Lexicon Learning for Few-Shot Neural Sequence Modeling

Total Page:16

File Type:pdf, Size:1020Kb

Lexicon Learning for Few-Shot Neural Sequence Modeling Lexicon Learning for Few-Shot Neural Sequence Modeling Ekin Akyürek Jacob Andreas Massachusetts Institute of Technology {akyurek,jda}@mit.edu Abstract Train Test dax r lug b wif g zup y zup fep y y y Sequence-to-sequence transduction is the core lexicon zup blicket lug ___? problem in language processing applications lug fep b b b dax blicket zup ___? as diverse as semantic parsing, machine trans- dax fep r r r zup kiki dax ___? lation, and instruction following. The neural lug blicket wif b g b wif kiki zup ___? network models that provide the dominant so- wif blicket dax g r g lution to these problems are brittle, especially lug kiki wif g b in low-resource settings: they fail to generalize dax kiki lug b r correctly or systematically from small datasets. Past work has shown that many failures of sys- Figure 1: A fragment of the Colors dataset from Lake tematic generalization arise from neural mod- et al.(2019), a simple sequence-to-sequence translation els’ inability to disentangle lexical phenomena task. The output vocabulary is only the colored circles from syntactic ones. To address this, we aug- r , g , b , y . Humans can reliably fill in the miss- ment neural decoders with a lexical transla- ing test labels on the basis of a small training set, but tion mechanism that generalizes existing copy standard neural models cannot. This paper describes a mechanisms to incorporate learned, decontex- neural sequence model that obtains improved general- tualized, token-level translation rules. We de- ization via a learned lexicon of token translation rules. scribe how to initialize this mechanism using a variety of lexicon learning algorithms, and show that it improves systematic generaliza- and small datasets (Lake and Baroni, 2018), pos- tion on a diverse set of sequence modeling ing a fundamental challenge for NLP tools in the tasks drawn from cognitive science, formal se- mantics, and machine translation.1 low-data regime. Pause for a moment to fill in the missing labels 1 Introduction in Fig.1. While doing so, which training exam- ples did you pay the most attention to? How many Humans exhibit a set of structured and remarkably times did you find yourself saying means or maps consistent inductive biases when learning from lan- to? Explicit representations of lexical items and guage data. For example, in both natural language their meanings play a key role diverse models of acquisition and toy language-learning problems syntax and semantics (Joshi and Schabes, 1997; like the one depicted in Fig.1, human learners Pollard and Sag, 1994; Bresnan et al., 2015). But exhibit a preference for systematic and composi- one of the main findings in existing work on gener- tional interpretation rules (Guasti 2017, Chapter 4; alization in neural models is that they fail to cleanly Lake et al. 2019). These inductive biases in turn separate lexical phenomena from syntactic ones support behaviors like one-shot learning of new (Lake and Baroni, 2018). Given a dataset like the concepts (Carey and Bartlett, 1978). But in natural one depicted in Fig.1, models conflate (lexical) language processing, recent work has found that information about the correspondence between zup state-of-the-art neural models, while highly effec- and y with the (syntactic) fact that y appears tive at in-domain prediction, fail to generalize in only in a sequence of length 1 at training time. human-like ways when faced with rare phenomena Longer input sequences containing the word zup 1Our code is released under https://github.com/ in new syntactic contexts cause models to output ekinakyurek/lexical tokens only seen in longer sequences (Section5). 4934 Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 4934–4946 August 1–6, 2021. ©2021 Association for Computational Linguistics In this paper, we describe a parameterization human-like generalization about language data, es- for sequence decoders that facilitates (but does pecially in the low-data regime (e.g. Fodor et al., not enforce) the learning of context-independent 1988; Marcus, 2018). Our results suggest that the word meanings. Specifically, we augment decoder situation is more complicated: by offloading the output layers with a lexical translation mecha- easier lexicon learning problem to simpler models, nism which generalizes neural copy mechanisms neural sequence models are actually quite effec- (e.g. See et al., 2017) and enables models to gen- tive at modeling (and generalizing about) about erate token-level translations purely attentionally. syntax in synthetic tests of generalization and real While the lexical translation mechanism is quite translation tasks. general, we focus here on its ability to improve few-shot learning in sequence-to-sequence models. 2 Related Work On a suite of challenging tests of few-shot seman- Systematic generalization in neural sequence tic parsing and instruction following, our model models The desired inductive biases noted above exhibits strong generalization, achieving the high- are usually grouped together as “systematicity” but est reported results for neural sequence models on in fact involve a variety of phenomena: one-shot datasets as diverse as COGS (Kim and Linzen 2020, learning of new concepts and composition rules with 24155 training examples) and Colors (Lake (Lake and Baroni, 2018), zero-shot interpretation et al. 2019, with 14). Our approach also generalizes of novel words from context cues (Gandhi and to real-world tests of few-shot learning, improving Lake, 2020), and interpretation of known concepts BLEU scores (Papineni et al., 2002) by 1.2 on a in novel syntactic configurations (Keysers et al., low-resource English–Chinese machine translation 2020; Kim and Linzen, 2020). What they share is a task (2.2 on test sentences requiring one-shot word common expectation that learners should associate learning). specific production or transformation rules with In an additional set of experiments, we explore specific input tokens (or phrases), and generalize effective procedures for initializing the lexical to use of these tokens in new contexts. translation mechanism using lexicon learning al- Recent years have seen tremendous amount of gorithms derived from information theory, statis- modeling work aimed at encouraging these gener- tical machine translation, and Bayesian cognitive alizations in neural models, primarily by equipping modeling. We find that both mutual-information- them with symbolic scaffolding in the form of pro- and alignment- based lexicon initializers perform gram synthesis engines (Nye et al., 2020), stack ma- well across tasks. Surprisingly, however, we show chines (Grefenstette et al., 2015; Liu et al., 2020), that both approaches can be matched or outper- or symbolic data transformation rules (Gordon formed by a rule-based initializer that identifies et al., 2019; Andreas, 2020). A parallel line of work high-precision word-level token translation pairs. has investigated the role of continuous representa- We then explore joint learning of the lexicon and tions in systematic generalization, proposing im- decoder, but find (again surprisingly) that this gives proved methods for pretraining (Furrer et al., 2020) only marginal improvements over a fixed initializa- and procedures for removing irrelevant contex- tion of the lexicon. tual information from word representations (Arthur In summary, this work: et al., 2016; Russin et al., 2019; Thrush, 2020). The latter two approaches proceed from similar intu- • Introduces a new, lexicon-based output mech- ition to ours, aiming to disentangle word meanings anism for neural encoder–decoder models. from syntax in encoder representations via alterna- • Investigates and improves upon lexicon learn- tive attention mechanisms and adversarial training. ing algorithms for initialising this mechanism. Our approach instead focuses on providing an ex- plicit lexicon to the decoder; as discussed below, • Uses it to solve challenging tests of generaliza- this appears to be considerably more effective. tion in instruction following, semantic parsing and machine translation. Copying and lexicon learning In neural encoder–decoder models, the clearest example A great deal of past work has suggested that of benefits from special treatment of word-level neural models come equipped with an inductive production rules is the copy mechanism. A great bias that makes them fundamentally ill-suited to deal of past work has found that neural models 4935 Inputs Outputs Lexicon Entries blessed 7! bless A crocodile blessed William . crocodile(x_1) AND bless.agent (x_2, x_1) AND bless.theme (x_2, William) needed 7! need William needed to walk . need.agent (x_1 , William) AND need.xcomp(x_1, x_3) AND walk.agent (x_3, William) William 7! William saturn 7! 土星 Many moons orbit around Saturn 1多 [星 ^W 土星 KL. earth 7! 0球 Earth is a planet . 0球 / 一個 L星. moon 7! [星 walk around left LTURN IWALK LTURN IWALK LTURN IWALK LTURN IWALK walk 7! IWALK turn right RTURN jump 7! IJUMP turn left LTURN right 7! RTURN jump IJUMP left 7! LTURN jump opposite right after look left LTURN ILOOK RTRUN IJUMP RTURN IJUMP look 7! ILOOK Table 1: We present example (input,output) pairs from COGS, English-to-Chinese machine translation and SCAN datasets. We also present some of the lexicon entries which can be learned by proposed lexicon learning methods and that are helpful to make generalizations required in each of the datasets. benefit from learning a structural copy operation based generative model; like the present work, that that selects output tokens directly from the input model effectively disentangles syntactic and lexical sequence without requiring token identity to be information by using training examples as implicit carried through all neural computation in the representations of lexical correspondences. encoder and the decoder. These mechanisms We generalize and extend this previous work in a are described in detail in Section3, and are number of ways, providing a new parameterization widely used in models for language generation, of attentive token-level translation and a detailed summarization and semantic parsing.
Recommended publications
  • Lexical Semantics I
    Semantic Theory: Lexical Semantics I Summer 2007 M.Pinkal/ S. Thater • 03-07-07 Lecture: Lexical Semantics I • 05-07-07 Lecture: Lexical Semantics II • 10-07-07 Lecture: Everything else • 12-07-07 Exercise: lexical Semantics • 17-07-07 Question time, Sample exam • 19-07-07 Individual question time • 25-07-07 Final exam, 11:00 (s.t.) Semantic Theory, SS 2007 © M. Pinkal, S. Thater 2 Structure of this course • Sentence semantics • Discourse semantics • Lexical semantics Semantic Theory, SS 2007 © M. Pinkal, S. Thater 3 Dolphins in First-order Logic Dolphins are mammals, not fish. !d (dolphin'(d)"mammal'(d) #¬fish'(d)) Dolphins live-in pods. !d (dolphin'(d)" $x (pod'(p)live-in'(d,p)) Dolphins give birth to one baby at a time. !d (dolphin'(d)" !x !y !t (give-birth-to' (d,x,t)give-birth-to' (d,y,t) " x=y) Semantic Theory, SS 2007 © M. Pinkal, S. Thater 4 Dolphins in First-order Logic Dolphins are mammals, not fish. !d (dolphin'(d)"mammal'(d) #¬fish'(d)) Dolphins live-in pods. !d (dolphin'(d)" $x (pod'(p)live-in'(d,p)) Dolphins give birth to one baby at a time. !d (dolphin'(d)" !x !y !t (give-birth-to' (d,x,t)give-birth-to' (d,y,t) " x=y) Semantic Theory, SS 2007 © M. Pinkal, S. Thater 5 The dolphin text Dolphins are mammals, not fish. They are warm blooded like man, and give birth to one baby called a calf at a time. At birth a bottlenose dolphin calf is about 90-130 cms long and will grow to approx.
    [Show full text]
  • LEXADV - a Multilingual Semantic Lexicon for Adverbs
    LEXADV - a multilingual semantic Lexicon for Adverbs Sanni Nimb Center for Sprogteknologi (CST) – Københavns Universitet Njalsgade 80, Copenhagen, Denmark [email protected] Abstract The LEXADV-project is a Scandinavian research project (2004-2006, financed by Nordplus Sprog) with the aim of extending three Scandinavian semantic lexicons building on the SIMPLE lexicon model (Lenci et al., 2000) with the word class of adverbs. In the lexicons of approx. 400 Danish, Norwegian and Swedish adverbs the different senses are described with a semantic type and a set of semantic features. A classification covering the many meanings that adverbs can have has been established and integrated in the original SIMPLE ontology. Similarly new features have been added to the model in order to describe the adverb senses. The working method of the project builds on the fact that the vocabularies of Danish, Norwegian and Swedish are closely related. An encoding tool has been developed with the special purpose of permitting easy transfer of semantic types and features between entries in the three languages. The Danish adverb senses have been described first, based on the definition in a modern, comprehensive Danish dictionary. Afterwards the lemmas have been translated and the semantic data have been copied into the Swedish as well as into the Norwegian equivalent entry. Finally these copies have been evaluated and when necessary adjusted by native speakers. ’Telic’, ’Agentive’ and ’Constitutive’ has been extended 1. The Scandinavian SIMPLE Lexicons with three more semantic types. The first one, ’Functional’, covers adverb senses of 1.1. Background which the meaning must be described by logic expressions referring to sets, either in the context or in The Danish and the Swedish SIMPLE lexicons are the discourse.
    [Show full text]
  • A New Semantic Lexicon and Similarity Measure in Bangla
    A New Semantic Lexicon and Similarity Measure in Bangla Manjira Sinha, Abhik Jana, Tirthankar Dasgupta, Anupam Basu Indian Institute of Technology Kharagpur {manjira87, abhikjana1, iamtirthankar, anupambas}@gmail.com ABSTRACT The Mental Lexicon (ML) refers to the organization of lexical entries of a language in the human mind.A clear knowledge of the structure of ML will help us to understand how the human brain processes language. The knowledge of semantic association among the words in ML is essential to many applications. Although, there are works on the representation of lexical entries based on their semantic association in the form of a lexicon in English and other languages, such works of Bangla is in a nascent stage. In this paper, we have proposed a distinct lexical organization based on semantic association between Bangla words which can be accessed efficiently by different applications. We have developed a novel approach of measuring the semantic similarity between words and verified it against user study. Further, a GUI has been designed for easy and efficient access. KEYWORDS : Bangla Lexicon, Synset, Semantic Similarity, Hierarchical Graph 1 Introduction The lexicon of a language is a collection of lexical entries consisting of information regarding words and expressions, comprising both form and meaning (Levelt,). Form refers to the orthography, phonology and morphology of the lexical item and Meaning refers to its syntactic and semantic information. The term Mental Lexicon refers to the organization and interaction of lexical entries of a language in the human mind. Depending on the definition of word , an adult knows and uses around 40000 to 150000 words.
    [Show full text]
  • The Diachronic Semantic Lexicon of Dutch As Linked Open Data
    The Diachronic Semantic Lexicon of Dutch as Linked Open Data Katrien Depuydt, Jesse de Does Instituut voor de Nederlandse Taal Rapenburg 61, 2311GJ Leiden, The Netherlands [email protected], [email protected] Abstract This paper describes the Linked Open Data (LOD) model for the diachronic semantic lexicon DiaMaNT, currently under development at the Instituut voor de Nederlandse Taal (INT; Dutch Language Institute). The lexicon is part of a digital historical language infrastructure for Dutch at INT. This infrastructure, for which the core data is formed by the four major historical dictionaries of Dutch covering Dutch language from ca. 500 - ca 1976, currently consists of three modules: a dictionary portal, giving access to the historical dictionaries, a computational lexicon GiGaNT, providing information on words, their inflectional and spelling variation, and DiaMaNT, aimed at providing information on diachronic lexical variation (both semasiological and onomasiological). The DiaMaNT lexicon is built by adding a semantic layer to the word form lexicon GiGaNT, using the semantic information in the historical dictionaries. Ontolex-Lemon is a good point of departure for the LOD model, but we need extensions to be able to deal with the historical dictionary content incorporated in our lexicon. Keywords: Linked Open Data, Ontolex, Diachronic Lexicon, Semantic Lexicon, Historical Lexicography, Language Resources 1. Background digital, based on a closed corpus, in digital format. Hav- Even though Dutch lexicography1 can be dated back to the ing four scholarly dictionaries of Dutch in digital format 13th century with the glossarium Bernense, a Latin-Middle opened up opportunities for further exploitation of the con- Dutch word list, we had to wait until the 19th century for tents of these dictionaries.
    [Show full text]
  • Natural Language Processing
    Chowdhury, G. (2003) Natural language processing. Annual Review of Information Science and Technology, 37. pp. 51-89. ISSN 0066-4200 http://eprints.cdlr.strath.ac.uk/2611/ This is an author-produced version of a paper published in The Annual Review of Information Science and Technology ISSN 0066-4200 . This version has been peer-reviewed, but does not include the final publisher proof corrections, published layout, or pagination. Strathprints is designed to allow users to access the research output of the University of Strathclyde. Copyright © and Moral Rights for the papers on this site are retained by the individual authors and/or other copyright owners. Users may download and/or print one copy of any article(s) in Strathprints to facilitate their private study or for non-commercial research. You may not engage in further distribution of the material or use it for any profitmaking activities or any commercial gain. You may freely distribute the url (http://eprints.cdlr.strath.ac.uk) of the Strathprints website. Any correspondence concerning this service should be sent to The Strathprints Administrator: [email protected] Natural Language Processing Gobinda G. Chowdhury Dept. of Computer and Information Sciences University of Strathclyde, Glasgow G1 1XH, UK e-mail: [email protected] Introduction Natural Language Processing (NLP) is an area of research and application that explores how computers can be used to understand and manipulate natural language text or speech to do useful things. NLP researchers aim to gather knowledge on how human beings understand and use language so that appropriate tools and techniques can be developed to make computer systems understand and manipulate natural languages to perform the desired tasks.
    [Show full text]
  • Linking Wordnet to the SIL Semantic Domains
    Bringing together over- and under-represented languages: Linking Wordnet to the SIL Semantic Domains Muhammad Zulhelmy bin Mohd Rosman Francis Bond and Frantisekˇ Kratochv´ıl Linguistics and Multilingual Studies, Nanyang Technological University, Singapore [email protected], [email protected], [email protected] Abstract English, Japanese and Finnish (Bond and Paik, 2012). We have created an open-source mapping SD is designed for rapid construction and in- between the SIL’s semantic domains (used tuitive organization of lexicons, not primarily for for rapid lexicon building and organiza- the analysis of the resulting data. As a result, tion for under-resourced languages) and many potentially interesting relationships are only WordNet, the standard resource for lexical implicitly realized. By linking SD to WN we semantics in natural language processing. can take advantage of the relationships modeled in We show that the resources complement WN to make more of these explicit. For example, each other, and suggest ways in which the the semantic relations in WN would be a useful mapping can be improved even further. input into SD while the domains hierarchy would The semantic domains give more general enforce the existing WN relations. This will al- domain and associative links, which word- low more quantitative computational modeling of net still has few of, while wordnet gives under-resourced languages. explicit semantic relations between senses, It is currently an exciting time for field lexicog- which the domains lack. raphy with better tools and hardware allowing for rapid digitization of lexical resources. Typically, 1 Introduction linguists tag text soon after they collect it.
    [Show full text]
  • Semanticnet-Perception of Human Pragmatics
    SemanticNet-Perception of Human Pragmatics Amitava Das 1 and Sivaji Bandyopadhyay 2 Department of Computer Science and Engineering Jadavpur University [email protected] 1 [email protected] 2 technical. It is often used in ordinary language Abstract to denote a problem of understanding that comes down to word selection or connotation. SemanticNet is a semantic network of We studied with various Psycholinguistics ex- lexicons to hold human pragmatic periments to understand how human natural knowledge. So far Natural Language intelligence helps to understand general se- Processing (NLP) research patronized mantic from nature. Our study was to under- much of manually augmented lexicon stand the human psychology about semantics resources such as WordNet. But the beyond language. We were haunting for the small set of semantic relations like intellectual structure of the psychological and Hypernym, Holonym, Meronym and neurobiological factors that enable humans to Synonym etc are very narrow to cap- acquire, use, comprehend and produce natural ture the wide variations human cogni- languages. Let’s come with an example of tive knowledge. But no such informa- simple conversation about movie between two tion could be retrieved from available persons. lexicon resources. SemanticNet is the Person A: Have you seen the attempt to capture wide range of con- movie ‘ No Man's Land’? How text dependent semantic inference is it? among various themes which human Person B: Although it is beings perceive in their pragmatic good but you should see knowledge, learned by day to day cog- ‘The Hurt Locker’? nitive interactions with the surrounding May be the conversation looks very casual, physical world.
    [Show full text]
  • Ten Choices for Lexical Semantics
    1 Ten Choices for Lexical Semantics Sergei Nirenburg and Victor Raskin1 Computing Research Laboratory New Mexico State University Las Cruces, NM 88003 {sergei,raskin}@crl.nmsu.edu Abstract. The modern computational lexical semantics reached a point in its de- velopment when it has become necessary to define the premises and goals of each of its several current trends. This paper proposes ten choices in terms of which these premises and goals can be discussed. It argues that the central questions in- clude the use of lexical rules for generating word senses; the role of syntax, prag- matics, and formal semantics in the specification of lexical meaning; the use of a world model, or ontology, as the organizing principle for lexical-semantic descrip- tions; the use of rules with limited scope; the relation between static and dynamic resources; the commitment to descriptive coverage; the trade-off between general- ization and idiosyncracy; and, finally, the adherence to the “supply-side” (method- oriented) or “demand-side” (task-oriented) ideology of research. The discussion is inspired by, but not limited to, the comparison between the generative lexicon ap- proach and the ontological semantic approach to lexical semantics. It is fair to say that lexical semantics, the study of word meaning and of its representation in the lexicon, experienced a powerful resurgence within the last decade.2 Traditionally, linguistic se- mantics has concerned itself with two areas, word meaning and sentence meaning. After the ascen- dancy of logic in the 1930s and especially after “the Chomskian revolution” of the late 1950s-early 1960s, most semanticists, including those working in the transformational and post-transforma- tional paradigm, focused almost exclusively on sentence meaning.
    [Show full text]
  • Filling Knowledge Gaps in a Broad-Coverage Machine Translation System*
    Filling Knowledge Gaps in a Broad-Coverage Machine Translation System* Kevin Knight, Ishwar Chander, Matthew Haines, Vasileios Hatzivassiloglou, Eduard Hovy, Masayo Lda, Steve K Luk, Richard Whitney, Kenji Yamada USC/Infonnation Sciences Institute 4676 Admiralty Way Manna del Rey, CA 90292 Abstract may be an unknown word, a missing grammar rule, a missing piece of world knowledge, etc A system can be Knowledge-based machine translation (KBMT) designed to respond to knowledge gaps in any number of techniques yield high quabty in domuoH with de• ways It may signal an error It may make a default or tailed semantic models, limited vocabulary, and random decision It may appeal to a human operator controlled input grammar Scaling up along these It may give up and turn over processing to a simpler, dimensions means acquiring large knowledge re• more robust MT program Our strategy is to use pro• sources It also means behaving reasonably when grams (sometimes statistical ones) that can operate on definitive knowledge is not yet available This pa more readily available data and effectively address par per describes how we can fill various KBMT know] ticular classes of KBMT knowledge gaps This gives us edge gap*, often using robust statistical techniques robust throughput and better quality than that of de• We describe quantitative and qualitative results fault decisions It also gives us a platform on which to from JAPANGLOSS, a broad-coverage Japanese- build and test knowledge bases that will product further English MT system improvements in quality
    [Show full text]
  • Approaches for Natural Language Processing (NLP)
    Approaches for Natural Language Processing (NLP) Thierry Hamon Institut Galil´ee- Universit´eParis 13,Villetaneuse, France & LIMSI-CNRS, Orsay, France [email protected] http://perso.limsi.fr/hamon/ March 2014 ERASMUS Mobility - M¨alardalen University (MDH) - V¨aster˚as- Sweden Thierry Hamon (LIMSI & Paris Nord) NLP Approaches March 2014 1 / 126 Introduction Plan Word and sentence segmentation Morphological analysis Parsing of texts Semantic analysis Thierry Hamon (LIMSI & Paris Nord) NLP Approaches March 2014 2 / 126 Segmentation Word and sentence segmentation Thierry Hamon (LIMSI & Paris Nord) NLP Approaches March 2014 3 / 126 Segmentation Introduction (1) Text: a set of characters (string) Segmentation of textual data: identification of a sublist of characters (substring) as a linguistic unit (word, sentence, phrase, term) But it is (very) difficult to define precisely a linguistic unit Linguistic unit identification: the first main step for the natural language processing Several further tasks depend on its quality: content analysis, indexing, part-of-speech tagging, multilingual alignment, etc. Thierry Hamon (LIMSI & Paris Nord) NLP Approaches March 2014 4 / 126 Segmentation Introduction (2) Why word and sentence segmentation? sentences: most of the grammars describe sentences words: basic information provided by dictionaries NB: Words can be simple (book) or complex/compound (French fries, spare time, one-way) Identification of two types of units: units with regular structure (punctuation, number, date, bibliographical references, etc. Units requiring a morphological analysis Thierry Hamon (LIMSI & Paris Nord) NLP Approaches March 2014 5 / 126 Segmentation Example 1: Biosci Biotechnol Biochem. 2003 Aug;67(8):1825-7. Related Articles, Links Comparative Analyses of Hairpin Substrate Recognition by Escherichia coli and Bacillus subtilis Ribonuclease P Ribozymes.
    [Show full text]
  • Towards Acquiring Case Indexing Taxonomies from Text
    Towards Acquiring Case Indexing Taxonomies From Text Kalyan Moy Gupta1, 2 and David W. Aha2 1ITT Industries; AES Division; Alexandria, VA 22303 2Navy Center for Applied Research in Artificial Intelligence; Naval Research Laboratory (Code 5515); Washington, DC 20375 [email protected] Abstract this feature acquisition problem has not been addressed Taxonomic case-based reasoning is a conversational case- previously. The problem is further compounded by based reasoning methodology that employs feature Taxonomic CBR's need to organize features into subsumption taxonomies for incremental case retrieval. subsumption taxonomies (Gupta 2001). Fortunately, this Although this approach has several benefits over standard feature acquisition and organization problem can be retrieval approaches, methods for automatically acquiring addressed by knowledge extraction techniques (Cowie and these taxonomies from text documents do not exist, which Lehnert 1996), which aim to deduce knowledge artifacts limits its widespread implementation. To accelerate and such as rules, cases, and domain models from text. simplify feature acquisition and case indexing, we introduce However, knowledge extraction involves significantly more FACIT, a domain independent framework that combines deep natural language processing techniques and generative complex natural language processing (NLP) methods than lexicons to semi-automatically acquire case indexing do traditional information extraction (IE) approaches. taxonomies from text documents. FACIT employs a novel In this paper, we introduce knowledge extraction method to generate a logical form representation of text, and methodologies in the form of a domain independent uses it to automatically extract and organize features. In framework for feature acquisition and case indexing from contrast to standard information extraction approaches, text (FACIT).
    [Show full text]
  • A Semantic Approach for Extracting Domain Taxonomies from Text
    A Semantic Approach for Extracting Domain Taxonomies from Text Kevin Meijer, Flavius Frasincar∗, Frederik Hogenboom Erasmus University Rotterdam, P.O. Box 1738, NL-3000 DR, Rotterdam, the Netherlands Abstract In this paper we present a framework for the automatic building of a domain taxonomy from text corpora, called Automatic Taxonomy Construction from Text (ATCT). This framework comprises four steps. First, terms are extracted from a corpus of documents. From these extracted terms the ones that are most relevant for a specific domain are selected using a filtering approach in the second step. Third, the selected terms are disambiguated by means of a word sense disambiguation technique and concepts are generated. In the final step, the broader-narrower relations between concepts are determined using a subsumption technique that makes use of concept co-occurrences in text. For evaluation, we assess the performance of the ATCT framework using the semantic precision, semantic recall, and the taxonomic F-measure that take into account the concept semantics. The proposed framework is evaluated in the field of economics and management as well as the medical domain. Keywords: Taxonomy learning, word sense disambiguation, term extraction, subsumption method, semantic taxonomy evaluation 1. Introduction edge is required [5, 13]. Even if the required knowledge is available, it remains a tedious task to organize a high num- In a world where the amount of digital data grows over more ber of concepts in a proper manner. Therefore it is interest- than 50% per year, any means to structure this data becomes in- ing to find ways to automatically build taxonomies [40].
    [Show full text]