Assembling Actor-based Mind-Maps from Text Stream
AAssembling Actor-based Mind-Maps from Text Streams
Claudine Brucks and Christoph SchommerUniversity of Luxembourg, Campus Kirchberg.Dept. of Computer Science and Communication, ILIAS Laboratory6, Rue Richard Coudenhove-Kalergi, L-1359 LuxembourgEmail: { claudine.brucks, christoph.schommer } @ uni.luOctober 22, 2018 Abstract
For human beings, the processing of text streams of unknown size leads generallyto problems because e.g. noise must be selected out, information be tested for itsrelevance or redundancy, and linguistic phenomenon like ambiguity or the resolutionof pronouns be advanced. Putting this into simulation by using an artificial mind-mapis a challenge, which offers the gate for a wide field of applications like automatic textsummarization or punctual retrieval. In this work we present a framework that is afirst step towards an automatic intellect. It aims at assembling a mind-map based onincoming text streams and on a subject-verb-object strategy, having the verb as aninterconnection between the adjacent nouns. The mind-map’s performance is enrichedby a pronoun resolution engine that bases on the work of [9].
A text stream is a data flow that is lost once it is read. Such a stream occurs very oftenin practice, for example while reading a text or listening to a story, probably told bysomeone else. In both cases, human beings store the major incidents rather associative.First, they remove noise and then extract information out of it, which can either berelevant or redundant/obvious. Then, relevant information is connected very adaptively,meaning that if the same information is read or listened again, the association betweenco-occurred words increases (or decreases, in case it is not). With such a performance,inconsiderable information gets lost whereas important facts can be kept. This is quiteimportant, because a constructive processing - like the generation of a summarisation ofthe text and a retrieve of contents - becomes manageable.Incremental-adaptive mind-maps serve in a similar way as they simulate such a human per-formance: through their associative, incremental, and adaptive architecture they processincoming data streams, adapt internal structures depending on the given input, strengthenor weaken internal connections, and send longer-established connections to a simulating1 a r X i v : . [ c s . C L ] O c t hort- and/or long-term memory. In this respect, we base on a work given by [14] thatargues for a real-time approach for finding associative relationships between categoricalentities from transactional data streams. Technically, these categorical entities are rep-resented as connectionist cells while associations are represented by links between them.These links may become stronger over time or degrade, according to whether the associa-tion re-occurs after a while or not is observed for a while. The work suggests a three-layerarchitecture: in the first layer, the short-term memory treats the incoming signals and con-structs the associations. The second layer, which is called the long-term memory , storesassociations that have a strong connection and that may be useful for a further analysis.The last layer, the action layer serves as a communication interface with the user overwhich he can consult the actual state of the system and interact with it.The generation of such a mind-map becomes complicated by the fact that the incoming textcan be corrupt or even ambiguous. For example, pronouns produce an ambiguity betweenexisting/referenced persons in the text: having The President of United States has saidthat . . . and a succeeding
Furthermore, he has mentioned that . . . leads undoubtedlyto the same person but the recognition of such relationships is not natural. If we keepsuch relationships unsolved, the mind-map can become ineffective or even wrong. In thisrespect, a meaningful part of the intended mind-map described in this work concerns withthe resolution of pronouns. For this, we are inspired by some earlier work, notably asyntax based approach [11]. All possible candidates for a pronoun are evaluated on a setof salience factors, as for example recency or subject emphasis. The candidate with thehighest salience weight will be chosen as antecedent. [12] presents a similar approach wherethe candidates are evaluated on indicators, but no syntactic or semantic information on thesentence are needed. Furthermore, the mind-map concerns with a temporal managementof text streams to construct an actor-based structure.
The motivation of pronoun resolution for the semantic network learning is to find thecorrect antecedent for each pronoun. This is important to construct complete mind-mapsfor each actor in a text. For this, the text stream is treated by a sliding window, whichfirst buffers and processes a certain number of sentences with the consequence that theinformation - once it is read - gets lost. For each sentence that is in the sliding window,a pre-defined subject-verb-object structure is instantiated and arranged in a semantic net-work structure, having concepts and connections between them. The connections becomestronger or weaker according to the underlying text stream, i.e., the occurrence of thesubject-verb-object instantiation.Figure 1 shows the general architecture of the mind-map. First, the complete text, i.e.,each sentence, is preprocessed, which is done in order to get syntactic and semantic infor-mation out of the text to further treat the input. In fact, pronouns are used as substitutesfor nouns in a text. As an example, the pronoun he refers back to Harry in a sentence like2igure 1: Architecture
Harry goes to the zoo where he looks at the beautiful animals . Then, a predefined structureof subject(s), verb(s) and object(s) is extracted from each sentence as well as the adjacentadjective(s) of both subjects and objects. All these extracted elements are in fact theessence of the sentence. Finally, the co-reference resolution focuses on merging conceptsthat relate to the same content. As an example, the concepts
President Washington and
George Washington relate both to the same person. However, the co-reference resolutionis limited to the actors of the text.
Following our experiences and looking back at the most important concepts for each cat-egory of the text - where most important refers to those that have the most outgoingedges - we have observed that these concepts are generally the actors of the stories (thisis in respect to stories) whereas for biographies and news articles, the most importantconcept is the person the biography or news text is about. In scientific texts, the actorsare often not the most occurring actors. In respect to the structures that occur multipletimes inside a text stream, one can observe that most of all subject-verb structures reoccurmore often than subject-verb-object structures. Those subject-verb structures that occurmultiple times mostly contain a verb of cognition or communication as for instance: say , think or explain .In concern of the accuracy of the pronoun resolution - that is how many pronouns are3orrectly or wrongly resolved and even remain unresolved (see Table 1) - we have observedthat the resolution results applied to pronouns given in third person singular are rathersuccessful. For this, we have used texts from different domains, i.e., fairy tales, newsarticles, biographies and scientific articles. Only the resolution of it and they lead to aninsufficient accuracy, which demand for an alternative method. pronoun correct (%) false (%) ? (%)he his him himself
100 0 0 she her herself
100 0 0 it its
44 56 0 itself - - - they their them themselves
100 0 0 I my
60 40 0 me
75 25 0 myself
100 0 0Table 1: Resolving the pronouns: correct, wrong, and unresolved.
In concern of the implementation, we use a graphical user interface, on which the user canoperate, for example to fix the window size, to fix the actors in the text, and to look atthe different outputs of the program - as for example the different sub-mind-map relatedto each actor, diverse actor statistics. For the preprocessing of the text streams, we stillneed • the tagged text, which permits to filter out all the nouns, proper nouns and pronouns. • the parse tree, which gives more information about the constituents of each sentence,as for example the clauses. 4 the grammatical relations between the single words of a sentence, relating for exam-ple a subject noun with its corresponding verb.Figure 2: Mind-map for John sees the yellow lion .With this, the pronoun resolution works as illustrated in the selective examples: • he/she : we take the last male/female noun or name occurring before the pronounthat acts as a subject in the sentence. If there is none, we take the last male/femalenoun or name before the pronoun. • they : we look back at the last two sentences and take the last plural before thepronoun. Plurals remain either plural nouns (e.g. the women , the children , the cars )or noun phrases containing nouns connected by and or , (e.g. John and Paul ). • it : we detect if it is pleonastic or not. If it is pleonastic, it has no antecedent as forexample in the phrase: It can be seen that ... ). This is done with the help of a setof some fixed sentence structure patterns (taken from [3]). If it is not pleonastic,we take the last non-living object occurring before the pronoun which is part of anon-prepositional phrase.To extract the structure of subject-verb-object from each sentence, the grammatical rela-tions described in [9] are used: John sees the yellow lion with nsubj(sees-2, John-1)det(lion-5, the-3)amod(lion-5, yellow-4)dobj(sees-2, lion-5)
The relation nsubj (nominal subject) relates the noun
John with the corresponding verb sees , whereas the relation dobj (direct object) relates this verb with the object lion . In this5igure 3: The workbench.6ay, the sentence elements are extracted and the sentence structure can be translated intothe mind-map. All subjects and objects take over the roles of the concepts, whereas theverbs serve as connections between the concepts. The adjectives represent sub-conceptsof both subjects and objects. From a graphical point of view, actors are representedas double-circles, while concepts that represent no actors are drawn as boxes. The sub-concepts (adjective) are drawn as diamonds. Concepts are linked by a directed arrow,labeled with the verb that relates the subject with the object. An example can be seen inFigure 2, representing the sentence
John sees the yellow lion .In order to merge concepts - that refer to the same actors - we use an incremental actor-based thesaurus. Sine the user can enter different information about the actor - for examplethe first name, the last name, nicknames, etc. in advance - we use this external informationto establish the thesaurus. Following the spirit of [14], the concepts are then matched.Figure 3 presents the implemented user-interface consisting of different components, forexample the technical (left) part (including processing information, graph options, andactor statistics), the monitoring part (below, including the last parsed sentences and in-formation about each node), and the notes part (to do and save own comments). Theworkbench is enriched by help buttons.
The following text is taken from an extract of the children story
Malcolm the Scotty Dog . Inthis example, the focus is on an actor called
Malcolm . The text is processed sentence-wide.With that, we start with
Malcolm picked the bone up and ran overto the other side of the garden.
The mind-map for the actor Malcolm after this sentence can be seen in Figure 4. The actor
Malcolm is centralized pointing to the concepts bone and side of garden . The last conceptis characterised by a sub-concept called other . After the next sentence, the mind-map of
Malcolm has evolved in the way as represented in Figure 5.
He set the bone down and looked around.
We observe that he has been resolved to Malcolm . An empty concept is added since lookedaround does not imply an object. The concept bone is stimulated again by set down (newconcept) and connected to it. With
He picked it up and could not wait to taste it. both occurrences of it have been replaced by bone (Figure 6). The negative verb could notwait is specially marked in the mind-map by an inhibitating arrow. The phrase he picked p the bone has re-occurred in the text stream. To mark this re-occurrence in the mind-map, the structure Malcolm - picked up - bone has been enforced (by means of a straightline). Here, it is possible for the reader to display the mind-maps in certain depths . Byselecting a depth of 1, only the concepts directly related to the actor will be represented,while for a depth of 2, all the concepts at a distance of two nodes will be displayed. Thiscan be illustrated by processing the following sentence.
The bone was big and it tasted delicious.
By displaying a depth of 1, the mind-map of
Malcolm will be as in Figure 6. But whendisplaying a depth of 2, the mind-map will look as in Figure 7. Here, we notice thatthe concept bone is explained in a more detailed way. And in fact, the user decides howdetailed the mind-map should be. Figure 8 shows the mind-map after the processing of alarger amount of sentences.Figure 4: Mind-map of
Malcolm after first sentenceFigure 5: Mind-map of
Malcolm after two sentences
The mind-map is a knowledge structure that continuously actualises itself as long as textis read. The representation of the mind-map as a semantic network structure permits to8igure 6: Mind-map of
Malcolm after three sentencesFigure 7: Mind-map of
Malcolm after four sentences, with a depth of 2gather all the actions, thoughts and states of being of one actor in a graphical represen-tation. Through the temporal consolidation, changes over time can easily be captured inthe mind-map. Currently, we work on two mind-map extensions that concern with animproved interaction. First, and since a main application is the support of a textual sum-marization of read text streams, we currently build an automatic text-based summariser.The first (prototypical) version simply outputs the concepts related to an actor, includingthe sub-concepts and the connections. As the connections are syntactically unchanged, itis easy to generate sentences out of it. Secondly, a selective information retrieval engineis currently done through the extension of the user/mind-map communication through aSQL-like interface. With that, we aim at queries like the following: select sub-concepts, concepts from mind-mapwith depth=1where concept = "Malcolm"
This leads to a result set where all concepts, sub-concepts, and associations are retrieved.The operation depth says that only the neighbor elements are considered. In case thatdepth is set to ≥
2, all components of the over-next level are retrieved. A second retrievalthen results in a set where only all sub-concepts of
Harry are retrieved. select sub-concepts, name from mind-map here name = "Harry"
To be more precise, the following commands are currently under implementation: • select : the projection that gives the concepts, sub-concepts, and associations toother concepts. • from : the selection to a mind-map; alternatively, several mind-maps can be ad-dressed. • with depth : the depth around a concept. • where : the where clause allows a diverse condition setting.However, a disadvantage of the mind-map is currently that it grows fast and becomes verylarge. With this implementation, texts with ≥
500 sentences are still an overkill. In thisrespect, the optimization of the existing solution is a future concern as well. Furthermore,sentences can be composed of not only one single clause, but of several clauses. Theseclauses are either independent or dependent clauses. Independent clauses can stand as asimple sentence and express a complete thought. Dependent clauses on the other handcan not express a complete thought by standing alone as a sentence. They simply make nosense when standing alone. This is why dependent clauses are connected to an independentclause. This connection is lost in the mind-maps so that some branches going off one actordo not make a lot of sense. Also, the application depends highly on the accuracy of theparser used during the preprocessing. If the parser can not identify the subject(s), verb(s)and object(s) of the sentence, errors or gaps will occur in the mind-maps. Also, as theresolution of some pronouns depends on the correct processing of the sentences, somepronouns may be wrongly resolved due to mistakes of the parser.
Acknowledgement
This work has been done within a Master Thesis at the MINE Research Group, ILIASLaboratory, University of Luxembourg.
References [1] C. Brucks, M. Hilker, C. Schommer, R. Weires, and C. Wagner. Symbolic Computingwith Incremental Mind-maps to Manage and Mine Data Streams - Some Applications.4 th International Workshop on Neural-Symbolic Learning and Reasoning. ECAI 2008.Patras, Greece. 2008.[2] I. Dagan, and A. Itai. Automatic processing of large corpora for the resolution ofanaphora references. In Proceedings of the 13 th International Conference on Compu-tational Linguistics (COLING ’90), pp. 330-332, 1990.113] M. Dimitrov. A Light-weight Approach to Coreference Resolution for Named Entitiesin Text. Proceedings of the Fourth Discourse Anaphora and Anaphor ResolutionColloquium (DAARC). 2002.[4] S. Lappin, H.J. Leass. An algorithm for pronomial anaphora resolution. Associationfor Computer Linguistics, 1994.[5] R. Mitkov. Anaphora Resolution: The State of the art. 1999. California State Uni-versity, Northridge.[6] R. Mitkov. Towards a more consistent, comprehensive evaluation of anaphora reso-lution algorithms, systems. 2000. Proceedings of the Discourse Anaphora, AnaphorResolution Colloquium, 96-107, Lancaster, UK.[7] R. Mitkov, C. Barbu. Using corpora to improve pronoun resolution. 2002. Languagesin context, 4(1).[8] S. Muthukrishnan. Data streams - Algorithms, applications. 2003. In Proceedings ofthe 14 thth