This study demonstrates that the modern generation of Large Language Models (LLMs, such as GPT-4) suffers from the same out-of-domain (OOD) performance gap observed in prior research on pre-trained Language Models (PLMs, such as BERT). We demonstrate this across two non-topical classification tasks: (1) genre classification and (2) generated text detection. Our results show that when demonstration examples for In-Context Learning (ICL) come from one domain (e.g., travel) and the system is tested on another domain (e.g., history), classification performance declines significantly. To address this, we introduce a method that controls which predictive indicators are used and which are excluded during classification. For the two tasks studied here, this ensures that topical features are omitted, while the model is guided to focus on stylistic rather than content-based attributes. This approach reduces the OOD gap by up to 20 percentage points in a few-shot setup. Straightforward Chain-of-Thought (CoT) methods, used as the baseline, prove insufficient, while our approach consistently enhances domain transfer performance.

Automatically recognizing an existing semantic relation (e.g. “is a”, “part of”, “property of”, “opposite of” etc.) between two words (phrases, concepts, etc.) is an important task affecting many NLP applications and has been subject of extensive experimentation and modeling. Current approaches to automatically telling if a relation exists between two given concepts X and Y can be grouped into two types: 1) those modeling word-paths connecting X and Y in text and 2) those modeling distributional properties of X and Y separately, not necessary in the proximity to each other. Here, we investigate how both types can be improved and combined. We suggest a distributional approach that is based on an attention-based transformer. We have also developed a novel word path model that combines useful properties of a convolutional network with a fully connected language model. While our transformer-based approach works better, both our models significantly outperform the state-of-the-art within their classes of approaches. We also demonstrate that combining the two approaches results in additional gains since they use somewhat different data sources.

I have shown that the presence of difficult query aspects that are revealed only implicitly (e.g. exploration, opposition, achievements, cooperation, risks) can be improved by taking advantage of the known presence of other, easier to verify query aspects. The approach proceeds by mining a large external corpus and results in substantial improvements in re-ranking the subset of the top retrieved documents.

This paper presents a new learning algorithm for multitask pattern recognition (MTPR) problems. We consider learning multiple multiclass classification tasks online where no information is ever provided about the task category of a training example. The algorithm thus needs an automated task recognition capability to properly learn the different classification tasks. The learning mode is "online" where training examples for different tasks are mixed in a random fashion and given sequentially one after another. We assume that the classification tasks are related to each other and that both the tasks and their training examples appear in random during "online training." Thus, the learning algorithm has to continually switch from learning one task to another whenever the training examples change to a different task. This also implies that the learning algorithm has to detect task changes automatically and utilize knowledge of previous tasks for learning new tasks fast. The performance of the algorithm is evaluated for ten MTPR problems using five University of California at Irvine (UCI) data sets. The experiments verify that the proposed algorithm can indeed acquire and accumulate task knowledge and that the transfer of knowledge from tasks already learned enhances the speed of knowledge acquisition on new tasks and the final classification accuracy. In addition, the task categorization accuracy is greatly improved for all MTPR problems by introducing the reorganization process even if the presentation order of class training examples is fairly biased.

Searching for words on a watchlist is one way in which large-scale surveillance of communication can be done, for example in intelligence and counterterrorism settings. One obvious defense is to replace words that might attract attention to a message with other, more innocuous, words. For example, the sentence the attack will be tomorrow" might be altered to the complex will be tomorrow", since 'complex' is a word whose frequency is close to that of 'attack'. Such substitutions are readily detectable by humans since they do not make sense. We address the problem of detecting such substitutions automatically, by looking for discrepancies between words and their contexts, and using only syntactic information. We define a set of measures, each of which is quite weak, but which together produce per-sentence detection rates around 90% with false positive rates around 10%. Rules for combining persentence detection into per-message detection can reduce the false positive and false negative rates for messages to practical levels. We test the approach using sentences from the Enron email and Brown corpora, representing informal and formal text respectively.

Many artificial intelligence tasks, such as automated question answering, reasoning or heterogeneous database integration, involve verification of a semantic category (e.g. "coffee" is a drink, "red" is a color, while "steak" is not a drink and "big" is not a color). We present a novel algorithm to automatically validate a semantic category. Contrary to the methods suggested earlier, our approach does not rely on any manually codified knowledge but instead capitalizes on the diversity of topics and word usage on the World Wide Web. We have tested our approach within our online fact-seeking (question answering) environment. When tested on the TREC questions that expect the answer to belong to a specific semantic category, our approach has improved the accuracy by up to 14% depending on the model and metrics used.

Those who want to conceal the content of their communications can do so by replacing words that might trigger attention. For example, instead of writing "The bomb is in position", a terrorist may chose to write "The flower is in position." The substituted sentence would sound a bit "odd" for a human reader and it has been shown in prior research that such oddity is detectable by text mining approaches. However, the importance of each component in the suggested oddity detection approach has not been thoroughly investigated. Also, the approach has not been compared with such an obvious candidate for the task as Hidden Markov Models (HMM). In this work, we explore further oddity detection algorithms reported earlier, specifically those based on pointwise mutual information (PMI) and Hidden Markov Models (HMM).

We present the architecture of our web question answering (fact seeking) system and introduce a novel algorithm to validate semantic categories of the expected answers. When tested on the questions used by the prior research, our system demonstrated the performance comparable to the current state of the art systems. Our semantic verification algorithm has improved the accuracy of answers of the affected questions by 30%.

We have designed a representation scheme, which is based on the discrete representation of a document ranking function, which is capable of reproducing and enhancing the properties of such popular ranking functions as tf.idf, BM25 or those based on language models. Our tests have demonstrated the capability of our approach to achieve the performance of the best known scoring functions solely through training, without using any known heuristic or analytic formulas.

While being successful in providing keyword based access to web pages, commercial search portals still lack the ability to answer questions expressed in a natural language. We present a probabilistic approach to automated question answering on the Web, based on trainable patterns, answer triangulation and semantic filtering. In contrast to the other “shallow ” approaches, our approach is entirely self-learning. It does not require any manually created scoring and filtering rules while still performing comparably. It also performs better than other fully trainable approaches.

No abstract available.

No abstract available.

In this article, we report our implementation and comparison of two text clustering techniques. One is based on Ward's clustering and the other on Kohonen's Self-organizing Maps. We have evaluated how closely clusters produced by a computer resemble those created by human experts. We have also measured the time that it takes for an expert to “clean up” the automatically produced clusters. The technique based on Ward's clustering was found to be more precise. Both techniques have worked equally well in detecting associations between text documents. We used text messages obtained from group brainstorming meetings.