UWB at SemEval-2016 Task 2: Interpretable Semantic Textual Similarity with Distributional Semantics for Chunks

Proceedings of SemEval-2016, pages 803–808, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics UWB at SemEval-2016 Task 2: Interpretable Semantic Textual Similarity with Distributional Semantics for Chunks Miloslav Konop´ık and Ondˇrej Praˇz´ak and David Steinberger and Tom´aˇs Brychc´ın NTIS – New Technologies for the Information Society, Department of Computer Science and Engineering, Faculty of Applied Sciences, University of West Bohemia, Technick´a 8, 306 14 Plzeˇn Czech Republic {konopik,brychcin}@ntis.zcu.cz, {ondfa,fenic}@kiv.zcu.cz Abstract We introduce a system focused on solving SemEval 2016 Task 2 – Interpretable Seman- tic Textual Similarity. The system explores machine learning and rule-based approaches to the task. We focus on machine learning and experiment with a wide variety of ma- chine learning algorithms as well as with sev- eral types of features. The core of our system consists in exploiting distributional semantics to compare similarity of sentence chunks. The system won the competition in 2016 in the “Gold standard chunk scenario”. We have not participated in the “System chunk scenario”. 1 Introduction The goal of the Interpretable Semantic Textual Sim- ilarity task is to go deeper with the assessment of semantic textual similarity of sentence pairs. It is requested to add an explanatory layer that offers a deeper insight into the sentence similarities. The sentences are split into chunks and the first goal is to find corresponding chunks (with respect to their meanings) among the compared sentences. When the corresponding chunks are known, the chunks are annotated with their similarity scores and their rela- tion types (e.g. equivalent, more specific, etc). The task follows a pilot task from the preced- ing SemEval 2015 competition (Agirre et al., 2015). The best performing systems adopted various ap- proaches, (Banjade et al., 2015) relied on hand- crafter rules, (Karumuri et al., 2015) employed a classifier for relation types and they associated each relation with a precomputed similarity score and (H¨anig et al., 2015) extended their word alignment algorithm for the task. 1.1 Math notation The data consist of sentence pairs Sa i and Sb i, where a denotes the first item of the pair, b denotes the sec- ond item of the pair and i indexes the sentences (for simiplicity we further omit i for sentences). We per- ceive a sentence Sa to be an ordered set of chunks CHa j ∈ Sa and the chunks to be ordered sets of words wk ∈ CHa j (and analogically for sentence Sb). Next we define two functions: sim(CHa i , CHb j) ∈ {0, 1, 2, 3, 4, 5} for chunk similarity and rel(CHa i , CHb j) ∈ TYPE for chunk relation type. The possible types are: TYPE = {EQUI, OPPO, SPE1, SPE2, SIMI, REL}. These are the main types. All these types can have two modifiers (FACT, POL). The modifiers are optionally attached to the main types. For example, you can generate SPE1 FACT. For more information, please see the annotation guidelines1. 2 System Overview 2.1 Preprocessing As a first step of our approach we perform the fol- …

Content truncated to 500 words