to 1, 0.75, and 0.15, respectively) that control the influence of each component; negative weights are typically clipped to zero to maintain non-negativity. This formulation effectively shifts the query centroid toward relevant documents while repelling it from non-relevant ones, optimizing for higher similarity scores with desired results.
Variants of the algorithm include blind feedback (using top-ranked documents as pseudo-relevant without user input) and ide dec-hi (focusing on a single high-ranked non-relevant document), which reduce user burden while preserving effectiveness in domains like web search and recommender systems. Despite its simplicity, the Rocchio algorithm remains influential, with extensions integrating it into probabilistic models and pipelines for tasks beyond traditional retrieval, such as and in large-scale corpora.

Overview

Definition and Purpose

The Rocchio algorithm is a classic method for in , designed to refine an initial user query by incorporating judgments on retrieved documents. It modifies the query representation to increase similarity with documents deemed relevant by the user while decreasing similarity with those judged non-relevant, thereby improving the overall effectiveness of the search process. This approach assumes binary relevance assessments from users, classifying documents simply as relevant or non-relevant without nuanced grading. The primary purpose of the Rocchio algorithm is to boost search through iterative , allowing the system to adapt more closely to the user's intent over multiple interactions. By leveraging (relevant documents) to emphasize useful terms and (non-relevant documents) to suppress misleading ones, it shifts the query in the toward clusters of pertinent content. This technique is particularly valuable in scenarios where initial queries are ambiguous or incomplete, enabling progressive refinement without requiring users to manually reformulate terms. For example, consider a user querying "," which could refer to the or the automobile brand, yielding mixed results. If the user marks documents about the animal as relevant and car-related ones as non-relevant, the algorithm adjusts the query by boosting weights for terms like "" or "predator" associated with the relevant set, while diminishing those linked to "" or "," thus yielding more targeted animal-focused results in subsequent searches.

Historical Development

The Rocchio algorithm originated in the mid-1960s as a method for in , developed by J.J. Rocchio Jr. during his work at Harvard University's Computation Laboratory. It was first detailed in Rocchio's 1965 technical report, which explored techniques to iteratively refine search queries based on user judgments of document relevance. This work laid the groundwork for automatic query modification, addressing limitations in early systems where initial searches often failed to capture precisely. By 1971, the algorithm had been integrated into the (Salton's Magical Automatic Retriever of Text) system at , under the leadership of Gerard Salton. Rocchio's contribution appeared as a in Salton's edited volume on SMART experiments, formalizing the vector-based approach that became central to the system's design. Initial evaluations were conducted on benchmark collections such as the (aeronautics documents) and MEDLARS (medical abstracts), demonstrating improvements in retrieval through using relevant and non-relevant examples. These tests highlighted the algorithm's potential in handling small-to-medium corpora typical of the era's computational constraints. The algorithm emerged amid the 1960s-1970s surge in IR research, influenced by the burgeoning pioneered by Salton, which represented documents and queries as vectors in a high-dimensional space. This period saw a shift from rigid retrieval to probabilistic and iterative methods, with Rocchio's technique advancing automatic as a core innovation. During the 1980s and 1990s, it gained widespread adoption in academic and experimental IR systems, often combined with term weighting schemes like tf-idf to enhance performance on larger datasets, as evidenced in evaluations using TREC benchmarks. In the 2000s, adaptations of the Rocchio algorithm extended to web-scale search environments, incorporating user profiles and query logs for personalized retrieval. For instance, variants were applied in reinforcement learning frameworks for web document filtering, adjusting queries dynamically to reflect browsing behavior and improving relevance in hyperlinked collections. This evolution underscored its enduring influence, transitioning from batch-oriented academic tools to components of interactive, large-scale search engines.

Theoretical Foundations

Vector Space Model

The (VSM) represents documents and user queries as vectors in a high-dimensional space, where each dimension corresponds to a distinct term from the vocabulary of the document collection. This geometric framework allows for the algebraic manipulation of text, treating retrieval as a problem of finding vectors closest to the query vector based on their spatial proximity. Introduced by Salton and colleagues, the VSM provides a foundational method for automatic indexing and search in systems. Vector components are computed using term weighting schemes to capture the importance of each term. A common approach is the tf-idf (term frequency-inverse document frequency) weighting, which balances local term prominence with global rarity. Term frequency (tf) measures how often a term t appears in a specific document d, while inverse document frequency (idf) downweights terms that occur frequently across the entire collection. The tf-idf weight for term t in document d is calculated as: w_{t,d} = \mathrm{tf}(t,d) \times \log \left( \frac{N}{\mathrm{df}(t)} \right) where N is the total number of documents and \mathrm{df}(t) is the document frequency of term t. This scheme, with idf originating from Spärck Jones' work on term specificity, assigns higher values to terms that are both frequent in the document and rare in the collection, thereby enhancing retrieval discrimination. Relevance between a query vector \mathbf{q} and a document vector \mathbf{d} is quantified using , which measures the angle between the vectors and normalizes for document length: \cos \theta = \frac{\mathbf{q} \cdot \mathbf{d}}{||\mathbf{q}|| \ ||\mathbf{d}||} Here, \mathbf{q} \cdot \mathbf{d} denotes the , and ||\cdot|| the Euclidean norm. Documents are ranked by descending cosine values, with higher scores indicating greater similarity. This measure is particularly effective in high-dimensional spaces as it focuses on directional alignment rather than magnitude. The VSM relies on key assumptions to simplify text representation: terms are treated as orthogonal, implying statistical independence between different terms, which facilitates vector operations but overlooks semantic relationships like synonymy. Additionally, it adopts a bag-of-words approach, representing documents as unordered collections of terms and ignoring syntactic structure, word order, or proximity, which prioritizes content over form for efficient retrieval.

Relevance Feedback Concept

Relevance feedback is a technique in systems where users evaluate the initial set of retrieved documents by marking them as relevant or non-relevant, enabling the system to iteratively refine the search query for better alignment with the user's . This process incorporates user judgments to adjust query parameters, such as expanding or reweighting terms, thereby enhancing the of subsequent retrievals. There are two primary types of relevance feedback: explicit and implicit. Explicit feedback involves direct user input, where individuals manually indicate document relevance, often in binary terms (relevant or non-relevant) or graded scales (e.g., somewhat relevant, relevant, highly relevant). Implicit feedback, in contrast, infers from observable user behaviors, such as click-through rates, on documents, or save actions, without requiring explicit judgments. The benefits of relevance feedback include addressing vocabulary mismatches between queries and documents, such as synonyms, , or related terms not captured in the original query, which improves overall retrieval effectiveness. It also supports iterative search processes, allowing systems to adapt dynamically to user needs and reduce the effort required for satisfactory results across multiple rounds of interaction. Historically, emerged in the 1960s as part of early interactive experiments, drawing from cybernetic principles of adaptive systems, and was formalized within the field through systems like developed by Gerard Salton and colleagues. It played a pivotal role in advancing retrieval performance in test collections such as (from the late 1950s) and became a cornerstone of model-based approaches by the 1970s.

Algorithm Description

Mathematical Formulation

The Rocchio algorithm operates within the , where the original query is represented as a \vec{q} and documents as \vec{d}, with components corresponding to term weights such as tf-idf scores. The algorithm modifies this query using feedback from a set of relevant documents R and non-relevant documents D. The core of the algorithm is the updated query vector \vec{q}', formulated as a linear combination that incorporates the original query and the average vectors of the relevant and non-relevant sets: \vec{q}' = \alpha \vec{q} + \beta \left( \frac{1}{|R|} \sum_{\vec{d} \in R} \vec{d} \right) - \gamma \left( \frac{1}{|D|} \sum_{\vec{d} \in D} \vec{d} \right) Here, \alpha, \beta, and \gamma are positive weighting parameters that balance the contributions of the original query, the centroid of relevant documents, and the centroid of non-relevant documents, respectively. Typical values are \alpha = 1, \beta = 0.75, and \gamma = 0.15, which emphasize a positive bias toward relevant feedback while moderately penalizing non-relevant information. This formulation intuitively derives from optimizing query-document similarity in : the term with \beta pulls the query toward the average of relevant vectors (their ), enhancing alignment with desired documents, while the subtracted term with \gamma pushes it away from the average of non-relevant vectors. If any component of \vec{q}' becomes negative—indicating a term more prominent in non-relevant documents than in the query or relevant set—that weight is set to zero, as negative term weights lack interpretive value in the model.

Implementation Steps

The implementation of the Rocchio algorithm proceeds through a series of practical steps that integrate user feedback into the initial query within a framework. These steps transform the original query vector into a refined version that better aligns with user judgments of relevance, enabling improved retrieval performance. The process begins with retrieving an initial set of documents using the original query vector q. In a typical system, this involves computing the similarity (e.g., ) between q and the document vectors in the , then ranking and presenting the top-k results to the user. Next, obtain user feedback by identifying subsets of relevant documents R and non-relevant documents D from the initial results. Users manually select or mark these subsets, often examining a small number (e.g., 5–20) of top-ranked documents to provide judgments; at least a few relevant documents are recommended for stable updates. If R or D is empty, the algorithm may skip the corresponding adjustment to avoid instability. Then, compute the average vectors, or centroids, for the relevant and non-relevant sets. The relevant centroid is given by \frac{1}{|R|} \sum_{d \in R} d, where each d is a document vector, and similarly the non-relevant centroid is \frac{1}{|D|} \sum_{d \in D} d. These averages are calculated term-wise across the vector dimensions, typically using term frequency-inverse document frequency (tf-idf) weights. For empty sets, the centroid is set to the zero vector. Apply the Rocchio formula to obtain the updated query q' = \alpha q + \beta \left( \frac{1}{|R|} \sum_{d \in R} d \right) - \gamma \left( \frac{1}{|D|} \sum_{d \in D} d \right), where \alpha, \beta, and \gamma are tunable weights (commonly \alpha = 1, \beta = 0.75, \gamma = 0.15) that balance the original query and feedback. Negative weights in q' are set to zero, as they do not contribute positively to relevance. Normalization of q' to unit length may be applied if the retrieval system uses normalized cosine similarity, ensuring consistent scaling. This step draws directly from the mathematical formulation of the algorithm. Finally, re-execute the search using q' against the to retrieve and present a new ranked list of documents, which should more closely match due to the incorporation. This iterative process can be repeated with additional rounds if needed. The following outlines the core implementation in a procedural form, assuming operations are supported and documents are pre-indexed in tf-idf :
function RocchioUpdate(original_query q, relevant_docs R, nonrelevant_docs D, weights α, β, γ):
    if |R| == 0 and |D| == 0:
        return q  // No feedback; return original query
    
    // Initialize centroids as zero vectors ([dimension](/page/Dimension) = [vocabulary](/page/Vocabulary) size)
    relevant_centroid = zero_vector()
    nonrelevant_centroid = zero_vector()
    
    // Compute relevant centroid
    if |R| > 0:
        for each document d in R:
            relevant_centroid += d
        relevant_centroid /= |R|
    
    // Compute non-relevant centroid
    if |D| > 0:
        for each document d in D:
            nonrelevant_centroid += d
        nonrelevant_centroid /= |D|
    
    // Apply Rocchio formula term-wise
    updated_query = α * q + β * relevant_centroid - γ * nonrelevant_centroid
    
    // Clip negative weights to zero
    for each term i in updated_query:
        if updated_query[i] < 0:
            updated_query[i] = 0
    
    // Optional normalization (for unit length in cosine similarity)
    // norm = sqrt(sum over i of (updated_query[i])^2)
    // if norm > 0:
    //     updated_query /= norm
    
    return updated_query

// Usage in retrieval loop
initial_results = retrieve_documents(corpus, q, top_k=20)
feedback_R, feedback_D = get_user_feedback(initial_results)
q_new = RocchioUpdate(q, feedback_R, feedback_D, 1.0, 0.75, 0.15)
refined_results = retrieve_documents(corpus, q_new, top_k=20)
present_results(refined_results)
This pseudocode handles empty sets by defaulting to the original query or omitting the adjustment, includes loops for vector averaging, and applies weight clipping for robustness.

Performance Analysis

Computational Complexity

The computational complexity of the Rocchio algorithm is primarily determined by the operations involved in updating the query vector based on within the . The core step of averaging the term-weight vectors for relevant and non-relevant documents incurs a of O(|V| \cdot |R| + |V| \cdot |D|), where |V| is the size of the vocabulary, |R| is the number of relevant feedback documents, and |D| is the number of non-relevant feedback documents; this accounts for summing and the components across all dimensions in a dense vector representation. In the context of an , where an initial set of n documents is retrieved and ranked (typically via , which is O(n \cdot |V|) assuming pre-computed vectors), the feedback update adds an overhead linear in the feedback set size, resulting in an overall per- of O(n \cdot |V|). Practical implementations benefit from the sparsity of document vectors in text corpora, where most terms have zero weights, reducing the effective time to O(\sum_{d \in R \cup D} |d|), the total number of non-zero terms across the feedback documents, rather than the full dense cost. This efficiency assumes that term frequency-inverse document frequency (tf-idf) weights are pre-computed and stored in an , avoiding redundant feature extraction during feedback; without such preprocessing, an additional O(|R| \cdot L + |D| \cdot L) cost arises for tokenization and weighting, where L is the average length in tokens. Regarding space complexity, the algorithm requires O(|V|) storage for the original query and the updated centroid-like representation, as these are typically maintained as sparse s with non-zero entries far fewer than |V|. If document s are not pre-indexed, temporary space of O(|R| \cdot |V| + |D| \cdot |V|) is needed in the worst case, though sparsity again lowers this to O(\sum_{d \in R \cup D} |d|); in standard retrieval systems, pre-existing indexes mitigate this entirely, confining auxiliary space to the subset selection. Compared to naive vector space search without feedback, which operates in O(n \cdot |V|) for retrieval alone, the Rocchio method introduces only linear additional overhead proportional to the feedback set size, making it computationally while enabling iterative improvements in query .

Theoretical Properties

The Rocchio algorithm relies on several key assumptions rooted in the of . It presupposes linear separability of classes in the high-dimensional , where documents and queries are represented as weighted vectors, allowing a to effectively distinguish relevant from non-relevant items. Additionally, the algorithm assumes the effectiveness of tf-idf in capturing discriminativeness, which aligns with probabilistic interpretations of importance in text representations. A further assumption is that small feedback sets—typically a handful of relevant and non-relevant examples—suffice to refine prototype vectors or queries, leveraging the redundancy in textual data to achieve meaningful adjustments without exhaustive labeling. Theoretical properties of the Rocchio algorithm include potential monotonic improvement in retrieval or performance under ideal conditions, such as increasing examples within the bag-of-words framework, where accuracy rises predictably due to better prototype estimation. The algorithm's behavior is highly sensitive to the parameters α, β, and γ, which weight the original , , and , respectively; for example, configuring β > γ imparts a that prioritizes toward relevant documents over repulsion from non-relevant ones. In cases of linear separability, its Perceptron-like updates ensure to a stable separating , though this assumes bounded to prevent . A notable probabilistic analysis by Joachims ties the to tf-idf weighting in text categorization, reformulating it as an approximation to Bayesian posterior estimation under word and bag-of-words assumptions. This derivation, yielding a probabilistic variant (PrTFIDF), elucidates the heuristic's strengths by linking term weights to conditional probabilities, with performance bounds emerging under specific distributional assumptions like multinomial word occurrences. However, iterative applications risk over-expansion, where repeated feedback may inflate the query vector excessively, incorporating noise and destabilizing the representation if negative examples are underrepresented.

Applications

In Information Retrieval

The Rocchio algorithm serves as a foundational method for explicit in (), enabling users to refine search queries by marking retrieved documents as relevant or non-relevant, thereby improving subsequent retrievals. Developed within the system, it was designed to address vocabulary mismatches and expand queries with terms from user-identified relevant documents, enhancing both in early automated library search environments. Similarly, the system incorporated Rocchio-inspired mechanisms alongside its BM25 weighting scheme to boost retrieval effectiveness in probabilistic setups during the late 20th century. In practical integration, the algorithm operates within a post-retrieval loop: an initial query retrieves a set of documents, users provide judgments (typically on 10–20 items), and the system iteratively updates the query using weighted contributions from relevant and non-relevant examples to re-rank or expand the search. This process proved effective in evaluations like the Text REtrieval Conference (TREC) during the 1990s, where Rocchio-based applied to systems such as / yielded notable improvements; for instance, in TREC-7 (1998), it enhanced baseline performance by up to 12% on ad-hoc tasks by better capturing query through reweighting. Such gains were particularly evident in domains with sparse or ambiguous queries, demonstrating the algorithm's role in bridging the gap between initial and optimal retrieval sets. Modern adaptations of the Rocchio algorithm persist in enterprise search platforms, where explicit feedback is implemented via plugins or custom handlers in tools like Apache Solr and Lucene to manage web-scale data efficiently. For example, Solr's "More Like This" component draws on Rocchio principles for query expansion, while extensions like relevancy feedback plugins allow sampling from large result sets (e.g., top-k documents) to mitigate computational overhead in distributed environments, maintaining relevance in high-volume corporate or e-commerce searches. These implementations often combine Rocchio with inverted indexes for scalability, ensuring feedback loops remain viable without full re-indexing. A notable case study of its application appears in medical IR, where the algorithm facilitates precise term expansion for specialized queries in systems like clinical decision support tools. In one , an adaptive proximity-based Rocchio model was applied to biomedical , incorporating pseudo-relevant feedback to refine queries for electronic health records and literature databases; this approach improved mean average precision by 8.5% over PRoc2, 12.24% over TF-PRF, and up to 32.8% over BM25+Rocchio baselines in tasks involving rare medical terms, such as disease-symptom associations, by emphasizing term co-occurrences and user-validated expansions akin to those in historical systems like precursors. Recent applications as of 2022 include enhancements in pseudo-relevance feedback, such as proximity-based models combined with BM25 for improved ranking in clinical decision support.

In Text Classification

The Rocchio algorithm, originally developed for relevance feedback in information retrieval, has been adapted for text classification by treating each category as a "pseudo-query" represented by a centroid vector. In this approach, a centroid for a class c is computed as the average of the term-frequency inverse-document-frequency (TF-IDF) weighted vectors of all documents assigned to that class during training: \vec{\mu}_c = \frac{1}{N_c} \sum_{d \in D_c} \vec{v}_d, where N_c is the number of documents in class c and \vec{v}_d is the vector for document d. A new document is then classified to the category whose centroid it is most similar to, typically measured by cosine similarity, enabling effective handling of both binary and multi-class problems in sparse high-dimensional spaces. Hybrid models incorporating the Rocchio algorithm with other techniques have shown improved performance in text categorization tasks. For instance, a 2017 approach combines Rocchio's vector-based with 's ensemble decision trees, where Rocchio generates initial category prototypes and refines predictions through bagging and feature randomness, achieving better F1-scores and lower Hamming loss compared to standalone methods on benchmark datasets. Such integrations leverage Rocchio's simplicity for prototype formation while addressing its limitations in handling complex decision boundaries via more robust learners. The algorithm performs well on sparse datasets like news articles, as demonstrated in evaluations on the Reuters-21578 corpus, where it achieved accuracies around 90% for categories such as "acq" (88.9%), "earn" (90.5%), and "crude" (90.2%), often comparable to or outperforming Naive Bayes in small sets due to its probabilistic TF-IDF variant. Unlike its iterative use in for , text employs a static phase to build fixed centroids, with determined by similarity thresholds or the maximum similarity score across classes, making it suitable for supervised, non-interactive scenarios.

Limitations and Extensions

Key Limitations

The Rocchio algorithm relies on explicit from users, who must judge and mark retrieved documents as relevant or irrelevant, which imposes a significant burden and limits its practicality in large-scale or automated systems. Users are often reluctant to provide such feedback, as it prolongs and requires additional effort beyond the initial query. This dependency on input restricts , particularly in environments where or without user involvement is preferred, leading to the development of pseudo-relevance feedback as an alternative. Furthermore, when the number of relevant documents (|R|) is small relative to the total document collection (|D|), the feedback introduces noise, degrading query reformulation quality and resulting in poor retrieval performance, especially in sparse relevance scenarios common to broad corpora. The algorithm's effectiveness is highly sensitive to the choice of weighting parameters α, β, and γ, which control the influence of the original query and the relevant/irrelevant document sets in the updated query vector. Poorly tuned values can lead to query drift, where the reformulated query shifts away from the user's intent due to overemphasis on feedback terms, or over-generalization, diluting the query's specificity and retrieving irrelevant results. Experimental tuning is typically required, with common settings like α=1, β=0.75, and γ=0.15. As a method rooted in the , the Rocchio algorithm treats documents and queries as bags of independent terms, ignoring semantic relationships such as synonyms, , or , which limits its ability to capture nuanced meaning. The single centroid prototype assumes spherical clusters, which does not handle contextual shifts effectively. This term independence also causes issues with , where negative weights are simply clipped to zero, preventing effective exclusion of irrelevant concepts. Additionally, the algorithm performs poorly on short queries, which produce sparse vectors lacking sufficient context for meaningful feedback incorporation. Empirically, the Rocchio algorithm has shown in modern tasks, particularly with web-scale data post-2000, where low proportions of relevant documents exacerbate noise and reduce effectiveness compared to more advanced models. Its bag-of-words foundation struggles with the semantic complexity of contemporary corpora, leading to suboptimal performance against methods incorporating latent semantics or neural embeddings, as evidenced by evaluations on diverse datasets like 20Newsgroups where overlapping vocabularies cause high misclassification rates.

Variants and Improvements

Pseudo-relevance feedback automates the relevance judgment process in Rocchio by assuming the top-k retrieved documents from an initial query are relevant, thus generating positive feedback without user intervention. This variant, integrated into systems like early web search engines, expands the query by averaging term weights from these pseudo-relevant documents while downweighting non-relevant ones. Seminal implementations, such as those in the , showed consistent gains in for ad hoc retrieval tasks by iteratively refining queries in this manner. Modern applications in search engines continue to employ pseudo-relevance feedback with Rocchio for efficient, user-independent . Semantic variants of the Rocchio algorithm enhance its foundation by incorporating latent factors or dense representations to capture term relationships beyond exact matches. Integration with Latent Semantic Indexing (LSI) projects queries and documents into a reduced-dimensional space via , allowing Rocchio to operate on semantically related terms and improve handling of synonymy and . This approach, formalized in the late , yielded notable improvements in test collections by aligning with underlying topic structures. Post-2010 extensions leverage word embeddings, such as those from or , to represent terms in continuous vector spaces, enabling Rocchio-style averaging over embedding neighborhoods for that preserves semantic proximity. For instance, embedding-based Rocchio has been applied to math retrieval, where weighted centroids of relevant embeddings refine searches more effectively than bag-of-words variants. Recent advancements, such as the 2024 MSRoc framework, extend Rocchio by incorporating BERT-based multidimensional semantic information for pseudo-relevance , achieving notable improvements on benchmarks like TREC collections (e.g., 14.69% MAP gain on AP90). Hybrid approaches combine Rocchio with classifiers to bolster its discriminative power, particularly in tasks. Joachims (1997) provided a probabilistic reinterpretation of Rocchio using TF-IDF weighting, framing it as a naive Bayes approximation for text categorization and achieving competitive accuracy on benchmark datasets like Reuters-21578. Later hybrids Rocchio prototypes with support vector machines (SVMs), using Rocchio for initial feature weighting and SVM for boundary refinement, which improved F1-scores in multi-class settings by 5-10% over standalone Rocchio. In the 2020s, such hybrids have extended to multilingual , adapting Rocchio for cross-lingual feedback by aligning embeddings across languages, as seen in systems processing English, , and texts with enhanced cross-lingual precision.

References

  1. [1]
    [PDF] sigir - xxiii. relevance feedback in information retrieval
    Relevance feedback involves user interaction with the system, allowing communication after each retrieval operation to improve search requests.
  2. [2]
    [PDF] Relevance feedback and query expansion - Stanford NLP Group
    The Rocchio Algorithm is the classic algorithm for implementing relevance feedback. It models a way of incorporating relevance feedback information into the ...
  3. [3]
    [PDF] Refining and Personalizing Searches - Princeton CS
    – jaguar car versus jaguar animal. • Can modify rankings for future searches. • Algorithms usually based on machine learning. – Learn ranking function that ...
  4. [4]
    The SMART Retrieval System: Experiments in Automatic Document ...
    Compilation of papers on technical aspects of fully automatic computer-based information retrieval systems, with particular reference to the experimental smart ...
  5. [5]
    [PDF] The History of Information Retrieval Research - Publication
    Abstract. This paper describes a brief history of the research and development of information retrieval systems starting.
  6. [6]
    [PDF] The State of Retrieval System Evaluation - Cornell eCommons
    [19-21]. Although the Smart and Cranfield test results were complementary, the conclusions were not accepted by everyone. The main objection was related to the ...
  7. [7]
    [PDF] Personalized Web-Document Filtering Using Reinforcement Learning
    A classical approach is the Rocchio method [Rocchio, 1971] and its variants. This is a batch algorithm that modifies the original query vector by the vectors of ...<|control11|><|separator|>
  8. [8]
    A vector space model for automatic indexing - ACM Digital Library
    Nov 1, 1975 · An approach based on space density computations is used to choose an optimum indexing vocabulary for a collection of documents.
  9. [9]
    [PDF] A statistical interpretation of term specificity and its application in ...
    It is suggested that specificity should be interpreted statistically, as a function of term use rather than of term meaning. The effects on retrieval of.
  10. [10]
    [PDF] Introduction to Information Retrieval - Stanford University
    Aug 1, 2006 · ... Rocchio algorithm for relevance feedback. 178. 9.1.2. Probabilistic ... SMART notation for tf-idf variants. 128. 6.16. Pivoted document ...Missing: Gerald | Show results with:Gerald
  11. [11]
    Using graded relevance assessments in IR evaluation
    To gain insight into the retrieval process, one should use both graded relevance assessments and effectiveness measures that enable one to observe the ...
  12. [12]
    Relevance feedback revisited - ACM Digital Library
    Relevance feedback is effective in interactive information retrieval, involving query expansion and reweighting, with multiple iterations improving performance.
  13. [13]
    The #rocchio71### algorithm. - Stanford NLP Group
    This was the relevance feedback mechanism introduced in and popularized by Salton's SMART system around 1970. In a real IR query context, we have a user ...Missing: origin Gerald Stephen
  14. [14]
    [PDF] Rocchio's Algorithm
    Rocchio's algorithm. • Relevance Feedback in Information Retrieval, SMART Retrieval System. Experiments in Automatic Document Processing, 1971, Prentice Hall.Missing: original | Show results with:original
  15. [15]
    Rocchio classification - Stanford NLP Group
    The best-known way of computing good class boundaries is Rocchio classification, which uses centroids to define the boundaries.Missing: seminal | Show results with:seminal
  16. [16]
    [PDF] A study on optimal parameter tuning for Rocchio Text Classifier
    3.3 P RC complexity. The evaluation of Rocchio classifier time complexity can be divided in to three steps: pre-processing, learning and classification. The ...
  17. [17]
    [PDF] Extending the Rocchio Relevance Feedback Algorithm to Provide ...
    Using characteristics about the user or specific feedback from them to modify the search query provides a context for those queries which can help remove ...
  18. [18]
    [PDF] 14 Vector space classification - Stanford NLP Group
    In fact, if linear separability holds,. LINEAR SEPARABILITY then there is an infinite number of linear separators (Exercise 14.4) as illus- trated by Figure ...
  19. [19]
    [PDF] A Probabilistic Analysis of the Rocchio Algorithm with TFIDF for Text ...
    Originally developed for information retrieval, the algorithm returns a ranking of documents without providing a threshold to de ne a decision rule for class ...Missing: original | Show results with:original
  20. [20]
    [PDF] Relevance Feedback for Best Match Term Weighting Algorithms in ...
    [15] J.J. Rocchio, Relevance feedback in information retrieval, In: G. Salton (ed.), The Smart retrieval system: experiments in automatic document processing ...
  21. [21]
    Improved biomedical term selection in pseudo relevance feedback
    Jul 2, 2018 · Z. (. 2012. ) Proximity-based rocchio's model for pseudo relevance. ... Apache Solr Lucene. http://lucene.apache.org/solr/ (4 April 2018 ...
  22. [22]
    An adaptive term proximity based rocchio's model for clinical ...
    Dec 12, 2019 · The Rocchio's model is a classical framework to realize pseudo relevance feedback representation, which incorporates the information of pseudo ...Missing: Medlars | Show results with:Medlars
  23. [23]
    Text categorization using Rocchio algorithm and random forest ...
    This research paper proposes a hybrid text categorization model that combines both Rocchio algorithm and Random Forest algorithm to perform Multi-label text ...Missing: classification | Show results with:classification
  24. [24]
    [PDF] A Hybrid Relevance-Feedback Approach to Text Retrieval
    A recent study [3] revealed that. Rocchio's algorithm has a poor performance when the proportion of relevant documents in the whole corpus is quite low. 2.2 ...
  25. [25]
    [PDF] Utilizing User-input Contextual Terms for Query Disambiguation
    Query drift occurs when the feedback documents contain few or no relevant ones. In this case, search results will be routed farther away from the search intent,.
  26. [26]
    [PDF] ReQ-ReC: High Recall Retrieval with Query Pooling and Interactive ...
    Note that in Rocchio updating, the parameter that inter- polates the new query vector with the original query is quite sensitive. This is because when one ...
  27. [27]
    [PDF] Towards a Supervised Rocchio-based Semantic Classification of ...
    Results illustrate some limitations mainly related to ignoring text semantics. In order to overcome these limitations, this work proposes to extend the original ...
  28. [28]
    A multi-dimensional semantic pseudo-relevance feedback ... - Nature
    Dec 30, 2024 · Pseudo-relevance feedback (PRF) is an essential technical branch of IR that enhances the efficiency of the retrieval system, producing search ...
  29. [29]
    Math-word embedding in math search and semantic extraction
    Jun 9, 2020 · Although both the Rocchio algorithm and our weighted centroid are vector based, they differ in several respects. Our approach is embedding based ...
  30. [30]
    Text Categorization Using SVMs with Rocchio Ensemble for Internet ...
    In this paper, a novel text categorization method based on multi-class Support Vector Machines (SVMs) with Rocchio ensemble is proposed for Internet information ...<|separator|>
  31. [31]
    Classifying Multilingual Text in Quality Assurance
    Apr 5, 2021 · Rocchio method is based on the nearest centroid that calculates the similarity between test documents and each of the prototype vector through ...