This paper briefly reports on an agent team doing ontology-based information extraction (OBIE, IE) for summarization in clinical Bone Marrow Transplantation (BMT). The SummIt-BMT agents contribute to OBIE through their flexible use of ontological knowledge. They assess input text passages from web retrieval with respect to a user query. They use an ontology that supports IE in particular with concepts, propositions, unifiers and paraphrases. Sentences with IE hits are annotated with the IDs of ontology propositions that recognize an instance of their content in the sentence. The agents are beginners, but they perform. Distributing ontology-based IE to agents has some promise: it enables parallel processing, it eases tracking of decisions and their explanation to users.
The agents are Java classes that extend the jade.core.Agent1. They run in a JADE container and use standard ACL (Agent Communication Language)2 means of interaction. All agents share a set of simple calls. Most calls consist of the name of the addressed agent and a simple German codeword: los (go), mehr (more), fertig (done). Only the tidy-up agent Paula is also assigned a more sophisticated command when she has to reorganize results for presentation: sumup (sum up). When broadcasting the close-down message to all agents, Kurt says schluss (finish) to make the agents delete.
The system blackboards serve data communication. The ScenarioBoard stores the query specification and the findings of the agents. While they interpret a sentence, the agents exchange data via the TextBlackBoard. External input comes from the text passage retrieval result. At the end of a session, the retrieval result (organized in documents, paragraphs and sentences) is augmented with the agents’ relevance judgements. They mark the relevant text clips, which are presented to the user.
The agents and the system users share an ontology of the application domain Bone Marrow Transplantation (BMT). The ontology was developed by human experts from a corpus of US BMT papers and user queries of physicians at Hanover Medical School. It is stored in a MySQL3 database. In the ontology the agents find the types of knowledge they need for IE (see Table 1): concepts, propositions, proposition syntax records, unifiers, paraphrases, scenarios, and some technical help tables.
We use a Prolog style first order predicate logic representation. Inside the MySQL database, all knowledge items are split into separate tables. Propositions
comprise a head and a set of arguments allocated to the propositionhead and
propositionargument tables, respectively. Their proposition ID keeps them together.
Every proposition obeys a syntax record that states its argument roles. Syntax tables
are built like proposition tables. Unifiers are lists of concepts provided by domain
experts. They unify ontology propositions and text-based candidate propositions: a
concept of the accredited unifier adapts the ontology proposition so that it matches a
candidate proposition from input. This expands the coverage of the ontology
propositions. Paraphrases map ontology propositions to possible surface formulations.
They are macropropositions
The agents specialize in different IE subtasks. They produce a summarization effect by extracting only propositions that match the query and by throwing away doubles. All agents activate each other as often as needed.
Scenario interpretation. Peter accepts a user query scenario and the user’s start signal. Into the scenario form, the user has entered what is known about the current situation and what knowledge is missing. The agent parses this organized query, deposits the resulting propositions on the ScenarioBoard and activates Kurt. Input. Kurt fetches the query and obtains results from outside web retrieval and text passage retrieval. He submits good input sentences one by one to the parser (the
Connexor5 FDG parser) and feeds wording and parser output into the agents’ production line by putting it onto the TextBlackBoard. He calls Frieda.
Let us assume for the sake of a demo that Kurt comes up with the sentence “All patients at cmv risk were administered high-dose ganciclovir.“ It will be hit by 9 ontology propositions (see table 2). We follow proposition 17685.
She finds candidate propositions in a parsed sentence and annotates them (see table 3, columns 8 – 13). She distinguishes verbal, prepositional and attributive candidate propositions. As soon as her annotation is done, Frieda activates the agent Heini.
selects the propositions one by one from the current sentence and initiates their verification. As long as he has input, he submits it to Hugo. When all propositions of the current sentence are done, he calls Paula, the tidy-up agent.
Finding ontology propositions. Hugo checks whether the current proposition shares at least two ontology concepts with any of the ontology propositions. As soon as he is done with a concept pair and has some results, Hugo passes the text-based proposition with the IDs of selected ontology propositions - and possibly some add-ons due to unifiers - to the TextBlackbord. He activates Rudi. If Hugo cannot find matching ontology propositions, he returns to Heini and asks for new supplies.
When Hugo begins to treat a new proposition, he puts the IDs of occurring ontology
concept IDs into its record. Using them he may find several concept pairs that call ontology propositions. Eventually he selects proposition 17685 (see figure 2): administer (, patient, methotrexate).
His pick takes the direct and the unifier pathway. According to the unifier in the proposition, any drug of the ontology may be put in. As ganciclovir is needed, Hugo adds an ersatz argument that contains ganciclovir. He puts his results into the package for Rudi.
Concept subsumption. Rudi tries to subsume text-based propositions under ontology propositions with at least two concepts in agreement. If the subsumption works, the proposition from text may be a legitimate instance of the subsuming ontology proposition, as far as ontology concepts are concerned. If so, Rudi passes it to Herta. She will inspect the verbal relation.
When Rudi fetches proposition 17685 that Hugo proposed, he looks for ersatz arguments, finds one and puts it in. Now his version of proposition 17685 says: administer (, patient, ganciclovir).
As figure 2 shows, the proposition has some open slots. Rudi tries to fill them, subsuming concepts from the text-based proposition. He succeeds once: he subsumes high dose under concept 43174 (quantity qualifier) in position 5. Now his proposition reads:
administer (, patient, ganciclovir, , high dose).
As all obligatory arguments are satisfied, Rudi passes his result (see table 4) to Herta. Verification of the verbal relation. Herta’s task is to check the verbal tie that keeps the ontology concepts together. She verifies the relation information against sets of paraphrases. If a paraphrase provides a relation wording that is compatible with the ontology proposition chosen by Rudi and the relation wording of the text surface, Herta has found the last missing link. She states the recognition success for that proposition by assigning the ID of the subsuming ontology proposition to the text sentence. She asks Rudi for fresh input.
In the test case, Herta receives Rudi’s reworked ontology proposition 17685. She writes the ontology concept’s IDs into the parse of proposition 1 (cf. table 3).
Herta procures her paraphrase set of proposition 17685. She will find the test paraphrase 14068437 (see table 5) that will fit. Herta seizes the concepts found in the text-based proposition via their hypercids and attaches them to the hypercids / argument roles of the paraphrase. Then she checks in three passes: From satisfied roles she goes towards the root of the dependency tree and checks all items on her way to ok.
Bcr-abl1 Bcr-abl2 Bcr-abl3 Childhood ALL1 Childhood ALL2 Childhood ALL3 CMVganciclovir1 CMVganciclovir2
CMVganciclovir3 She compares the verbal chain of the proposition and paraphrase. There should be a reasonable fit, depending on word classes. If so, Herta places her controls. At the end, Herta starts from the ontology proposition arguments without fillers.
Again she goes up the dependency hierarchy and sets all words on her way to optional.
If Herta obtains all ticks as needed, she has verified the verbal relation. In the present case, she has found
“patient is administered ganciclovir”.
She writes the hit ID to the TextBB. Paula will reorganize all results. depend relation subj: v-ch: main: obj: hypercid 38811 0 0 39375 deptarget 3 4 1 4 Cleaning up. Paula is the organizer. When processing of a sentence is finished and has brought some results, Paula stores the sentence with the recognition results to the ScenarioBoard. She tells Kurt to provide new input. When all input is done, Paula reorganizes the ScenarioBoard. She sorts the recognition IDs of individual sentences so that she obtains orderly recognition profiles. Based on the profiles and the wording of the sentences, Paula weeds out doubles. Surviving hits are added to their text clips in the retrieval result. Paula asks Kurt to close down the agent community. A biochemist and the author evaluated the agents’ performance in a testbed with a small sample of Medline abstracts. Methods were adapted from qualitative field research. The agents ran twice. Between their two runs, the judges improved the ontology, and results became much better. Ontology quality matters. Often the agents stumble over simple human errors, sloppy categorizations or into ontology gaps. In overcrowded areas, they are obstructed by too many chances to derive the same recognition result.
In their second run, the agents achieved fair scores. They are still beginners, but they come up with results. Table 6 shows their marks on a familiar 5-score scale.
SummIt-BMT integrates knowledge from many sources. Ontology and agents are
based on empirical observation of human summarizers
Ontology-based IE (for summarization) can be distributed to an agent team. This has advantages: Agents’ decisions can be tracked more easily. The agents may explain them. New agents are easily integrated, so that the community “learns”. If running in parallel, agent teams may be fast and scale up well.