@inproceedings{lang-lapata-11_unsupervised-semantic-role-discovery,
  title={Unsupervised semantic role induction with graph partitioning},
  author={Lang, Joel and Lapata, Mirella},
  booktitle={Proceedings of the Conference on Empirical Methods in Natural Language Processing},
  pages={1320--1331},
  year={2011},
  annote = {
  \textbf{\Large Overview }\\
  Semantic roles describe the relation that hold between a predicate and its arguments. Most of the previous work has focused on supervised methods which rely on human-annotated corpus of semantic roles like PropBank for model training. The problem is typically addressed in two parts, argument identification and argument classification. Current supervised approaches are limited to specific domains and do not perform well when tested on out-of-domain data.\\
  The idea in this paper is to design an unsupervised system by formalizing semantic role induction as graph partition problem. \\\\
  
  \textbf{\Large Architecture}\\
  Standard architecture of treating argument identification and argument classification separately has been adopted.\\
  It is assumed that input is syntactically analyzed in the form of dependency trees. The task of \textbf{argument identification} is reformulated as task of discarding as many non-semantic arguments as possible by using a set of rules. \\
  
  \textbf{Argument Classification}\\
  An undirected, weighted graph is constructed for each verb with verb argument instances as vertices. Edge weights are assigned based on some similarity measure which takes into account lexical similarity, syntactic position and occurence in same phrase as main criteria. A variant of chinese Whispers (Biemann, 2006) is used to realize graph partitioning.\\ Initially each vertex is assigned its own new label. Subsequently for each label, a score is computed by summming together the weights of edges to neighbouring vertices with that label and the label with maximal score is selected.\\ The algorithm is run for several iterations. While updating any vertex to a specific label, confidence of the update is assigned as the measure of average similarity to the neighbours with that label.\\ 
  The high conidence updates are prioritized to promote reliable updates in the initial phrases of algorithm.\\\\
  
  \textbf{\Large {Evaluation Metrics}}\\
  Cluster purity, collocation and their harmonic mean(F1) are used as evaluation metrics. The model is compared to the competitive supervised approach- \textbf{Syntactic Function Baseline} and other unsupervised approaches.\\\\
  
  \textbf{\Large {Results}}\\
  Experimental results on the \textbf{CoNLL 2008} benchmark dataset demonstrate that the model is competitive with other unsupervised approaches in terms of F1 whilst attaining significantly higher cluster purity.
  Compared to the Syntactic Function baseline,the Graph Partitioning algorithm has higher F1 on the auto/auto and auto/gold datasets but lags behind by 0.5 points on the gold/auto dataset and by 0.9 points on the gold/gold dataset.\\\\

  \textbf{\Large {Discussion}}\\
  The model is able to assign semantic roles in a conceptually and algorithmically simple unsupervised apporach. The model is domain independent and uses language information only to calculate similarity score which is used to assign weights to the edges in the graph. The approach is general and amenable to other graph partitioning algorithms.\\\\
  
  \textbf{\Large Future Work}\\
  The similarity function chosen in the paper is by necessity rudimentary because it cannot be estimated from data. Performance could be improved by developing a better similarity function. One such way can be using a weakly supervised setting, where the similarity function is estimated from a small amount of labeled instances. 
}}