Samsung Poland NLP Team at SemEval-2016 Task 1: Necessity for diversity; combining recursive autoencoders, WordNet and ensemble methods to measure semantic similarity.

Proceedings of SemEval-2016, pages 602–608, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics Samsung Poland NLP Team at SemEval-2016 Task 1: Necessity for diversity; combining recursive autoencoders, WordNet and ensemble methods to measure semantic similarity. Barbara Rychalska, Katarzyna Pakulska, Krystyna Chodorowska, Wojciech Walczak and Piotr Andruszkiewicz Samsung R&D Institute Poland Polna 11 Warsaw, Poland {b.rychalska, k.pakulska, k.chodorowsk, w.walczak, p.andruszki2}@samsung.com Abstract This paper describes our proposed solutions designed for a STS core track within the Se- mEval 2016 English Semantic Textual Sim- ilarity (STS) task. Our method of similar- ity detection combines recursive autoencoders with a WordNet award-penalty system that ac- counts for semantic relatedness, and an SVM classifier, which produces the final score from similarity matrices. This solution is further supported by an ensemble classifier, combin- ing an aligner with a bi-directional Gated Re- current Neural Network and additional fea- tures, which then performs Linear Support Vector Regression to determine another set of scores. 1 Introduction The tasks from the Semantic Textual Similarity (STS) contest have always attracted vivid interest from the NLP community. The goal is to measure the semantic similarity between two given sentences on a scale from 0 to 5, trying to emulate the idea of similarity degrees, thus replicating human language understanding. After processing two pieces of text, semantic tex- tual similarity software captures degrees of semantic equivalence. One of the goals of the STS task is to create a unified framework for extracting and mea- suring semantic similarity. Improvements achieved in the course of this task can be useful in many research areas, such as question answering (Marsi and Krahmer, 2005), machine translation (Callison- Burch, 2008), and plagiarism detection (Clough et al., 2002). We present a solution designed to detect both the similarity between single words and longer, multi- word phrases. It employs two important compo- nents: the unfolding recursive autoencoder (RAE) (Socher et al., 2011) and the penalty-award weight system based on WordNet (Miller et al., 2002). First, RAE is used to perform unsupervised learning on parse trees, then the WordNet module adjusts the distances of RAE vectors using awards and penal- ties based on semantic similarities of words. The complete pipeline includes a deep net (RAE) mod- ule, a WordNet module, a normalization module and a sentence similarity matrices computing module. Another solution that ran in parallel to the RAE pipeline, was the monolingual word aligner (in some cases we used its corrected version with ad- ditional features, including a bag-of-words). Fi- nally an ensemble classifier was used to perform Linear Support Vector Regression (Drucker et al., 1996) over the results from all the other classifiers. This included: the base word aligner (Sultan et al., 2015), bi-directional Gated Recurrent Neural Net- work (Cho et al., 2014; Chung et al., 2014), the RAE with WordNet features and the corrected aligner. 2 System Overview This section describes the modules that constitute our three runs. Detailed information about the con- figuration of these runs can be found in Section 3. 2.1 RAE with …

Content truncated to 500 words