Information Extraction Information Extraction generally relates to automatic approaches to locate important facts in large collections of documents aiming at highlighting specific information to be used for enriching other texts and documents while populating summaries, feeding reports, filling in forms or storing information for further processing (e.g. data mining); the extracted information is usually structured in the form of “templates”
Information Extraction The process of Information Extraction consists of two major steps: To extract individual “facts” from the text of a document through local text analysis To integrate extracted facts producing larger facts or new facts (through inference)
Information Extraction Short history (1) IE originated in the natural language processing community under the MUC conferences (starting at 1987 and sponsored by DARPA) with the definition of a task: inside a specific application domain and corpus, a template with the relevant information has to be filled for every event of each foreseen class.
Information Extraction Short history (2) In 1995 further goals for IE were proposed: To identify processing tasks largely domain independent (e.g. NE Named Entity Recognition) To focus on portability in the IE tasks to new event classes To add three new tasks: co- reference resolution, word-sense-disambiguation, predicate-argument syntactic structuring
Information Extraction Terminology A template is a sort of linguistic pattern (a set of attribute-value pairs with the values being texts string) described by experts to represent the structure of a specific event in a given domain. The template relates to the final output format of selected information The scenario identifies the specification of the particular events or relations to be extracted.
Information Extraction General Architecture for an IE system “An IE system is a cascade of transducers or modules that at each step add structure and often lose information, hopefully irrelevant, by applying rules that are acquired manually and/or automatically”. (by J. Hobbs).
Information Extraction General Architecture for an IE system Each system could be characterized by its own set of modules belonging to the following set: Text zoner, pre-processing, filter, preparser, parser, fragment combiner, semantic interpreter, lexical disambiguation, coreference resolution / discourse processing, template generator.
Information Extraction Text zoner This module turns a text into a set of text segments. As a minimum results it would separate the formatted from the unformatted regions.
Information Extraction Pre-processing This module: locates sentence boundaries in the text producing for each sentence a sequence of lexical items (words together with their possible POS). It recognizes also multiword (lexical lookup methods) recognizes and normalizes certain basic types that occur in the genre, such as dates, times, personal and company names, locations, currency amounts, and so on.
Information Extraction Filter For speeding processing time this module uses superficial techniques to filter out (from previously recognized ones) the sentences that are likely to be irrelevant. In any application, subsequent modules will be looking for patterns of words that signal relevant events. If a sentence has none of this words, then there is no reason to process it further.
Information Extraction Preparser This module recognizes very common small-scale structures, simplifying the task of the parser. A few systems at this level recognize noun groups (noun phrases up through the head noun) as well as verb groups (verbs together with their auxiliaries). Appositives can be attached to their head nouns with high reliability (e.g. Prime Minister, President of the Republic, etc.).
Information Extraction Parser This module takes a sequence of lexical items (fragments) and tries to produce a parse tree for the entire sentence. Recently more and more systems are abandoning full-sentence parsing in information extraction applications being interested just in recognizing fragments, then they try only to locate within the sentence various patterns that are of interest for the application.
Information Extraction Fragment combiner This module provides indication on how to combine the previously obtained parse tree fragments
Information Extraction Semantic interpreter This module translates the parse tree or parse tree fragments into any of: a semantic structure, a logical form or event frame. Often lexical disambiguation takes place at this level as well. The method for semantic interpretation is function application or an equivalent process that matches predicates with their arguments.
Information Extraction Lexical disambiguation Lexical disambiguation allows translating a semantic structure with general or ambiguous predicates into a semantic structure with specific, unambiguous predicates. More generally, lexical disambiguation generally happens by constraining the interpretation by the context in which the ambiguous word occurs, perhaps together with the “a priori” probabilities of each word sense.
Information Extraction Coreference resolution / discourse processing This module revolves: co-reference for basic entities such as pronouns, definite noun phrases, and anaphora. the reference for more complex entities like events identified either with an event that was found previously or as a consequence of a previously found event, or it may fill a role in a previous event.
Information Extraction Template generator Semantic structures generated by the natural language processing modules are used to produce the template as described by the final user only in the case events pass the defined threshold of interest.
Information Extraction There is an agreement also on a number of features: named entity recognition, co- reference resolution, template production, scenario template production.
Information Extraction Named entity recognition It refers to named entities (NE) identification (inside the text) and extraction. NEs generally relate to domain concepts and are associated to semantic classes such as person, organization, place, date, amount, etc. The accuracy in NE recognition is very high (more than 90%) and comparable with those of humans.
Information Extraction Co-reference resolution It allows identifying identity relations between previously extracted NEs. Anaphora resolution is widely used to recognize relevant information about either concepts (NE) or events sparse in the text: this activity constitutes an important source of information enabling the system to assign a statistical relevance to recognized events.
Information Extraction Template production As a result of the previous activities, an IE system becomes aware of NEs and their descriptions. This represents a first level of template (called TE – “Template Element”). The TEs collections may be considered as a basic knowledge base to which the system accesses for getting information on main domain concepts, as they have been recognized in the text.
Information Extraction Scenario template production It results in a synthesis of several tasks, mainly the identification of Template Elements that relate among them: it represents an event (scenario) related to the domain under analysis; recognized values are used to fill in a scenario template
Information Extraction Adaptive IE systems As an example, several big companies have millions of documents, stored in different parts of the world, available via intranets, where the knowledge of their employees is stored. Textual documents cannot be queried in a traditional fashion and therefore the stored knowledge can neither be used by automatic systems, nor be easily managed by humans. Knowledge is difficult to capture, share and reuse among employees, reducing the company's efficiency and competitiveness.
Information Extraction Adaptive IE systems IE is the perfect support for knowledge identification and extraction from Web documents as it can provide support in documents analysis either in an automatic approach (unsupervised extraction of information) or in a semi-automatic one (e.g. as support for human annotators in locating relevant facts in documents, via information highlighting). Machine-learning approach may be helpful.
Information Extraction Adaptive IE systems Machine learning (ML) techniques has been successfully applied to some lower level NLP tasks. NE recognition, chunking, co-reference and anaphora resolution, are interesting examples of such approaches.
Question / Answering A Q/A system accepts questions in natural language form, searches for answer over a collection of documents extracts relevant information for the question formulates concise answers.
Question / Answering Short history TREC Conferences Q/A tracks has supported the definition of a common approach to the matter. Q/A systems are open domain, then their performances are tightly coupled with the complexity of the questions asked and the difficulty of answer extraction.
Question / Answering Taxonomy of Q/A systems 1.Linguistic and knowledge resources 2.Natural language processing involved 3.Document processing 4.Reasoning methods 5.Wheather or not answer is explicitely stated in a document 6. Wheather or not answer fusion is necessary
Question / Answering Questions classes 1.Q/A systems capable of processing factual questions 2.Q/A systems enabling simple reasoning mechanisms 3.Q/A systems capable of answer fusion from different documents 4.Interactive Q/A systems 5.Speculative questions
Question / Answering Question analysis The question is analyzed for subsequent processing. The question may be interpreted in the context of an on-going dialogue and in the light of a model which the system has of the user. The user could be asked to clarify his question before processing
Question / Answering Document collection processing The reference document collection is the knowledge source for answering questions. It requires to be preprocessed.
Question / Answering Candidate document selection A subset of documents collection is selected, comprising those documents deemed most likely to contain an answer to the question.
Question / Answering Candidate document analysis Additional detailed analysis of the candidates selected at the preceding stage could be required.
Question / Answering Answer extraction Candidate answers are extracted from the the documents and ranked in terms of probale correctness.
Question / Answering Response generation A response is returned to the user. It may be affected by the dialogue context and user model, if present, and may in turn lead to this neing updated.