Informer - BCS

InformerInternet Agent System), named after theGreek Messenger God Hermes’s wingedsandals, is a multi-agent system thatintegrates information gathering and filteringin the context of supporting a user to manageher/his Web interests.Research Update:New Light on Information AgentsBy Milan Stankovic and Uros KrcadinacAgent-based systems are one of the mostimportant and exciting areas of research anddevelopment that emerged in informationtechnology (IT) in the past two decades. In anutshell, an agent is a computer program thatis capable of performing a flexible,autonomous action in typically dynamic andunpredictable domains. In particular, aninformation agent is an agent that may accessone or multiple, distributed, andheterogeneous information sources available,and pro-actively acquires, mediates, andmaintains relevant information on behalf of itsuser(s) or other agents, preferably just-intime.The system provides each user with a personalagent, which periodically monitors the Websites that the user expressed interest in. Theagent informs its user about relevant changes,filtered by assumed user preferences anddefault relevance factors. One of the mostimportant characteristics of this agent-basedsoftware are its usability advantages. Humanagent interaction is completely realized viaemail, so that a user can converse with her/hisagent in plain natural language, whereas theagent heuristically interprets concreteinstructions from the mail text (e.g., “changesite list” or “kill yourself”). In order to improvecharacteristics of Web sites which do notprovide RSS feeds, one of the features of theT a l a r i a System is the creation of RSSfeeds based on the new and presumablyrelevant content of the Web sites that arebeing inspected.EMERGING TRENDSInformation agents emerged as a response tolarge and ever-increasing quantities ofavailable data. Blogs, Wikis, and other socialsoftware have added a new value to the Webby making it a source of information onvarious topics. New content is being addedcontinuously by users who actively publish newtexts and multimedia.Many Web users express need for gettingrelevant information from various Web sites,but do not have time to inspect them inregular basis. It is fair to say that RSS andmany available RSS readers enabled users toeasily track changes on constantly changingweb resources, and new tools, like YahooPipes, enabled them to filter and combinesources in order to retrieve information of theirinterest. However, there are many sites thatdo not provide a RSS feed, and among themthere are sites that contain valuableinformation.Figure 1 - TALARIA SystemT a l a r i a System is being mainly tested foracademic proposes, inspecting Web sites ofvarious faculty subjects and informing usersabout possible changes of schedules and examdates, and it proved to be of vital importanceto students. Human-like NLP-based interaction,autonomy-related aspects of this system, andacting on behalf of the user emphasize theusability advantages of this agent-basedsoftware.INFORMATION GATHERING MULTIAGENT SYSTEMIn order to solve this common problem ofuser-oriented information gathering, ourresearch group at the Department ofInformation Systems, FON - School of BusinessAdministration, University of Belgrade, Serbia,has developed a multi-agent system, calledT a l a r i a System(http://iis.fon.bg.ac.yu/talaria/). T a l a r i aSystem (The Autonomous Lookup And ReportInformerConclusionConcerning the fast development of the WorldWide Web, it is clear that there is anincreasing need for information gatheringservices, in order to help Web users find theright information just-in-time. We havedeveloped our system in order to meet someof those requirements, but it is just a small-4–Autumn 2007

Informerstep, and there is a lot of research to be done.Future research in the field of informationagents must take into an account all major ITtrends, such as the Semantic Web, Webservices, grid computing, autonomoussystems, and many more. It is our opinion thatall of these visions share the common need foragent-based technologies, and thatinformation agents (or something very similarto them) will continue to be one of the mostvivid areas of the research in the years tocome.Research Update:Scalable semantic personalisedsearch of spoken and writtencontents on the Semantic WebBy Javier TejedorThe continuous growthof contents in text andother media on theWeb and intranetsdemands a parallelprogress of informationaccess technologies. Inthis project wepropose:Uros Krcadinac is anundergraduate teachingassistant at the Departmentof Information Systems andTechnologies, FON - Schoolof Business Administration,University of Belgrade,Serbia. He is a member ofthe GOOD OLD AI researchgroup(http://goodoldai.org.yu/). His researchinterests include Intelligent Agents, Multimediaand Semantic Retrieval. He can be contactedat uros@krcadinac.com.(a) the development of a novel semanticsearch model, with ontology-based contentranking algorithms, improving andcomplementing the achievements in thisdirection in the Semantic Web field, andachieving measurable improvements in searcheffectiveness with respect to current keywordbased techniques;(b) the combination of personalisationtechniques with the semantic-basedrepresentation models, to achieveimprovements in the relative precision andrelevance of search results with respect to theparticular interests of individual users;Milan Stankovic is anundergraduate teachingassistant at theDepartment ofInformation Systems andTechnologies, FON School of BusinessAdministration, Universityof Belgrade, Serbia. He is a member of theGOOD OLD AI research group(http://goodoldai.org.yu/). His researchinterests include Intelligent Agents,Information Retrieval, and Social Software. Hecan be contacted viamilan.stankovic@gmail.com.Informer(c) the integration of text and voice contentsin a single access platform for large-scalerepositories; and(d) an experience in the realisation ofSemantic Web proposals, to contribute to itsadvancement, including the definition ofontologies, the semi-automatic construction ofknowledge bases, and a semi-automaticcontent annotation system, incorporatingspeech recognition and human languagetechnologies.A demonstrative application will be built in thenews domain, for the evaluation andrefinement of the developed techniques, in areal-world application setting. In this article, Iwill give a global vision of the project modulesand an overview of the most important speechrecognition techniques applied nowadays andwithin this project. Finally a conclusion will bealso presented.-5–Autumn 2007

Informerrelevant keywords contained within it.Normally, Large Vocabulary Continuous SpeechRecognition (LVCSR) systems outperform anyother kind of techniques, but this approachhas three main drawbacks, a) it is necessary avery high amount of data in order to train thesystem, b) it has a high computationally costand c) it is vocabulary dependant. In additionto this, a high knowledge about the languageis needed. In this project, not only LVCSRbased techniques are going to be applied butalso keyword spotting-based techniques arebeing investigated in order to search for areduced set of keywords which commonly donot appear in the vocabulary of LVCSRsystems such as proper names, acronyms, etc.GLOBAL VISION OF THE PROJECTMODULESThe scenery of the project will be as follows: Auser makes a query. This query is passed overa semantic query processing in which it isexecuted against a knowledge base (KB). Inthe KB, the concepts related to the applicationdomain are formally described by means ofontologies. The concepts which satisfy thequery are retrieved by this process and serveas an index to search for the keywords in anXML file annotated previously from the speechbased techniques. The concept-keywordmapping makes possible to transform theconcepts in the ontologies into keywords in thespeech recognition vocabulary. Finally, aranking strategy proposed to order thedocuments where the keywords appear willallow us to show the user the final output, inorder to he can listen to the speech contentand read the text documents according tohis/her query. A friendly user interface is alsobeing investigated within this project. Thewhole architecture can be seen in the nextFigure 1:These techniques take advantage of the littleamount of data necessary to train the systemand of a faster recognition process, as well asmay be applied over an opened vocabulary incontrast with the LVCSR system. Most of thekeyword spotters developed for years usedfiller or garbage models in order to deal withthe non-keyword segments (the Out-OfVocabulary (OOV) words) which appear in thespeech. These filler models vary from phonesto whole words. In this project, our fillermodels are composed by phones, phonemesand classes phonemes and no word modelsgarbage are required. Hidden Markov Models(HMM) are used as acoustic modelling andHTK toolkit is used to perform the recognitionprocesses. New confidence measures based onthe scores of the recognizer output and on alexical access module to build the keywordsfrom the sequence of phones in the output ofa phone recognition process are being used inorder to achieve the best possible systemperformance and a good compromise betweencorrect keywords and false acceptances. Thearchitecture which achieves the best resultsnowadays in the keyword spotting system isshown in next Figure 2:Figure 1. Global system architectureMore information about this project can befound in (http://nets.ii.uam.es/s5t).SPEECH RELATED TECHNIQUESAs it is well known, dealing with the audiocontent implies the development of speechtechnology-based techniques in order to findInformer-6–Autumn 2007

InformerCONCLUSIONSIn this project we aim to develop a systemwhich allows the user to make a query andretrieve the text and sound files which satisfyhis/her query. In this way, semantic web areaby means of ontologies, speech recognition bymeans of keyword spotting and LVCSRsystems and speech segmentation are going tobe investigated within it. It will also beinvestigated the user-friendly interface overwhich the final results will be presented to theuser.Figure 2. Keyword Spotting system architectureAs preliminary results, we have presented anarchitecture for the keyword spotting systemin http://jth2006.unizar.es/finals/4jth 106.pdf.The approach in this architecture performs aKeyword Spotting process to retrieve thesequence of keywords and filler models. At thesame time, the Phonetic Decoding retrievesthe allophones string related to the speechfiles. The Substring selection extracts theallophones string in the time intervals wherethe keyword spotting proposes a keyword.This string is passed through a Lexical accessmodule, which, from a previous confusionmatrix trained from the substitution, insertionand deletion allophones errors in this stringand from a dynamic programming algorithmproposes the most probable keyword relatedto this string. The final Confidence measureeliminates from the keywords proposed by thekeyword spotting module those which do notcoincide with the one proposed by the lexicalaccess and those whose score retrieved by thelexical access is higher than a threshold. Boththe threshold and the confusion matrix weretrained from a development set, different fromthe one used in the system evaluation.Update research outperforms the rateachieved with that architecture and presents a93% of keywords (concepts) detectedcorrectly and 11% of false acceptances withthe architecture presented in Figure 2, over 80keywords appearing 1672 times extracted fromthe clean speech Spanish Albayzin databasetest geographic corpus.Javier Tejedor Noguerales is a 28-old PhDstudent in the University Autonoma in Madrid(UAM). I am also an assistant professor in thatuniversity in the “Theory of signal andcommunications” department. I belong toHuman Computer Technology Laboratory(HCTLab) and my main research is focused oninformation retrieval of audio content (Audiomining) and my PhD-thesis is based on it. Thecontact e-mail is: javier.tejedor@uam.esBy other hand, normally the sound filesrecorded from the radio and televisionprograms, with which we have to deal withinthis project, do not only contain clean speechbut also noises, music, etc. In this way, therewill be lots of samples in which phenomenasuch as overlapping, noises, breathes andother effects occur. It causes errors in therecognition process. Therefore, the project willalso investigate speech segmentationtechniques and robust speech recognitiontechniques, e.g. those which try to separatethe clean speech of the noise, in order toimprove the search of the keywords within thisnews domain.Informer-7–Autumn 2007

Informercorpora. Since Spanish is one of the mostspoken languages in the world, it seemsappropriated to invest in research anddevelopment of resources and systems for thislanguage.Research Update:An approach to the extraction ofchronological information:Temporal Expression Recognitionand Normalization in SpanishTERN TaskBy María Teresa Vicente-DíezTemporal Expressions Recognition andNormalization (TERN) systems process inputdata from documents in a specific language (inthis case Spanish) identifying dates, durations,anchors and intervals in them. The processingof each document includes several stages as isdepicted in Figure 1, from the pre-processingof the source data to the return of the resultsin a standard format.Temporal informationextraction frommultimedia data sourcessupposes importantlinguistic challenges atthe time of identifyingand representingtemporal expressions.For instance, it isfrequently neededdealing with big amounts of informationobtained from news and electronic documents.Since it is usual that this kind of documents isscarcely temporally annotated, it is difficult tobe able to determine the time when the eventsof their contents occur. Dating of contents iscrucial in the Information Retrieval (IR) field.The extraction of chronological informationallows improving the results obtained bytypical IR applications, such as QuestionAnswering (QA) or Automatic TextSummarization. In QA it is fundamental tosolve references that can help finding theanswer to temporary questions (“What yeardid Mozart die?”) or to questions with timerestrictions (“Who was the president of theUSA in 2005?”).The TERN system in which I am currentlyworking on identifies dates and intervals in aRecognition phase carried out by aRecognizer module. Both absolute (April 3,2005) and deictic (yesterday) expressions aredetected. Afterwards, this system managesand returns them in a standard format thatavoids semantic ambiguities in their retrieval(Normalization phase).One of the aims of my PhD research is thedevelopment of technologies for the extractionof temporal information from different datasources, in order to apply them to eventordering of the contents, as well as to theintegration of a time reasoning system in a QAenvironment. In such a platform, it is expectedthat the introduction of inference rules allowsthe improvement of the question analysis aswell as the quality of the extracted answers.Figure 1 A TERN System Architecture ProposalSince there are different types of temporalexpressions, each one (absolute, deictic,intervals or those that directly suppose a pointin time (i.e.: Christmas, Halloween)), has to betreated independently. The Normalizer moduletakes into account this fact, so it counts on aDispatcher sub-module that identifies thedifferent cases and deals with them in a properway. An example of this classification, as wellas the obtained normalized output is presentedin Table 1.Spanish ResourcesThe scientific community counts with severalresources in the field of temporal expressionsmanagement but, in their majority, only theEnglish language is considered. Among them,there are several guides and methods ofannotation, as well as temporally annotatedInformer-8–Autumn 2007

InformerTable 1: An Example of the TemporalExpressions Classification OUTPUTABSOLUTEDD-MM-YYYY31-12-20052005-12-31[DET] DD ”de” MONTH ”de” YYYY[el] 31 dediciembre de2005([the] 31th ofDecember of2005)diciembre de2005(December of2005)hoy (today)2005-12-31MONTH “of” YYYYDEICTICPRESENTREFERENCEMODAL VERB“HACER” QUANTITY TIME UNIThace un mes(one monthago)(errors) objects is quite acceptable in themajority of the documents analyzed.ConclusionsCompletely integrating the TERN system into aQA system is the next step that must be donein order to check the improvement that itsupposes at the time of the IR.In the meanwhile, and in order to enlarge thenumber of temporal expressions properlyrecognized by the system the firstenhancement that should be carried out is thetagging of the recognized expressions(durations, sets of times, etc) in order tocapture as much semantic information aspossible. Likewise, the grammar coverage forthe recognition of temporal expressions that isconsidered by the system might be increased,adding other types of expressions that are notcurrently taken into account. Moreover, futurework includes the implementation ofdictionaries with a broader coverage of directlytranslatable temporal expressions, such asholidays, festivities, etc.2005-122007-06-012007-05-01This proposal was presented to the AutomaticContent Extraction (ACE) 2007 EvaluationWorkshop, participating in the TERN task forSpanish language. The task using Spanish datasupposed a pilot experience this year. Thiswork constitutes a preliminary approximationwhere simple manual rule-based techniquesare implemented both in recognition andnormalization phases. The system focuses onthe treatment of temporary expressionsfrequently appearing in the Spanish language,which should be labelled according to TIDESIn the same way, it is high-priority consideredthe research concerning context extractionmechanisms that facilitate the recognition ofdeictic temporal expressions.2005 Standard for the Annotation of TemporalExpressionsFinally, we are studying the introduction ofmachine learning techniques in temporalexpressions recognizing and classifying stages,so that they complement the current rulebased approach.[http://timex2.mitre.org/annotation guidelines/2005 timex2 standard v1.1.pdf]Testing the systemSome runs with several corpora, composed ofnews in Spanish, were made in order toevaluate the system. These results gave us aglobal estimation of the amount of expressionsdetected as well as the quality of the markeddetections. Results obtained are consideredquite successful since the number of spuriousexpressions (also know as false alarms)detected by thesystemrepresents alow percentageof the totaldetectionsmade. In thesame way, thetotal number ofmissing andMaría Teresa Vicente-Díez is a PhD student atthe Advanced Databases Group of theUniversity Carlos III of Madrid. Her research inthe field of Natural Language Processing isfocused on the recognition of temporalexpressions, and the resolution of theirtemporal reference, applied to event orderingand time reasoning, with the aim of improvingQuestion Answering Systems. She can becontacted by e-mail ��Autumn 2007

Informermultimedia content. In the case of videoannotation the terms of the ontologies areassociated to the individual elements of thevideo either manually of automatically,exploiting the resultResearch Update:Enriched Multimedia Ontologiesby using fuzzy conceptsBy Giuseppe SerraThe rapid increase of theavailable amount ofmultimedia informationhas revealed an urgentneed for developingefficient automatedmultimedia informationretrieval algorithms. Infact user increasingly desire to search oncomplex queries such as “Show me the videoclip where both George Bush and Tony Blairenter in a car” or “Show me the video clipwhere Maradona bounces a ball”.Moreover such annotation would not only helphuman user to query and manage their digit

review of "Natural Language Processing and Text Mining", by Kao and Poteet. By the way, we still have a copy of Martin White's "Making Search Work" available for review - if you're interested, just drop us a line at irsg@bcs.org . All the best, Tony Tony Rose, PhD MBCS CEng Editor, Informer Vice chair, IRSG Email: irsg@bcs.org.uk