This paper describes an approach for source retrieval task of PAN 2015 competition. We apply two methods to extract important terms, namely weighted noun phrases and keyword phrases which are extracted from long sentences in terms of word count. Queries are constructed from top marked sentences. The prepared system tries to gather a complete dataset of downloaded sources and employ it in query filtering operations. The ChatNoir search API is used for submitted queries. Each query is split into two sub-queries and the system extract one snippet for each of sub-queries and exploits them in downloading operation. The evaluation results show high scores for three measures: recall, total queries number and no detection.
The advent and rapid development of the World Wide Web facilitate public access to
digital information, so that everybody can easily read and alter the content of web
pages for personal use. There can be considered two aspect of information generation.
First, using the existing information with their references as input and extending them
with new innovation as output, and second, taking information and altering the
content while maintaining the concept without any reference to source which is called
plagiarism. In simple terms, plagiarism is the unreferenced use of other's ideas or text.
Plagiarism detection in PAN [
The rest of this paper is structured as follows: Section 2 describes our approach that includes document segmentation, keyword extraction methods, query formulation and query and document filtering. Section 3 presents the results of our method and a discussion on the results. Conclusion and steps for future works will be explained in section 4. 2
Our approach has been divided into five steps as follows:
Suspicious Document Chunking,
Noun phrase and keyword phrase Extraction,
Query Formulation,
Search Control,
Document Filtering and Downloading
These steps are equal to those described in [
After the preprocessing step, the documents are prepared for suspicious document chunking. Each document is segmented into some parts called chunks. These chunks are separately used for keyword phrase and noun phrase extraction and also query construction. Therefore, their length should be long enough to extract meaningful queries. On the other hand, these chunks may contain unknown numbers of plagiarism cases from source documents. Suspicious documents are divided into chunks of 500 words length and then each chunk is tokenized into individual sentences. As a result, we have some sentences that are used to extract appropriate keywords. 2.2
This step has the main role in source retrieval task. Extracting appropriate keywords
help us to efficiently perform the next steps. There are many previous studies that
have tried to extract the keywords by investigating the contents [
The input to this step is a series of sentences that are extracted from a specific chunk at the chunking step. Before starting extraction process, sentences with low information content are discarded: We rank the input sentences based on their length and the number of nouns, and then discard the lower 20% of the sentences in the ranking. The resulting sentences are long enough and have rich content for keyword extraction.
In this stage, two different types of tf.idf weighting scheme are used for rating important words in the sentences: Type #1 (tf.idf1): In this type, Tf = term frequency in the chunk, and Idf = inverse term frequency in the whole suspicious document Type #2 (tf.idf2): in this type, Tf = term frequency in the chunk, and Idf = inverse document frequency in PAN 2011 corpus Keyword phrase Extraction. Keyword phrases are considered as a collection of keywords with high tf.idf weights in a sentence. The number of keywords that can be used as input to ChatNoir search engine is limited, so the number of query words should be restricted in such a way to maximize the information content. The process of keyword phrase extraction is done in four steps as follows: In the first step, for each type of tf.idf weighting scheme that mentioned above, ten words with highest tf.idf value are selected from the whole chunk. In the second step, the sentences that contain words with high values of tf.idf1 AND tfidf2 are selected. In the next step, among the selected sentences, those that contain words with highest tf.idf1 and tfidf2 are selected for keyword phrase extraction. As a result, three sentences are selected in this step. Finally, in the fourth step, the keywords are extracted from the resulting sentences as follows: Nouns with high tf.idf values Remaining nouns in the sentence Adjective and verbs with high tf.idf1 values It should be noted that the selection process is done based on the above mentioned priority. When we reach the maximum number of keywords in each stage, then the process would be stopped.
The four steps mentioned above are repeated for each chunk and the extracted keyword phrases are passed to the next step for query building.
Noun phrase Extraction. Noun phrase extraction is accomplished by processing the
remaining sentences. The formulation has been deployed based on the formal English
noun phrase structure [
As a result, the implemented system uses two different scenarios applied to sentences for keyword extraction. After dividing a suspicious document into some chunks, the following scenarios are used based on operations depicted in Table 1. Scenario1: Operation 1 Operation 2 Operation 3 for noun phrase extraction Scenario2: Operation 1 Operation 2 Operation 4 for keyword phrase extraction Operation number 1 2 3 4
The outputs of these scenarios are also passed through a filter that removes some terms with low weight to reach the word count limitation of ChatNoir API. The remaining terms formulate a query and in the next step we can select top weighted sentence for query formulation. 2.3
For top sentences selected from previous step, the extracted keywords are simply placed next to each other based on their order in sentence and are passed to next step as a query. According to ChatNoir limitations, the threshold for the number of words in each query is limited to 10. 2.4
In this step, we filtered the constructed queries based on the possibly previous downloaded documents. The input query is compared against the downloaded documents that are gathered from previous rounds of source retrieval steps. If there is at least one downloaded document that contains at least 60% words of the query, then the query is dropped from passing to the next step. This threshold was achieved based on experiments. 2.5
We have used ChatNoir API for applying the input queries into the search engine. Then 14 top ranked results returned for each query. Input query is divided into two sub-queries and for each of them, one snippet with the length of 500 characters is extracted per returned document. These snippets are combined with one another and make a passage. If the resulted passage contains at least 50% words of the query, then the related document is downloaded and maintained for search control operation. This threshold was achieved based on experiments.
We have implemented our approach using python programming language and NLTK
package for text processing operations [
For future works, we will try to decrease the number of downloaded source documents while keeping the complete set of related documents for query filtering. This work has been accomplished in ICT research Institute, ACECR, under the support of Vice Presidency for Science and Technology of Iran - grant No. 1164331. The authors gratefully acknowledge the support of aforementioned organizations. Special thanks go to the members of ITBM research group for their valuable collaboration.