Searching strategies for the Hungarian language

This paper reports on the underlying IR problems encountered when dealing with the complex morphology and compound constructions found in the Hungarian language. It describes evaluations carried out on two general stemming strategies for this language, and also demonstrates that a light stemming approach could be quite effective. Based on searches done on the CLEF test collection, we find that a more aggressive suffix-stripping approach may produce better MAP. When compared to an IR scheme without stemming or one based on only a light stemmer, we find the differences to be statistically significant. When compared with probabilistic, vector-space and language models, we find that the Okapi model results in the best retrieval effectiveness. The resulting MAP is found to be about 35% better than the classical tf idf approach, particularly for very short requests. Finally, we demonstrate that applying an automatic decompounding procedure for both queries and documents significantly improves IR performance (+10%), compared to word-based indexing strategies.     

Efficient stemmer generation

This paper presents an algorithm for generating stemmers from text stemmer specification files. A small study shows that the generated stemmers are computationally efficient, often running faster than stemmers custom written to implement particular stemming algorithms. The stemmer specification files are easily written and modified by non-programmers, making it much easier to create a stemmer, or tune a stemmer's performance, than would be the case with a custom stemmer program. Stemmer generation is thus also human-resource efficient.     

Stemming to improve translation lexicon creation form bitexts

Arabic is a morphologically rich language that presents significant challenges to many natural language processing applications because a word often conveys complex meanings decomposable into several morphemes (i.e. prefix, stem, suffix). By segmenting words into morphemes, we could improve the performance of English/Arabic translation pair’s extraction from parallel texts. This paper describes two algorithms and their combination to automatically extract an English/Arabic bilingual dictionary from parallel texts that exist in the Internet archive after using an Arabic light stemmer as a preprocessing step. Before using the Arabic light stemmer, the total system precision and recall were 88.6% and 81.5% respectively, then the system precision an recall increased to 91.6% and 82.6% respectively after applying the Arabic light stemmer on the Arabic documents.     

Stemming和Lemmatization对英文文本聚类的影响研究

词干化、词形还原是英文文本处理中的一个重要步骤。本文利用3种聚类算法对两个Stemming算法和一个Lemmatization算法进行较为全面的实验。结果表明,Stemming和Lemmatization都可以提高英文文本聚类的聚类效果和效率,但对聚类结果的影响并不显著。相比于Snowball Stemmer和Stanford Lemmatizer,Porter Stemmer方法在Entropy和Pu-rity表现上更好,也更为稳定。     

Does dictionary based bilingual retrieval work in a non-normalized index?

Many operational IR indexes are non-normalized, i.e. no lemmatization or stemming techniques, etc. have been employed in indexing. This poses a challenge for dictionary-based cross-language retrieval (CLIR), because translations are mostly lemmas. In this study, we face the challenge of dictionary-based CLIR in a non-normalized index. We test two optional approaches: FCG (Frequent Case Generation) and s-gramming. The idea of FCG is to automatically generate the most frequent inflected forms for a given lemma. FCG has been tested in monolingual retrieval and has been shown to be a good method for inflected retrieval, especially for highly inflected languages. S-gramming is an approximate string matching technique (an extension of n-gramming). The language pairs in our tests were English–Finnish, English–Swedish, Swedish–Finnish and Finnish–Swedish. Both our approaches performed quite well, but the results varied depending on the language pair. S-gramming and FCG performed quite equally in all the other language pairs except Finnish–Swedish, where s-gramming outperformed FCG.     

Bias–variance analysis in estimating true query model for information retrieval

The estimation of query model is an important task in language modeling (LM) approaches to information retrieval (IR). The ideal estimation is expected to be not only effective in terms of high mean retrieval performance over all queries, but also stable in terms of low variance of retrieval performance across different queries. In practice, however, improving effectiveness can sacrifice stability, and vice versa. In this paper, we propose to study this tradeoff from a new perspective, i.e., the bias–variance tradeoff, which is a fundamental theory in statistics. We formulate the notion of bias–variance regarding retrieval performance and estimation quality of query models. We then investigate several estimated query models, by analyzing when and why the bias–variance tradeoff will occur, and how the bias and variance can be reduced simultaneously. A series of experiments on four TREC collections have been conducted to systematically evaluate our bias–variance analysis. Our approach and results will potentially form an analysis framework and a novel evaluation strategy for query language modeling.     

Empirical studies on the impact of lexical resources on CLIR performance

In this paper, we compile and review several experiments measuring cross-lingual information retrieval (CLIR) performance as a function of the following resources: bilingual term lists, parallel corpora, machine translation (MT), and stemmers. Our CLIR system uses a simple probabilistic language model; the studies used TREC test corpora over Chinese, Spanish and Arabic. Our findings include:

• •One can achieve an acceptable CLIR performance using only a bilingual term list (70–80% on Chinese and Arabic corpora).
• •However, if a bilingual term list and parallel corpora are available, CLIR performance can rival monolingual performance.
• •If no parallel corpus is available, pseudo-parallel texts produced by an MT system can partially overcome the lack of parallel text.
• •While stemming is useful normally, with a very large parallel corpus for Arabic–English, stemming hurt performance in our empirical studies with Arabic, a highly inflected language.

     

HPS: High precision stemmer

Research into unsupervised ways of stemming has resulted, in the past few years, in the development of methods that are reliable and perform well. Our approach further shifts the boundaries of the state of the art by providing more accurate stemming results. The idea of the approach consists in building a stemmer in two stages. In the first stage, a stemming algorithm based upon clustering, which exploits the lexical and semantic information of words, is used to prepare large-scale training data for the second-stage algorithm. The second-stage algorithm uses a maximum entropy classifier. The stemming-specific features help the classifier decide when and how to stem a particular word.     

Character contiguity in N-gram-based word matching: the case for Arabic text searching

This work assesses the performance of two N-gram matching techniques for Arabic root-driven string searching: contiguous N-grams and hybrid N-grams, combining contiguous and non-contiguous. The two techniques were tested using three experiments involving different levels of textual word stemming, a textual corpus containing about 25 thousand words (with a total size of about 160KB), and a set of 100 query textual words. The results of the hybrid approach showed significant performance improvement over the conventional contiguous approach, especially in the cases where stemming was used. The present results and the inconsistent findings of previous studies raise some questions regarding the efficiency of pure conventional N-gram matching and the ways in which it should be used in languages other than English.     

Towards understanding and mitigating unintended biases in language model-driven conversational recommendation

Conversational Recommendation Systems (CRSs) have recently started to leverage pretrained language models (LM) such as BERT for their ability to semantically interpret a wide range of preference statement variations. However, pretrained LMs are prone to intrinsic biases in their training data, which may be exacerbated by biases embedded in domain-specific language data (e.g., user reviews) used to fine-tune LMs for CRSs. We study a simple LM-driven recommendation backbone (termed LMRec) of a CRS to investigate how unintended bias — i.e., bias due to language variations such as name references or indirect indicators of sexual orientation or location that should not affect recommendations — manifests in substantially shifted price and category distributions of restaurant recommendations. For example, offhand mention of names associated with the black community substantially lowers the price distribution of recommended restaurants, while offhand mentions of common male-associated names lead to an increase in recommended alcohol-serving establishments. While these results raise red flags regarding a range of previously undocumented unintended biases that can occur in LM-driven CRSs, there is fortunately a silver lining: we show that train side masking and test side neutralization of non-preferential entities nullifies the observed biases without significantly impacting recommendation performance.     

Leveraging relevance cues for language modeling in speech recognition

Language modeling (LM), providing a principled mechanism to associate quantitative scores to sequences of words or tokens, has long been an interesting yet challenging problem in the field of speech and language processing. The n-gram model is still the predominant method, while a number of disparate LM methods, exploring either lexical co-occurrence or topic cues, have been developed to complement the n-gram model with some success. In this paper, we explore a novel language modeling framework built on top of the notion of relevance for speech recognition, where the relationship between a search history and the word being predicted is discovered through different granularities of semantic context for relevance modeling. Empirical experiments on a large vocabulary continuous speech recognition (LVCSR) task seem to demonstrate that the various language models deduced from our framework are very comparable to existing language models both in terms of perplexity and recognition error rate reductions.     

A second look at Egghe’s universal IR surface and a simple derivation of a complete set of universal IR evaluation points

Egghe’s three papers regarding the universal IR surface (2004, 2007, 2008) clearly represent an original and significant contribution to the IR evaluation literature. However, Egghe’s attempt to find a complete set of universal IR evaluation points (P,R,F,M) fell short of his goal: his universal IR surface equation did not suffice in and of itself, and his continuous extension argument was insufficient to find all the remaining points (quadruples). Egghe found only two extra universal IR evaluation points, (1,1,0,0) and (0,0,1,1), but it turns out that a total of 15 additional, valid, universal IR evaluation points exist. The gap first appeared in Egghe’s earliest paper and was carried into subsequent papers. The mathematical method used here for finding the additional universal IR evaluation points involves defining the relevance metrics P,R,F,M in terms of the Swets variables a,b,c,d. Then the maximum possible number of additional quadruples is deduced, and finally, all the invalid quadruples are eliminated so that only the valid, universal IR points remain. Six of these points may be interpreted as being continuous extensions of the universal IR surface, while the other nine points may be interpreted as being “off the universal IR surface.” This completely solves the problem of finding the maximum range possible of universal IR evaluation points.     

BERT-SMAP: Paying attention to Essential Terms in passage ranking beyond BERT

Passage ranking has attracted considerable attention due to its importance in information retrieval (IR) and question answering (QA). Prior works have shown that pre-trained language models (e.g. BERT) can improve ranking performance. However, these simple BERT-based methods tend to focus on passage terms that exactly match the question, which makes them easily fooled by the overlapping but irrelevant (distracting) passages. To solve this problem, we propose a self-matching attention-pooling mechanism (SMAP) to highlight the Essential Terms in the question-passage pairs. Further, we propose a hybrid passage ranking architecture, called BERT-SMAP, which combines SMAP with BERT to more effectively identify distracting passages and downplay their influence. BERT-SMAP uses the representations obtained through SMAP to enhance BERT’s classification mechanism as an interaction-focused neural ranker, and as the inputs of a matching function. Experimental results on three evaluation datasets show that our model outperforms the previous best BERT${}_{base}$-based approaches, and is comparable to the state-of-the-art method that utilizes a much stronger pre-trained language model.     

Probabilistic co-relevance for query-sensitive similarity measurement in information retrieval

Interdocument similarities are the fundamental information source required in cluster-based retrieval, which is an advanced retrieval approach that significantly improves performance during information retrieval (IR). An effective similarity metric is query-sensitive similarity, which was introduced by Tombros and Rijsbergen as method to more directly satisfy the cluster hypothesis that forms the basis of cluster-based retrieval. Although this method is reported to be effective, existing applications of query-specific similarity are still limited to vector space models wherein there is no connection to probabilistic approaches. We suggest a probabilistic framework that defines query-sensitive similarity based on probabilistic co-relevance, where the similarity between two documents is proportional to the probability that they are both co-relevant to a specific given query. We further simplify the proposed co-relevance-based similarity by decomposing it into two separate relevance models. We then formulate all the requisite components for the proposed similarity metric in terms of scoring functions used by language modeling methods. Experimental results obtained using standard TREC test collections consistently showed that the proposed query-sensitive similarity measure performs better than term-based similarity and existing query-sensitive similarity in the context of Voorhees’ nearest neighbor test (NNT).     

Transitive dictionary translation challenges direct dictionary translation in CLIR

The paper reports on experiments carried out in transitive translation, a branch of cross-language information retrieval (CLIR). By transitive translation we mean translation of search queries into the language of the document collection through an intermediate (or pivot) language. In our experiments, queries constructed from CLEF 2000 and 2001 Swedish, Finnish and German topics were translated into English through Finnish and Swedish by an automated translation process using morphological analyzers, stopword lists, electronic dictionaries, n-gramming of untranslatable words, and structured and unstructured queries. The results of the transitive runs were compared to the results of the bilingual runs, i.e. runs translating the same queries directly into English. The transitive runs using structured target queries performed well. The differences ranged from −6.6% to +2.9% units (or −25.5% to +7.8%) between the approaches. Thus transitive translation challenges direct translation and considerably simplifies global CLIR efforts.     

DeASCIIfication approach to handle diacritics in Turkish information retrieval

The absence of diacritics in text documents or search queries is a serious problem for Turkish information retrieval because it creates homographic ambiguity. Thus, the inappropriate handling of diacritics reduces the retrieval performance in search engines. A straightforward solution to this problem is to normalize tokens by replacing diacritic characters with their American Standard Code for Information Interchange (ASCII) counterparts. However, this so-called ASCIIfication produces either synthetic words that are not legitimate Turkish words or legitimate words with meanings that are completely different from those of the original words. These non-valid synthetic words cannot be processed by morphological analysis components (such as stemmers or lemmatizers), which expect the input to be valid Turkish words. By contrast, synthetic words are not a problem when no stemmer or a simple first-n-characters-stemmer is used in the text analysis pipeline. This difference emphasizes the notion of the diacritic sensitivity of stemmers. In this study, we propose and evaluate an alternative solution based on the application of deASCIIfication, which restores accented letters in query terms or text documents. Our risk-sensitive evaluation results showed that the diacritics restoration approach yielded more effective and robust results compared with normalizing tokens to remove diacritics.     

入选PubMed数据库对提升医学期刊国际影响力的作用

[目的]探讨入选PubMed数据库对提升我国医学期刊国际影响力的作用,分析加快医学期刊国际化的途径和策略。[方法]采用文献资料法、文献计量法,以2011版《中文核心期刊要目总览》中6种入选PubMed数据库和6种未入选PubMed数据库的综合性医药卫生类核心期刊为研究对象,以《中国科技期刊引证报告(核心版)》和SCI数据库为数据来源,利用SPSS20.0分别对期刊的国内被引频次、影响因子和期刊在SCI数据库的被引频次做两独立样本的非参数检验。[结果]入选PubMed的期刊在SCI数据库中的被引频次显著高于未入选PubMed的期刊(P<0.05),而两组期刊的国内被引频次和影响因子均无统计学差异(P>0.05)。[结论]入选PubMed数据库对提升医学期刊国际影响力具有重要作用。加快医学期刊的国际化除加入SCI外还可考虑加入PubMed数据库等。     

Percent perfect performance (PPP)

An information retrieval performance measure that is interpreted as the percent of perfect performance (PPP) can be used to study the effects of the inclusion of specific document features or feature classes or techniques in an information retrieval system. Using this, one can measure the relative quality of a new ranking algorithm, the result of incorporating specific types of metadata or folksonomies from natural language, or determine what happens when one makes modifications to terms, such as stemming or adding part-of-speech tags. For example, knowledge that removing stopwords in a specific system improves the performance 5% of the way from the level of random performance to the best possible result is relatively easy to interpret and to use in decision making; using this percent based measure also allows us to simply compute and interpret that there remains 95% of the possible performance to be obtained using other methods. The PPP measure as used here is based on the average search length, a measure of the ordering quality of a set of data, and may be used when evaluating all the documents or just the first N documents in an ordered list of documents. Because the ASL may be computed empirically or may be estimated analytically, the PPP measure may also be computed empirically or performance may be estimated analytically. Different levels of upper bound performance are discussed.