<?xml version="1.0"?><!DOCTYPE article SYSTEM "/project/take/software/searchbench_offline_processing/paperxml_generator/aclextractor/src/python/../resource/dtd/paperxml.dtd"><article><header><firstpageheader><page local="1" global="592"/><title>A Hybrid Approach for Event Extraction and Event Actor Identification</title><pubinfo>Proceedings of Recent Advances in Natural Language Processing,pages 592-597, Hissar, Bulgaria, 12-14 September 2011.</pubinfo><author surname="Kolya" givenname="Anup Kumar"><org  name="Jadavpur University" country="India" city="Calcutta"/></author><author surname="Ekbal" givenname="Asif"><org  name="Jadavpur University" country="India" city="Calcutta"/></author><author surname="Bandyopadhyay" givenname="Sivaji"><org  name="Jadavpur University" country="India" city="Calcutta"/></author></firstpageheader><frontmatter><p><b>A Hybrid Approach for Event Extraction and Event Actor</b></p><p><b>Identification</b></p><p><b>Anup Kumar Kolya<footnote anchor="1"/>     Asif Ekbal<footnote anchor="2"/>    Sivaji Bandyopadhyay<footnote anchor="1"/></b></p><p>Computer Science and Engineering Department, Jadavpur University, India</p><p><footnote anchor="2"/> Patna (IITP), India</p><p>anup.kolya@gmail.com,  asif.ekbal@gmail.com,</p><p>sivaj i_cse_ju@yahoo.com</p></frontmatter><abstract>This paper, we propose an approach for <i>event extrac­tion </i>and corresponding <i>event actor </i>identification-within the TimeML framework. Firstly, for <i>event ex­traction, </i>we develop SVM based hybrid approach and for <i>event actor </i>identification the <i>baseline </i>model is developed based on the <i>subject </i>information of the dependency-parsed event sentences. Then we develop an unsupervised syntax based model that is based on the relationship of the event verbs with their argument structure extracted from the <i>head </i>information of the chunks in the parsed sentences. Evaluation on a col­lection of TempEval-2 corpus shows the precision, recall and F-measure values for the <i>baseline </i>model as 64.31%, 67.74% and 65.98%, respectively and the syntax based model as 69.12%, 66.90% and 67.99%, respectively. </abstract></header><body><section number="1" title="Introduction"><p>New sources of textual information, rich in events, grow significantly, such as social net­works, blogs, and wikis. They are added to old sources like the informative web sites, emails and forums, which shows the importance to manage these data automatically. One of the im­portant tasks of text analysis clearly requires identifying events described in a text and locat­ing these in time. Event extraction has emerged to be very important in improving complex nat­ural language processing (NLP) applications such as automatic summarization (Daniel et al., 2003) and question answering (QA). TimeML (Pustejovsky et al., 2003) presented a rich speci­fication for annotating events in NL text extend­ing the features of the previous one.</p><p>This paper is focused on the TimeML view of events. TimeML defines events as situations that <i>happen or occur, </i>or elements describing <i>states or circumstances </i>in which something ob­tains or holds the truth. These events are gener­ally expressed by tensed or un-tensed verbs, nominalizations, adjectives, predicative clauses or prepositional phrases. The 2007 TempEval challenge attempted to address this question event extraction task was introduced as task B. Let us consider a sentence like, <i>BAGHDAD, Iraq (AP) _ an American leader of a U.</i><i>N. weapons inspection team <b>resumed work </b>in Iraq Friday, nearly two months after his team was effectively <b>blocked.</b></i></p><doubt alpha="50.0" length="44" tooSmall="False" monospace="0.0">(Boguraev et al, 2005). In 2010, TempEval-2,</doubt><p>Sentence 1 has three events, namely <b><i>'re­sumed, 'work </i></b>and <b>'blocked'. </b>In this sentence <i>resumed </i>and <i>blocked </i>can be considered as verbal events but <i>work </i>is an nonverbal event. Generally, verbal or non-verbal event are executed by some abstract entities, directly or indirectly. Entities are basically person, organization or location.</p></section><section number="2" title="Event Extraction"><p>Below we present our hybrid approach for event extraction. The system is based on a supervised machine learner, Support Vector Machine (SVM). It makes use of the various features ex­tracted from the TimeML corpus. In order to im­prove the performance of the system, we incor­porate the knowledge of semantic role labeling, WordNet and several heuristics.</p><subsection number="2.1" title="SVM based Approach"><p>Initially, we started with the development of an event extraction method based on SVM. This is used as the <i>baseline </i>model. The SVM system is developed based on (Valdimir, 1995), which per­form classification by constructing a N-dimensional hyperplane that optimally separates data into two categories. We use <i>YamCha </i>tool-kit<footnote anchor="1"/>, a SVM-based tool for detecting classes in documents and formulating the event extraction task as a sequential labeling problem.<page local="2" global="593"/> Here, the <i>pair wise </i>multi-class decision method and <i>poly­nomial kernel function </i>are used. We use Ti-nySVM-0.0<footnote anchor="2"/> classifier for classification.</p><footnote label="1"> http://chasenorg/~taku/software/yamcha</footnote><p>We extract the gold-standard TimeBank fea­tures for events in order to train/test the SVM model. We mainly use the various combinations of <i>part of speech </i>(PoS), <i>event tense, event as­pect, event polarity, event modality, event stem </i>and <i>event class </i>features.</p></subsection><subsection number="2.2" title="Use of Semantic Roles for Event Extrac­tion"><p>We use Semantic Role Label (SRL) (Gildea et el, 2002; Sameer et al, 2004) to identify different features of the sentences of a document. These features help us to extract the events from the text. In the present work, we use predicate as an event. Semantic roles can be used to detect the events that are the nominalizations of verbs such as <i>agreement </i>for <i>agree </i>or <i>construction </i>for <i>con­struct. </i>Event nominalisations (or, <i>deverbal nouns) </i>are commonly defined as nouns, morpho­logically derived from verbs, usually by suffixa­tion (Quirk et al., 1985). Let us consider the fol­lowing example sentence to understand how se­mantic roles can be used for event extraction. The output of SRL for this sentence is as fol­lows:</p><p><i>[ARG1 All sites] were [TARGET inspected] to the satisfaction of the inspection team and with full cooperation of Iraqi authorities, [ARG0 Da-cey] [TARGET said]</i></p><p>A sentence is scanned as many times as the number of target words in the sentence. In the first traversal, <i>inspected </i>is identified as the event. In the second pass, <i>said </i>is identified as an event. All the extracted target words are treated as the event words. We observed that many of these target words are identified as the event expres­sions by the SVM model. But, there exists many nominalised event expressions (i.e., <i>deverbal nouns) </i>that are not identified as events by the supervised SVM. These nominalised expressions are correctly identified as events by SRL. We observe performance improvement with the in­clusion of this module.</p></subsection><subsection number="2.3" title="Use of WordNet for Event Extraction"><p>WorldNet (Miller, 1990) is mainly used to iden­tify <i>non-deverbal event nouns. </i>We observed from the outputs of SVM and SRL that the event entities like 'war', <i>'attempt^, Ltouf </i>etc. are not properly identified. These words have noun PoS categories, and the SVM along with SRL can only identify those event words that are verbs. We know from the lexical information of Word­Net that the words like <i>'war' </i>and <i>'tour' </i>are gen­erally used as both <i>noun </i>and <i>verb </i>forms in the sentence. We design two following rules based on the WordNet:</p><footnote label="2">( http://cl.aist-nara.ac.jp/~taku ku/software/TinySVM)</footnote><p><b>Rule 1: </b>The word (for example <i>war) </i>tokens hav­ing noun PoS categories are looked into the WordNet. If it appears in the WordNet with noun and verb senses, then that word token is also considered as an event.</p><p><b>Rule 2: </b>The <i>stems </i>of the noun word tokens are looked into WordNet. If one of the WordNet senses is verb then the token will be identified as verb.</p><p>We observe significant performance im­provement on event extraction with the above mentioned two rules.</p></subsection><subsection number="2.4" title="Use of Rules for Event Extraction"><p>We used WordNet to extract the event expres­sions that appear in the WordNet with both noun and verb senses. Here, we mainly concentrate to identify the specific lexical classes like ' <i>inspec­tion </i>and <i><footnote anchor="1"/> resignation'. </i>These can be identified by the suffixes such as ('-ciôn'), ('-tion') or ('­<i>ion'), </i>i.e. the morphological markers of deverbal derivations.</p><p>Initially, we run the SVM based Stanford Named Entity (NE) tagger<footnote anchor="3"/> on the TempEval-2 test dataset. The output of the system is tagged with <i>Person, Location, Organization </i>and <i>Other </i>classes. The words starting with the capital let­ters are also considered as NEs. Thereafter, we came up with the following rules for event ex­traction:</p><p><b>Rule-1: </b>The morphologically deverbal nouns are usually identified by the suffixes like <i>'-tion\ </i>'­<i>ion', </i>' <i>-ing' </i>and <i>'-ed' </i>etc. The non NE nouns but ends with these suffixes are considered as the event words.</p><p><b>Rule-2: </b>After searching verb-noun combination from the test set, non-NE noun words are consid­ered as the events.</p><p><b>Rule- 3: </b>The non-NE nouns occurring after ( i) the complements of aspectual PPs headed by prepositions (ii) any time-related verbs (iii) cer­tain expressions are considered as events.</p><footnote label="3"> http://nlp . stanford. edu/software/CRF-NER. shtml</footnote><page local="3" global="594"/></subsection><subsection number="2.5" title="Evaluation Results"><p>We use the TempEval-2010 datasets to report the evaluation results. We develop a number of SVM models depending upon the various features in­cluded into it. We have a training data in the form <i>(Wi, Ti) </i>, where, <i>Wi</i><i> </i>is the <i>ith </i>pair along with its feature vector and <i>Ti </i>is its correspond­ing output label (i.e., <i>Event </i>or <i>Other).</i><i> </i>Models are built based on the training data and the feature template. We used different feature combinations within the context of previous 3 and next 3 words. The test data had 373 verbal and 125 non-deverbal event nouns. Overall evaluation results are reported in Table 1. The SVM based system shows the precision, recall and F-measure values of 75.8%, 78.5% and 77.13%, respectively. The performance increases by almost 1.59 percentage F-measure points with the use of semantic roles. Table 1 shows very high performance improve­ment (i.e., 10.98%) with the use of WordNet. The rule-based component also shows the effec­tiveness with the improvement of 5.37 F-measure percentage points. Finally, the system achieves the precision, recall and F-measure val­ues of 93.00%, 96.00% and 94.47%, respectively. This is actually an improvement of approxi­mately 12% F-measure value over the best re­ported system.</p></subsection></section><section number="3" title="Event Actor Identification"><p>In this section, we detail our method for event actor identification.</p><subsection number="3.1" title="Subject Based Baseline Model"><p>We have preprocessed the TempEval-2 corpus for identifying the actors of the events. We have previously (Kolya et al. 2010a; Kolya et al 2010b) worked on the various problems of event and temporal relation identification such as (i). Event-time and event-documentation creation time (DCT) temporal relation (TE) identification in the same sentence, (ii). Even-event temporal relation identification in two consecutive sen­tences and (iii). Subevent-subevent temporal re­lation identification in the same sentence on the Tempeval-1 and Temeval-2 corpus. We have observed from this experience that almost all events are involved with the actors, either active or passive. Actually, event actions are done by someone or somebody is doing this kind of ac­tion. Event actions involve with person, organi­zation and sometimes with location also. In the present attempt, we consider the approaches that were conducted for identifying emotion holders (Das and Bandyopadhyay, 2010). Thereafter, we came up with the following heuristics for actor identification, (i) we discard the non-event sen­tences, i.e. those sentences that don't contain any event entity. (ii) If multiple events exist in any sentence, then all the events will have the same actors. Once an actor is identified for any event, it is assigned to the other event as well. (iii) If there are multiple actors and events, then &lt;event, actor&gt; pairs are formed by considering an event and its closest possible actor in the sentence. All the events may not have an active actor. The ac­tor may be passive also. For example, consider the following sentences:</p><p><b>Table 1. </b>Evaluation results of event extraction <i>1.</i><i> This time a &lt;bomb/&gt; at an abortion clinic.</i></p><p>2. <i>Plates &lt;recovered/&gt; at the Olympic park bombing &lt;appear/&gt; to &lt;match/&gt; those &lt;found/&gt; at the abortion clinic &lt;bombing/&gt; in Atlanta.</i></p><p><b>Corpus Preparation: </b>We did not have any gold standard corpus for event actor identifica­tion. We have used the Temeval-2 corpus as a gold standard event actor corpus by manually annotating event actors in each sentence. The gold corpus looks as follows:</p><p>&lt;eActor&gt; People &lt;/eActor&gt; have &lt;predicted/&gt; his &lt;demise/&gt; so many times , and the &lt;eAc-tor&gt;US&lt;eActor&gt; has &lt;tried/&gt; to &lt;hasten/&gt; it on sev­eral occasions .</p><p>Here, a <i>"People " </i>is the event actor of both events <b>predicted </b>and <b>demise, </b>and "US'" is the event actor of the events, <b>tried </b>and <b>hasten. </b>This corpus has 11 documents, 156 sentences and 459 events.</p><p><b>3.2. Baseline Model based on Dependency Parsing and Subject Extraction</b></p><p>Stanford Parser (de Marneffe et al,2006), a prob­abilistic lexicalized parser containing 45 different PoS tags of Pen Tree bank is used to get the parsed sentences with dependency relations.<page local="4" global="595"/> The input event sentences are passed through the parser. The dependency relationships extracted from the parsed data are checked for predicates <i>"nsubj" </i>and <i>"xsubj" </i>so that the <i>subject </i>related information in the <i>"nsubj" </i>and <i>"xsubj" </i>predicate are considered as the probable candidate for identifying the event actor. Other dependency relations are filtered out from the parsed output. The present system is developed based on the filtered subject information only. An example sentence is noted below whose parsed output and dependency relations are shown. Here, the <i>"nsubj" </i>relations containing the event word "en­dures" tags "eActor" as an event actor. <i>"Time and again, he <b>endures."</b></i></p><table class="main" frame="box" rules="all" border="1" regular="False"><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>Model</p></td><td class="cell"><p>preci­sion</p></td><td class="cell"><p>recall</p></td><td class="cell"><p>F-measure</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>SVM</p></td><td class="cell"><p>75.80</p></td><td class="cell"><p>78.50</p></td><td class="cell"><p>77.13</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>SVM+SRL</p></td><td class="cell"><p>77.20</p></td><td class="cell"><p>80.30</p></td><td class="cell"><p>78.72</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>SVM+SRL+W ordNet</p></td><td class="cell"><p>89.30</p></td><td class="cell"><p>90.10</p></td><td class="cell"><p>89.70</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"><p>SVM + SRL +</p><p>WordNet + Rules</p></td><td class="cell"><p>93.50</p></td><td class="cell"><p>96.70</p></td><td class="cell"><p>95.07</p></td><td class="cell"></td></tr><tr class="row"><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td><td class="cell"></td></tr></table><doubt alpha="41.4" length="128" tooSmall="False" monospace="0.0">(ROOT   (S     (S       (UCP     (NP (NNP Time))(CC and)(ADVP  (RB again))))(, ,)   (NP  (PRP he))   (VP  (VBZ endures)) (. .)))</doubt><doubt alpha="56.1" length="139" tooSmall="False" monospace="0.0">ccomp       (endures-6, Time-1),advmod (Time-1,   again-3),conj   and (Time-1, again-3),ccomp (endures-6, again-3)nsubj   (endures-6, he-5)</doubt><p>This <i>baseline </i>model is evaluated on the gold standard holder annotated an emotional sentence that has been extracted from the VerbNet. Total 156 sentences are evaluated and evaluation re­sults are presented in Table 2. So, the next step is to explore the syntactical way for identifying argument structure of the sentences for their cor­responding emotional verbs and to capture the emotion holder as a <i>thematic role </i>respectively.</p></subsection><subsection number="3.3." title="Syntax Based Model"><p>The syntax of a sentence is an important clue to capture the event actor inscribed in text. More specifically, the argument structure or subcatego-rization information for a verb plays an impor­tant role to identify the event actor from an event sentence. A subcategorization frame is a state­ment of what types of syntactic arguments a verb (or an adjective) takes, such as objects, infini­tives, that-clauses, participial clauses, and sub-categorized prepositional phrases (Manning et al. 1993). VerbNet (Kipper-Schuler et al, 2005) is the largest online verb lexicon with explicitly stated syntactic and semantic information based on Levin's verb classification (Levin et al 1993). It is a hierarchical domain-independent, broad-coverage verb lexicon with mappings to other lexical resources such as WordNet (Miller et al, 1990), XTAG (2001) and FrameNet (Baker et al, 1998). We use VerbNet throughout this experi­ment for identifying the event actors. The exist­ing syntax for each event verb is extracted from VerbNet and a separate rule based argument structure acquisition system is developed in the present task for identifying the event actor. The acquired argument structures are compared against the extracted VerbNet frame syntaxes. If the acquired argument structure matches with any of the extracted frame syntaxes, the event actor corresponding to each event verb is tagged with the actor information in the appropriate slot in the sentence.</p><p><b>Syntax Acquisition from VerbNet</b>:VerbNet associates the semantics of a verb with its syntac­tic frames and combines traditional lexical se­mantic information such as thematic roles and semantic predicates, with syntactic frames and selectional restrictions. Verb entries in the same VerbNet class share common syntactic frames, and thus they are believed to have the same syn­tactic behavior. The VerbNet files containing the verbs with their possible subcategorization frames and membership information are stored in XML file format.</p><p>&lt;THEMROLES/&gt; &lt;FRAMES&gt; &lt;FRAME&gt; &lt;DESCRIPTION descriptionNumloved to write.&lt;/EXAMPLE&gt; &lt;SEMANTICS&gt; &lt;PRED value<b>="event_state"&gt;</b> &lt;ARGS&gt; &lt;ARG type="Event" value="E"/&gt; &lt;ARG type<b>="VerbSpecific" </b>value<b>="Event"/&gt; </b>&lt;ARG type="ThemRole" val-ue="Passive"/&gt; ...</p><doubt alpha="65.5" length="29" tooSmall="False" monospace="0.0">ber="8.1" primary=MTO-INF-SCM</doubt><doubt alpha="53.8" length="39" tooSmall="False" monospace="0.0">secondary="" xtag="0.1"/&gt;....&lt;EXAMPLE&gt;I</doubt><doubt alpha="64.9" length="37" tooSmall="False" monospace="0.0">&lt;SYNTAX&gt;     &lt;NP value="Experiencer"&gt;</doubt><doubt alpha="61.1" length="18" tooSmall="False" monospace="0.0">&lt;SYNRESTRS/&gt; &lt;/NP&gt;</doubt><doubt alpha="61.5" length="26" tooSmall="False" monospace="0.0">&lt;VERB/&gt; &lt;NP value="Theme"&gt;</doubt><p>&lt;/ARGS&gt; &lt;/PRED&gt; &lt;/SEMANTICS&gt; &lt;/FRAME&gt;..</p><p>The XML files of VerbNet are preprocessed to build up a general list that contains all member verbs and their available syntax information re­trieved from VerbNet. This preprocessed list is searched to acquire the syntactical frames for each event verb. One of the main criteria consid­ered for selecting the frames is the presence of <i>"event state" </i>type predicate associated with the frame semantics.</p><p><b>Argument Structure Acquisition Frame­work: </b>To acquire the argument structure for a sentence, two separate approaches, Methods A and B, have been used, one (Method A) is from the parsed result directly and another (Method B) is from the PoS tagged and chunked sentences accordingly.<page local="5" global="596"/> The parsed event sentences are passed through a rule based <i>phrasal-head </i>extrac­tion process to identify the phrase level argument structure of the sentences corresponding to the event verbs. The extracted <i>head part </i>of every phrase from the well-structured bracketed parsed data is considered as the component of the argu­ment structure. For example, the <i>head </i>parts of the phrases are extracted to make the phrase level pattern or argument structures of the following sentences.</p><p>Sentence1: <i>"Ram killed Shyam with a knife." </i>Parsed Output:</p><doubt alpha="57.4" length="94" tooSmall="False" monospace="0.0">(ROOT (S(NP(NNP Ram))(VP(VBD killed)(NP(NNS Shyam))(PP(INwith)(NP (DT a) (NN knife)))) (. .)))</doubt><p>Acquired Argument Structure: [NP VP NP PP-with]</p><p>Simplified Extracted VerbNet Frame Syntax: [&lt;NP value="Actor"&gt; &lt;VERB/&gt; &lt;NP <i>patient&gt;</i></p><doubt alpha="65.0" length="20" tooSmall="False" monospace="0.0">&lt;PREP value="with"&gt;]</doubt><p><b>Event Actor for Event Verbs-The role of Subject and Syntax: </b>It is to be mentioned that the phrases headed by "S" (sentential comple­ment), "PP" (Preposition Phrase), "NP" (Noun</p><p>Phrase) followed by the event verb phrase con­tribute in structuring the syntactical argument. One tag conversion routine has been developed to transform the POS information of the system­generated argument structure for comparison with the POS categories of the VerbNet syntax. It has been observed that the phrases that start with ADJP, ADVP (adjective, adverbial phrases) tags generally do not contribute towards valid argument selection strategy. But, the entities in the slots of active frame elements are added if they construct a frame that matches with any of the extracted frames from VerbNet. The <i>head </i>part of each phrase with its component attributes (e.g. "with" component attribute for "PP" phrase) in the parsed result helps in identifying the maximum matching possibilities. Another alter­native way to identify the argument structure from a sentence is carried out based on the PoS tagged and chunked data. The PoS tagged sen­tences are passed through a Conditional Random acquire chunked data where each component of the chunk is marked with <i>beginning </i>or <i>interme­diate </i>or <i>end </i>corresponding to the elements slot in that chunk. The POS of the <i>beginning </i>part of every chunk are extracted and frames are devel­oped to construct the argument structure of the sentence corresponding to the event verb. The acquired argument structure of a sentence is mapped to all of the extracted VerbNet frames. If a single match is found, the slot devoted for the actor in VerbNet frame is used to tag in the ap­propriate slot in the acquired frame. For exam­ple, the argument structure acquired from the following chunked sentence is "NP-VP-NP".</p><doubt alpha="63.8" length="47" tooSmall="False" monospace="0.0">Field (CRF) based chunker (Phan et al, 2006) to</doubt><p>But, it has been observed that this second sys­tem suffers from the inability to recognize argu­ments from adjuncts as the system blindly cap­tures <i>beginning </i>parts as arguments whereas they are adjuncts in real. So, this system is biased to the <i>b ginning </i>chunk.</p></subsection><subsection number="3.4." title="Evaluation"><p>The evaluation of the baseline system is straight­forward. The event actor annotated sentences are extracted from the VerbNet and the sentences are passed through the baseline system to annotate the sentences with their <i>subject </i>based actor tag accordingly. Evaluation with 156 sentences is shown in Table 2. It is observed that the <i>subject </i>information helps in identifying event actor with high <i>recall. </i>But, the actor identification task for passive sentences fails in this <i>baseline </i>method and hence there is a fall in <i>precision </i>value. Two types of unsupervised rule based methods have been adopted to acquire the argument structure from the event sentences. It has been observed that, the Method-A that acquires argument struc­ture from parsed result directly outperforms the Method-B that acquires these structures from PoS tagged and chunked data. The <i>recall </i>value has decreased in Method-B as it fails to distin­guish the arguments from the adjuncts. The event actor identification system based on argument structure directly from parsed output gives satis­factory performance.</p><p><b>Table 2. </b>Evaluation results of actor identification Type    Baseline Model     Syntactic Model _<u>Method A   Method B</u><page local="6" global="597"/></p><doubt alpha="25.7" length="35" tooSmall="False" monospace="0.0">Precision    64.31      69.12 64.05</doubt><doubt alpha="25.5" length="55" tooSmall="False" monospace="0.0">Recall 67.74 66.90 65.52 F-measure 65.98    67.99 64.78</doubt></subsection></section><section number="4" title="Conclusion"><p>In this paper, we have reported our work on event extraction under the TempEval -2010 eval­uation exercise. Initially, we developed a SVM based supervised system in conjunction with number of techniques based on SRL, WordNet and handcrafted rules for event extraction. We then identify the actors for the events based on the roles associated to <i>subject </i>information. The syntactic way of developing the actor extraction module by focusing on the role of arguments of the event verbs improves the result significantly.</p><p>Future works include the identification of more precise rules for event identification and multiword events. The actor-annotated corpus preparation from VerbNet especially for event verbs followed by the argument extraction mod­ule can be further explored through the help of machine learning approach.</p></section><section title="Acknowledgments"><p>The work is partially supported by a grant from English to Indian language Machine Transla-tion(EILMT) funded by Department of Informa­tion and Technology (DIT), Government of In­dia.</p></section><references><p>Boguraev, B., Ando, R-K. 2005. <i>TimeBank-DrivenTimeML Analysis. Annotating, Extracting and Reasoning about Time and Events </i>2005.</p><doubt alpha="61.2" length="98" tooSmall="False" monospace="0.0">Baker, C.F., Fillmore, C.J., Lowe, J.B.:The Berkeley FrameNet project.COLING/ACL, pp. 86-90 (1998)</doubt><p>Daniel, Naomi, Dragomir Radev, and Timothy Alli-son.2003. <i>Sub-event based multi-document summa-riza-tion. </i>HLT-NAACL Text summarization work­shop, pages 9-16.</p><doubt alpha="66.3" length="178" tooSmall="False" monospace="0.0">Das Dipankar and Sivaji Bandyopadhyay. 2010. Emo­tion Holder for Emotional Verbs-The role of Subject and Syntax. CICLing- 2010), A. Gelbukh (Ed.), LNCS 6008, pp. 385-393, Romania</doubt><doubt alpha="64.0" length="50" tooSmall="False" monospace="0.0">De Marneffe, M.-C., MacCartney, B., Manning, C.D.:</doubt><p><i>Generating Typed Dependency Parses from Phrase</i> <i>Structure Parses.</i><i> </i>LREC (2006)</p><p>Gildea, D. and D. Jurafsky. 2002. <i>Automatic Labeling of Semantic Roles. </i>Computational Linguistics, 28(3):245-288.</p><p>Kipper-Schuler, K.: VerbNet: <i>A broad-coverage, comprehensive verb lexicon. </i>Ph.D. the-sis,Computer and Information Science Dept., Uni­versity of Pennsylvania, Philadelphia,PA (2005)</p><p>Kolya, A., Ekbal, A. and Bandyopadhyay, S. 2010. JU_CSE_TEMP: <i>A First Step towards Evaluating Events, Time Expressions and Temporal Relations. </i>SemEval, ACL, July 15-16, Sweden, pp. 345-350.</p><p>Kolya, A., Ekbal, A. and Bandyopadhyay, S. (2010a). <i>Event-Time Relation Identification using Machine Learning and Rules. </i>In Proceedings of 13th Inter­national Conference on Text, Speech and Dialogue, 2010, pp. 114-120.</p><p>Kolya, A., Ekbal, A. and Bandyopadhyay, S. (2010b). <i>Identification of Event-Time Relation: </i>A CRF based approach. ICCPOL 2010, USA, PP.63-66.</p><p>Levin, B.: English Verb Classes and Alternation: <i>A Preliminary Investigation. </i>The University of Chi­cago Press, Chicago (1993)</p><p>Manning, C.D.: <i>Automatic Acquisition of a Large Subcategorization Dictionary from Corpora.In: </i>31st Meeting of the ACL, Columbus, Ohio, pp.</p><doubt alpha="0.0" length="14" tooSmall="False" monospace="0.0">235-242 (1993)</doubt><doubt alpha="66.7" length="108" tooSmall="False" monospace="0.0">Miller, G.A.: WordNet:An on-line lexical database.International Journal of Lexicography 3(4), 235­312 (1990)</doubt><p>Pustejovsky, James, Jos'e M. Casta~no, Robert In-gria,Roser Saur'?, Robert Gaizauskas, Andrea Setzer, andGraham Katz. 2003. <i>TimeML: Robust Specification of Event and Temporal Expressions</i> <i>in Text.</i><i> </i>In IWCS-5.</p><doubt alpha="64.4" length="73" tooSmall="False" monospace="0.0">Phan, X.-H.: CRFChunker:CRF English Phrase Chunker.In: PACLIC 2006 (2006)</doubt><p>Quirk, R., S. Greenbaum, G. Leech, and J. Svart-vik.1985. <i>A Comprehensive Grammar of the Eng­lish Language. </i>Longman.</p><p>Sameer S. Pradhan, Wayne Ward, Kadri Hacioglu, James H. Martin, Daniel Jurafsky, <i>Shallow Seman­tic Parsing using Support  Vector Machines.</i></p><p>HLT/NAACL-2004, Boston, MA, May 2-7, 2004.</p></references></body></article>