Association for Computational Linguistics

Scene context is well known to facilitate humans' perception of visible objects.In this paper, we investigate the role of context in Referring Expression Generation (REG) for objects in images, where existing research has often focused on distractor contexts that exert pressure on the generator.We take a new perspective on scene context in REG and hypothesize that contextual information can be conceived of as a resource that makes REG models more resilient and facilitates the generation of object descriptions, and object types in particular.We train and test Transformer-based REG models with target representations that have been artificially obscured with noise to varying degrees.We evaluate how properties of the models' visual context affect their processing and performance.Our results show that even simple scene contexts make models surprisingly resilient to perturbations, to the extent that they can identify referent types even when visual information about the target is completely missing. 1TRFtgt red van (A) noise 0.0 TRFvis red truck (A) TRFsym red truck (A) TRFtgt left elephant (F) noise 1.0 TRFvis white truck (A) TRFsym car on left (A)

Recently more attention has been given to adversarial attacks on neural networks for natural language processing (NLP). A central research topic has been the investigation of search algorithms and search constraints, accompanied by benchmark algorithms and tasks. We implement an algorithm inspired by zeroth order optimization-based attacks and compare with the benchmark results in the TextAttack framework. Surprisingly, we find that optimizationbased methods do not yield any improvement in a constrained setup and slightly benefit from approximate gradient information only in unconstrained setups where search spaces are larger. In contrast, simple heuristics exploiting nearest neighbors without querying the target function yield substantial success rates in constrained setups, and nearly full success rate in unconstrained setups, at an order of magnitude fewer queries. We conclude from these results that current TextAttack benchmark tasks are too easy and constraints are too strict, preventing meaningful research on black-box adversarial text attacks.

Text readability is the problem of determining whether a text is suitable for a certain group of readers, and thus building a model to assess the readability of text yields great significance across the disciplines of science, publishing, and education. While text readability has attracted attention since the late nineteenth century for English and other popular languages, it remains relatively underexplored in Vietnamese. Previous studies on this topic in Vietnamese have only focused on the examination of shallow word-level features using surface statistics such as frequency and ratio. Hence, features at higher levels like sentence structure and meaning are still untapped. In this study, we propose the most comprehensive analysis of Vietnamese text readability to date, targeting features at all linguistic levels, ranging from the lexical and phrasal elements to syntactic and semantic factors. This work pioneers the investigation on the effects of multi-level linguistic features on text readability in the Vietnamese language.

This study explores some of the factors that influence the probability that a coherence relation will be explicitly marked with a connective. In a corpus of 80 Dutch texts from expository and persuasive genres with annotations of the coherence structure (using Rhetorical Structure Theory, RST) and the genre-specific structure (using move analysis), we investigate the roles and interactions of genre, level in the hierarchical discourse structure, and type of coherence relation.

A frequently replicated finding is that the frequency of words affects their phonetic shape. In English, high frequency words have been shown to contain more centralized vowels than low frequency words. By contrast, a recent study on vowel articulation in German has shown a contrary finding. At the gestural level, tongue movements in HF words showed more extensive vowel targets and less coarticulationwith consonants. This paper further evaluates the later finding by taking into account a large set of verbs covering the continuum between high and low frequency. In addition to frequency the effects of two factors were analyzed: inflection (sagt vs. sagen) and speech rate (normal vs. fast). Our results imply that language experience increases the proficiency with which words are articulated: speakers are able to plan and target tongue movements earlier.

While various approaches to domain adaptation exist, the majority of them requires knowledge of the target domain, and additional data, preferably labeled. For a language like English, it is often feasible to match most of those conditions, but in low-resource languages, it presents a problem. We explore the situation when neither data nor other information about the target domain is available. We use two samples of Danish, a low-resource language, from the consumer review domain (film vs. company reviews) in a sentiment analysis task. We observe dramatic performance drops when moving from one domain to the other. We then introduce a simple offline method that makes models more robust towards unseen domains, and observe relative improvements of more than 50%.

Tokenization is widely regarded as a solved problem due to the high accuracy that rulebased tokenizers achieve. But rule-based tokenizers are hard to maintain and their rules language specific. We show that highaccuracy word and sentence segmentation can be achieved by using supervised sequence labeling on the character level combined with unsupervised feature learning. We evaluated our method on three languages and obtained error rates of 0.27 ‰ (English), 0.35 ‰ (Dutch) and 0.76 ‰ (Italian) for our best models.

Incongruity resolution is one of the most widely accepted theories of humour, suggesting that humour is due to the mixing of two disparate interpretation frames in one statement. In this paper, we explore several computational models for incongruity resolution. We introduce a new data set, consisting of a series of ‘set-ups’ (preparations for a punch line), each of them followed by four possible coherent continuations out of which only one has a comic effect. Using this data set, we redefine the task as the automatic identification of the humorous punch line among all the plausible endings. We explore several measures of semantic relatedness, along with a number of joke-specific features, and try to understand their appropriateness as computational models for incongruity detection.

Abstract This paper proposes a new training method for the phoneme identification neural network called “neural fuzzy training.” In the proposed training, nondeterministic (fuzzy) class information is assigned to the training signal, in contrast to the traditional method where a deterministic class information is assigned.This study aims at the realization of a robust neural network, thereby improving the cumulative recognition rate of the phoneme identification and avoiding overtraining. The proposed neural fuzzy training is realized by backpropagation. In the conventional training, a deterministic phoneme class information is assigned to the training signal of the neural network as the value 1 or 0. However, in the proposed training, the fuzzy class information is assigned to the training signal for each training sample as the likelihood value between 0 and 1.In the proposed training method, the likelihood is calculated by the monotonically decreasing function (such as exp(−α · d2)) of the distance between the training sample and the closest sample belonging to each phoneme class. The proposed neural fuzzy training method has a problem in that a large amount of computation cost is required since the training signal is determined by calculating the distances to all training samples. To solve this problem, the representative samples in each phoneme class are defined and the likelihood to the phoneme classes are determined by calculating the distance between the representative sample and the training sample.By this simplification of the likelihood calculation, the computational cost to determine the training signal is reduced considerably. To demonstrate the usefulness of the neural fuzzy training, an experiment is conducted: /b, d, g, m, n, N/ identification, 18 consonant identification and phrase recognition using TDNN‐LR. The ATR database is used in the experiment. In the phoneme identification experiment, the speech samples which are extracted using the hand‐label is used. The TDNN is trained using speed samples uttered in word style, and the evaluation is performed using speech samples uttered in phrase style and in sentence style.In the phrase recognition experiment using TDNN‐LR, the TDNN is trained using speed samples uttered word style using a hand label. The evaluation is performed using speech samples uttered in phrase style. In either experiment, an improvement of using the fuzzy training can be observed. Especially, in the phrase recognition experiment using TDNN‐LR, the top recognition rate is improved from 71.2 percent to 80.9 percent, and the top 5th recognition rate is improved from 92.8 percent to 96.O percent. Furthermore, it appeared also that the neural fuzzy training is a high‐speed training method.