Text reuse identification has become a challenging problem due to the presence of enormous amounts of digital data, more so because of obfuscated text reuse. The result of obfuscation is a modified version of the original text. The modification can be at the level of words, phrases, sentences, or even whole texts by applying a random sequence of text operations such as change of tense, alteration of voice (active to passive, and vice versa), change in treatment of direct speech, abbreviations, shuffling a word or a group of words, deleting a word, inserting a word from an external source, or replacing a word with a synonym, antonym, hypernym, or hyponym. These alterations may or may not modify the original meaning of the source text.
Obfuscation in text reuse can be at different levels of degree. It may be from no obfuscation to a low or high level of obfuscation. The range from word to sentence level defines the level of obfuscation from low to high. Translation of text from the language of source document to another language of suspicious target document is also a kind of high level obfuscation which is extremely difficult to deal with, more so when the reuse is local in nature. Local reuse occurs when sentences, facts, or passages, rather than whole documents, are reused and modified.
Therefore, techniques that can be applied in order to identify the different levels of obfuscation and their local nature may also vary due to the complexity of the problem. In a large corpus analyzing complete sets of source documents for local text reuse is an expensive affair; therefore, it is better to retrieve a subset of documents and then at the time of post processing, a retrieved set of source documents may be analyzed for local text reuse. Initial filtering is known as heuristic retrieval (Barrón-Cedeño, 2010). Heuristic retrieval, also called pre-retrieval, can therefore be defined as the process of retrieving a small sub-set of a potential reused document for any particular source-document from a large set of documents (corpora), with a view to minimize time and space requirements.
Heuristic retrieval is important due to the (i) enormous size of a typical corpus; (ii) presence of large numbers of irrelevant documents for a particular set of suspicious documents; and (iii) the processing cost involved in processing the dataset. Also, while for small document collections it is practicable to perform a complete comparison against every document, this is obviously not possible when the collection size is enormous. So it is better to go for heuristic retrieval before post processing.
In this paper an attempt is made to find key terms from a target set of suspicious documents to retrieve an initial set of source documents for further post processing. In case of local obfuscated text reuse, generating keywords from the whole text can drift the query and may fetch many unwanted documents. The main focus is, therefore, to formulate effective queries to retrieve a subset of documents that bears more closeness to any given suspicious document.
In the proposed work three methods have been studied for formation of query, especially when the content of the documents are obfuscated or translated and the text reuse is local in nature. The results are derived from the PAN CLEF 2012 training corpus.
In Section 2 existing work in pre-retrieval query performance is discussed, Section 3 discusses proposed methodology, and Section 4 and 5 discuss experimental results and conclusions, respectively.
Hauff, Hiemstra, and Jong (2008) assessed the performance of 22 pre-retrieval predictors on three different TREC collections. As most predictors exploit inverse term/document frequencies in some way, they hypothesize that the amount of smoothing influences the quality of predictors.
Cummins, Jose, and O’Riordan (2011) developed a new predictor based on standard deviation of scores in a variable length ranked list, and showed that this new predictor outperforms state-of-the-art approaches without the need of tuning.
Possas et al. (2005) worked on TREC-8 test collection and proposed a technique for automatically structuring web queries as a set of smaller sub queries. To select representative sub queries, information of distributions is used and a concept of maximal term sets derived from formalism of association rules theory is used for modelling.
Many kinds of text reuse detection techniques have been proposed from time to time by different authors, including: Potthast et al. (2013); Gustafson et al. (2008); Mittelbach et al. (2010); Palkovskii, Muzyka, and Belov (2010); Seo and Croft (2008); Clough et al. (2002); Gupta and Rosso (2012); Bar, Zesch, and Gurevych (2012); and Potthast et al. (2013a, b).
Gipp et al. (2013) proposed Citation-based Plagiarism Detection. Compared to existing approaches, CbPD does not consider textual similarity alone, but uses the citation patterns within scientific documents as a unique, language independent fingerprint to identify semantic similarity.
Vogel, Hey, and Tillmann (1996) presented an HMM-based approach for modelling word alignments in parallel texts in English and French. The characteristic feature of this approach is to make the alignment probabilities explicitly dependent on the alignment position of the previous word. The HMM-based approach produces translation probabilities.
Barrón-Cedeño (2012) compared two recently proposed cross-language plagiarism detection methods: CL-CNG, based on character n-grams, and CL-ASA, based on statistical translation, to their new approach based on machine translation and monolingual similarity analysis (T+MA). Barrón-Cedeño explores the effectiveness of his approach for less related languages. CL-CNG is not appropriate for this kind of language pairs, whereas T+MA performs better than the previously proposed models. The study investigated Basque, a language where, due to lack of resources, cross language plagiarism is often committed from texts in Spanish and English.
Grozea and Popescu (2009) evaluated cross-language similarity among suspected and original documents using a statistical model which finds the relevance probability between suspected and source documents, regardless of the order in which the terms appear in the suspected and original documents. Their method is combined with a dictionary corpus of text in English and Spanish to detect similarity in cross language.
A plagiarism detection technique based on Semantic Role Labeling was introduced by Osmana et al. (2012). They improved the similarity measure using argument weighting with an aim to studying the argument behaviour and effect in plagiarism detection.
Pouliquen et al. (2003) have worked on European languages and have presented a working system that can identify translations and other very similar documents among a large number of candidates, by representing the document content with a vector of thesaurus terms from multilingual thesaurus, and then by measuring the semantic similarity between the vectors.
The approach used by Palkovskii and Belov (2011) implied the usage of automatic language translation (Google Translate web service) to normalize one of the input texts to the target comparison language, and applies a model that includes several filters, each of which adds ranking points to the final score.
Ghosh, Pal, and Bandyopadhyay (2011) treated cross-language text re-use detection as a problem of Information Retrieval, and it is solved with the help of Nutch, an open source Information Retrieval (IR) system. Their system contains three phases – knowledge preparation, candidate retrieval, and cross-language text reuse detection.
Gupta and Singhal (2011) tried to see the impact of available resources like Bi-lingual Dictionary, WordNet, and Transliteration, mapping Hindi-English text reuse document pairs and using the Okapi BM25 model to calculate the similarity between document pairs.
The approach used by Aggarwal et al. (2012) in journalistic text reuse consists of two major steps, the reduction of search space by using publication date and vocabulary overlap, and then ranking of the news stories according to their relatedness scores. Their approach uses Wikipedia-based Cross-Lingual Explicit Semantic Analysis (CLESA) to calculate the semantic similarity and relatedness score between two news stories in different languages.
Arora, Foster, and Jones (2013) used an approach consisting of two steps: (1) the Lucene search engine was used with varied input query formulations using different features and heuristics designed to identify as many relevant documents as possible to improve recall; and (2) merging of document list and re-ranking was performed with the incorporation of a date feature.
Pal and Gillam (2013) converted English documents to Hindi using Google Translate and compared them to the potential Hindi sources based on five features of the documents: title, content of the article, unique words in content, frequent words in content, and publication date using Jaccard similarity. A weighted combination of the five individual similarity scores provides an overall value for similarity.
Tholpadi and Param (2013) describe a method that leverages the structure of news articles, especially the title, to achieve good performance on the focal news event linking task. They found that imposing date constraints did not improve precision.
IDF, Reference Monotony, and Extended Contextual N-grams were used by Torrejon and Ramos (2013) to link English and Hindi News.
Haiduc et al. (2013) have proposed a recommender, Refoqus, which automatically recommends a reformulation strategy for a given query to improve its retrieval performance in Text Retrieval. Refoqus is based on Machine Learning and its query reformulation strategy is based on the properties of the query.
Carmel et al. (2006), while trying to find a solution to the question “what makes a query difficult,” have devised a model to predict query difficulty and number of topic aspects expected to be covered by the search results and to analyze the findability of a specific domain.
The Capacity Constrained Query Formulation method was devised by Hagen and Stein (2010). They focused on the query formulation problem as the crucial first step in the detection process and have presented this strategy, which achieves better results than maximal term set query formulation strategy.
Most of the authors listed have worked upon whole corpora, which consume valuable resources with respect to space and time. Also, most query formulation and retrieval strategies fail in the case of highly obfuscated and translated reused documents. In such cases, even the most popular TF-IDF strategy is no exception. Use of thesaurus in query formulation aids in query drifting and results in fetching unwanted irrelevant documents. Devising a simple strategy which can reduce the size of the corpora by retrieving potential reused documents in the pre-retrieval stage, and before going for state-of-the-art text reuse or plagiarism detection techniques, can render the process more efficient in terms of system resources and would produce more accurate results. Our strategy is straightforward and simple and tackles this important pre-computation step that finds promising candidate documents for in-depth analysis.
In this preliminary study, an attempt has been made to analyse proposed strategies on the training corpus of PAN CLEF 2012 which is 1.29 GB in size and contains 12,024 files divided into 8 folders.
The suspicious documents that fall under the categories of low and high obfuscation, simulated paraphrasing, and translation are studied in this paper. Simulated paraphrasing is intentional obfuscation done by humans with an intention to hide plagiarism attempts.
The steps followed for formulating a query for heuristic retrieval are as follows:
As we are dealing with local text reuse and obfuscation which can be in any part of the document, the document is divided into units. The document unit that we have taken is ‘paragraph,’ with the assumption that even if sentences or a block is reused and obfuscated, that will most likely be a paragraph or a sentence within any paragraph. Therefore, the suspicious document is divided into paragraphs before tokenizing the text. The document is normalized by removing the punctuation. Stop-words, verbs, and adverbs are removed using a pre-compiled list of stop-words,
Queries have been formulated paragraph-wise for each suspicious document. So the number of paragraphs in the suspicious document decides the number of queries for that document. Relying on a given document structure like paragraphs bears the risk of failing for some unseen documents that are not well formatted (Potthast et al., 2013); still, the idea behind generating queries for each paragraph is that if reuse is local, then at least keywords from the paragraph which have been reused will maximize the chance of fetching the required source document. The query is generated from each paragraph by selecting a) proper nouns; b) unique words or Hapax; or c) most frequent words.
Along with this we have also tried many other strategies for heuristic retrieval in obfuscated corpora, none of which showed fruitful results. Therefore we are not discussing them in our work which we have presented here.
Three pre-retrieval strategies lead to the formulation of these three different runs:
3.2.1. Run 1: Query Formulation Using Proper Nouns
It has been observed that rarely are nouns or proper nouns ever changed while obfuscating or translating the text. This feature of obfuscation is highly visible in translated text reuse (Fig. 1) which shows only the unchanged proper noun “Goethe” when the whole text in German was translated into English for reuse (suspicious doc-id:01277, source doc-id:02637, susp_language=“en” susp_offset=“29575” susp_length=“675” source_language=“de” source_offset=“176202” source_length=“787” of PAN CLEF).
This prompted us to select proper nouns for formulating queries. The assumption that if the same scripts are being used, only the proper nouns—that is, the names of persons, locations, and organizations, etc.—do not change, led to the formulation of queries with proper nouns. The grammar rule that proper nouns begin with a capital letter has been used to identify them.
Before formulation of query, stop-words/function words were removed using a precompiled list of stopwords, verbs, and adverbs acquired from the web. Cases of the words were taken care of. Thus what were left were mainly nouns and adjectives. Mostly the adjectives do not start any sentence without the support of an article and therefore, adjectives could not have been the ones with a starting uppercase letter. Any noun may have contained a starting uppercase letter and could have become the only source of introducing noise and contributing to a few false positives.
3.2.2. Run 2: Query Formulation Using Unique Words (Hapax)
The terms which appear only once in the document are also known as
The reason behind taking Hapax for query formulation is that even if one or two words are left out at the time of obfuscation, then these words shall help in identifying the local text reuse.
3.2.3. Run 3: Query Formulation Using Most Frequent Words
In this run the terms which were most frequent in the pre-processed paragraphs were used for query formulation. This is the most common strategy used in query formulation. In our approach the most frequent words for query formulation have been included just for comparison purposes and for analysing the efficiency of this strategy in obfuscated and translated locally reused text.
All of the above-mentioned pre-retrieval query formulation strategies are prompted by a set of source documents retrieved during initial experimental investigations. Query words in italics are proper nouns and underlined query words are Hapax under different obfuscation (Table 1).
Query Formulation using Proper Nouns and Hapax under Different Kinds of Obfuscation
The corpus of the source documents is indexed using the Indri retrieval engine.
The performance of the two strategies of query formulation for a) proper nouns, b) unique words or Hapax, is analysed against the performance of the most commonly used method, i.e. queries formed using, c) most frequent words.
The complete process of heuristic retrieval is shown in Fig. 2.
In this preliminary study, an attempt has been made to analyse proposed strategies on the training corpus of PAN CLEF 2012, which is 1.29 GB in size and contains 12,024 files divided into 8 sub-folders.
The dataset is comprised of six different sub-sets contained in folders named no-plagiarism, no-obfuscation, artificial-low, artificial-high, translation, and simulated paraphrasing. We dealt with the later four categories. We tried to formulate simple querying strategies with a view to retrieving a subset of potential documents. Formulating any querying strategies from the terms of English target documents with an aim to retrieve the original German and Spanish documents would fail, because although the English translated documents use the same script they have an altogether different vocabulary when compared to German or Spanish. This would require translation of both texts into one single language which is a tedious task for large corpora. In the proposed work we are not dealing with the complete task of text reuse detection. Our aim is only to reduce the size of the corpora on which state-of-the-art techniques can be used for retrieving the reused portions with further refined processes.
The PAN-CLEF training corpus was divided into 8 sub-folders (Fig. 3):
The corpus consists of:
/susp: suspicious documents as plain text. /src : source documents as plain text.
The suspicious documents contain passages ‘plagiarized’ from the source documents, obfuscated with one of five different obfuscation techniques.
Furthermore, the corpus contains 6,000 XML files each of which report, for a pair of suspicious and source documents, the exact locations of the plagiarized passages (Fig. 4). The XML files are split into six datasets:
/01_no_plagiarism: XML files for 1,000 document pairs without any plagiarism. /02_no_obfuscation: XML files for 1,000 document pairs where the suspicious document contains exact copies of passages in the source document. /03_artificial_low: XML files for 1,000 document pairs where the plagiarized passages are obfuscated by the means of moderate word shuffling. /04_artificial_high: XML files for 1,000 document pairs where the plagiarized passages are obfuscated by the means of not-so moderate word shuffling. /05_translation: XML files for 1,000 document pairs where the plagiarized passages are obfuscated by translation into a different language. /06_simulated_paraphrase: XML files for 1,000 document pairs where the plagiarized passages are obfuscated by humans via Amazon Mechanical Turk.
The experiments were performed on low obfuscated, high obfuscated, simulated paraphrasing, and translation corpus texts.
It was observed that reused text in low obfuscation showed only minor changes in the text, and only a few of the words, mainly adverbs, were replaced in suspicious text. The reused text with high obfuscation had a comparatively larger number of words and even phrases replaced by synonyms, antonyms, and other similar phrases.
Simulated paraphrasing, although comparable, still was a difficult case to deal with, as whole texts were completely and intentionally paraphrased to hide the signs of reuse.
None of the authors are native speakers of any of the languages used in reused translated texts like German, Spanish, etc. but as the script was the same, the only action authors could do is to observe proper nouns appearing in both source and reused text. A snapshot of different kinds of obfuscation in the PAN CLEF 2012 training data set is shown in Table 2.
Snapshot of Different Kinds of Obfuscation in PAN CLEF 2012 Training Data Set
It is observed that independent authors can create the same short sentences rather than long, similar ones, which are less likely to be similar by chance (Gustafson et al., 2008). Therefore, to avoid false positives queries were formed and posed to retrieval engine only if three or more than three words of each type were extracted from each paragraph.
The comparison was made based on two kinds of result sets of retrieval: 1) when the number of retrieved documents per query is 5; and 2) when the number of retrieved documents per query is 1.
In the training corpus of PAN-CLEF, in addition to the XML files, each folder contains a text file called ‘pairs.’ For the 1,000 document-pairs (XML files) in the folder, this file lists the filename of the suspicious and the source document in a row, separated by a blank (Fig. 5):
e.g. suspicious-document00086.txt source-document00171. txt
A list of all of the relevant documents for each category was compiled using the relevance data provided by CLEF and the result of the experimentation was compared against this. As far as translated texts are concerned, they were in either German or Spanish, and none of the authors(s) are either native speakers of these languages or familiar with the vocabulary of these languages. So for the analysis of our results we only relied on the relevant source document list provided by CLEF for each suspicious document and have made that list our judgment criteria for pre-retrieval. Table 3 shows one such list for suspicious documents and source documents under translation dataset.
Compiled List of Source Documents for Each Target Document under Translation Dataset
Table Showing Average Retrieval Percentage Summary
When query is formulated using proper nouns and results are recorded for five documents, the per query percentage of correct source documents retrieved is 88.34%, 71.50%, 52.85%, and 79.17%, respectively, in the cases of low, high obfuscation, simulated paraphrasing, and translation. In case of Hapax the percentage is 92.21%, 84.17%, 59.68%, and 52.50% (Table 5 and Fig. 6).
In the case of one document retrieved per query, the percentage is 81.55%, 66%, 36.18%, and 77.50%, respectively, for low obfuscation, high obfuscation, simulated paraphrasing, and translation when query is formulated using proper nouns, and the same is 86.43%, 76.83%, 41.24%, and 42.92%, respectively, for Hapax (Table 5 and Fig. 7).
Queries using most frequent words showed the worst performance in all cases. In the case of five documents retrieved per query it is 55.45%, 30.17%, 18.45%, and 36.25% for low obfuscation, high obfuscation, simulated paraphrasing, and translation, respectively (Table 5 and Fig. 6), and in the case of one document per query performance is further degraded with only 47.35%, 19.17%, 6.14%, and 31.67% retrieval of source documents, respectively, for different obfuscation levels (Table 5 and Fig. 7).
The Indri retrieval results show that queries formed with proper nouns and Hapax outperformed the most frequent words query formulation strategy (Table 4). In the cases of low and high obfuscation and simulated paraphrasing, keywords with Hapax proved to be slightly more efficient. Queries formed using proper nouns performed exceptionally well in the case of Translated Local Text Reuse. In some of the cases the latter strategy retrieved all documents whereas the former could not retrieve even a single document. Similarity scores of proper noun queries were higher than those of the other two methods in most of the cases, even when the source document ranking was the same (Table 4). Most of the words of proper noun queries are also found in Hapax queries for the same paragraph in a query (Table 4). Proper noun queries are crisp and concise whereas Hapax queries are long (Table 4). The criterion that any word which starts with an uppercase letter is considered to be a proper noun has been applied, so a few other terms, mainly nouns which start a sentence, are mistakenly included in the query.
Comparison of Queries Formulated and used (Suspicious Document id:00000 Source Document ids: 01882, 01883)
Query performances are comparable in the cases of proper nouns and Hapax (Figs. 6 and 7), and retrieval scores are higher in the case of proper noun queries; therefore proper noun queries may be preferred over Hapax queries as those formed using proper nouns also reduce the chances of getting query drift.
The results of these experiments show that a pre-retrieval strategy of proper nouns and Hapax outperformed a most frequent words strategy. Initial study reveals that level of obfuscation may also have an influence on pre-retrieval strategy. Whereas Hapax was observed to be slightly more efficient than other strategies in the cases of low obfuscation, high obfuscation, and simulated paraphrasing, queries formulated using proper nouns were definitely the most efficient in the case of heuristic retrieval for local reuse in translated texts. The Heuristic retrieval strategies that were somewhat more efficient require further study on mono- and cross-lingual text reuse with different scripts. Further work is in progress as these are intermediate results of the experiments performed on the PAN CLEF 2012 training data set.