Parth Gupta

Query expansion for mixed-script information retrieval

For many languages that use non-Roman based indigenous scripts (e.g., Arabic, Greek and Indic languages) one can often find a large amount of user generated transliterated content on the Web in the Roman script. Such content creates a monolingual or multi-lingual space with more than one script which we refer to as the Mixed-Script space. IR in the mixed-script space is challenging because queries written in either the native or the Roman script need to be matched to the documents written in both the scripts. Moreover, transliterated content features extensive spelling variations. In this paper, we formally introduce the concept of Mixed-Script IR, and through analysis of the query logs of Bing search engine, estimate the prevalence and thereby establish the importance of this problem. We also give a principled solution to handle the mixed-script term matching and spelling variation where the terms across the scripts are modelled jointly in a deep-learning architecture and can be compared in a low-dimensional abstract space. We present an extensive empirical analysis of the proposed method along with the evaluation results in an ad-hoc retrieval setting of mixed-script IR where the proposed method achieves significantly better results (12% increase in MRR and 29% increase in MAP) compared to other state-of-the-art baselines.

Methods for cross-language plagiarism detection

Three reasons make plagiarism across languages to be on the rise: (i) speakers of under-resourced languages often consult documentation in a foreign language, (ii) people immersed in a foreign country can still consult material written in their native language, and (iii) people are often interested in writing in a language different to their native one. Most efforts for automatically detecting cross-language plagiarism depend on a preliminary translation, which is not always available. In this paper we propose a freely available architecture for plagiarism detection across languages covering the entire process: heuristic retrieval, detailed analysis, and post-processing. On top of this architecture we explore the suitability of three cross-language similarity estimation models: Cross-Language Alignment-based Similarity Analysis (CL-ASA), Cross-Language Character n-Grams (CL-CNG), and Translation plus Monolingual Analysis (T+MA); three inherently different models in nature and required resources. The three models are tested extensively under the same conditions on the different plagiarism detection sub-tasks-something never done before. The experiments show that T+MA produces the best results, closely followed by CL-ASA. Still CL-ASA obtains higher values of precision, an important factor in plagiarism detection when lesser user intervention is desired.

Cross-Language high similarity search using a conceptual thesaurus

This work addresses the issue of cross-language high similarity and near-duplicates search, where, for the given document, a highly similar one is to be identified from a large cross-language collection of documents. We propose a concept-based similarity model for the problem which is very light in computation and memory. We evaluate the model on three corpora of different nature and two language pairs English-German and English-Spanish using the Eurovoc conceptual thesaurus. Our model is compared with two state-of-the-art models and we find, though the proposed model is very generic, it produces competitive results and is significantly stable and consistent across the corpora.

PAN@FIRE: Overview of the Cross-Language !ndian News Story Search (CL!NSS) Track

The automatic alignment of documents in a quasi-comparable corpus is an important research problem for a resource poor cross-language technologies. News stories form one of the most prolific and abundant language resource. The PAN@FIRE task, cross-language !ndia news story search (CL!NSS), aimed to address the news story linking task across languages English and Hindi. We present the overview of the track with results and analysis.

Cross-language plagiarism detection over continuous-space- and knowledge graph-based representations of language

We study the combination of knowledge graph and continuous space representations for cross-language plagiarism detection.We also compare methods that only make use of continuous-space representations of text.We present the continuous word alignment-based similarity analysis, a model to estimate similarity between text fragments.We obtain state-of-the-art performance compared to several state-of-the-art models. Cross-language (CL) plagiarism detection aims at detecting plagiarised fragments of text among documents in different languages. The main research question of this work is on whether knowledge graph representations and continuous space representations can complement to each other and improve the state-of-the-art performance in CL plagiarism detection methods. In this sense, we propose and evaluate hybrid models to assess the semantic similarity of two segments of text in different languages. The proposed hybrid models combine knowledge graph representations with continuous space representations aiming at exploiting their complementarity in capturing different aspects of cross-lingual similarity. We also present the continuous word alignment-based similarity analysis, a new model to estimate similarity between text fragments. We compare the aforementioned approaches with several state-of-the-art models in the task of CL plagiarism detection and study their performance in detecting different length and obfuscation types of plagiarism cases. We conduct experiments over Spanish-English and German-English datasets. Experimental results show that continuous representations allow the continuous word alignment-based similarity analysis model to obtain competitive results and the knowledge-based document similarity model to outperform the state-of-the-art in CL plagiarism detection.

Mapping Hindi-English Text Re-use Document Pairs

An approach to find the most probable English source document for the given Hindi suspicious document is presented. The approach does not involve complex method of Machine Translation as a language normalization step, rather relies on standard cross-language resources available between Hindi-English and calculates the similarity using the Okapi BM25 model. We also present the further improvements in the system after the analysis and discuss the challenges involved. The system is developed as a part of CLiTR competition and uses the CLiTR-Dataset for the experimentation. The approach achieves the recall of 0.90 - the highest and F-measure of 0.79 - the 2 nd highest reported on the Dataset.

Knowledge Graphs as Context Models: Improving the Detection of Cross-Language Plagiarism with Paraphrasing

Cross-language plagiarism detection attempts to identify and extract automatically plagiarism among documents in different languages. Plagiarized fragments can be translated verbatim copies or may alter their structure to hide the copying, which is known as paraphrasing and is more difficult to detect. In order to improve the paraphrasing detection, we use a knowledge graph-based approach to obtain and compare context models of document fragments in different languages. Experimental results in German-English and Spanish-English cross-language plagiarism detection indicate that our knowledge graph-based approach offers a better performance compared to other state-of-the-art models.

Continuous space models for CLIR

We present and evaluate a novel technique for learning cross-lingual continuous space models to aid cross-language information retrieval (CLIR). Our model, which is referred to as external-data composition neural network (XCNN), is based on a composition function that is implemented on top of a deep neural network that provides a distributed learning framework. Different from most existing models, which rely only on available parallel data for training, our learning framework provides a natural way to exploit monolingual data and its associated relevance metadata for learning continuous space representations of language. Cross-language extensions of the obtained models can then be trained by using a small set of parallel data. This property is very helpful for resource-poor languages, therefore, we carry out experiments on the English-Hindi language pair. On the conducted comparative evaluation, the proposed model is shown to outperform state-of-the-art continuous space models with statistically significant margin on two different tasks: parallel sentence retrieval and ad-hoc retrieval.

Squeezing bottlenecks: Exploring the limits of autoencoder semantic representation capabilities

We present a comprehensive study on the use of autoencoders for modelling text data, in which (differently from previous studies) we focus our attention on the following issues: i) we explore the suitability of two different models bDA and rsDA for constructing deep autoencoders for text data at the sentence level; ii) we propose and evaluate two novel metrics for better assessing the text-reconstruction capabilities of autoencoders; and iii) we propose an automatic method to find the critical bottleneck dimensionality for text language representations (below which structural information is lost).

English-to-Hindi system description for WMT 2014: Deep Source-Context Features for Moses

This paper describes the IPN-UPV participation on the English-to-Hindi translation task from WMT 2014 International Evaluation Campaign. The system presented is based on Moses and enhanced with deep learning by means of a source-context feature function. This feature depends on the input sentence to translate, which makes it more challenging to adapt it into the Moses framework. This work reports the experimental details of the system putting special emphasis on: how the feature function is integrated in Moses and how the deep learning representations are trained and used.