Variants of Persian (Farsi, Dari, and Tajiki) are spoken by more than 110 million people worldwide. Despite the growing interest in broadening the coverage of Natural Language Processing (NLP) methods, Persian has received limited attention. This survey addresses this gap by reviewing over 200 peer-reviewed studies published across the last two decades, documenting more than 40 publicly available corpora, and analysing linguistic challenges, pipelines, and applications across approaches ranging from rule-based methods to Large Language Models (LLMs). Given the rapid advancement of LLMs, we study the state of Persian-specific models across different tasks, review the available LLMs for Persian, assess their performance and limitations, and provide a comprehensive discussion of their strengths and gaps. Our analysis shows that although attention to Persian NLP has increased and notable progress has been made, critical gaps remain—particularly the scarcity of standardised benchmarks and gold corpora in tasks such as toxicity and safety, bias and fairness, and explainability and reasoning. Future progress requires expanding culturally grounded datasets, establishing robust evaluation frameworks, and fostering open-source collaboration and shared-task participation to accelerate the development of Persian NLP.

Toxic language detection is crucial for creating safer online environments and limiting the spread of harmful content. While toxic language detection has been under-explored in Persian, the current work compares different methods for this task, including fine-tuning, data enrichment, zero-shot and few-shot learning, and cross-lingual transfer learning. What is especially compelling is the impact of cultural context on transfer learning for this task: We show that the language of a country with cultural similarities to Persian yields better results in transfer learning. Conversely, the improvement is lower when the language comes from a culturally distinct country.

Valentina Pyatkin, Bonnie Webber, Ido Dagan, Reut Tsarfaty. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Over 15 years ago, Ward and Birner (2006) suggested that non-canonical constructions in English can serve both to mark information status and to structure the information flow of discourse.One such construction is preposing, where a phrasal constituent appears to the left of its canonical position, typically sentenceinitially.But computational work on discourse has, to date, ignored non-canonical syntax.We take account of non-canonical syntax by providing quantitative evidence relating NP/PP preposing to discourse relations.The evidence comes from an LLM mask-filling task that compares the predictions when a mask is inserted between the arguments of an implicit intersentential discourse relation -first, when the right-hand argument (Arg2) starts with a preposed constituent, and again, when that constituent is in canonical (post-verbal) position.Results show that (1) the top-ranked maskfillers in the preposed case agree more often with "gold" annotations in the Penn Discourse TreeBank (Webber et al., 2019) than they do in the latter case, and (2) preposing in Arg2 can affect the distribution of discourse-relational senses.

Abstract Annotated data is an essential ingredient in natural language processing for training and evaluating machine learning models. It is therefore very desirable for the annotations to be of high quality. Recent work, however, has shown that several popular datasets contain a surprising number of annotation errors or inconsistencies. To alleviate this issue, many methods for annotation error detection have been devised over the years. While researchers show that their approaches work well on their newly introduced datasets, they rarely compare their methods to previous work or on the same datasets. This raises strong concerns on methods’ general performance and makes it difficult to assess their strengths and weaknesses. We therefore reimplement 18 methods for detecting potential annotation errors and evaluate them on 9 English datasets for text classification as well as token and span labeling. In addition, we define a uniform evaluation setup including a new formalization of the annotation error detection task, evaluation protocol, and general best practices. To facilitate future research and reproducibility, we release our datasets and implementations in an easy-to-use and open source software package.1

Implicit discourse relation recognition is a challenging task that involves identifying the sense or senses that hold between two adjacent spans of text, in the absense of an explicit connective between them.In both PDTB-2 (Prasad et al., 2008) and PDTB-3 (Webber et al., 2019), discourse relational senses are organized into a three-level hierarchy ranging from four broad top-level senses, to more specific senses below them.Most previous work on implicitf discourse relation recognition have used the sense hierarchy simply to indicate what sense labels were available.Here we do more -incorporating the sense hierarchy into the recognition process itself and using it to select the negative examples used in contrastive learning.With no additional effort, the approach achieves stateof-the-art performance on the task.Our code is released in https://github.com/wanqiulong0923/Contrastive_IDRR.

Lexical cohesion is a fundamental mechanism for text which requires a pair of words to be interpreted as a certain type of lexical relation (e.g., similarity) to understand a coherent context; we refer to such relations as the contextual lexical relation. However, work on lexical cohesion has not modeled context comprehensively in considering lexical relations due to the lack of linguistic resources. In this paper, we take initial steps to address contextual lexical relations by focusing on the contrast relation, as it is a well-known relation though it is more subtle and relatively less resourced. We present a corpus named Cont 2 Lex to make Contextual Lexical Contrast Recognition a computationally feasible task. We benchmark this task with widely-adopted semantic representations; we discover that contextual embeddings (e.g. BERT) generally outperform static embeddings (e.g. Glove), but barely go beyond 70% in accuracy performance. In addition, we find that all embeddings perform better when CLC occurs within the same sentence, suggesting possible limitations of current computational coherence models. Another intriguing discovery is the improvement of BERT in CLC is largely attributed to its modeling of CLC word pairs co-occurring with other word repetitions. Such observations imply that the progress made in lexical coherence modeling remains relatively primitive even for semantic representations such as BERT that have been empowering numerous standard NLP tasks to approach human benchmarks. Through presenting our corpus and benchmark, we attempt to seed initial discussions and endeavors in advancing semantic representations from modeling syntactic and semantic levels to coherence and discourse levels.

Many NLG tasks such as summarization, dialogue response, or open domain question answering focus primarily on a source text in order to generate a target response.This standard approach falls short, however, when a user's intent or context of work is not easily recoverable based solely on that source texta scenario that we argue is more of the rule than the exception.In this work, we argue that NLG systems in general should place a much higher level of emphasis on making use of additional context, and suggest that relevance (as used in Information Retrieval) be thought of as a crucial tool for designing user-oriented text-generating tasks.We further discuss possible harms and hazards around such personalization, and argue that value-sensitive design represents a crucial path forward through these challenges.

The availability of corpora has led to significant advances in training semantic parsers in English.Unfortunately, for languages other than English, annotated data is limited and so is the performance of the developed parsers.Recently, pretrained multilingual models have been proven useful for zero-shot cross-lingual transfer in many NLP tasks.What else does it require to apply a parser trained in English to other languages for zero-shot cross-lingual semantic parsing?Will simple language-independent features help?To this end, we experiment with six Discourse Representation Structure (DRS) semantic parsers in English, and generalize them to Italian, German and Dutch, where there are only a small number of manually annotated parses available.Extensive experiments show that despite its simplicity, adding Universal Dependency (UD) relations and Universal POS tags (UPOS) as model-agnostic features achieves surprisingly strong improvement on all parsers.We have publicly released our code at https://github.com/GT-SALT/

Abstract Because the 2020 ACL Lifetime Achievement Award presentation could not be done in person, we replaced the usual LTA talk with an interview between Professor Kathy McKeown (Columbia University) and the recipient, Bonnie Webber. The following is an edited version of the interview, with added citations.

As different genres are known to differ in their communicative properties and as previously, for Chinese, discourse relations have only been annotated over news text, we have created the TED-CDB dataset.TED-CDB comprises a large set of TED talks in Chinese that have been manually annotated according to the goals and principles of Penn Discourse Treebank, but adapted to features that are not present in English.It serves as a unique Chinese corpus of spoken discourse.Benchmark experiments show that TED-CDB poses a challenge for state-of-the-art discourse relation classifiers, whose F1 performance on 4way classification is <60%.This is a dramatic drop of 35% from performance on the news text in the Chinese Discourse Treebank.Transfer learning experiments have been carried out with the TED-CDB for both same-language cross-domain transfer and same-domain crosslanguage transfer.Both demonstrate that the TED-CDB can improve the performance of systems being developed for languages other than Chinese and would be helpful for insufficient or unbalanced data in other corpora.The dataset and our Chinese annotation guidelines has been made freely available.1

Text corpora annotated with language-related properties are an important resource for the development of Language Technology. The current work contributes a new resource for Chinese Language Technology and for Chinese-English translation, in the form of a set of TED talks (some originally given in English, some in Chinese) that have been annotated with discourse relations in the style of the Penn Discourse TreeBank, adapted to properties of Chinese text that are not present in English. The resource is currently unique in annotating discourse-level properties of planned spoken monologues rather than of written text. An inter-annotator agreement study demonstrates that the annotation scheme is able to achieve highly reliable results.

Jun Quan, Deyi Xiong, Bonnie Webber, Changjian Hu. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019.

Abstract This paper serves as a short overview of the JNLE special issue on representation of the meaning of the sentence, bringing together traditional symbolic and modern continuous approaches. We indicate notable aspects of sentence meaning and their compatibility with the two streams of research and then summarize the papers selected for this special issue.

Theories of discourse coherence posit relations between discourse segments as a key feature of coherent text.Our prior work suggests that multiple discourse relations can be simultaneously operative between two segments for reasons not predicted by the literature.Here we test how this joint presence can lead participants to endorse seemingly divergent conjunctions (e.g., but and so) to express the link they see between two segments.These apparent divergences are not symptomatic of participant naïveté or bias, but arise reliably from the concurrent availability of multiple relations between segments -some available through explicit signals and some via inference.We believe that these new results can both inform future progress in theoretical work on discourse coherence and lead to higher levels of performance in discourse parsing.

When a reader is first introduced to an entity, its referring expression must describe the entity.For entities that are widely known, a single word or phrase often suffices.This paper presents the first study of how expressions that refer to the same entity develop over time.We track thousands of person and organization entities over 20 years of New York Times (NYT).As entities move from hearernew (first introduction to the NYT audience) to hearer-old (common knowledge) status, we show empirically that the referring expressions along this trajectory depend on the type of the entity, and exhibit linguistic properties related to becoming common knowledge (e.g., shorter length, less use of appositives, more definiteness).These properties can also be used to build a model to predict how long it will take for an entity to reach hearer-old status.Our results reach 10-30% absolute improvement over a majority-class baseline.

It might surprise some young researchers that the first recipient of the ACL LifeTime Achievement award,1 Aravind Joshi, was so often compared to “Yoda,” one of the oldest and most powerful of the Jedi Masters in the Star Wars universe. But Aravind was also one of the kindest, wisest, and most justly celebrated people that one was fortunate to know.When Aravind received the award in 2002, at the age of 73, he said that he hoped his lifetime wasn’t over. Fortunately, it wasn’t: For the next 15 years, Aravind continued to enjoy time spent on research; advising students and younger colleagues; attending ACL conferences both at home in the United States and in far-flung places such as Sydney, Singapore, and Jeju Island; and enjoying the company of his extraordinary wife, the embryologist Susan Heyner, his daughters, Meera and Shyamala Joshi, and his grandchildren, Marco and Ava. Then on 31 December 2017, Aravind died peacefully at home in Philadelphia, sitting in his favorite chair, at the age of 88.Aravind Joshi was born in Pune, India, on 5 August, 1929. He sailed to the United States in 1954 to study electrical engineering (EE) at the University of Pennsylvania, after he was rejected by Harvard because his application, mailed from India, arrived a day late. While completing his M.Sc. in EE, he worked as an engineer at RCA (Camden, NJ), and then while completing his Ph.D. in EE, as a research assistant at the University of Pennsylvania’s Department of Linguistics. After being awarded his doctorate, Aravind joined the Penn faculty, remaining in EE until the brilliant and prescient Saul Gorn, who chaired Penn’s Graduate Group in Computer and Information Science, convinced the University to establish a new academic department of Computer and Information Science (CIS). Aravind joined this new department as a full professor and Chair, grateful that Saul Gorn had argued so forcefully that the new department should embrace the science of information as well as the practical study of computers and computing. Allowing for such broad intellectual content allowed the evolving CIS Department to constantly reach out to researchers in other disciplines—including those at Penn’s Wharton School of Business, as well as at the Departments of Linguistics, Psychology, Philosophy and Bioinformatics.Aravind remained Chair of CIS for an incredible 13 years, until 1985—continuing throughout this time to carry out cutting-edge research, to serve leadership roles in both ACL (as President in 1975 and then as Book Series Editor from 1982) and IJCAI (as General Chair of the 1985 IJCAI Conference), while at the same time giving generously to his students and colleagues. As General Chair of IJCAI, Aravind reached out to invite Soviet refusenik computer scientists, linguists, and mathematicians to attend. Several of the invitees had been arrested and were serving long sentences in labor camps. Although Aravind knew that they would never be able to attend, the invitation reassured them that their academic accomplishments would be recognized internationally, despite the political environment.One of Aravind’s major achievements during his time as Chair of CIS was co-founding, with psycholinguist Lila Gleitman, Penn’s famous Cognitive Science Program. Funded initially by the Sloan Foundation, the Program received further funding from the National Science Foundation in 1991, to become Penn’s world-famous Institute for Research in Cognitive Science (IRCS).2 Aravind and Lila co-directed IRCS until 2001, contributing a stream of over 100 postdocs from linguistics, psychology, computer science, philosophy, neuroscience, and mathematics that it hosted and nurtured before closing its doors in 2016.Meanwhile, Aravind’s five decades at Penn saw his research and inventions span much of what we know as computational natural language processing, including• parsing using finite state transducers, with Aravind’s early FST parser reimplemented as “A Parser from Antiquity” (Joshi and Hopely 1996). Aravind always encouraged his students and colleagues to go back and re-examine earlier work. In fact, he once suggested that students in introductory CL courses be asked to look at the literature and reconstruct some old system as a way of both shortening the period to re-discovery and giving students a better historical grounding in the field.• grammatical formalisms, most notably the development and detailed characterization of the “mildly context-sensitive” Tree Adjoining Grammar (TAG) in both its original and lexicalized forms (LTAG), providing enough power to handle the range of phenomena in human language syntax while remaining computationally tractable (Joshi and Schabes 1997).• cooperative Question Answering and the range of inference it requires (Joshi, Webber, and Weischedel 1984, 1986)• prominence in discourse (Grosz, Joshi, and Weinstein 1995; Walker, Joshi, and Prince 1998), in the form of work on “Centering,” which was meant to account for ease of inference and the use of anaphoric expressions, linked by the observation that an entity that can be accessed with an expression as small as a pronoun must also be prominent.• discourse and syntax (Webber et al. 1999, 2003), where Aravind reconceptualized discourse connectives in the framework of LTAG, culminating in development of the NSF-funded Penn Discourse TreeBank3 and similarly annotated corpora in Chinese (Zhou and Xue 2012, 2015), Hindi (Oza et al. 2009), Turkish (Zeyrek et al. 2010), and biomedicine (Prasad et al. 2011).Here it is worth saying a bit more about two aspects of Aravind’s work: His work on grammar formalisms and his work on discourse. In the early 1980s, Aravind identified a set of computational properties that provided an informal definition of a class of Mildly Context Sensitive (MCS) languages that he claimed properly included all human languages and was properly included among the much vaster class of context-sensitive languages. These properties were: (a) polynomial parsability; (b) the constant growth property (which excludes languages with unbounded gaps in the length of sentences, of which an artificial example is the indexed language a2n, made up of strings of 2n a’s); and (c) a limit on crossing dependency of the kind seen in the artificial tree-adjoining language (TAL) anbncn, and hence on permutation-completeness (Joshi and Levy 1982; Joshi 1988; Joshi, Vijay-Shanker, and Weir 1991). TAG was the first fully formalized theory of grammar to be proved to characterize only languages within the MCS class, and provided the basis for further proofs of MCS expressive power for several other constrained grammar formalisms that were developed around the same time via their (weak) equivalence to TAG.One natural generalization of TAG was to the Linear Context-free Rewriting Systems (LCFRS) or Multiple Context Free Grammars (MCFG). These are considerably more expressive than TAG, and were for a while conjectured to provide a formal definition of MCS languages. As a result, other formalisms that were considerably more expressive than TAG have laid claim to the Joshian mantle of Mild Context Sensitivity, showing that it has become a highly influential meme in the field.However, it has since been shown that the artificial permutation-complete language MIX3, consisting of all permutations over the strings of the TAL anbncn, is a Multiple Context Free Language (MCFL). Thus, the formal characterization of the MCS class is currently a matter for debate. Nevertheless, TAG itself remains among the least more expressive formalisms than CFG that is known. Now the more important question is whether TAG or one of the other weakly equivalent formalisms is expressive enough to capture the full range of phenomena actually exhibited by natural languages, as argued in Frank (2004).Aravind’s interest in discourse semantics has been long-standing, going back at least to the mid-1970s and the publication by Academic Press of Subject and Topic (Li 1976). The collection contained two articles that Aravind annotated extensively—Li and Thompson’s article on topic-prominent languages and Lehmann’s article on the history of topic-prominence in Indo-European. Aravind’s interest in the notion of topic-prominence and the discourse semantics of information structure seems to have been stimulated by his knowledge of Sanskrit and of his native language, Marathi, both of which exhibit aspects of topic-prominence and ergativity. Together with the article in the same collection by Keenan and Schieffelin, this concern with prominence in discourse seems to have fuelled his later work on Centering and prominence in discourse.Later on, when Aravind started to look at discourse connectives (Webber et al. 1999, 2003), his concerns went deeper than the obvious parallels between lexically anchored trees for sentence-level syntactic analysis and trees lexically-anchored on discourse connectives that could be used in going beyond sentences to small units of discourse. Rather, his concerns were grounded in his growing belief that too many constructions found between the start of a sentence and its final punctuation didn’t really belong to syntax. In particular, Aravind described parentheticals, epithets, extraposed predicates, and sentential relatives as “constructions that require a skilled tree surgeon to force a single tree over a sentence.” Whereas syntactic analysis may simply punt by attaching a parenthetical such as “John thinks,” in “Mary, John thinks, will win the race,” to the root node of its parse tree (just as “Mary” is attached to the root node as its subject), for Aravind, the two attachments (of the subject and of the parenthetical) were completely different, with the attachment of the parenthetical belonging to an orthogonal dimension, in a “paratactic” relation, another characteristic of many constructions in Sanskrit and Marathi. He felt the same about epithets like “damn” in “I finished the damn book”: “damn” should be attached along an orthogonal dimension because it bore a different semantic relation to “book” than say “thick book” or “book about insects.” For Aravind, discourse provided this orthogonal dimension, reducing the number of “no-win” decisions that followed from insisting on a single parse tree over a sentence.Having spent so much time with sentences from the Penn TreeBank (constructed from articles from the ACL/DCI Wall Street Journal corpus), Aravind found the examples that made the most convincing demand for distributing the burden between syntax and discourse to be attribution phrases (Dinesh et al. 2005). He felt that sometimes an attribution phrase like “the company says” belongs to sentential syntax, feeding its semantics up to that of the sentence as a whole (as in the contrast expressed in “Observers say negotiations have halted, while the company says it is talking with several prospects”). In other cases, however, such as the concession expressed in “There have been no orders for the Cray-3 so far, though the company says it is talking with several prospects,” he felt that the attribution phrase belongs to an orthogonal discourse dimension, because what is contrary to expectations associated with the lack of orders for the Cray-3 is not the company (Cray) saying something, but rather the existence of several prospects that it is talking with. Both analyses become possible if syntax and discourse can provide distinct bases for analysis. Hence Aravind’s interest in both.Although Aravind’s ingenuity was all his own, the inventions he was involved with were possible only because of his unprecedented inclusion of linguists, psychologists, philosophers, and mathematicians, as well as computer scientists and engineers, in his work. In recognition of Aravind’s inclusionary spirit and many achievements, Penn’s CIS Department hosted JoshiFest in Fall 2012, an all-day symposium of talks and encomia in his honor. The complete program of talks and presentations is available for viewing at https://www.cis.upenn.edu/about-cis/events/joshi-fest/program.php.JoshiFest was neither the first nor the last time that Aravind’s achievements were publicly recognized. Besides the ACL Lifetime Achievement Award in 2002, Aravind was the recipient of the 1997 IJCAI Award for Research Excellence; the 2003 David E. Rumelhart Prize of the Cognitive Science Society; the 2005 Benjamin Franklin Medal in Computer and Cognitive Science (awarded by the Franklin Institute in Philadelphia); and the Henry Salvatori Chair in Cognitive and Computer Science. He was elected to the National Academy of Engineering in 1999 and named a Fellow of IEEE in 1976 and the Association for Computing Machinery (ACM) in 1998. In 1990, he became a Founding Fellow of the Association for the Advancement of Artificial Intelligence (AAAI), and in 2011, a Founding Fellow of the ACL. Most recently, the Charles University (Prague) recognized Aravind’s accomplishments—including his joint work with Professor Eva Hajicova’s group at Charles University—with the award of Doctor Honoris Causa in physics and mathematics.In his obituary for Aravind Joshi in Language Log,4 Mark Liberman quoted posts from Bob Frank (whose Ph.D. thesis Aravind supervised, and who is now Chair of Linguistics at Yale) and Julia Hockenmaier (a postdoc at IRCS following her Ph.D. from the University of Edinburgh, now an Associate Professor at the University of Illinois (Champaign-Urbana). Because both posts express their author’s thoughts and feelings so well, they seem an appropriate way to close this obituary.I just heard the crushing news that Aravind Joshi passed away yesterday. It’s hard for me to overstate how profoundly Aravind influenced my career and my life, since he took me on as his PhD student 30 years ago. The content of his work laid the foundations for so much of what I have worked on over the years, and his vision of interdisciplinary interaction shaped how I see the field. I will never forget his insatiable curiosity and intellectual energy, his remarkable ability to identify good problems and insightful solutions, and his gentle kindness and humanity. And I will so much miss the boyish excitement he exuded whenever he would share his latest ideas with me. Thank you for everything, Aravind. You will be missed. [Bob Frank]I can’t begin to describe how much I owe to Aravind’s advice and mentorship, his intellect, his curiosity, his kindness, and his great sense of humor. It was such a privilege to work so closely with him, even as one of his last postdocs. His impact on our field and our community can simply not be overstated. We’ve lost one of our founding fathers. Not just because he was one of the few, or probably even the only one, still around from the very early days of NLP. We’ve also lost someone who has really shaped the intellectual and social culture of our community in fundamental ways. If you are among those that feel at home in our field because of its intellectual richness and diversity, and also because you never felt out of place because you are a woman, you should know how much you owe to Aravind and the legacy of his very many distinguished students, and the culture he and his colleagues created at Penn and in the community as a whole. Rest in peace, Aravind. In sorrow, and gratitude. [Julia Hockenmaier]In writing this article, I drew in part from text by Meera and Shyamala Joshi, John Nerbonne, Mark Liberman and Mark Steedman, all of whom have written eloquently about Aravind’s life and accomplishments. All errors, however, are my own.

Automatic negation scope detection is a task that has been tackled using different classifiers and heuristics.Most systems are however 1) highly-engineered, 2) English-specific, and 3) only tested on the same genre they were trained on.We start by addressing 1) and 2) using a neural network architecture.Results obtained on data from the *SEM2012 shared task on negation scope detection show that even a simple feed-forward neural network using word-embedding features alone, performs on par with earlier classifiers, with a bi-directional LSTM outperforming all of them.We then address 3) by means of a specially-designed synthetic test set; in doing so, we explore the problem of detecting the negation scope more in depth and show that performance suffers from genre effects and differs with the type of negation considered.

Although the performance of SMT systems has improved over a range of different linguistic phenomena, negation has not yet received adequate treatment.Previous works have considered the problem of translating negative data as one of data sparsity (Wetzel and Bond (2012)) or of structural differences between source and target language with respect to the placement of negation (Collins et al. (2005)).This work starts instead from the questions of what is meant by negation and what makes a good translation of negation.These questions have led us to explore the use of semantics of negation in SMTspecifically, identifying core semantic elements of negation (cue, event and scope) in a source-side dependency parse and reranking hypotheses on the n-best list produced after decoding according to the extent to which an hypothesis realises these elements.The method shows considerable improvement over the baseline as measured by BLEU scores and Stanford's entailmentbased MT evaluation metric (Padó et al. (2009)).

We present a French to English transla-tion system for Wikipedia biography ar-ticles. We use training data from out-of-domain corpora and adapt the system for biographies. We propose two forms of domain adaptation. The first biases the system towards words likely in biogra-phies and encourages repetition of words across the document. Since biographies in Wikipedia follow a regular structure, our second model exploits this structure as a sequence of topic segments, where each segment discusses a narrower subtopic of the biography domain. In this structured model, the system is encouraged to use words likely in the current segment’s topic rather than in biographies as a whole. We implement both systems using cache-based translation techniques. We show that a system trained on Europarl and news can be adapted for biographies with 0.5 BLEU score improvement using our mod-els. Further the structure-aware model out-performs the system which treats the entire document as a single segment. 1

The Penn Discourse Treebank (PDTB) was released to the public in 2008. It remains the largest manually annotated corpus of discourse relations to date. Its focus on discourse relations that are either lexically-grounded in explicit discourse connectives or associated with sentential adjacency has not only facilitated its use in language technology and psycholinguistics but also has spawned the annotation of comparable corpora in other languages and genres. Given this situation, this paper has four aims: (1) to provide a comprehensive introduction to the PDTB for those who are unfamiliar with it; (2) to correct some wrong (or perhaps inadvertent) assumptions about the PDTB and its annotation that may have weakened previous results or the performance of decision procedures induced from the data; (3) to explain variations seen in the annotation of comparable resources in other languages and genres, which should allow developers of future comparable resources to recognize whether the variations are relevant to them; and (4) to enumerate and explain relationships between PDTB annotation and complementary annotation of other linguistic phenomena. The paper draws on work done by ourselves and others since the corpus was released.

Discourse is coming in from the cold. After years of being ignored by researchers in other areas of computational linguistics and language technology, many of these same researchers are beginning to think that their own work could benefit from treating text as more than just a bag of sentences. That is, they are beginning to think that discourse offers some low-hanging fruit—achievable improvements in system performance that exploit either aspects of text structure or the context that text establishes and uses for efficient referring and/or predicational expressions.This new monograph on Discourse Processing by Manfred Stede both reflects this new zeitgeist and provides an introduction to discourse for researchers in computational linguistics or language technology with little or no background in the area. This clear and timely monograph consists of a brief introduction to discourse, a meaty chapter on each of the three aspects of discourse processing that hold most promise for language technology, and a brief conclusion on where discourse research might go in the future. I will go through the three major chapters, and then make some general remarks.Chapter 2Chapter 2 addresses two distinct types of large-scale discourse structure: structure that follows from a text belonging to a particular genre, and structure that follows from the topic (or topic mix) of a text. The genre of a text affects features such as style and register. What is relevant here is structure that genre may confer on a text. Stede suggests that some, but not all, texts inherit large-scale structure from their genre, calling some unstructured, some structured, and some semi-structured. As a reader, I did not find this distinction useful, because all text that belongs to a genre seems to get some large-scale structure from it. On the other hand, all or part of this structure might simply not be manifest in the kind of lexico-syntactic features that automated systems regularly rely on for text segmentation. As a case in point, although Stede offers the text Suffering (used as a running example throughout the book) as an example of unstructured text, like other instances of Comments in the Talk of the Town section of the New Yorker magazine, its large-scale structure comprises a “hook” aimed at getting the reader's attention, followed by a short essay that concludes with a serious point. Although ways of attracting a reader's attention may not have specific lexico-syntactic features, it might still be possible to recognize the transition between “hook” and essay, and essay structure itself is what ETS's eRater system (Burstein and Chodorow 2010) aims to recognize and evaluate.This first half of Chapter 2 focuses on the genre-based structure of scientific texts and of film reviews. Here researchers have already shown that language technologies such as information extraction and sentiment analysis benefit from taking such structure into account, so this is entirely appropriate for the book's target audience. More on genre-based functional structure and its use in producing structured biomedical abstracts can be found in the recent survey of research on discourse structure and language technology by Webber, Egg, and Kordoni 2012.The second half of Chapter 2 discusses large-scale discourse structure associated with patterns of topics. Such structure is often found in expository writing such as encyclopedia articles and travel pieces. Here, changing patterns of content words correlate well with changes in topic, rendering them useful for the many approaches to text segmentation that are well-described in this half of the chapter. Because the discussion here of probabilistic models for topic segmentation is rather short, the reader who wants to know more should consult the excellent survey of topic segmentation methods by Purver (2011).Chapter 3Chapter 3, entitled Coreference Resolution, addresses more than this, dealing with the resolution of other expressions whose reduction is licensed by the discourse context, such as bridging reference and “other” reference, which Halliday and Hasan 1976 call comparative reference because it occurs with comparative forms such as “larger fish” and “a more impressive poodle,” as well as with “other,” “another,” and “such.” Stede justifies inclusion of this chapter for two reasons—the close connection between coreference resolution and topic segmentation and the benefits to text analysis provided by having its pronouns resolved. But another reason must be the link mentioned earlier between text and context: Discourse creates the context in which context-reduced expressions make sense, so it falls naturally within the tasks of discourse processing to resolve them, either through modeling context explicitly or through the use of proxies.The chapter starts with an overview of coreference and anaphora that covers both their forms and their functions. This is followed by an important section on corpus annotation (Section 3.2), included because (as Stede notes) what has been annotated and why it has been annotated strongly determines what expressions are resolved and how. This section identifies many of the problems in coreference annotation that have been raised in the literature, but recognizes that research has to make use of the resources that exist and not just the resources it wants. Several of these are indicated at the end of the section, reminding one that it would have been useful to have some pointers in Chapter 2 to corpora available for genre-based segmentation (such as Liakata's ART corpus)1 or for topic-based segmentation.Stede then links the current chapter to the previous one through a discussion of entity-based coherence (Section 3.3) and then discusses how to identify when a pronoun or definite noun phrase should be treated as anaphoric (Section 3.4) as groundwork for discussion of anaphora resolution (Sections 3.5–3.7). Missing from the discussion of detecting non-anaphoric (pleonastic) pronouns is mention of Bergsma's recent system NADA for doing this (Bergsma and Yarowsky 2011).2The discussion of anaphora resolution covers rule-based approaches to resolving nominal anaphora (Section 3.5) and then supervised machine learning methods for anaphora resolution (Section 3.6). The latter follows the structure (albeit not the content) of Ng's survey 2010, in discussing mention-pair models, and then entity-mention models. Whereas Ng then discusses ranking models, including his cluster ranker (Rahman and Ng 2009), which is conceptually similar to the Lappin and Leass 1994 approach described in Section 3.5, Stede discusses a range of more recent models, most of which are subsequent to Ng's survey.Section 3.8 surveys methods evaluating coreference resolution and some of the known problems in doing so. A good complement to this is Byron's too-little-known discussion of problems in the consistent reporting of such results (Byron 2001). Chapter 3 concludes with a section on Recent Trends, which would also have been useful in Chapter 2.Chapter 4The fourth and longest chapter deals with semantic or pragmatically oriented coherence relations that hold between adjacent text spans or discourse units. Whereas the previous two chapters were essentially theory-neutral, the presentation in Chapter 4 largely reflects the perspective of Rhetorical Structure Theory (Mann and Thompson 1988). RST takes a text to be a sequence of elementary discourse units that comprise the leaves of a tree structure of coherence relations between recursively defined discourse units. RST also assumes that one of the arguments to a coherence relation may be more important to the speaker's purpose than the other, calling the former the nucleus and the latter, the satellite.This RST framework dictates the structure of the chapter: Following an introductory section that explains and motivates coherence relations, each subsequent section considers the next task in an RST analysis—segmenting a text into elementary discourse units (Section 4.2), recognizing which (adjacent) units stand in a coherence relation and what (single) relation holds between them (Section 4.3), and finally, inducing the overall tree structure of coherence relations that hold between recursively defined discourse units (Section 4.4). All these tasks are well described, both from a theoretical perspective and in terms of automated procedures for carrying them out. Coverage of relevant work is very high.Where the reader may get confused, however, is that a good proportion of the more recent work on identifying coherence relations does not fall within the framework of RST, and thus doesn't adhere to several of its assumptions—in particular, that a text is divisible into a covering sequence of elementary discourse units, that only one relation can hold between discourse units, that the arguments to a coherence relation must be adjacent, that one argument to a coherence relation may intrinsically convey information that is more important to the speaker's purpose than the other, and that coherence relations impose an overall tree structure on a text in terms of recursively defined discourse units.Although Chapter 4 discusses the Penn Discourse TreeBank (Prasad et al. 2008) and its “somewhat modest annotations” (page 126), the discussion is framed in terms of RST tasks, whereas the assumptions underlying the Penn Discourse TreeBank reflect its concerns with a quite different set of tasks involved in recognizing coherence relations. The first task requires finding evidence for a coherence relation (in the form of a discourse connective such as a coordinating or subordinating conjunction or a discourse adverbial, or in the form of sentence adjacency) and then determining (1) if the evidence does indeed signal a coherence relation, given that evidence is often ambiguous; (2) if it does, what constitutes its arguments; and (3) what is its sense. Although Chapter 4 covers some of this work (Dinesh et al. 2005; Wellner and Pustejovsky 2007; Elwell and Baldridge 2008; Pitler and Nenkova 2009; Prasad, Joshi, and Webber 2010), its appearance within the context of a discussion of RST-tasks may lead to some confusion.Chapter 4 concludes with a brief discussion of some important open issues regarding coherence relations, including problems with associating a large text span with a single recursive structure of coherence relations and problems with inter-annotator agreement.SummaryFor its intended audience, this monograph will serve as a compact, readable introduction to the subject of discourse processing. The relevant phenomena are presented clearly, as are many of the computational methods for dealing with them. What readers won't get is criteria for choosing among the methods or an understanding of what each method is good for. This problem may reflect the absence of comparable performance results and useful error analyses in the original publications, however.Also missing from the monograph is discussion of applications of discourse processing, and pointers to more of the resources available to researchers interested in discourse structure. This is where the additional resources I have mentioned may prove complementary.Finally, a plea to the series editor: Monographs such as this one really need an index. Some monographs in the series have one, whereas others (like this one) don't. Because the series appears in both electronic and physical format, one could excuse the former not having an explicit index, since in most cases, one can get away with the basic search facility in the Adobe Reader. Nothing similar is available for the nicely sized physical monographs. Their authors should be strongly encouraged to provide them.

Abstract An increasing number of researchers and practitioners in Natural Language Engineering face the prospect of having to work with entire texts, rather than individual sentences. While it is clear that text must have useful structure, its nature may be less clear, making it more difficult to exploit in applications. This survey of work on discourse structure thus provides a primer on the bases of which discourse is structured along with some of their formal properties. It then lays out the current state-of-the-art with respect to algorithms for recognizing these different structures, and how these algorithms are currently being used in Language Technology applications. After identifying resources that should prove useful in improving algorithm performance across a range of languages, we conclude by speculating on future discourse structure-enabled technology.

Every text has at least one topic and at least one genre. Evidence for a text's topic and genre comes, in part, from its lexical and syntactic features—features used in both Automatic Topic Classification and Automatic Genre Classification (AGC). Because an ideal AGC system should be stable in the face of changes in topic distribution, we assess five previously published AGC methods with respect to both performance on the same topic–genre distribution on which they were trained and stability of that performance across changes in topic–genre distribution. Our experiments lead us to conclude that (1) stability in the face of changing topical distributions should be added to the evaluation critera for new approaches to AGC, and (2) Part-of-Speech features should be considered individually when developing a high-performing, stable AGC system for a particular, possibly changing corpus.

Abstract In this introduction, we present our overview of interactive question answering (IQA). We contextualize IQA in the wider field of question answering, and establish connections to research in Information Retrieval and Dialogue Systems. We highlight the development of QA as a field, and identify challenges in the present research paradigm for which IQA is a potential solution. Finally, we present an overview of papers in this special issue, drawing connections between these and the challenges they address.

A recent posting to the BioNLP mailing list notes that the past few months of 2008 have seen the appearance of over fifty papers on biomedical natural language processing/text mining (BioNLP). This number (which included medical, as well as genomic work) represents about as many papers on genomic language processing as existed in all of PubMed at the end of 2003 [1] – just five years ago, and the current supplement in BMC Bioinformatics presents another ten! These papers have in common the fact that they are follow-on work to papers originally published in the proceedings of the BioNLP 2008 workshop at the annual meeting of the Association for Computational Linguistics (ACL). All have gone through a separate rigorous review process and represent an advance beyond the work originally presented at the workshop. Like the annual BioNLP workshop itself, they represent a wide cross-section of the type of work that goes on in BioNLP today.

No abstract available.

Standard practice in our field has been to announce research results at our annual conference or one of

COBrA is a Java-based ontology editor for bio-ontologies that distinguishes itself from other editors by supporting the linking of concepts between two ontologies, and providing sophisticated analysis and verification functions. In addition to the Gene Ontology and Open Biology Ontologies formats, COBrA can import and export ontologies in the Semantic Web formats RDF, RDFS and OWL.

The process of microplanning in natural language generation (NLG) encompasses a range of problems in which a generator must bridge underlying domain‐specific representations and general linguistic representations. These problems include constructing linguistic referring expressions to identify domain objects, selecting lexical items to express domain concepts, and using complex linguistic constructions to concisely convey related domain facts. In this paper, we argue that such problems are best solved through a uniform, comprehensive, declarative process. In our approach, the generator directly explores a search space for utterances described by a linguistic grammar. At each stage of search, the generator uses a model of interpretation, which characterizes the potential links between the utterance and the domain and context, to assess its progress in conveying domain‐specific representations. We further address the challenges for implementation and knowledge representation in this approach. We show how to implement this approach effectively by using the lexicalized tree‐adjoining grammar (LTAG) formalism to connect structure to meaning and using modal logic programming to connect meaning to context. We articulate a detailed methodology for designing grammatical and conceptual resources which the generator can use to achieve desired microplanning behavior in a specified domain. In describing our approach to microplanning, we emphasize that we are in fact realizing a deliberative process of goal‐directed activity. As we formulate it, interpretation offers a declarative representation of a generator's communicative intent. It associates the concrete linguistic structure planned by the generator with inferences that show how the meaning of that structure communicates needed information about some application domain in the current discourse context. Thus, interpretations areplansthat the microplanner constructs and outputs. At the same time, communicative intent representations provide arich and uniform resourcefor theprocessof NLG. Using representations of communicative intent, a generator can augment the syntax, semantics, and pragmatics of an incomplete sentence simultaneously, and can work incrementally toward solutions for the various problems of microplanning.

We argue in this article that many common adverbial phrases generally taken to signal a discourse relation between syntactically connected units within discourse structure instead work anaphorically to contribute relational meaning, with only indirect dependence on discourse structure. This allows a simpler discourse structure to provide scaffolding for compositional semantics and reveals multiple ways in which the relational meaning conveyed by adverbial connectives can interact with that associated with discourse structure. We conclude by sketching out a lexicalized grammar for discourse that facilitates discourse interpretation as a product of compositional rules, anaphor resolution, and inference.

OBJECTIVE: To ascertain whether three-dimensional geometric and probabilistic reasoning methods can be successfully combined for computer-based assessment of conditions arising from ballistic penetrating trauma to the chest and abdomen. DESIGN: The authors created a computer system (TraumaSCAN) that integrates three-dimensional geometric reasoning about anatomic likelihood of injury with probabilistic reasoning about injury consequences using Bayesian networks. Preliminary evaluation of TraumaSCAN was performed via a retrospective study testing performance of the system on data from 26 cases of actual gunshot wounds. MEASUREMENTS: Areas under the receiver operating characteristics (ROC) curve were calculated for each condition modeled in TraumaSCAN that was present in the 26 cases. The comprehensiveness and relevance of the TraumaSCAN diagnosis for the 26 cases were used to assess the overall performance of the system. To test the ability of TraumaSCAN to handle limited findings, these measurements were calculated both with and without input of observed findings into the Bayesian network. RESULTS: For the 11 conditions assessed, the worst area under the ROC curve with no observed findings input into the Bayesian network was 0.542 (95% CI, 0.146-0.937), the median was 0.883 (95% CI, 0.713-1.000), and the best was 1.00 (95% CI, 1.000-1.000). The worst area under the ROC curve with all observed findings input into the Bayesian network was 0.835 (95% CI, 0.602-1.000), the median was 0.941 (95% CI, 0.827-1.000), and the best was 0.992 (95% CI, 0.965-1.000). A comparison of the areas under the curve obtained with and without input of observed findings into the Bayesian network showed that there were significant differences for 2 of the 11 conditions assessed. CONCLUSION: A computer-based method that combines geometric and probabilistic reasoning shows promise as a tool for assessing ballistic penetrating trauma to the chest and abdomen.

Describes a method for diagnostic reasoning under uncertainty that is used in TraumaSCAN, a computer-based system for assessing penetrating trauma. Uncertainty in assessing penetrating injuries arises from two different sources: the actual extent of damage associated with a particular injury mechanism may not be easily discernable, and there may be incomplete information about patient findings (signs, symptoms and test results) which provide clues about the extent of the injury. Bayesian networks are used in TraumaSCAN for diagnostic reasoning because they provide a mathematically sound means of making probabilistic inferences about the injury in the face of uncertainty. We also present a comparison of TraumaSCAN's results in assessing 26 actual gunshot wound cases with those of TraumAID, a validated rule-based expert system for the diagnosis and treatment of penetrating trauma.

We show that discourse structure need not bear the full burden of conveying discourse relations by showing that many of them can be explained nonstructurally in terms of the grounding of anaphoric presuppositions (Van der Sandt, 1992). This simplifies discourse structure, while still allowing the realisation of a full range of discourse relations. This is achieved using the same semantic machinery used in deriving clause-level semantics.

Examines an approach for integrating 3D structural reasoning, using computer models of the human anatomy, with diagnostic reasoning based on Bayesian networks in order to probabilistically predict injuries to anatomic structures from gunshot wounds. An interactive 3D graphical system has been created which allows the user to visualize different bullet path hypotheses and computes the probability that an anatomical structure associated with a given penetration path is injured. The probabilities derived are essential for mediating between structural reasoning and diagnostic reasoning.

The way in which discourse features express connections back to the previous discourse has been described in the literature in terms of adjoining at the right frontier of discourse structure. But this does not allow for discourse features that express expectations about what is to come in the subsequent discourse. After characterizing these expectations and their distribution in text, we show how an approach that makes use of substitution as well as adjoining on a suitably defined right frontier, can be used to both process expectations and constrain discouse processing in general.

The TraumAID system has been designed to provide on-line decision support throughout the initial definitive management of injured patients. Here we describe its retrospective evaluation and the use we subsequently made of judges comments on the validation data to evaluate TraumaTIQ, a new critiquing interface for TraumAID, investigating the question of whether, with timely recording of information, a system could produce commentary in line with that of human experts. Our results show that (1) comparable commentary can be produced, and (2) validation studies, which take great time and effort to conduct, can produce useful data beyond their original design goals.

Good communication is vital in health care, both among health care professionals, and between health care professionals and their patients. And well-written documents, describing and/or explaining the information in structured databases may be easier to comprehend, more edifying, and even more convincing than the structured data, even when presented in tabular or graphic form. Documents may be automatically generated from structured data, using techniques from the field of natural language generation. These techniques are concerned with how the content, organization and language used in a document can be dynamically selected, depending on the audience and context. They have been used to generate health education materials, explanations and critiques in decision support systems, and medical reports and progress notes.

Based on an ongoing attempt to integrate Natural Language instructions with human figure animation, we demonstrate that agents' understanding and use of instructions can complement what they can derive from the environment in which they act. We focus on two attitudes that contribute to agents' behavior—their intentions and their expectations—and shown how Natural Language instructions contribute to such attitudes in ways that complement the environment. We also show that instructions can require more than one context of interpretation and thus that agents' understanding of instructions can evolve as their activity progresses. A significant consequence is that Natural Language understanding in the context of behavior cannot simply be treated as “front end” processing, but rather must be integrated more deeply into the processes that guide an agent's behavior and respond to its perceptions.

TraumAID is a consultation system for the diagnosis and treatment of multiple trauma. It integrates diagnostic reasoning, planning, and action. Its reasoner identifies diagnostic and therapeutic goals appropriate to the physician's knowledge of the patient's state, while its planner advises on beneficial actions to next perform. The physician's lack of complete knowledge of the situation and the time limitations of emergency medicine constrain the ability of any planner to identify what would be the best thing to do. TraumAID's Progressive Horizon Planner has been designed to create a plan for patient care that is in keeping with the standards of managing trauma.>

Two independent mechanisms of context change have been discussed separately in the literaturecontext change by entity introduction and context change by event simulation.Here we discuss their integration.The effectiveness of the integration depends in part on a representation of events that captures people's uncertainty about their outcome -in particular, people's incomplete expectations about the changes effected by events.We propose such a representation and a process of accommodation that makes use of it, and discuss our initial implementation of these ideas.

Computational approaches to discourse understanding have a two-part goal: (1) to identify those aspects of discourse understanding that require process-based accounts, and (2) to characterize the processes and data structures they involve. To date, in the area of reference, process-based accounts have been developed for subsequent reference via anaphoric pronouns and reference via definite descriptors. In this paper, I propose and argue for a process-based account of subsequent reference via deictic expressions. A significant feature of this account is that it attributes distinct mental reality to units of text often called discourse segments, a reality that is distinct from that of the entities described therein.

This paper gives an account of the role tense plays in the listener's reconstruction of the events and situations a speaker has chosen to describe. Several new ideas are presented: (a) that tense is better viewed by analogy with definite NPs than with pronouns; (b) that a narrative has a temporal focus that grounds the context-dependency of tense; and (c) that focus management heuristics can be used to track the movement of temporal focus.

The aim of this paper is to justify why Natural Language (NL) interaction, of a very rich functionality, is critical to the effective use of Expert Systems and to describe what is needed and what has been done to support such interaction. Interactive functions discussed here include defining terms, paraphrasing, correcting misconceptions, avoiding misconceptions, and modifying questions.

this paper, we investigate this revi.-ed principle as applied to question answering. In particular the goals of the research described here are to: 1. characterize tractable cases in which the system as respondent (R) can anticipate the possibility of the user/questioner (Q} drawing false conclusions from its response and can hence alter or expand its response so as to prevent it happening; 2. develop a formal method for computing the projected inferences that Q may draw from a particular response, identifylug those 1This work is partially supported by NSF Grants MCS 81-07290, MCS 8.3-05221, and [ST 88-1 2At present visiting the Department of Computer and Information Science, University of Pemsylvania, Phi|adelphia, PA 19104. factors whose presence or absence catalyzes the inferences; 3. enable the system to generate modifications of its response that can defuse possible false inferences and that may provide addi6oual useful information as well

Language is a system for encoding and transmitting ideas. A theory that seeks to explain linguistic phenomena in terms of this fact is a functional theory. One that does not misses the point. [10]

No abstract available.