We initially describe a feature-rich discriminative Conditional Random Field (CRF) model for Information Extraction in the workshop announcements domain, which offers good baseline performance in the PASCAL shared task. We then propose a method for leveraging domain knowledge in Information Extraction tasks, scoring candidate document labellings as one-value-per-field templates according to domain feasibility after generating sample labellings from a trained sequence classifier. Our relational models evaluate these templates according to our intuitions about agreement in the domain: workshop acronyms should resemble (...) their names, workshop dates occur after paper submission dates. These methods see a 5% f-score improvement in fields retrieved when sampling labellings from a Maximum-Entropy Markov Model, however we do not observe improvement over a CRF model. We discuss reasons for this, including the problem of recovering all field instances from a best template, and propose future work in adapting such a model to the CRF, a better standalone system. (shrink)

1 Boolean retrieval 1 2 The term vocabulary and postings lists 19 3 Dictionaries and tolerant retrieval 49 4 Index construction 67 5 Index compression 85 6 Scoring, term weighting and the vector space model 109 7 Computing scores in a complete search system 135 8 Evaluation in information retrieval 151 9 Relevance feedback and query expansion 177 10 XML retrieval 195 11 Probabilistic information retrieval 219 12 Language models for information retrieval 237 13 Text classification and Naive Bayes 253 (...) 14 Vector space classification 289 15 Support vector machines and machine learning on documents 319 16 Flat clustering 349 17 Hierarchical clustering 377 18 Matrix decompositions and latent semantic indexing 403 19 Web search basics 421 20 Web crawling and indexes 443 21 Link analysis 461.. (shrink)

assumptions latent in a vanilla treebank grammar. Indeed, its performance of 86.36% (LP/LR F1) is..

This paper describes a system for extracting typed dependency parses of English sentences from phrase structure parses. In order to capture inherent relations occurring in corpus texts that can be critical in real-world applications, many NP relations are included in the set of grammatical relations used. We provide a comparison of our system with Minipar and the Link parser. The typed dependency extraction facility described here is integrated in the Stanford Parser, available for download.

“Everyone knows that language is variable.” This is the bald sentence with which Sapir (1921:147) begins his chapter on language as an historical product. He goes on to emphasize how two speakers’ usage is bound to differ “in choice of words, in sentence structure, in the relative frequency with which particular forms or combinations of words are used”. I should add that much sociolinguistic and historical linguistic research has shown that the same speaker’s usage is also variable (Labov 1966, Kroch (...) 2001:722). However, the tradition of most syntacticians has been to ignore this thing that everyone knows. (shrink)

separated from evaluating entailment. Current approaches to semantic inference in question answer-.

first-order HMM, the current tag t0 is predicted based on the previous tag t−1 (and the current word).1 The back- We present a new part-of-speech tagger that ward interaction between t0 and the next tag t+1 shows demonstrates the following ideas: (i) explicit up implicitly later, when t+1 is generated in turn. While unidirectional models are therefore able to capture both use of both preceding and following tag con-.

This paper presents the first use of a computational model of natural logic—a system of logical inference which operates over natural language—for textual inference. Most current approaches to the PAS- CAL RTE textual inference task achieve robustness by sacrificing semantic precision; while broadly effective, they are easily confounded by ubiquitous inferences involving monotonicity. At the other extreme, systems which rely on first-order logic and theorem proving are precise, but excessively brittle. This work aims at a middle way. Our system finds (...) a low-cost edit sequence which transforms the premise into the hypothesis; learns to classify entailment relations across atomic edits; and composes atomic entailments into a top-level entailment judgment. We provide the first reported results for any system on the FraCaS test suite. We also evaluate on RTE3 data, and show that hybridizing an existing RTE system with our natural logic system yields significant performance gains. (shrink)

I wish to present a codi cation of syntactic approaches to dealing with ergative languages and argue for the correctness of one particular approach, which I will call the Inverse Grammatical Relations hypothesis.1 I presume familiarity with the term `ergativity', but, brie y, many languages have ergative case marking, such as Burushaski in (1), in contrast to the accusative case marking of Latin in (2). More generally, if we follow Dixon (1979) and use A to mark the agent-like argument of (...) a transitive verb, O to mark the patient-like argument of a transitive verb, and S to mark the single argument of an intransitive verb, then we can call ergative any subsystem of a language that groups S and O in contrast to A, as shown in (3). (shrink)

Most current statistical natural language processing models use only local features so as to permit dynamic programming in inference, but this makes them unable to fully account for the long distance structure that is prevalent in language use. We show how to solve this dilemma with Gibbs sam- pling, a simple Monte Carlo method used to perform approximate inference in factored probabilistic models. By using simulated annealing in place of Viterbi decoding in sequence models such as HMMs, CMMs, and CRFs, (...) it is possible to incorporate non-local structure while preserving tractable inference. We use this technique to augment an existing CRF-based information extraction system with long-distance dependency models, enforcing label consistency and extraction template consistency constraints. This technique results in an error reduction of up to 9% over state-of-the-art systems on two established information extraction tasks. (shrink)

We propose a model of natural language inference which identifies valid inferences by their lexical and syntactic features, without full semantic interpretation. We extend past work in natural logic, which has focused on semantic containment and monotonicity, by incorporating both semantic exclusion and implicativity. Our model decomposes an inference problem into a sequence of atomic edits linking premise to hypothesis; predicts a lexical semantic relation for each edit; propagates these relations upward through a semantic composition tree according to properties of (...) intermediate nodes; and joins the resulting semantic relations across the edit sequence. A computational implementation of the model achieves 70% accuracy and 89% precision on the FraCaS test suite. Moreover, including this model as a component in an existing system yields significant performance gains on the Recognizing Textual Entailment challenge. (shrink)

Kristina Toutanova Christopher D. Manning Dept of Computer Science Depts of Computer Science and Linguistics Gates Bldg 4A, 353 Serra Mall Gates Bldg 4A, 353 Serra Mall Stanford, CA 94305–9040, USA Stanford, CA 94305–9040, USA [email protected] manning@cs.stanford.edu..

This paper proposes a new architecture for textual inference in which finding a good alignment is separated from evaluating entailment. Current approaches to semantic inference in question answering and textual entailment have approximated the entailment problem as that of computing the best alignment of the hypothesis to the text, using a locally decomposable matching score. While this formulation is adequate for representing local (word-level) phenomena such as synonymy, it is incapable of representing global interactions, such as that between verb negation (...) and the addition/removal of qualifiers, which are often critical for determining entailment. We propose a pipelined approach where alignment is followed by a classification step, in which we extract features representing high-level characteristics of the entailment problem, and give the resulting feature vector to a statistical classifier trained on development data. (shrink)

Marie-Catherine de Marneffe, Anna N. Rafferty and Christopher D. Manning Linguistics Department Computer Science Department Stanford University Stanford University Stanford, CA 94305 Stanford, CA 94305 {rafferty,manning}@stanford.edu [email protected]..

Nathanael Chambers, Daniel Cer, Trond Grenager, David Hall, Chloe Kiddon Bill MacCartney, Marie-Catherine de Marneffe, Daniel Ramage Eric Yeh, Christopher D. Manning Computer Science Department Stanford University Stanford, CA 94305.

We propose an approach to natural language inference based on a model of natural logic, which identifies valid inferences by their lexical and syntactic features, without full semantic interpretation. We greatly extend past work in natural logic, which has focused solely on semantic containment and monotonicity, to incorporate both semantic exclusion and implicativity. Our system decomposes an inference problem into a sequence of atomic edits linking premise to hypothesis; predicts a lexical entailment relation for each edit using a statistical classifier; (...) propagates these relations upward through a syntax tree according to semantic properties of intermediate nodes; and composes the resulting entailment relations across the edit sequence. We evaluate our system on the FraCaS test suite, and achieve a 27% reduction in error from previous work. We also show that hybridizing an existing RTE system with our natural logic system yields significant gains on the RTE3 test suite. (shrink)

We present a system for deciding whether a given sentence can be inferred from text. Each sentence is represented as a directed graph (extracted from a dependency parser) in which the nodes represent words or phrases, and the links represent syntactic and semantic relationships. We develop a learned graph matching model to approximate entailment by the amount of the sentence’s semantic content which is contained in the text. We present results on the Recognizing Textual Entailment dataset (Dagan et al., 2005), (...) and show that our approach outperforms Bag- Of-Words and TF-IDF models. (shrink)

Pi-Chuan Chang, Michel Galley, and Christopher D. Manning Computer Science Department, Stanford University Stanford, CA 94305 pichuan,galley,manning@cs.stanford.edu..

I examine what would be necessary to move part-of-speech tagging performance from its current level of about 97.3% token accuracy (56% sentence accuracy) to close to 100% accuracy. I suggest that it must still be possible to greatly increase tagging performance and examine some useful improvements that have recently been made to the Stanford Part-of-Speech Tagger. However, an error analysis of some of the remaining errors suggests that there is limited further mileage to be had either from better machine learning (...) or better features in a discriminative sequence classifier. The prospects for further gains from semisupervised learning also seem quite limited. Rather, I suggest and begin to demonstrate that the largest opportunity for further progress comes from improving the taxonomic basis of the linguistic resources from which taggers are trained. That is, from improved descriptive linguistics. However, I conclude by suggesting that there are also limits to this process. The status of some words may not be able to be adequately captured by assigning them to one of a small number of categories. While conventions can be used in such cases to improve tagging consistency, they lack a strong linguistic basis. (shrink)

This paper presents a new method for producing a dictionary of subcategorization frames from unlabelled text corpora. It is shown that statistical filtering of the results of a finite state parser running on the output of a stochastic tagger produces high quality results, despite the error rates of the tagger and the parser. Further, it is argued that this method can be used to learn all subcategorization frames, whereas previous methods are not extensible to a general solution to the problem.

Discriminative feature-based methods are widely used in natural language processing, but sentence parsing is still dominated by generative methods. While prior feature-based dynamic programming parsers have restricted training and evaluation to artificially short sentences, we present the first general, featurerich discriminative parser, based on a conditional random field model, which has been successfully scaled to the full WSJ parsing data. Our efficiency is primarily due to the use of stochastic optimization techniques, as well as parallelization and chart prefiltering. On WSJ15, (...) we attain a state-of-the-art F-score of 90.9%, a 14% relative reduction in error over previous models, while being two orders of magnitude faster. On sentences of length 40, our system achieves an F-score of 89.0%, a 36% relative reduction in error over a generative baseline. (shrink)

The same categorical phenomena which are attributed to hard grammatical constraints in some languages continue to show up as statistical preferences in other languages, motivating a grammatical model that can account for soft constraints. The effects of a hierarchy of person (1st, 2nd 3rd) on grammar are categorical in some languages, most famously in languages withError: Illegal entry in bfrange block in ToUnicode CMap inverse systems, but also in languages with person restrictions on passivization. In Lummi, for example, the person (...) of the subject argument cannot be lower than the person of a nonsubject argument. If this would happen in the active, passivization is obligatory; if it would happen in the passive, the active is obligatory (Jelinek and Demers 1983). These facts follow from the theory of harmonic alignment in OT: constraints favoring the harmonic association of prominent person (1st, 2nd) with prominent syntactic function (subject) are hypothesized to be present as subhierarchies of the grammars of all languages, but to vary in their effects across languages depending on their interactions with other constraints (Aissen 1999). There is a statistical reflection of these hierarchies in English. The same disharmonic person/argument associations which are avoided categorically in languages like Lummi by making passives either impossible or obligatory, are avoided in the SWITCHBOARD corpus of spoken English by either depressing or elevating the frequency of passives relative to actives. The English data can be grammatically analyzed within the stochastic OT framework (Boersma 1998, Boersma and Hayes 2001) in a way which provides a principled and unifying explanation for their relation to the crosslinguistic categorical person effects studied by Aissen (1999). (shrink)

Grammatical theory has long wrestled with the fact that causative constructions exhibit properties of both single words and complex phrases. However, as Paul Kiparsky has observed, the distribution of such properties of causatives is not arbitrary: ‘construal’ phenomena such as honorification, anaphor and pronominal binding, and quantifier ‘floating’ typically behave as they would if causatives were syntactically complex, embedding constructions; whereas case marking, agreement and word order phenomena all point to the analysis of causatives as single lexical items.1 Although an (...) analysis of causatives in terms of complex syntactic structures has frequently been adopted in an attempt to simplify the mapping to semantic structure, we believe that motivating syntactic structure based on perceived semantics is questionable because in general a syntax/semantics homomorphism cannot be maintained without vitiating syntactic theory (Miller 1991). Instead, we sketch a strictly lexical theory of Japanese causatives that deals with the evidence offered for a complex phrasal analysis. Such an analysis makes the phonology, morphology and syntax parallel, while a mismatch occurs with the semantics. The conclusions we will reach are given in (1). (shrink)

We present a novel generative model for natural language tree structures in which semantic (lexical dependency) and syntactic (PCFG) structures are scored with separate models. This factorization provides conceptual simplicity, straightforward opportunities for separately improving the component models, and a level of performance comparable to similar, non-factored models. Most importantly, unlike other modern parsing models, the factored model admits an extremely effective A* parsing algorithm, which enables efficient, exact inference.

Michel Galley, Pi-Chuan Chang, Daniel Cer, Jenny R. Finkel, and Christopher D. Manning Computer Science and Linguistics Departments Stanford University..

The alignment problem—establishing links between corresponding phrases in two related sentences—is as important in natural language inference (NLI) as it is in machine translation (MT). But the tools and techniques of MT alignment do not readily transfer to NLI, where one cannot assume semantic equivalence, and for which large volumes of bitext are lacking. We present a new NLI aligner, the MANLI system, designed to address these challenges. It uses a phrase-based alignment representation, exploits external lexical resources, and capitalizes on (...) a new set of supervised training data. We compare the performance of MANLI to existing NLI and MT aligners on an NLI alignment task over the well-known Recognizing Textual Entailment data. We show that MANLI significantly outperforms existing aligners, achieving gains of 6.2% in F1 over a representative NLI aligner and 10.5% over GIZA++. (shrink)

The correct locus (or loci) of binding theory has been a matter of much discussion. Theories can be seen as varying along at least two dimensions. The rst is whether binding theory is con gurationally determined (that is, the theory exploits the geometry of a phrase marker, appealing to such purely structural notions as c-command and government) or whether the theory depends rather on examining the relations between items selected by a predicate (where by selection I am intending to cover (...) everything from semantic dependencies to syntactic subcategorization). The second is the level of grammar on which binding is de ned. Attempting to roughly equate levels across di erent theories, suggestions have included the semantics/lexical conceptual structure (Jackendo 1992), thematic structure (Jackendo 1972, Wilkins 1988), argument structure/D-structure/initial grammatical relations (Manning 1994, Belletti and Rizzi 1988, Perlmutter 1984), surface syntax/grammatical relations, logical form, linear order, pragmatics (Levinson 1991), and discourse (Iida 1992). The data is su ciently varied and complex that many theories end up as mixtures, variously employing a combination of elements along both dimensions (for instance, Chomsky (1986) relies purely on con gurational notions for the relationship between an anaphor and its antecedent, but uses concepts from selection in the de nition of the binding domain of an anaphor). LFG has always rejected a con gurational account of binding. For instancError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMapError: Illegal entry in bfrange block in ToUnicode CMape, Simpson (1991) argues that a con gurational theory of binding in Warlpiri cannot be maintained, among other reasons because nite clauses lack a VP.. (shrink)

Abeill´e and Godard (1994) seek to show that the rightward branching analysis of French tense auxiliaries shown in (1b), that I argued for in Manning (1992) and which is widely adopted in general, is wrong, and that rather we should adopt a flat analysis for this construction as shown in (1c), and they show how such an analysis can be realized within HPSG (Pollard and Sag 1994).

Technology for local textual inference is central to producing a next generation of intelligent yet robust human language processing systems. One can think of it as Information Retrieval++. It is needed for a search on male fertility may be affected by use of cell phones to match a document saying Startling new research into mobile phones suggests they can reduce a man’s sperm count up to 30%, despite the fact that the only word overlap is phones. But textual inference is (...) useful more broadly. It is an enabling technology for applications of document interpretation, such as customer response management, where one would like to conclude from the message My Squeezebox regularly skips during music playback that Sender has set up Squeezebox and Sender can hear music through Squeezebox, and information extraction, where from the text Jorma Ollila joined Nokia in 1985 and held a variety of key management positions before taking the helm in 1992, one wants to extract that Jorma Ollila has served as the CEO of Nokia, a relation that might be more formally denoted as role(CEO, Nokia, Jorma Ollila). Textual inference is a difficult problem (as the results from early evaluations have shown): current systems do statistically better than random guessing, but not by very much. Nevertheless, it is also an area where there is promising developing technology and a good deal of natural language community interest. In other words, it is an ideal research problem. To further this research agenda, data sets have been constructed to assess textual inference systems. This paper examines how the task of textual inference has been and should be defined and discusses what kind of evaluation data is appropriate for the task.1.. (shrink)

fundamental rule” in an order-independent manner, such that the same basic algorithm supports top-down and Most PCFG parsing work has used the bottom-up bottom-up parsing, and the parser deals correctly with CKY algorithm (Kasami, 1965; Younger, 1967) with the difficult cases of left-recursive rules, empty elements, Chomsky Normal Form Grammars (Baker, 1979; Jeand unary rules, in a natural way.

We present a Chinese word segmentation system submitted to the closed track of Sighan bakeoff 2005. Our segmenter was built using a conditional random field sequence model that provides a framework to use a large number of linguistic features such as character identity, morphological and character reduplication features. Because our morphological features were extracted from the training corpora automatically, our system was not biased toward any particular variety of Mandarin. Thus, our system does not overfit the variety of Mandarin most (...) familiar to the system's designers. Our final system achieved a F-score of.. (shrink)

This paper shows that a simple two-stage approach to handle non-local dependencies in Named Entity Recognition (NER) can outperform existing approaches that handle non-local dependencies, while being much more computationally efficient. NER systems typically use sequence models for tractable inference, but this makes them unable to capture the long distance structure present in text. We use a Conbel.

Determining the semantic role of sentence constituents is a key task in determining sentence meanings lying behind a veneer of variant syntactic expression. We present a model of natural language generation from semantics using the FrameNet semantic role and frame ontology. We train the model using the FrameNet corpus and apply it to the task of automatic semantic role and frame identification, producing results competitive with previous work (about 70% role labeling accuracy). Unlike previous models used for this task, our (...) model does not assume that the frame of a sentence is known, and is able to identify null- instantiated roles, which commonly occur in our corpus and whose identification is crucial to natural language interpretation. (shrink)

adjacent phrases, but they typically lack the ability to perform the kind of long-distance reorderings possible with syntax-based systems. In this paper, we present a novel hierarchical phrase reordering model aimed at improving non-local reorderings, which seamlessly integrates with a standard phrase-based system with little loss of computational efficiency. We show that this model can successfully handle the key examples often used to motivate syntax-based systems, such as the rotation of a prepositional phrase around a noun phrase. We contrast our (...) model with reordering models commonly used in phrase-based systems, and show that our approach provides statistically significant BLEU point gains for two language pairs: Chinese-English (+0.53 on MT05 and +0.71 on MT08) and Arabic-English (+0.55 on MT05). (shrink)

We present a maximum-entropy based system for identifying Named Entities (NEs) in biomedical abstracts and present its performance in the only two biomedical Named Entity Recognition (NER) comparative evaluations that have been held to date, namely BioCreative and Coling BioNLP. Our system obtained an exact match f-score of 83.2% in the BioCreative evaluation and 70.1% in the BioNLP evaluation. We discuss our system in detail including its rich use of local features, attention to correct boundary identification, innovative use of external (...) knowledge resources including parsing and web searches, and rapid adaptation to new NE sets. We also discuss in depth problems with data annotation in the evaluations which caused the final performance to be lower than the optimal. (shrink)

We present a novel algorithm for the fast computation of PageRank, a hyperlink-based estimate of the “importance” of Web pages. The original PageRank algorithm uses the Power Method to compute successive iterates that converge to the principal eigenvector of the Markov matrix representing the Web link graph. The algorithm presented here, called Quadratic Extrapolation, accelerates the convergence of the Power Method by periodically subtracting off estimates of the nonprincipal eigenvectors from the current iterate of the Power Method. In Quadratic Extrapolation, (...) we take advantage of the fact that the first eigenvalue of a Markov matrix is known to be 1 to compute the nonprincipal eigenvectors using successive iterates of the Power Method. Empirically, we show that using Quadratic Extrapolation speeds up PageRank computation by 50- 300% on a Web graph of 80 million nodes, with minimal overhead. (shrink)

We present a novel technique for speeding up the computation of PageRank, a hyperlink-based estimate of the “importance” of Web pages, based on the ideas presented in [7]. The original PageRank algorithm uses the Power Method to compute successive iterates that converge to the principal eigenvector of the Markov matrix representing the Web link graph. The algorithm presented here, called Power Extrapolation, accelerates the convergence of the Power Method by subtracting off the error along several nonprincipal eigenvectors from the current (...) iterate of the Power Method, making use of known nonprincipal eigenvalues of the Web hyperlink matrix. Empirically, we show that using Power Extrapolation speeds up PageRank computation by 30% on a Web graph of 80 million nodes in realistic scenarios over the standard power method, in a way that is simple to understand and implement. (shrink)

Automatically clustering web pages into semantic groups promises improved search and browsing on the web. In this paper, we demonstrate how user-generated tags from largescale social bookmarking websites such as del.icio.us can be used as a complementary data source to page text and anchor text for improving automatic clustering of web pages. This paper explores the use of tags in 1) K-means clustering in an extended vector space model that includes tags as well as page text and 2) a novel (...) generative clustering algorithm based on latent Dirichlet allocation that jointly models text and tags. We evaluate the models by comparing their output to an established web directory. We find that the naive inclusion of tagging data improves cluster quality versus page text alone, but a more principled inclusion can substantially improve the quality of all models with a statistically signifi- cant absolute F-score increase of 4%. The generative model outperforms K-means with another 8% F-score increase. (shrink)

In Pollard and Sag (1987) and Pollard and Sag (1994:Ch. 1–8), the subcategorized arguments of a head are stored on a single ordered list, the subcat list. However, Borsley (1989) argues that there are various defi- ciencies in this approach, and suggests that the unified list should be split into separate lists for subjects, complements, and specifiers. This proposal has been widely adopted in what is colloquially known as HPSG3 (Pollard and Sag (1994:Ch. 9) and other recent work in HPSG). (...) Such a move provides in HPSG an analog of the external/internal argument distinction generally adopted in GB, solves certain technical problems such as allowing prepositions to take complements rather than things identical in subcat list position to subjects, and allows recognition of the special features of subjects which have been noted in the LFG literature, where keyword grammatical relations are used. In HPSG3, it is these valence features subj, comps and spr whose values are ‘cancelled off’ (in a Categorial Grammar-like manner) as a head projects a phrase. A lexical head combines with its complements and subject or specifier (if any) according to the lexically inherited specification, as in (1). (shrink)

Linking constructions involving dሇ (DE) are ubiquitous in Chinese, and can be translated into English in many different ways. This is a major source of machine translation error, even when syntaxsensitive translation models are used. This paper explores how getting more information about the syntactic, semantic, and discourse context of uses of dሇ (DE) can facilitate producing an appropriate English translation strategy. We describe a finergrained classification of dሇ (DE) constructions in Chinese NPs, construct a corpus of annotated examples, and (...) then train a log-linear classifier, which contains linguistically inspired features. We use the DE classifier to preprocess MT data by explicitly labeling dሇ (DE) constructions, as well as reordering phrases, and show that our approach provides significant BLEU point gains on MT02 (+1.24), MT03 (+0.88) and MT05 (+1.49) on a phrasedbased system. The improvement persists when a hierarchical reordering model is applied. (shrink)

We describe a machine learning system for the recognition of names in biomedical texts. The system makes extensive use of local and syntactic features within the text, as well as external resources including the web and gazetteers. It achieves an F- score of 70% on the Coling 2004 NLPBA/BioNLP shared task of identifying five biomedical named entities in the GENIA corpus.

translation. We present a method for using part of speech tag information to improve alignment accu-.

A desirable quality of a coreference resolution system is the ability to handle transitivity constraints, such that even if it places high likelihood on a particular mention being coreferent with each of two other mentions, it will also consider the likelihood of those two mentions being coreferent when making a final assignment. This is exactly the kind of constraint that integer linear programming (ILP) is ideal for, but, surprisingly, previous work applying ILP to coreference resolution has not encoded this type (...) of constraint. We train a coreference classifier over pairs of mentions, and show how to encode this type of constraint on top of the probabilities output from our pairwise classifier to extract the most probable legal entity assignments. We present results on two commonly used datasets which show that enforcement of transitive closure consistently improves performance, including improvements of up to 3.6% using the b3 scorer, and up to 16.5% using cluster f-measure. (shrink)

This paper examines feature selection for log linear models over rich constraint-based grammar (HPSG) representations by building decision trees over features in corresponding probabilistic context free grammars (PCFGs). We show that single decision trees do not make optimal use of the available information; constructed ensembles of decision trees based on different feature subspaces show signifi- cant performance gains (14% parse selection error reduction). We compare the performance of the learned PCFG grammars and log linear models over the same features.

Jackendoff (1987, 1990) has brought up various problems with the current use of thematic roles (Kiparsky, 1987; Bresnan & Kanerva, 1989 and references cited therein) and suggested a different way of thinking of thematic roles as structural configurations in his semantic Lexical Conceptual Structures (LCSs). Conversely, Joshi (1989) has claimed that Jackendoff’s LCSs alone are insufficient, and that an analysis of certain facts in Marathi additionally requires the existence of a level of predicate-argument structure (PAS). Below we will mention a (...) few of Jackendoff’s arguments against the current conception of thematic roles. We will then look at Joshi’s arguments about the necessity of a level of PAS in addition to LCS and conclude that providing Jackendoff’s LCSs are integrated into a suitable syntactic theory, neither of her points are problematic to Jackendoff.1 From there we will go on to re-examine some of the facts of Marathi, and show that certain facts that have merely been stipulated or left unanalyzed when using thematic roles, receive a rather elegant treatment when described via a combination of their syntax and LCS. (shrink)

In Sebastian Thrun, Lawrence K. Saul, and Bernhard Schölkopf (eds), Advances in Neural Information Processing Systems 16 (NIPS 2003). Cambridge, MA: MIT Press, pp. 497-504.

Many NLP tasks rely on accurately estimating word dependency probabilities P(w1|w2), where the words w1 and w2 have a particular relationship (such as verb-object). Because of the sparseness of counts of such dependencies, smoothing and the ability to use multiple sources of knowledge are important challenges. For example, if the probability P(N |V ) of noun N being the subject of verb V is high, and V takes similar objects to V , and V is synonymous to V , then (...) we want to conclude that P(N |V ) should also be reasonably high—even when those words did not cooccur in the training data. (shrink)

Part-of-speech tagging, like any supervised statistical NLP task, is more difficult when test sets are very different from training sets, for example when tagging across genres or language varieties. We examined the problem of POS tagging of different varieties of Mandarin Chinese. An analytic study first showed that unknown words were a major source of difficulty in cross-variety tagging. Unknown words in English tend to be proper nouns. By contrast, we found that Mandarin unknown words were mostly common nouns and (...) verbs. We showed these results are caused by the high frequency of morphological compounding in Mandarin; in this sense Mandarin is more like German than English. Based on this analysis, we propose a variety of new morphological unknown-word features for POS tagging, extending earlier work by others on unknown-word tagging in English and German. Our features were implemented in a maximum entropy Markov model. Our system achieves state-of-the-art performance in Mandarin tagging, including improving unknown-word tagging performance on unseen varieties in Chinese Treebank 5.0 from 61% to 80% correct. (shrink)