Large language models (LLMs) can, in principle, bootstrap language technologies for longtail languages due to their pattern recognition capabilities.Yet in practice, without structured guidance, they produce narrow, unrepresentative samples that fail to cover the morphosyntactic space of typologically underrepresented languages.We propose Modular Typology-Informed Generation (mTIG), a prompting framework that transforms descriptive grammars into explicit control mechanisms that guide LLMs to generate typologically balanced synthetic data for downstream training.mTIG decomposes grammars into modular grammar slices, each targeting a specific morphosyntactic phenomenon (e.g., passive voice, causative morphology).Across three low-resource languages, mTIG improves typological entropy by up to 19% and yields a "student-beats-teacher" effect, where distilled models outperform the source LLM by up to +20 chrF in machine translation.These findings show that grammar-as-control can construct training corpora wherever formal linguistic descriptions exist.

Multilingual models often treat language diversity as a problem of data imbalance, overlooking structural variation.We introduce the Morphological Index (MoI), a typologically grounded metric that quantifies how strongly a language relies on surface morphology for noun classification.Building on MoI, we propose MoI-MoE, a Mixture of Experts model that routes inputs based on morphological structure.Evaluated on 10 Bantu languages-a large, morphologically rich and underrepresented family-MoI-MoE outperforms strong baselines, improving Swahili accuracy by 14 points on noun class recognition while maintaining performance on morphology-rich languages like Zulu.These findings highlight typological structure as a practical and interpretable signal for multilingual model adaptation.

Yutong Shao, Ndapa Nakashole. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

We study the task of spoken natural language to SQL parsing (speech-to-SQL), where the goal is to map a spoken utterance to the corresponding SQL. A simple way to develop a speech-to-SQL parser is to pass the speech to an automatic speech recognition (ASR) system, and pass the transcription to a text-to-SQL parser. However, ASR is still error-prone. We propose an ASR correction method, DBATI (DataBase-Aware TaggerILM). The method first detects erroneous spans in the input, and rewrites each span. Our method leverages a novel joint representation of text and the database (DB). Our experiments show that our method yields better performance on both text quality and downstream SQL accuracy, compared to existing ASR error correction methods.

Bosung Kim, Hayate Iso, Nikita Bhutani, Estevam Hruschka, Ndapa Nakashole, Tom Mitchell. Proceedings of the 13th International Joint Conference on Natural Language Processing and the 3rd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics (Volume 1: Long Papers). 2023.

Current medical question answering systems have difficulty processing long, detailed and informally worded questions submitted by patients, called Consumer Health Questions (CHQs). To address this issue, we introduce a medical question understanding and answering system with knowledge grounding and semantic self-supervision. Our system is a pipeline that first summarizes a long, medical, user-written question, using a supervised summarization loss. Then, our system performs a two-step retrieval to return answers. The system first matches the summarized user question with an FAQ from a trusted medical knowledge base, and then retrieves a fixed number of relevant sentences from the corresponding answer document. In the absence of labels for question matching or answer relevance, we design 3 novel, self-supervised and semantically-guided losses. We evaluate our model against two strong retrieval-based question answering baselines. Evaluators ask their own questions and rate the answers retrieved by our baselines and own system according to their relevance. They find that our system retrieves more relevant answers, while achieving speeds 20 times faster. Our self-supervised losses also help the summarizer achieve higher scores in ROUGE, as well as in human evaluation metrics.

Khalil Mrini, Franck Dernoncourt, Seunghyun Yoon, Trung Bui, Walter Chang, Emilia Farcas, Ndapa Nakashole. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

Khalil Mrini, Emilia Farcas, Ndapa Nakashole. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2021.

This paper presents the problem of conversational plotting agents that carry out plotting actions from natural language instructions.To facilitate the development of such agents, we introduce CHARTDIALOGS, a new multi-turn dialog dataset, covering a popular plotting library, matplotlib.The dataset contains over 15, 000 dialog turns from 3, 200 dialogs covering the majority of matplotlib plot types.Extensive experiments show the bestperforming method achieving 61% plotting accuracy, demonstrating that the dataset presents a non-trivial challenge for future research on this task.

This paper presents computational approaches for automatically detecting critical plot twists in reviews of media products.First, we created a large-scale book review dataset that includes fine-grained spoiler annotations at the sentence-level, as well as book and (anonymized) user information.Second, we carefully analyzed this dataset, and found that: spoiler language tends to be book-specific; spoiler distributions vary greatly across books and review authors; and spoiler sentences tend to jointly appear in the latter part of reviews.Third, inspired by these findings, we developed an end-to-end neural network architecture to detect spoiler sentences in review corpora.Quantitative and qualitative results demonstrate that the proposed method substantially outperforms existing baselines.• This was a perfect, albeit bloody, end to the series.• Though there were deaths that were definitely unwarranted: Fred Hedwig Moody Tonks Lupin Dobby, there were some really heartfelt and memorable moments: Narcissa saving Harry, Ron coming back, Hermione and Ron, Harry and Ginny, Molly killing Bellatrix, etc. • I wish we could have spent more time at Hogwarts, as one of my favorite characters, the amazing Minerva McGonagall, resides there, and we couldn't see more of her amazingness in the Battle of Hogwarts.• Harry Potter was a really, really great series that I think will be (and is) timeless.

Inducing multilingual word embeddings by learning a linear map between embedding spaces of different languages achieves remarkable accuracy on related languages.However, accuracy drops substantially when translating between distant languages.Given that languages exhibit differences in vocabulary, grammar, written form, or syntax, one would expect that embedding spaces of different languages have different structures especially for distant languages.With the goal of capturing such differences, we propose a method for learning neighborhood sensitive maps, NORMA.Our experiments show that NORMA outperforms current state-of-the-art methods for word translation between distant languages.

Leveraging zero-shot learning to learn mapping functions between vector spaces of different languages is a promising approach to bilingual dictionary induction.However, methods using this approach have not yet achieved high accuracy on the task.In this paper, we propose a bridging approach, where our main contribution is a knowledge distillation training objective.As teachers, rich resource translation paths are exploited in this role.And as learners, translation paths involving low resource languages learn from the teachers.Our training objective allows seamless addition of teacher translation paths for any given low resource pair.Since our approach relies on the quality of monolingual word embeddings, we also propose to enhance vector representations of both the source and target language with linguistic information.Our experiments on various languages show large performance gains from our distillation training objective, obtaining as high as 17% accuracy improvements.

In this paper we describe approaches for discovering acoustic concepts and relations in text. The first major goal is to be able to identify text phrases which contain a notion of audibility and can be termed as a sound or an acoustic concept. We also propose a method to define an acoustic scene through a set of sound concepts. We use pattern matching and parts of speech tags to generate sound concepts from large scale text corpora. We use dependency parsing and LSTM recurrent neural network to predict a set of sound concepts for a given acoustic scene. These methods are not only helpful in creating an acoustic knowledge base but in the future can also directly help acoustic event and scene detection research.

Whereas people learn many different types of knowledge from diverse experiences over many years, most current machine learning systems acquire just a single function or data model from just a single data set. We propose a never-ending learning paradigm for machine learning, to better reflect the more ambitious and encompassing type of learning performed by humans. As a case study, we describe the Never-Ending Language Learner (NELL), which achieves some of the desired properties of a never-ending learner, and we discuss lessons learned. NELL has been learning to read the web 24 hours/day since January 2010, and so far has acquired a knowledge base with over 80 million confidence-weighted beliefs (e.g., servedWith(tea, biscuits)). NELL has also learned millions of features and parameters that enable it to read these beliefs from the web. Additionally, it has learned to reason over these beliefs to infer new beliefs, and is able to extend its ontology by synthesizing new relational predicates. NELL can be tracked online at http://rtw.ml.cmu.edu, and followed on Twitter at @CMUNELL.

Ndapandula Nakashole, Tom M. Mitchell. Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing (Volume 1: Long Papers). 2015.

Learning to determine when the timevarying facts of a Knowledge Base (KB) have to be updated is a challenging task.We propose to learn state changing verbs from Wikipedia edit history.When a state-changing event, such as a marriage or death, happens to an entity, the infobox on the entity's Wikipedia page usually gets updated.At the same time, the article text may be updated with verbs either being added or deleted to reflect the changes made to the infobox.We use Wikipedia edit history to distantly supervise a method for automatically learning verbs and state changes.Additionally, our method uses constraints to effectively map verbs to infobox changes.We observe in our experiments that when state-changing verbs are added or deleted from an entity's Wikipedia page text, we can predict the entity's infobox updates with 88% precision and 76% recall.One compelling application of our verbs is to incorporate them as triggers in methods for updating existing KBs, which are currently mostly static.

This paper introduces FactChecker, language-aware approach to truth-finding.FactChecker differs from prior approaches in that it does not rely on iterative peer voting, instead it leverages language to infer believability of fact candidates.In particular, FactChecker makes use of linguistic features to detect if a given source objectively states facts or is speculative and opinionated.To ensure that fact candidates mentioned in similar sources have similar believability, FactChecker augments objectivity with a co-mention score to compute the overall believability score of a fact candidate.Our experiments on various datasets show that FactChecker yields higher accuracy than existing approaches.

Temporal scope adds a time dimension to facts in Knowledge Bases (KBs).These time scopes specify the time periods when a given fact was valid in real life.Without temporal scope, many facts are underspecified, reducing the usefulness of the data for upper level applications such as Question Answering.Existing methods for temporal scope inference and extraction still suffer from low accuracy.In this paper, we present a new method that leverages temporal profiles augmented with context-Contextual Temporal Profiles (CTPs) of entities.Through change patterns in an entity's CTP, we model the entity's state change brought about by real world events that happen to the entity (e.g, hired, fired, divorced, etc.).This leads to a new formulation of the temporal scoping problem as a state change detection problem.Our experiments show that this formulation of the problem, and the resulting solution are highly effective for inferring temporal scope of facts.

We propose a demonstration of PATTY, a system for learning semantic relationships from the Web. PATTY is a collection of relations learned automatically from text. It aims to be to patterns what WordNet is to words. The semantic types of PATTY relations enable advanced search over subject-predicate-object data. With the ongoing trends of enriching Web data (both text and tables) with entity-relationship-oriented semantic annotations, we believe a demo of the PATTY system will be of interest to the database community.

Harvesting relational facts from Web sources has received great attention for automatically constructing large knowledge bases. Stateof-the-art approaches combine pattern-based gathering of fact candidates with constraint-based reasoning. However, they still face major challenges regarding the trade-offs between precision, recall, and scalability. Techniques that scale well are susceptible to noisy patterns that degrade precision, while techniques that employ deep reasoning for high precision cannot cope with Web-scale data.This paper presents a scalable system, called PROSPERA, for high-quality knowledge harvesting. We propose a new notion of ngram-itemsets for richer patterns, and use MaxSat-based constraint reasoning on both the quality of patterns and the validity of fact candidates.We compute pattern-occurrence statistics for two benefits: they serve to prune the hypotheses space and to derive informative weights of clauses for the reasoner. The paper shows how to incorporate these building blocks into a scalable architecture that can parallelize all phases on a Hadoop-based distributed platform. Our experiments with the ClueWeb09 corpus include comparisons to the recent ReadTheWeb experiment. We substantially outperform these prior results in terms of recall, with the same precision, while having low run-times.

We present a robust method for gathering relational facts from the Web, based on matching generalized patterns which are automatically learned from seed facts for relations of interest. Our approach combines these generalized patterns for high recall information extraction with a rule-based, declarative reasoning approach to also ensure high precision. Newly extracted candidate facts are assigned statistical weights which reflect the strengths of the patterns used to extract them. For checking the plausibility of candidate facts with respect to existing knowledge and competing hypotheses, we use an efficient algorithm for weighted Max-Sat over propositional-logic clauses. In contrast to prior work on reasoning-based information extraction, we employ richer statistics and smart pruning to bound the number of grounded rules passed on to the Max-Sat solver.