We present Temporal Positive Pointwise Mutual Information (TPPMI) embeddings as a robust and data-efficient alternative for modeling temporal semantic change. Based on the assumption that the semantics of the most frequent words in a corpus are relatively stable over time, our model represents words as vectors of their PPMI similarities with a predefined set of such context words. We evaluate our method on the temporal word analogy benchmark of Yao et al. (2018) and compare it to the TWEC model (Di Carlo et al., 2019), demonstrating the competitiveness of the approach. While the performance of TPPMI stays below that of the state-of-the-art TWEC model, it offers a higher degree of interpretability and is applicable in scenarios where only a limited amount of data is available.

This paper discusses challenges in Natural Language Generation (NLG), specifically addressing neural networks producing output that is fluent but incorrect, leading to “hallucinations”. The SHROOM shared task involves Large Language Models in various tasks, and our methodology employs both model-agnostic and model-aware approaches for hallucination detection. The limited availability of labeled training data is addressed through automatic label generation strategies. Model-agnostic methods include word alignment and fine-tuning a BERT-based pretrained model, while model-aware methods leverage separate classifiers trained on LLMs’ internal data (layer activations and attention values). Ensemble methods combine outputs through various techniques such as regression metamodels, voting, and probability fusion. Our best performing systems achieved an accuracy of 80.6% on the model-aware track and 81.7% on the model-agnostic track, ranking 3rd and 8th among all systems, respectively.

We present DreamDrug, a crowdsourced dataset for detecting mentions of drugs in noisy user-generated item listings from darknet markets. Our dataset contains nearly 15,000 manually annotated drug entities in over 3,500 item listings scraped from the darknet market platform “DreamMarket” in 2017. We also train and evaluate baseline models for detecting these entities, using contextual language models fine-tuned in a few-shot setting and on the full dataset, and examine the effect of pretraining on in-domain unannotated corpora.

This paper describes our methods submitted for the GermEval 2021 shared task on identifying toxic, engaging and fact-claiming comments in social media texts (Risch et al., 2021). We explore simple strategies for semi-automatic generation of rule-based systems with high precision and low recall, and use them to achieve slight overall improvements over a standard BERT-based classifier.

Large datasets are essential for neural modeling of many NLP tasks. Current publicly available open-domain dialogue datasets offer a trade-off between quality (e.g., DailyDialog) and size (e.g., Opensubtitles). We narrow this gap by building a high-quality dataset of 14.8M utterances in English, and smaller datasets in German, Dutch, Spanish, Portuguese, Italian, and Hungarian. We extract and process dialogues from public-domain books made available by Project Gutenberg. We describe our dialogue extraction pipeline, analyze the effects of the various heuristics used, and present an error analysis of extracted dialogues. Finally, we conduct experiments showing that better response quality can be achieved in zero-shot and finetuning settings by training on our data than on the larger but much noisier Opensubtitles dataset. Our open-source pipeline (https://github.com/ricsinaruto/gutenberg-dialog) can be extended to further languages with little additional effort. Researchers can also build their versions of existing datasets by adjusting various trade-off parameters.

In this paper we present a novel rule-based, language independent method for determining lexical entailment relations using semantic representations built from Wiktionary definitions. Combined with a simple WordNet-based method our system achieves top scores on the English and Italian datasets of the Semeval-2020 task “Predicting Multilingual and Cross-lingual (graded) Lexical Entailment” (Glavaš et al., 2020). A detailed error analysis of our output uncovers future di- rections for improving both the semantic parsing method and the inference process on semantic graphs.

We study a typical intermediary task to Machine Translation, the alignment of NPs in the bitext. After arguing that the task remains relevant even in an end-to-end paradigm, we present simple, dictionary- and word vector-based baselines and a BERT-based system. Our results make clear that even state of the art systems relying on the best end-to-end methods can be improved by bringing in old-fashioned methods such as stopword removal, lemmatization, and dictionaries

We present a system for mapping Universal Dependency structures to raw text which learns to restore word order by training an Interpreted Regular Tree Grammar (IRTG) that establishes a mapping between string and graph operations. The reinflection step is handled by a standard sequence-to-sequence architecture with a biLSTM encoder and an LSTM decoder with attention. We modify our 2019 system (Kovács et al., 2019) with a new grammar induction mechanism that allows IRTG rules to operate on lemmata in addition to part-of-speech tags and ensures that each word and its dependents are reordered using the most specific set of learned patterns. We also introduce a hierarchical approach to word order restoration that independently determines the word order of each clause in a sentence before arranging them with respect to the main clause, thereby improving overall readability and also making the IRTG parsing task tractable. We participated in the 2020 Surface Realization Shared task, subtrack T1a (shallow, closed). Human evaluation shows we achieve significant improvements on two of the three out-of-domain datasets compared to the 2019 system we modified. Both components of our system are available on GitHub under an MIT license.

Current neural network-based conversational models lack diversity and generate boring responses to open-ended utterances. Priors such as persona, emotion, or topic provide additional information to dialog models to aid response generation, but annotating a dataset with priors is expensive and such annotations are rarely available. While previous methods for improving the quality of open-domain response generation focused on either the underlying model or the training objective, we present a method of filtering dialog datasets by removing generic utterances from training data using a simple entropy-based approach that does not require human supervision. We conduct extensive experiments with different variations of our method, and compare dialog models across 17 evaluation metrics to show that training on datasets filtered this way results in better conversational quality as chatbots learn to output more diverse responses.

The Surface Realization Shared Task involves mapping Universal Dependency graphs to raw text, i.e. restoring word order and inflection from a graph of typed, directed dependencies between lemmas. Interpreted Regular Tree Grammars (IRTGs) encode the correspondence between generations in multiple algebras, and have previously been used for semantic parsing from raw text. Our system induces an IRTG for simultaneously building pairs of surface forms and UD graphs in the SRST training data, then prunes this grammar for each UD graph in the test data for efficient parsing and generation of the surface ordering of lemmas. For the inflection step we use a standard sequence-to-sequence model with a biLSTM encoder and an LSTM decoder with attention. Both components of our system are available on GitHub under an MIT license.

We present the dict_to_4lang tool for processing entries of three monolingual dictionaries of English and mapping definitions to concept graphs following the 4lang principles of semantic representation introduced by (Kornai, 2010). 4lang representations are domain- and language-independent, and make use of only a very limited set of primitives to encode the meaning of all utterances. Our pipeline relies on the Stanford Dependency Parser for syntactic analysis, the dep to 4lang module then builds directed graphs of concepts based on dependency relations between words in each definition. Several issues are handled by construction-specific rules that are applied to the output of dep_to_4lang. Manual evaluation suggests that ca. 75% of graphs built from the Longman Dictionary are either entirely correct or contain only minor errors. dict_to_4lang is available under an MIT license as part of the 4lang library and has been used successfully in measuring Semantic Textual Similarity (Recski and Ács, 2015). An interactive demo of core 4lang functionalities is available at http://4lang.hlt.bme.hu.

We propose a novel method for detecting optional arguments of Hungarian verbs using only positive data. We introduce a custom variant of collexeme analysis that explicitly models the noise in verb frames. Our method is, for the most part, unsupervised: we use the spectral clustering algorithm described in Brew and Schulte in Walde (2002) to build a noise model from a short, manually verified seed list of verbs. We experimented with both raw count- and context-based clusterings and found their performance almost identical. The code for our algorithm and the frame list are freely available at http://hlt.bme.hu/en/resources/tade.

We describe, and make public, large-scale language resources and the toolchain used in their creation, for fifteen medium density European languages: Catalan, Czech, Croatian, Danish, Dutch, Finnish, Lithuanian, Norwegian, Polish, Portuguese, Romanian, Serbian, Slovak, Spanish, and Swedish. To make the process uniform across languages, we selected tools that are either language-independent or easily customizable for each language, and reimplemented all stages that were taking too long. To achieve processing times that are insignificant compared to the time data collection (crawling) takes, we reimplemented the standard sentence- and word-level tokenizers and created new boilerplate and near-duplicate detection algorithms. Preliminary experiments with non-European languages indicate that our methods are now applicable not just to our sample, but the entire population of digitally viable languages, with the main limiting factor being the availability of high quality stemmers.

Aligning the NPs of parallel corpora is logically halfway between the sentence- and word-alignment tasks that occupy much of the MT literature, but has received far less attention. NP alignment is a challenging problem, capable of rapidly exposing flaws both in the word-alignment and in the NP chunking algorithms one may bring to bear. It is also a very rewarding problem in that NPs are semantically natural translation units, which means that (i) word alignments will cross NP boundaries only exceptionally, and (ii) within sentences already aligned, the proportion of 1-1 alignments will be higher for NPs than words. We created a simple gold standard for English-Hungarian, Orwell’s 1984, (since this already exists in manually verified POS-tagged format in many languages thanks to the Multex and MultexEast project) by manually verifying the automaticaly generated NP chunking (we used the yamcha, mallet and hunchunk taggers) and manually aligning the maximal NPs and PPs. The maximum NP chunking problem is much harder than base NP chunking, with F-measure in the .7 range (as opposed to over .94 for base NPs). Since the results are highly impacted by the quality of the NP chunking, we tested our alignment algorithms both with real world (machine obtained) chunkings, where results are in the .35 range for the baseline algorithm which propagates GIZA++ word alignments to the NP level, and on idealized (manually obtained) chunkings, where the baseline reaches .4 and our current system reaches .64.