Search

This paper proposes an approach to cross-language sentence selection in a low-resource setting. It uses data augmentation and negative sampling techniques on noisy parallel sentence data to directly learn a cross-lingual embedding-based query relevance model. Results show that this approach performs as well as or better than multiple state-of-the-art machine translation + monolingual retrieval systems trained on the same parallel data. Moreover, when a rationale training secondary objective is applied to encourage the model to match word alignment hints from a phrase-based statistical machine translation model, consistent improvements are seen across three language pairs (English-Somali, English-Swahili and English-Tagalog) over a variety of state-of-the-art baselines., ACL 2021 main conference

Typical ASR systems segment the input audio in toutterances using purely acoustic information, which may not resemble the sentence-like units that are expected by conventional machine translation (MT) systems for Spoken Language Translation. In this work, we propose a model for correcting the acoustic segmentation of ASR models for low-resource languages to improve performance on down-stream tasks. We propose the use of subtitles as a proxy dataset for correcting ASR acoustic segmentation, creating synthetic acoustic utterances by modeling common error modes. We train a neural tagging model for correcting ASR acoustic segmentation and show that it improves downstream performance on MT and audio-document cross-language information retrieval (CLIR).

In cross-language information retrieval using probabilistic structured queries (PSQ), translation probabilities from statistical machine translation act as a bridge between the query and document vocabulary. These translation probabilities are typically estimated from a sentence-aligned corpus on a word to word basis without taking into account the context. Neural methods, by contrast, can learn to translate using the context around the words, and this can be used as a basis for estimating context-dependent translation probabilities. However, sparsity limits the accuracy of context-specific translation probabilities for rare words, which can be important in retrieval applications. This paper presents evidence that combining such context-dependent translation probabilities with context-independent translation probabilities learned from the same parallel corpus can yield improvements in the effectiveness of cross-language ranked retrieval.

Cross-language speech retrieval systems face a cascade of errors due to transcription and translation ambiguity. Using 1-best speech recognition and 1-best translation in such a scenario could adversely affect recall if those 1-best system guesses are not correct. Accurately representing transcription and translation probabilities could therefore improve recall, although possibly at some cost in precision. The difficulty of the task is exacerbated when working with languages for which limited resources are available, since both recognition and translation probabilities may be less accurate in such cases. This paper explores the combination of expected term counts from recognition with expected term counts from translation to perform cross-language speech retrieval in which the queries are in English and the spoken content to be retrieved is in Tagalog or Swahili. Experiments were conducted using two query types, one focused on term presence and the other focused on topical retrieval. Overall, the results show that significant improvements in ranking quality result from modeling transcription and recognition ambiguity, even in lower-resource settings, and that adapting the ranking model to specific query types can yield further improvements.

System combination has been shown to improve overall performance on many rank-based retrieval tasks, often by combining results from multiple systems into a single ranked list. In contrast, set-based retrieval tasks call for a technique to combine results in ways that require decisions on whether each document is in or out of the result set. This paper presents a set-generating unsupervised system combination framework that draws inspiration from evaluation techniques in sparse data settings. It argues for the existence of a duality between evaluation and system combination, and then capitalizes on this duality to perform unsupervised system combination. To do this, the framework relies on the consensus of the systems to estimate latent “goodness” for each system. An implementation of this framework using data programming is compared to other unsupervised system combination approaches to demonstrate its effectiveness on CLEF and MATERIAL collections.

In recent years, a large amount of multimedia data consisting of images and text have been generated in libraries through the digitization of physical materials into data for their preservation. When they are archived, appropriate cataloging metadata are assigned to them by librarians. Automatic annotations are helpful for reducing the cost of manual annotations. To this end, we propose a mapping system that links images and the associated text to entries on Wikipedia as a replacement for annotation by targeting images and associated text from photo-sharing sites. The uploaded images are accompanied by descriptive labels of contents of the sites that can be indexed for the catalogue. However, because users freely tag images with labels, these user-assigned labels are often ambiguous. The label “albatross”, for example, may refer to a type of bird or aircraft. If the ambiguities are resolved, we can use Wikipedia entries for cataloging as an alternative to ontologies. To formalize this, we propose a task called image label disambiguation where, given an image and assigned target labels to be disambiguated, an appropriate Wikipedia page is selected for the given labels. We propose a hybrid approach for this task that makes use of both user tags as textual information and features of images generated through image recognition. To evaluate the proposed task, we develop a freely available test collection containing 450 images and 2,280 ambiguous labels. The proposed method outperformed prevalent text-based approaches in terms of the mean reciprocal rank, attaining a value of over 0.6 on both our collection and the ImageCLEF collection.

In this paper, we survey different state-of-the-art visual processing methods and utilize them in hyperlinking. Visual information, calculated using Features Signatures, SIMILE descriptors and convolutional neural networks (CNN), is utilized as similarity between video frames and used to find similar faces, objects and setting. Visual concepts in frames are also automatically recognized and textual output of the recognition is combined with search based on subtitles and transcripts. All presented experiments were performed in the Search and Hyperlinking 2014 MediaEval task and Video Hyperlinking 2015 TRECVid task.

In this paper, we describe SHAMUS, our system for an easy search and navigation in multimedia archives. The system consists of three components. The Search component provides a text-based search in a multimedia collection, the Anchoring component determines the most important segments of videos, and segments topically related to the anchoring ones are retrieved by the Hyperlinking component. In the paper, we describe each component of the system as well as the online demo interface http://ufal.mff.cuni.cz/shamus which currently works with a collection of TED talks.

In this paper, we explore the use of audio information in the retrieval of multimedia content. Specifically, we focus on linking similar segments in a collection consisting of 4,000 hours of BBC TV programmes. We provide a description of our system submitted to the Hyperlinking Sub-task of the Search and Hyperlinking Task in the MediaEval 2014 Benchmark, in which it scored best. We explore three automatic transcripts and compare them to available subtitles. We confirm the relationship between retrieval performance and transcript quality. The performance of the retrieval is further improved by extending transcripts by metadata and context, by combining different transcripts, using the highest confident words of the transcripts, and by utilizing acoustic similarity.

This paper deals with Information Retrieval from audio-visual recordings. Such recordings are often quite long and users may want to find the exact starting points of relevant passages they search for. In Passage Retrieval, the recordings are automatically segmented into smaller parts, on which the standard retrieval techniques are applied. In this paper, we discuss various techniques for segmentation of audio-visual recordings and focus on machine learning approaches which decide on segment boundaries based on various features combined in a decision-tree model. Our experiments are carried out on the data used for the Search and Hyperlinking Task and Similar Segments in Social Speech Task of the MediaEval Benchmark 2013.

The importance of Information Retrieval (IR) in audio-visual recordings has been increasing with steeply growing numbers of audio-visual documents available on-line. Compared to traditional IR methods, this task requires specific techniques, such as Passage Retrieval which can accelerate the search process by retrieving the exact relevant passage of a recording instead of the full document. In Passage Retrieval, full recordings are divided into shorter segments which serve as individual documents for the further IR setup. This technique also allows normalizing document length and applying positional information. It was shown that it can even improve retrieval results. In this work, we examine two general strategies for Passage Retrieval: blind segmentation into overlapping regular-length passages and segmentation into variable-length passages based on semantics of their content. Time-based segmentation was already shown to improve retrieval of textual documents and audio-visual recordings. Our experiments performed on the test collection used in the Search subtask of the Search and Hyperlinking Task in MediaEval Benchmarking 2012 confirm those findings and show that parameters (segment length and shift) tuning for a specific test collection can further improve the results. Our best results on this collection were achieved by using 45-second long segments with 15-second shifts. Semantic-based segmentation can be divided into three types: similarity-based (producing segments with high intra-similarity and low inter-similarity), lexical-chain-based (producing segments with frequent lexically connected words), and feature-based (combining various features which signalize a segment break in a machine-learning setting). In this work, we mainly focus on feature-based segmentation which allows exploiting various features from all modalities of the data (including segment length) in a single trainable model and produces segments which can eventually overlap. Our preliminary results show that even simple semantic-based segmentation outperforms regular segmentation. Our model is a decision tree incorporating the following features: shot segments, output of TextTiling algorithm, cue words (well, thanks, so, I, now), sentence breaks, and the length of the silence after the previous word. In terms of the MASP, the relative improvement over regular segmentation is more than 19%.

International audience; Searching for relevant webpages and following hyperlinks to related content is a widely accepted and effective approach to information seeking on the textual web. Existing work on multimedia information retrieval has focused on search for individual relevant items or on content linking without specific attention to search results. We describe our research exploring integrated multimodal search and hyperlinking for multimedia data. Our investigation is based on the MediaEval 2012 Search and Hyperlinking task. This includes a known-item search task using the Blip10000 internet video collection, where automatically created hyperlinks link each relevant item to related items within the collection. The search test queries and link assessment for this task was generated using the Amazon Mechanical Turk crowdsourcing platform. Our investigation examines a range of alternative methods which seek to address the challenges of search and hyperlinking using multimodal approaches. The results of our experiments are used to propose a research agenda for developing effective techniques for search and hyperlinking of multimedia content.

This paper deals with evaluation of information retrieval from unsegmented speech. We focus on Mean Generalized Average Precision, the evaluation measure widely used for unsegmented speech retrieval. This measure is designed to allow certain tolerance in matching retrieval results (starting points of relevant segments) against a gold standard relevance assessment. It employs a Penalty Function which evaluates non-exact matches in the retrieval results based on their distance from the beginnings of their nearest true relevant segments. However, the choice of the Penalty Function is usually ad-hoc and does not necessary reflect users' perception of the speech retrieval quality. We perform a lab test to study satisfaction of users of a speech retrieval system to empirically estimate the optimal shape of the Penalty Function.