Papers reviewed and determined not to be word norm studies. Use the flag icon to report errors or suggest re-inclusion.
18625 papers
The poor grammatical output of Machine Translation (MT) systems appeals syntax-based approaches within language modeling. However, previous studies showed that syntax-based language modeling using (Context-Free) Treebank Grammars was not very helpful in improving BLEU scores for Chinese-English machine translation. In this article we further study this issue in the context of Chinese-English syntax-based Statistical Machine Translation (SMT) where Synchronous Tree Substitution Grammars (STSGs) are utilized to model the translation process. In particular, we develop a Tree Substitution Grammar-based language model for syntax-based MT, and present three methods to efficiently integrate the proposed language model into MT decoding. In addition, we design a simple and effective method to adapt syntax-based language models for MT tasks. We demonstrate that the proposed methods are able to benefit a state-of-the-art syntax-based MT system. On the NIST Chinese-English MT evaluation corpora, we finally achieve an improvement of 0.6 BLEU points over the baseline.
This paper considers how implemented grammars can enhance descriptive ones. An implemented grammar encodes the analyses in machine readable form, facilitating automatic annotation of morphological, syntactic and semantic structures. A grammar augmented with a collection of such structures would enable, for example, a reader to search for items in which a PP argument fills the third most prominent semantic role.
Therapist self-disclosure has been theorized and found to have both positive and negative effects. These effects depend, in part, on the nature of the disclosure. This study sought to examine the differential effects of therapist disclosures of more and less resolved countertransference issues on perceptions of therapists and therapy sessions. Using an analogue method, undergraduate participants (N = 116) were randomly assigned to watch one of two videos in which a therapist disclosed personal issues that were relatively resolved or relatively unresolved. As hypothesized, therapist disclosure of issues that were more resolved caused the therapist to be rated as more attractive and trustworthy and instilled greater hope than therapist disclosure of less resolved issues. The type of therapist disclosure, however, did not affect ratings of the expertness of the therapist, the depth or smoothness of the session, or the perceived universality between client and therapist. Implications of the results for the judicious use of self-disclosure are discussed.
The Internet has rarely been used in auditory perception studies due to concerns about standardisation and calibration across different systems and settings. However, not all auditory research is based on the investigation of fine-grained differences in auditory thresholds. Where meaningful ‘real-world’ listening, for instance the perception of speech, is concerned, the Internet may be a more appropriate and ecologically valid setting to collect data. This study compared affective ratings of low-pass-filtered infant-, foreigner- and British adult-directed speech obtained with traditional methods in the laboratory, with those obtained from an Internet sample. Dropout rates and demographic distribution of participants in the Internet condition were also assessed. The results show that affective ratings were similar for both the Internet and laboratory samples. These findings indicate the viability of Internet-based research into affective speech perception and suggest that precise acoustic environmental control may not always be necessary.
Manually performed treebanking is an expensive effort compared with automatic annotation.In return, manual treebanking is generally believed to provide higherquality/value syntactic annotation than automatic methods.Unfortunately, there is little or no empirical evidence for or against this belief, though arguments have been voiced for the high degree of subjectivity in other levels of linguistic analysis (e.g.morphological annotation).We report a double-blind annotation experiment at the level of dependency syntax, using a small Finnish corpus as the analysis data.The results suggest that an interannotator agreement can be reached as a result of reviews and negotiations that is much higher than the corresponding labelled attachment scores (LAS) reported for stateof-the-art dependency parsers.
Gastel, A., S. Schulze, Y. Versley & E. Hinrichs (2011). Annotation of implicit discourse relations in the TüBa-D/Z treebank. In T. Hedeland, T. Schmidt & K. Wörner (eds.), Multilingual Resources and Multilingual Applications, Proceedings of the German Society of Computational Linguistics and Language Technology (GSCL) 2011, 99–104. Hamburg.
This paper makes two contributions. First, we describe the Hindi Proposition Bank that contains annotations of predicate argument structures of verb predicates. Unlike PropBanks in most other languages, the Hind PropBank is annotated on top of dependency structure, the Hindi Dependency Treebank. We explore the similarities between dependency and predicate argument structures, so the PropBank annotation can be faster and more accurate. Second, we present a probabilistic rule-based system that maps syntactic dependents to semantic arguments. With simple rules, we classify about 47 % of the entire PropBank arguments with over 90 % confidence. These preliminary results are promising; they show how well these two frameworks are correlated. This can also be used to speed up our annotations. 1
We propose a method to improve the accuracy of parsing bilingual texts (bitexts) with the help of statistical machine translation (SMT) systems. Previous bitext parsing methods use human-annotated bilingual treebanks that are hard to obtain. Instead, our approach uses an auto-generated bilingual treebank to produce bilingual constraints. However, because the auto-generated bilingual treebank contains errors, the bilingual constraints are noisy. To overcome this problem, we use large-scale unannotated data to verify the constraints and design a set of effective bilingual features for parsing models based on the verified results. The experimental results show that our new parsers significantly outperform state-of-the-art baselines. Moreover, our approach is still able to provide improvement when we use a larger monolingual treebank that results in a much stronger baseline. Especially notable is that our approach can be used in a purely monolingual setting with the help of SMT.
Latent semantic analysis (LSA) is a statistical technique for representing word meaning that has been widely used for making semantic similarity judgments between words, sentences, and documents. In order to perform an LSA analysis, an LSA space is created in a two-stage procedure, involving the construction of a word frequency matrix and the dimensionality reduction of that matrix through singular value decomposition (SVD). This article presents LANSE, an SVD algorithm specifically designed for LSA, which allows extremely large matrices to be processed using off-the-shelf computer hardware.
The limbic system and especially the amygdala have been identified as key structures in emotion induction and regulation. Recently research has additionally focused on the influence of prefrontal areas on emotion processing in the limbic system and the amygdala. Results from fMRI studies indicate that the prefrontal cortex (PFC) is involved not only in emotion induction but also in emotion regulation. However, studies using fNIRS only report prefrontal brain activation during emotion induction. So far it lacks the attempt to compare emotion induction and emotion regulation with regard to prefrontal activation measured with fNIRS, to exclude the possibility that the reported prefrontal brain activation in fNIRS studies are mainly caused by automatic emotion regulation processes. Therefore this work tried to distinguish emotion induction from regulation via fNIRS of the prefrontal cortex. 20 healthy women viewed neutral pictures as a baseline condition, fearful pictures as induction condition and reappraised fearful pictures as regulation condition in randomized order. As predicted, the view-fearful condition led to higher arousal ratings than the view-neutral condition with the reappraise-fearful condition in between. For the fNIRS results the induction condition showed an activation of the bilateral PFC compared to the baseline condition (viewing neutral). The regulation condition showed an activation only of the left PFC compared to the baseline condition, although the direct comparison between induction and regulation condition revealed no significant difference in brain activation. Therefore our study underscores the results of previous fNIRS studies showing prefrontal brain activation during emotion induction and rejects the hypothesis that this prefrontal brain activation might only be a result of automatic emotion regulation processes.
Until recently, most systems performing temporal extraction and reasoning from text have focused on recognizing and normalizing temporal expressions alone, for which the TIDES annotation scheme has been adopted. Temporal awareness of a text, however, involves not only identifying the temporal expressions, but the events which these expressions anchor, as well as other events which must be ordered relative to them. Because of these broader concerns, TimeML has been developed as an annotation specification that encompasses not only temporal expressions, but all temporally relevant aspects of a text. The annotation schemes, however, are not interchangeable, resulting in incompatible corpora and accompanying extraction algorithms for each standard. In this paper, we describe an automatic migration process from the TIMEX2 tags of TIDES to the TIMEX3 tags of TimeML. This transformation procedure has been implemented and evaluated with two different corpora, obtaining 93.3 and 89.2% overall F-Measure respectively.
It has been shown that the human genome contains extensive copy number variations (CNVs). Investigating the medical and evolutionary impacts of CNVs requires the knowledge of locations, sizes and frequency distribution of them within and between populations. However, CNV study of Chinese minorities, which harbor the majority of genetic diversity of Chinese populations, has been underrepresented considering the same efforts in other populations. Here we constructed, to our knowledge, a first CNV map in seven Chinese populations representing the major linguistic groups in China with 1,440 CNV regions identified using Affymetrix SNP 6.0 Array. Considerable differences in distributions of CNV regions between populations and substantial population structures were observed. We showed that ∼35% of CNV regions identified in minority ethnic groups are not shared by Han Chinese population, indicating that the contribution of the minorities to genetic architecture of Chinese population could not)
Despite tremendous advances in artificial language synthesis, no machine has so far succeeded in deceiving a human. Most research focused on analyzing the behavior of "good" machine. We here choose an opposite strategy, by analyzing the behavior of "bad" humans, i.e., humans perceived as machine. The Loebner Prize in Artificial Intelligence features humans and artificial agents trying to convince judges on their humanness via computer-mediated communication. Using this setting as a model, we investigated here whether the linguistic behavior of human subjects perceived as non-human would enable us to identify some of the core parameters involved in the judgment of an agents' humanness. We analyzed descriptive and semantic aspects of dialogues in which subjects succeeded or failed to convince judges of their humanness. Using cognitive and emotional dimensions in a global behavioral characterization, we demonstrate important differences in the patterns of behavioral expressiveness of the)
In contrast to quantity processing, up to date, the nature of ordinality has received little attention from researchers despite the fact that both quantity and ordinality are embodied in numerical information. Here we ask if there are two separate core systems that lie at the foundations of numerical cognition: (1) the traditionally and well accepted numerical magnitude system but also (2) core system for representing ordinal information. We report two novel experiments of ordinal processing that explored the relation between ordinal and numerical information processing in typically developing adults and adults with developmental dyscalculia (DD). Participants made "ordered" or "non-ordered" judgments about 3 groups of dots (non-symbolic numerical stimuli; in Experiment 1) and 3 numbers (symbolic task: Experiment 2). In contrast to previous findings and arguments about quantity deficit in DD participants, when quantity and ordinality are dissociated (as in the current tasks), DD parti)
Although number words are common in everyday speech, learning their meanings is an arduous, drawn-out process for most children, and the source of this delay has long been the subject of inquiry. Children begin by identifying the few small numerosities that can be named without counting, and this has prompted further debate over whether there is a specific, capacity-limited system for representing these small sets, or whether smaller and larger sets are both represented by the same system. Here we present a formal, computational analysis of number learning that offers a possible solution to both puzzles. This analysis indicates that once the environment and the representational demands of the task of learning to identify sets are taken into consideration, a continuous system for learning, representing and discriminating set-sizes can give rise to effective discontinuities in processing. At the same time, our simulations illustrate how typical prenominal linguistic constructions (''the)
The human population history in Southeast Asia was shaped by numerous migrations and population expansions. Their reconstruction based on archaeological, linguistic or human genetic data is often hampered by the limited number of informative polymorphisms in classical human genetic markers, such as the hypervariable regions of the mitochondrial DNA. Here, we analyse housekeeping gene sequences of the human stomach bacterium Helicobacter pylori from various countries in Southeast Asia and we provide evidence that H. pylori accompanied at least three ancient human migrations into this area: i) a migration from India introducing hpEurope bacteria into Thailand, Cambodia and Malaysia; ii) a migration of the ancestors of Austro-Asiatic speaking people into Vietnam and Cambodia carrying hspEAsia bacteria; and iii) a migration of the ancestors of the Thai people from Southern China into Thailand carrying H. pylori of population hpAsia2. Moreover, the H. pylori sequences reflect iv) the migra)
We measured perceived depth from the optic flow (a) when showing a stationary physical or virtual object to observers who moved their head at a normal or slower speed, and (b) when simulating the same optic flow on a computer and presenting it to stationary observers. Our results show that perceived surface slant is systematically distorted, for both the active and the passive viewing of physical or virtual surfaces. These distortions are modulated by head translation speed, with perceived slant increasing directly with the local velocity gradient of the optic flow. This empirical result allows us to determine the relative merits of two alternative approaches aimed at explaining perceived surface slant in active vision: an ''inverse optics'' model that takes head motion information into account, and a probabilistic model that ignores extra-retinal signals. We compare these two approaches within the framework of the Bayesian theory. The ''inverse optics'' Bayesian model produces veridi)
Exogenous neurotrophin delivery to the deaf cochlea can prevent deafness-induced auditory neuron degeneration, however, we have previously reported that these survival effects are rapidly lost if the treatment stops. In addition, there are concerns that current experimental techniques are not safe enough to be used clinically. Therefore, for such treatments to be clinically transferable, methods of neurotrophin treatment that are safe, biocompatible and can support long-term auditory neuron survival are necessary. Cell transplantation and gene transfer, combined with encapsulation technologies, have the potential to address these issues. This study investigated the survival-promoting effects of encapsulated BDNF over-expressing Schwann cells on auditory neurons in the deaf guinea pig. In comparison to control (empty) capsules, there was significantly greater auditory neuron survival following the cell-based BDNF treatment. Concurrent use of a cochlear implant is expected to result in )
This paper deals with the ways in which minority students in the Danish public school system bring mono-lingually based norms into their poly-lingual peer group interaction. In sequential micro-analyses of interaction we show how the students use the voice of an authority in their reproduction and negotiation of linguistic norms. We base our analyses on the Bakhtinian concept of double-voicing, the Goffmanian concept of keying and Tholander's concept of subteaching. We discuss in detail the relation between the local practices of the students and the linguistic norms expressed in broader society, for instance in newspaper editorials and government papers. (PsycINFO Database Record (c) 2016 APA, all rights reserved)
Reviews the book, A Better Pencil: Readers, Writers, and the Digital Revolution by Dennis Baron (2009). This book addresses the impact of digital technologies on reading and writing practices by contextualizing these developments within the communication technologies that have preceded them. Though computers are often blamed for a perceived deterioration of culture and language, Baron notes that scorn and mistrust of communication technologies is not a new phenomenon; in fact, writing, the printing press, and the telegraph were all greeted at first with skepticism. And despite the pervasive anxiety that 'internet speech' is resulting in the destruction of grammar and spelling, Baron observes that online communities enforce their own linguistic norms. Rather than a linguistic free-for-all, Baron argues, 'spelling counts' in virtual spaces just as it does offline. In sum, Baron accounts for both the benefits and pitfalls of the digital revolution and analyzes it as merely another set of technological innovations, usefully situating them within a history of writing technologies. (PsycINFO Database Record (c) 2016 APA, all rights reserved)
Using the linguistic dependency treebanks and digitized texts created by the Perseus Digital Library, we are creating social networks for a collection of Greek tragedies that allow users to visualize the interactions between characters in the plays. Because the number of characters who appear on stage in Greek tragedy is limited, most of these social network diagrams fall into a few basic types. The most interesting aspect of these networks are, therefore, the edges that connect the nodes within the graphs. The linguistic data used to label or even create these edges becomes the jumping off point for visualizing and exploring the language of Greek tragedy.
Methods for evaluating dependency parsing using attachment scores are highly sensitive to representational variation between dependency treebanks, making cross-experimental evaluation opaque. This paper develops a robust procedure for cross-experimental evaluation, based on deterministic unificationbased operations for harmonizing different representations and a refined notion of tree edit distance for evaluating parse hypotheses relative to multiple gold standards. We demonstrate that, for different conversions of the Penn Treebank into dependencies, performance trends that are observed for parsing results in isolation change or dissolve completely when parse hypotheses are normalized and brought into the same common ground. 1
Phenomena in a variety of verbal tasks—for example, masked priming, lexical decision, and word naming—are typically explained in terms of similarity between word-forms. Despite the apparent commonalities between these sets of phenomena, the representations and similarity measures used to account for them are not often related. To show how this gap might be bridged, we build on the work of Hannagan, Dupoux, and Christophe, Cognitive Science 35:79-118, (2011) to explore several methods of representing visual word-forms using holographic reduced representations and to evaluate them on their ability to account for a wide range of effects in masked form priming, as well as data from lexical decision and word naming. A representation that assumes that word-internal letter groups are encoded relative to word-terminal letter groups is found to predict qualitative patterns in masked priming, as well as lexical decision and naming latencies. We then show how this representation can be integrated with the BEAGLE model of lexical semantics (Jones & Mewhort, Psychological Review 114:1–37, 2007) to enable the model to encompass a wider range of verbal tasks.
Statistical prediction of an outcome variable using multiple independent variables is a common practice in the social and behavioral sciences. For example, neuropsychologists are sometimes called upon to provide predictions of preinjury cognitive functioning for individuals who have suffered a traumatic brain injury. Typically, these predictions are made using standard multiple linear regression models with several demographic variables (e.g., gender, ethnicity, education level) as predictors. Prior research has shown conflicting evidence regarding the ability of such models to provide accurate predictions of outcome variables such as full-scale intelligence (FSIQ) test scores. The present study had two goals: (1) to demonstrate the utility of a set of alternative prediction methods that have been applied extensively in the natural sciences and business but have not been frequently explored in the social sciences and (2) to develop models that can be used to predict premorbid cognitive functioning in preschool children. Predictions of Stanford–Binet 5 FSIQ scores for preschool-aged children is used to compare the performance of a multiple regression model with several of these alternative methods. Results demonstrate that classification and regression treesprovided more accurate predictions of FSIQ scores than does the more traditional regression approach. Implications of these results are discussed.
”[Der findes ikke abstract til denne artikel]”
The accuracy and variability of response times (RTs) collected on stock Apple Macintosh computers using USB keyboards was assessed. A photodiode detected a change in the screen’s luminosity and triggered a solenoid that pressed a key on the keyboard. The RTs collected in this way were reliable, but could be as much as 100 ms too long. The standard deviation of the measured RTs varied between 2.5 and 10 ms, and the distributions approximated a normal distribution. Surprisingly, two recent Apple-branded USB keyboards differed in their accuracy by as much as 20 ms. The most accurate RTs were collected when an external CRT was used to display the stimuli and Psychtoolbox was able to synchronize presentation with the screen refresh. We conclude that RTs collected on stock iMacs can detect a difference as small as 5–10 ms under realistic conditions, and this dictates which types of research should or should not use these systems.
This work aims to improve an N-gram-based statistical machine translation system between the Catalan and Spanish languages, trained with an aligned Spanish–Catalan parallel corpus consisting of 1.7 million sentences taken from El Periódico newspaper. Starting from a linguistic error analysis above this baseline system, orthographic, morphological, lexical, semantic and syntactic problems are approached using a set of techniques. The proposed solutions include the development and application of additional statistical techniques, text pre- and post-processing tasks, and rules based on the use of grammatical categories, as well as lexical categorization. The performance of the improved system is clearly increased, as is shown in both human and automatic evaluations of the system, with a gain of about 1.1 points BLEU observed in the Spanish-to-Catalan direction of translation, and a gain of about 0.5 points in the reverse direction. The final system is freely available online as a linguistic resource.
In this paper, we present our attempts to design and implement a large-coverage computational grammar for the Persian language based on the Generalized Phrase Structured Grammar (GPSG) model. This grammatical model was developed for continuous speech recognition (CSR) applications, but is suitable for other applications that need the syntactic analysis of Persian. In this work, we investigate various syntactic structures relevant to the modern Persian language, and then describe these structures according to a phrase structure model. Noun (N), Verb (V), Adjective (ADJ), Adverb (ADV), and Preposition (P) are considered basic syntactic categories, and X-bar theory is used to define Noun phrases, Verb phrases, Adjective phrases, Adverbial phrases, and Prepositional phrases. However, we have to extend Noun phrase levels in X-bar theory to four levels due to certain complexities in the structure of Noun phrases in the Persian language. A set of 120 grammatical rules for describing different phrase structures of Persian is extracted, and a few instances of the rules are presented in this paper. These rules cover the major syntactic structures of the modern Persian language. For evaluation, the obtained grammatical model is utilized in a bottom-up chart parser for parsing 100 Persian sentences. Our grammatical model can take 89 sentences into account. Incorporating this grammar in a Persian CSR system leads to a 31% reduction in word error rate.
From two corpus studies into varieties of clausal coordination in English (Meyer, 1995 and Greenbaum & Nelson, 1999), it is known that the incidence of clausal coordinate ellipsis (CCE) is about two times higher in written than in spoken language. We present a treebank study into CCE in written and spoken Dutch and German which confirms this tendency. Moreover, we observe considerable differences between written and spoken language with respect to the incidence of four main types of clausal coordinate ellipsis—Gapping, Forward Conjunction Reduction (FCR), Backward Conjunction Reduction (BCR), and Subject Gap with Finite/Fronted Verb (SGF). We argue that the detailed data pattern cannot be accounted for in terms of audience design, and propose an explanation based on the assumption that during spontaneous speaking—but not during writing—, the scope of online grammatical planning is basically restricted to one (finite) clause.
We present an approach for Seman-tic Role Labeling (SRL) using Condi-tional Random Fields in a joint identifi-cation/classification step. The approach is based on shallow syntactic information (chunks) and a number of lexicalized fea-tures such as selectional preferences and automatically inferred similar words, ex-tracted using lexical databases and distri-butional similarity metrics. We use se-mantic annotations from the Proposition Bank for training and evaluate the system using CoNLL-2005 test sets. The addi-tional lexical information led to improve-ments of 15 % (in-domain evaluation) and 12 % (out-of-domain evaluation) on over-all semantic role classification in terms of F-measure. The gains come mostly from a better recall, which suggests that the addi-tion of richer lexical information can im-prove the coverage of existing SRL mod-els even when very little syntactic knowl-edge is available. 1
The paper describes the basic strategies behind the word and semantic level alignment in the Bulgarian-English treebank. The word level alignment has taken into consideration the experience within other NLP groups in the context of the Bulgarian language specific features. The semantic level alignment builds on the word level alignment and is represented in the framework of the Minimal Recursion Semantics. 1
Annotated corpora are essential for almost all NLP applications. Whereas they are expected to be of a very high quality because of their importance for the followup developments, they still contain a considerable number of errors. With this work we want to draw attention to this fact. Additionally, we try to estimate the amount of errors and propose a method for their automatic correction. Whereas our approach is able to find only a portion of the errors that we suppose are contained in almost any annotated corpus due to the nature of the process of its creation, it has a very high precision, and thus is in any case beneficial for the quality of the corpus it is applied to. At last, we compare it to a different method for error detection in treebanks and find out that the errors that we are able to detect are mostly different and that our approaches are complementary. 1
Recursive structure is commonly found in the inputs of different modalities such as natural scene images or natural language sentences. Discovering this recursive structure helps us to not only identify the units that an image or sentence contains but also how they interact to form a whole. We introduce a max-margin structure prediction architecture based on recursive neural networks that can successfully recover such structure both in complex scene images as well as sentences. The same algorithm can be used both to provide a competitive syntactic parser for natural language sentences from the Penn Treebank and to outperform alternative approaches for semantic scene segmentation, annotation and classification. For segmentation and annotation our algorithm obtains a new level of state-of-theart performance on the Stanford background dataset (78.1%). The features from the image parse tree outperform Gist descriptors for scene classification by 4%. 1.
We outline a method of detecting ad hoc, or anomalous, rules in treebank grammars, by exploiting the fact that such rules do not fit with the rest of the grammar. Ad hoc rules are rules used for specific constructions in one data set and unlikely to be used again. These include ungeneralizable rules, erroneous rules, rules for ungrammatical text, and rules which are not consistent with the rest of the annotation scheme. Based on the idea that valid rules should receive support from other rules in the grammar, we develop two methods for detecting ad hoc rules in flat treebanks and show they are successful in detecting such rules. Although one can put some linguistic knowledge into determining rule similarity and dissimilarity, the methods work best by using a simple, modified Levenshtein distance. We illustrate this on the English Wall Street Journal treebank and the German TIGER treebank. For the latter, we extend the method to formalisms incorporating discontinuous constituents, employing CFG-like rules for the comparisons.
In this paper, we first analyze and classify the empty categories in a Hindi dependency tree-bank and then identify various discovery procedures to automatically detect the existence of these categories in a sentence. For this we make use of lexical knowledge along with the parsed output from a constraint based parser. Through this work we show that it is possible to successfully discover certain types of empty categories while some other types are more difficult to identify. This work leads to the state-of-the-art system for automatic insertion of empty categories in the Hindi sentence.
We describe a method for training a semantic role labeler for CCG in the absence of gold-standard syntax derivations. Traditionally, semantic role labeling is performed by placing human-annotated semantic roles on gold-standard syntactic parses, identifying patterns in the syntaxsemantics relationship, and then predicting roles on novel syntactic analyses. The gold standard syntactic training data can be eliminated from the process by extracting training instances from semantic roles projected onto a packed parse chart. This process can be used to rapidly develop NLP tools for resource-poor languages of interest. 1
In this paper we present a user-centered approach for defining the dependency syntactic specification for a treebank. We show that by collecting information on syntactic interpretations from the future users of the treebank, we can model so far dependency-syntactically undefined syntactic structures in a way that corresponds to the users ’ intuition. By consulting the users at the grammar definition phase we aim at better usability of the treebank in the future. We focus on two complex syntactic phenomena: elliptical comparative clauses and participial NPs or NPs with a verbderived noun as their head. We show how the phenomena can be interpreted in several ways and ask for the users ’ intuitive way of modeling them. The results aid in constructing the syntactic specification for the treebank. 1